Microsoft’s Carbon Control Plane Explained

1
00:00:00,000 –> 00:00:02,620
Most organizations treat corporate social responsibility

2
00:00:02,620 –> 00:00:05,840
like a marketing layer, a glossy PDF, a donation,

3
00:00:05,840 –> 00:00:08,520
a press release, and a net zero badge for the website.

4
00:00:08,520 –> 00:00:09,280
They are wrong.

5
00:00:09,280 –> 00:00:11,560
CSR is a governance system under constraint,

6
00:00:11,560 –> 00:00:13,320
and the constraint is getting physical,

7
00:00:13,320 –> 00:00:15,160
power, water, land, and carbon.

8
00:00:15,160 –> 00:00:17,580
In this episode, Microsoft is the case study

9
00:00:17,580 –> 00:00:20,080
and will use three operational artifacts,

10
00:00:20,080 –> 00:00:22,880
the sustainability report, the internal carbon fee,

11
00:00:22,880 –> 00:00:24,880
and cloud for sustainability.

12
00:00:24,880 –> 00:00:25,960
Here’s the open loop.

13
00:00:25,960 –> 00:00:27,920
Microsoft promised carbon negative,

14
00:00:27,920 –> 00:00:29,480
yet its emissions still rose.

15
00:00:29,480 –> 00:00:31,320
That tension is the audit.

16
00:00:31,320 –> 00:00:33,400
CSR in one clean definition.

17
00:00:33,400 –> 00:00:36,080
CSR, stripped of slogans, is simple.

18
00:00:36,080 –> 00:00:37,640
A company makes business decisions

19
00:00:37,640 –> 00:00:40,520
that account for societal impact, not just profit,

20
00:00:40,520 –> 00:00:44,480
not donations, not awareness campaigns, decisions.

21
00:00:44,480 –> 00:00:47,080
And decisions only matter when they change trade-offs

22
00:00:47,080 –> 00:00:48,160
inside the system.

23
00:00:48,160 –> 00:00:50,340
The common framing is the triple bottom line.

24
00:00:50,340 –> 00:00:52,120
People, planet, profit.

25
00:00:52,120 –> 00:00:53,960
Most executives say it like a mantra,

26
00:00:53,960 –> 00:00:57,080
then go back to a single bottom line with nicer typography.

27
00:00:57,080 –> 00:00:58,560
The more accurate way to treat it

28
00:00:58,560 –> 00:00:59,960
is as an operating model.

29
00:00:59,960 –> 00:01:02,400
Three constraint domains that compete for priority

30
00:01:02,400 –> 00:01:04,640
every time a team ships a product,

31
00:01:04,640 –> 00:01:06,960
chooses a supplier or builds a data center.

32
00:01:06,960 –> 00:01:10,520
People means the company’s footprint on humans, employees,

33
00:01:10,520 –> 00:01:13,200
customers, communities, and the extended workforce

34
00:01:13,200 –> 00:01:14,200
in the supply chain.

35
00:01:14,200 –> 00:01:16,960
In practice, this includes wages, safety, diversity,

36
00:01:16,960 –> 00:01:18,960
accessibility, and how the company responds

37
00:01:18,960 –> 00:01:20,360
when incentives produce harm.

38
00:01:20,360 –> 00:01:22,600
Planet means the company’s physical footprint,

39
00:01:22,600 –> 00:01:25,320
emissions, water use, waste, land use,

40
00:01:25,320 –> 00:01:28,080
and the resource extraction hidden behind digital.

41
00:01:28,080 –> 00:01:30,880
Profit means the company stays solvent and scalable.

42
00:01:30,880 –> 00:01:32,200
CSR doesn’t remove profit.

43
00:01:32,200 –> 00:01:34,800
It removes the fantasy that profit is the only variable

44
00:01:34,800 –> 00:01:36,680
that deserves deterministic control.

45
00:01:36,680 –> 00:01:38,600
There’s also a useful vocabulary split.

46
00:01:38,600 –> 00:01:40,760
Four types of CSR you’ll hear repeatedly.

47
00:01:40,760 –> 00:01:43,560
Environmental, ethical, philanthropic, and economic

48
00:01:43,560 –> 00:01:44,800
environmental is the obvious one.

49
00:01:44,800 –> 00:01:46,760
Emissions, energy, water, waste.

50
00:01:46,760 –> 00:01:49,480
Ethical is how the company behaves when nobody is watching.

51
00:01:49,480 –> 00:01:51,960
Labor standards, privacy, procurement integrity,

52
00:01:51,960 –> 00:01:53,840
and avoiding the predictable exploitation

53
00:01:53,840 –> 00:01:56,520
that shows up wherever pressure meets opacity.

54
00:01:56,520 –> 00:01:58,880
Philanthropic is what most people mistake for CSR,

55
00:01:58,880 –> 00:02:01,080
giving money, volunteering, community programs.

56
00:02:01,080 –> 00:02:03,400
It can be good, it can also become a flush,

57
00:02:03,400 –> 00:02:06,240
and economic responsibility is the uncomfortable part,

58
00:02:06,240 –> 00:02:08,840
running the business in a way that doesn’t externalize costs

59
00:02:08,840 –> 00:02:12,240
onto society and then call the externalization innovation.

60
00:02:12,240 –> 00:02:14,880
Here’s the thing most people miss.

61
00:02:14,880 –> 00:02:17,760
CSR becomes real only when it becomes measurable.

62
00:02:17,760 –> 00:02:19,720
The moment you can measure something, you can budget it.

63
00:02:19,720 –> 00:02:21,480
The moment you can budget it, you can enforce it,

64
00:02:21,480 –> 00:02:24,160
and the moment you can enforce it, it stops being culture

65
00:02:24,160 –> 00:02:25,520
and becomes control.

66
00:02:25,520 –> 00:02:27,400
That’s why sustainability reports matter

67
00:02:27,400 –> 00:02:28,880
even when they’re annoying to read.

68
00:02:28,880 –> 00:02:30,360
They’re not inspirational literature,

69
00:02:30,360 –> 00:02:33,000
they’re a public interface to an internal accounting system.

70
00:02:33,000 –> 00:02:35,040
They declare boundaries, define scopes,

71
00:02:35,040 –> 00:02:36,680
and force a company to put numbers

72
00:02:36,680 –> 00:02:38,640
where it previously put adjectives.

73
00:02:38,640 –> 00:02:41,240
They also reveal where the company refuses measurement,

74
00:02:41,240 –> 00:02:43,600
which is usually where the real risk lives.

75
00:02:43,600 –> 00:02:45,240
Measurement changes everything.

76
00:02:45,240 –> 00:02:48,840
You get baselines, targets, deltas, and accountability cycles.

77
00:02:48,840 –> 00:02:50,800
Teams stop saying we care and start saying,

78
00:02:50,800 –> 00:02:53,800
we’re up 12% year over year, and here’s why.

79
00:02:53,800 –> 00:02:56,360
That isn’t virtue, that is operational reality.

80
00:02:56,360 –> 00:02:58,520
And it sets the stage for the only question that matters.

81
00:02:58,520 –> 00:03:00,360
Do the numbers change decisions?

82
00:03:00,360 –> 00:03:01,960
Or do they just decorate the narrative?

83
00:03:01,960 –> 00:03:03,520
This is the uncomfortable truth.

84
00:03:03,520 –> 00:03:06,920
In 2006, CSR exists because constraints arrived.

85
00:03:06,920 –> 00:03:09,240
Not because executives collectively became enlightened.

86
00:03:09,240 –> 00:03:11,760
Customers ask for disclosures, employees ask

87
00:03:11,760 –> 00:03:13,560
whether they’re building something defensible.

88
00:03:13,560 –> 00:03:16,040
Investors ask about risk exposure, regulators ask

89
00:03:16,040 –> 00:03:18,520
for auditable reporting, and physics asks for payment

90
00:03:18,520 –> 00:03:20,560
eventually because data centers are not metaphors.

91
00:03:20,560 –> 00:03:22,640
They’re industrial facilities that consume

92
00:03:22,640 –> 00:03:24,600
power, water, and hardware at a scale

93
00:03:24,600 –> 00:03:27,480
that makes sustainability an architecture problem.

94
00:03:27,480 –> 00:03:30,320
So when this episode calls CSR control,

95
00:03:30,320 –> 00:03:31,520
it’s not a rhetorical trick.

96
00:03:31,520 –> 00:03:33,720
It’s describing the system’s actual function.

97
00:03:33,720 –> 00:03:36,200
CSR is the control plane for non-financial risk,

98
00:03:36,200 –> 00:03:38,920
human risk, environmental risk, reputational risk,

99
00:03:38,920 –> 00:03:39,960
and regulatory risk.

100
00:03:39,960 –> 00:03:42,080
If the control plane is weak, policy drifts,

101
00:03:42,080 –> 00:03:43,520
exceptions accumulate.

102
00:03:43,520 –> 00:03:45,200
And you end up with conditional chaos,

103
00:03:45,200 –> 00:03:47,480
claims that sound coherent while the underlying flows

104
00:03:47,480 –> 00:03:48,400
do whatever they want.

105
00:03:48,400 –> 00:03:50,760
Now we can talk about Microsoft without getting hypnotized

106
00:03:50,760 –> 00:03:51,480
by branding.

107
00:03:51,480 –> 00:03:53,600
We’ll treat Microsoft’s sustainability story

108
00:03:53,600 –> 00:03:56,800
as what it is, and attempt to build a carbon control plane

109
00:03:56,800 –> 00:03:59,080
that makes infrastructure decisions uncomfortable

110
00:03:59,080 –> 00:04:00,080
in the right places.

111
00:04:00,080 –> 00:04:01,560
That distinction matters.

112
00:04:01,560 –> 00:04:04,680
Why environmental responsibility became a business imperative?

113
00:04:04,680 –> 00:04:07,600
Environmental responsibility didn’t become a business imperative

114
00:04:07,600 –> 00:04:10,120
because the corporate world discovered a conscience.

115
00:04:10,120 –> 00:04:12,600
It became imperative because stakeholders turned

116
00:04:12,600 –> 00:04:14,920
environmental impact into a requirement

117
00:04:14,920 –> 00:04:17,880
that shows up in deals, hiring, capital, and law.

118
00:04:17,880 –> 00:04:19,080
The system tightened.

119
00:04:19,080 –> 00:04:20,720
The margin for vibes disappeared.

120
00:04:20,720 –> 00:04:23,320
Start with customers because customers now run procurement

121
00:04:23,320 –> 00:04:24,680
like an audit committee.

122
00:04:24,680 –> 00:04:27,680
Large buyers ask suppliers for emissions, disclosures, targets,

123
00:04:27,680 –> 00:04:28,320
and progress.

124
00:04:28,320 –> 00:04:30,240
Sometimes it’s formal questionnaires.

125
00:04:30,240 –> 00:04:32,120
Sometimes it’s contractual language.

126
00:04:32,120 –> 00:04:34,360
Either way, it changes the sales motion.

127
00:04:34,360 –> 00:04:36,080
You’re no longer selling just capability.

128
00:04:36,080 –> 00:04:37,920
You’re selling capability under constraint.

129
00:04:37,920 –> 00:04:40,600
If your sustainability posture can’t survive a spreadsheet,

130
00:04:40,600 –> 00:04:42,240
you don’t just lose reputation.

131
00:04:42,240 –> 00:04:43,480
You lose revenue.

132
00:04:43,480 –> 00:04:45,360
And when the largest buyers behave that way,

133
00:04:45,360 –> 00:04:47,280
it cascades through the supply chain

134
00:04:47,280 –> 00:04:49,600
until environmental responsibility

135
00:04:49,600 –> 00:04:52,240
stops being optional and becomes table stakes.

136
00:04:52,240 –> 00:04:54,920
Then there’s employees which executives still underestimate

137
00:04:54,920 –> 00:04:57,800
because they treated as culture instead of labor economics.

138
00:04:57,800 –> 00:05:00,480
Talent markets, especially in the knowledge economy,

139
00:05:00,480 –> 00:05:01,960
don’t just price compensation.

140
00:05:01,960 –> 00:05:03,480
They price alignment and risk.

141
00:05:03,480 –> 00:05:05,600
People don’t want to spend five years building systems

142
00:05:05,600 –> 00:05:06,960
that will be politically toxic,

143
00:05:06,960 –> 00:05:09,320
regulatory, constrained, or morally indefensible

144
00:05:09,320 –> 00:05:10,040
on a bad day.

145
00:05:10,040 –> 00:05:12,640
Millennials and Gen Z don’t need to be romanticized here.

146
00:05:12,640 –> 00:05:15,000
The only relevant fact is that they ask different questions

147
00:05:15,000 –> 00:05:16,280
and they ask them earlier.

148
00:05:16,280 –> 00:05:18,120
They want to see evidence, not pledges.

149
00:05:18,120 –> 00:05:21,120
A sustainability narrative that can’t survive internal scrutiny

150
00:05:21,120 –> 00:05:22,840
becomes a retention problem.

151
00:05:22,840 –> 00:05:25,200
And retention problems become productivity problems,

152
00:05:25,200 –> 00:05:26,560
which turn into product problems,

153
00:05:26,560 –> 00:05:28,240
which then become financial problems.

154
00:05:28,240 –> 00:05:30,080
It’s not ideology, it’s throughput.

155
00:05:30,080 –> 00:05:31,680
Investors show up next and they bring

156
00:05:31,680 –> 00:05:33,800
the coldest logic of all, risk pricing.

157
00:05:33,800 –> 00:05:36,240
They don’t need to believe in sustainability

158
00:05:36,240 –> 00:05:39,360
to punish you for unmanaged exposure.

159
00:05:39,360 –> 00:05:42,640
Climate risk shows up as operational risk, supply chain risk,

160
00:05:42,640 –> 00:05:45,000
energy price volatility, insurance friction,

161
00:05:45,000 –> 00:05:46,360
and future compliance costs.

162
00:05:46,360 –> 00:05:48,920
A company that can’t articulate its emissions profile

163
00:05:48,920 –> 00:05:51,720
is signaling something worse than environmental negligence.

164
00:05:51,720 –> 00:05:53,800
It’s signaling that it can’t measure itself.

165
00:05:53,800 –> 00:05:55,680
And any system that can’t measure itself,

166
00:05:55,680 –> 00:05:56,680
can’t control itself.

167
00:05:56,680 –> 00:05:58,240
Capital hates that.

168
00:05:58,240 –> 00:06:00,440
Regulation is the final forcing function,

169
00:06:00,440 –> 00:06:02,880
because regulators don’t care about your brand story.

170
00:06:02,880 –> 00:06:05,480
They care about traceability and comparability.

171
00:06:05,480 –> 00:06:08,680
Reporting requirements turn narrative into liability.

172
00:06:08,680 –> 00:06:10,680
Once you publish a metric, you’ve created a promise

173
00:06:10,680 –> 00:06:12,720
that can be examined, challenged, and used

174
00:06:12,720 –> 00:06:13,880
against you later.

175
00:06:13,880 –> 00:06:15,640
That changes corporate behavior fast,

176
00:06:15,640 –> 00:06:17,160
not because companies fear shame,

177
00:06:17,160 –> 00:06:19,560
but because they fear enforcement and litigation.

178
00:06:19,560 –> 00:06:22,000
It’s the same reason security programs mature.

179
00:06:22,000 –> 00:06:24,240
The threat model eventually includes auditors.

180
00:06:24,240 –> 00:06:27,240
Now, here’s the pivot that most CSR conversations avoid.

181
00:06:27,240 –> 00:06:29,960
Environmental responsibility became a business imperative

182
00:06:29,960 –> 00:06:33,080
because sustainability moved into the core machinery,

183
00:06:33,080 –> 00:06:36,080
procurement, architecture, and finance.

184
00:06:36,080 –> 00:06:38,920
Those three domains decide what an organization actually does.

185
00:06:38,920 –> 00:06:40,000
Everything else is commentary.

186
00:06:40,000 –> 00:06:42,360
Procurement is where sustainability stops being a poster

187
00:06:42,360 –> 00:06:43,240
and becomes a filter.

188
00:06:43,240 –> 00:06:45,960
If you require suppliers to disclose emissions,

189
00:06:45,960 –> 00:06:47,600
you’re not just collecting data.

190
00:06:47,600 –> 00:06:49,520
You’re reshaping the market around you.

191
00:06:49,520 –> 00:06:51,960
Suppliers either adapt or they get replaced.

192
00:06:51,960 –> 00:06:54,040
That is governance via purchasing power.

193
00:06:54,040 –> 00:06:56,280
And it’s brutal in the way all incentives are brutal.

194
00:06:56,280 –> 00:06:58,520
It doesn’t care about intent, only compliance.

195
00:06:58,520 –> 00:07:01,160
Architecture is where sustainability becomes physical.

196
00:07:01,160 –> 00:07:04,160
Cloud systems feel abstract until you look at the data center.

197
00:07:04,160 –> 00:07:07,680
Power delivery, cooling, water, hardware life cycles, and land.

198
00:07:07,680 –> 00:07:10,200
The industry spent a decade selling digital transformation,

199
00:07:10,200 –> 00:07:11,200
like it was weightless.

200
00:07:11,200 –> 00:07:11,960
It isn’t.

201
00:07:11,960 –> 00:07:14,160
Cloud is an industrialization of compute,

202
00:07:14,160 –> 00:07:17,200
and AI is an intensification of that industrialization.

203
00:07:17,200 –> 00:07:20,160
When power becomes scarce or politically contested,

204
00:07:20,160 –> 00:07:22,520
the architecture stops being a purely technical choice

205
00:07:22,520 –> 00:07:24,560
and becomes a capacity planning problem

206
00:07:24,560 –> 00:07:26,080
with public consequences.

207
00:07:26,080 –> 00:07:28,160
Latency and resiliency still matter.

208
00:07:28,160 –> 00:07:30,560
But now they’re joined by something that doesn’t negotiate.

209
00:07:30,560 –> 00:07:31,160
Energy.

210
00:07:31,160 –> 00:07:35,360
Finance is the last piece because finance turns values into behavior.

211
00:07:35,360 –> 00:07:38,120
If sustainability lives in a report, it’s optional.

212
00:07:38,120 –> 00:07:40,720
If sustainability lives in a budget line, it’s real.

213
00:07:40,720 –> 00:07:42,680
The moment carbon has an internal price,

214
00:07:42,680 –> 00:07:44,840
engineers feel it without being asked to care.

215
00:07:44,840 –> 00:07:46,280
That’s the uncomfortable trick.

216
00:07:46,280 –> 00:07:47,960
You don’t scale ethical awareness.

217
00:07:47,960 –> 00:07:48,960
You scale incentives.

218
00:07:48,960 –> 00:07:51,080
And that’s why old school CSR marketing

219
00:07:51,080 –> 00:07:52,680
fails under modern constraints.

220
00:07:52,680 –> 00:07:54,800
It tells a story, but it doesn’t change decisions.

221
00:07:54,800 –> 00:07:56,600
It adds a layer of language over a system

222
00:07:56,600 –> 00:07:59,320
whose defaults still optimize for speed and revenue.

223
00:07:59,320 –> 00:08:01,800
Over time, that mismatch creates drift.

224
00:08:01,800 –> 00:08:05,400
The brand claims one thing, the operational graph does another.

225
00:08:05,400 –> 00:08:06,400
The world notices.

226
00:08:06,400 –> 00:08:08,440
Sometimes slowly, sometimes all at once.

227
00:08:08,440 –> 00:08:11,400
So the business imperative isn’t be sustainable.

228
00:08:11,400 –> 00:08:13,840
The imperative is remain governable in a world

229
00:08:13,840 –> 00:08:16,520
where sustainability is enforced by customers, employees,

230
00:08:16,520 –> 00:08:18,160
investors and regulators.

231
00:08:18,160 –> 00:08:20,160
Microsoft sits in the center of that pressure.

232
00:08:20,160 –> 00:08:21,680
It sells computer planetary scale,

233
00:08:21,680 –> 00:08:24,240
therefore it can’t treat sustainability as a side program.

234
00:08:24,240 –> 00:08:25,960
It has to treat it like a control plane.

235
00:08:25,960 –> 00:08:28,280
Next, we draw the fraud boundary,

236
00:08:28,280 –> 00:08:31,480
what counts as impact and what counts as theater.

237
00:08:31,480 –> 00:08:34,840
CSR marketing versus measurable impact, the fraud boundary.

238
00:08:34,840 –> 00:08:37,080
Most organizations don’t lie about CSR

239
00:08:37,080 –> 00:08:38,480
in the dramatic cartoon wave.

240
00:08:38,480 –> 00:08:40,400
They do something more operationally dangerous.

241
00:08:40,400 –> 00:08:43,280
They tell the story first and they negotiate measurement later.

242
00:08:43,280 –> 00:08:45,920
That ordering matters because once the story exists,

243
00:08:45,920 –> 00:08:48,240
everything underneath starts optimizing to protect it.

244
00:08:48,240 –> 00:08:51,120
Matrix become accessories, definitions become flexible,

245
00:08:51,120 –> 00:08:52,880
boundaries become negotiable.

246
00:08:52,880 –> 00:08:56,200
And the CSR program slowly turns into a distributed narrative

247
00:08:56,200 –> 00:08:58,800
engine that can survive almost any reality

248
00:08:58,800 –> 00:09:01,320
as long as the reality stays hard to audit.

249
00:09:01,320 –> 00:09:02,840
This is where greenwashing actually lives,

250
00:09:02,840 –> 00:09:05,960
not in one false claim, but in a system design choice.

251
00:09:05,960 –> 00:09:08,120
If messaging leads and measurement follows,

252
00:09:08,120 –> 00:09:10,280
the company builds a probabilistic truth model.

253
00:09:10,280 –> 00:09:12,400
It can always say we meant this other scope

254
00:09:12,400 –> 00:09:14,960
or we’re improving or we’ll get there by 203

255
00:09:14,960 –> 00:09:16,840
because the control plane never had enforcement.

256
00:09:16,840 –> 00:09:17,720
It had branding.

257
00:09:17,720 –> 00:09:18,920
So here’s the fraud boundary.

258
00:09:18,920 –> 00:09:20,600
It’s not did they run an ad?

259
00:09:20,600 –> 00:09:21,680
Everyone runs ads.

260
00:09:21,680 –> 00:09:24,080
The boundary is simpler, did they change the operating system

261
00:09:24,080 –> 00:09:27,120
of decisions or did they just change the wording around them?

262
00:09:27,120 –> 00:09:29,000
A useful test is to ignore the campaign

263
00:09:29,000 –> 00:09:31,960
and look for operational artifacts, budget mechanisms,

264
00:09:31,960 –> 00:09:34,720
procurement requirements, engineering gates, audit trails.

265
00:09:34,720 –> 00:09:37,080
The thing most people miss is that genuine CSR

266
00:09:37,080 –> 00:09:38,800
leaves fingerprints where it hurts.

267
00:09:38,800 –> 00:09:41,160
It shows up in the places that teams complain about.

268
00:09:41,160 –> 00:09:43,640
It creates friction, it introduces constraints,

269
00:09:43,640 –> 00:09:45,200
it makes someone’s life harder today

270
00:09:45,200 –> 00:09:47,320
so the organization can still function tomorrow.

271
00:09:47,320 –> 00:09:49,040
Marketing CSR does the opposite.

272
00:09:49,040 –> 00:09:50,280
It removes friction.

273
00:09:50,280 –> 00:09:51,760
It tells you the company is committed

274
00:09:51,760 –> 00:09:53,560
so you’ll stop asking how it’s enforced.

275
00:09:53,560 –> 00:09:55,840
That’s why the best fraud detector is governance language.

276
00:09:55,840 –> 00:09:56,600
Listen for verbs.

277
00:09:56,600 –> 00:09:58,800
We support and we care our atmosphere.

278
00:09:58,800 –> 00:10:01,680
We require, we charge, we disclose, we verify,

279
00:10:01,680 –> 00:10:03,040
we report our structure.

280
00:10:03,040 –> 00:10:05,480
And structure is the only thing that survives scale.

281
00:10:05,480 –> 00:10:07,680
Now annual reporting becomes the really important

282
00:10:07,680 –> 00:10:09,040
middle layer in this whole game.

283
00:10:09,040 –> 00:10:11,400
People treat sustainability reports like PR.

284
00:10:11,400 –> 00:10:12,080
They’re not.

285
00:10:12,080 –> 00:10:13,760
They’re report cards and a report card

286
00:10:13,760 –> 00:10:15,560
is an accountability interface.

287
00:10:15,560 –> 00:10:19,040
Once a company publishes a boundary, a baseline and a target,

288
00:10:19,040 –> 00:10:20,560
it has created a public contract.

289
00:10:20,560 –> 00:10:23,640
Even if nobody reads the report, it exists as an anchor

290
00:10:23,640 –> 00:10:26,080
that employees, customers, journalists and regulators

291
00:10:26,080 –> 00:10:27,880
can use later to measure drift

292
00:10:27,880 –> 00:10:30,120
but reporting also has its own failure mode.

293
00:10:30,120 –> 00:10:31,360
Disclosure theatre.

294
00:10:31,360 –> 00:10:34,920
A company can disclose endlessly while changing nothing.

295
00:10:34,920 –> 00:10:37,560
It can publish attractive charts, heroic goals

296
00:10:37,560 –> 00:10:41,120
and bold commitments while the underlying system stays untouched.

297
00:10:41,120 –> 00:10:42,840
It looks like transparency but it isn’t.

298
00:10:42,840 –> 00:10:44,760
It’s just higher resolution storytelling.

299
00:10:44,760 –> 00:10:47,360
So the fraud boundary isn’t reporting versus not reporting.

300
00:10:47,360 –> 00:10:49,440
It’s whether reporting connects to enforcement.

301
00:10:49,440 –> 00:10:52,400
If reporting feeds budgets and defaults, it’s governance.

302
00:10:52,400 –> 00:10:54,640
If reporting feeds slogans, it’s decoration.

303
00:10:54,640 –> 00:10:56,640
And this is where philanthropy becomes the favorite

304
00:10:56,640 –> 00:10:57,480
camouflage layer.

305
00:10:57,480 –> 00:10:59,280
Philanthropy can be real good done in the world.

306
00:10:59,280 –> 00:11:01,920
It can also be an entropy generator for accountability

307
00:11:01,920 –> 00:11:04,440
because it gives the organization an easy moral win

308
00:11:04,440 –> 00:11:06,560
that doesn’t require changing the core footprint.

309
00:11:06,560 –> 00:11:08,840
Donations don’t change data center design.

310
00:11:08,840 –> 00:11:11,360
Volunteering doesn’t change supply chain emissions.

311
00:11:11,360 –> 00:11:13,160
A charitable program can be meaningful

312
00:11:13,160 –> 00:11:16,280
and still be irrelevant to the company’s primary harm vector.

313
00:11:16,280 –> 00:11:18,720
The uncomfortable truth is that a lot of CSR programs

314
00:11:18,720 –> 00:11:21,000
are built around what is emotionally compelling,

315
00:11:21,000 –> 00:11:23,160
not what is operationally material.

316
00:11:23,160 –> 00:11:25,240
They focus on what looks good in a video montage

317
00:11:25,240 –> 00:11:26,920
because what’s material is boring.

318
00:11:26,920 –> 00:11:29,640
Procurement clauses, internal chargeback models,

319
00:11:29,640 –> 00:11:33,080
supplier reporting and teams arguing about boundaries.

320
00:11:33,080 –> 00:11:34,640
If you want a clean way to think about it,

321
00:11:34,640 –> 00:11:36,960
use the core footprint rule.

322
00:11:36,960 –> 00:11:39,120
Real CSR attacks the core footprint.

323
00:11:39,120 –> 00:11:41,840
The thing the company does at scale to generate revenue.

324
00:11:41,840 –> 00:11:44,160
If the company sells oil, the core footprint

325
00:11:44,160 –> 00:11:46,040
is extraction and combustion.

326
00:11:46,040 –> 00:11:47,720
If the company sells compute,

327
00:11:47,720 –> 00:11:50,720
the core footprint is power, hardware supply chains

328
00:11:50,720 –> 00:11:52,040
and physical infrastructure.

329
00:11:52,040 –> 00:11:54,760
CSR that doesn’t touch the core footprint is not CSR.

330
00:11:54,760 –> 00:11:56,000
It’s brand risk management.

331
00:11:56,000 –> 00:11:57,680
That doesn’t mean the intent is evil.

332
00:11:57,680 –> 00:11:59,360
It means the architecture is dishonest.

333
00:11:59,360 –> 00:12:01,560
Systems follow incentives, not feelings.

334
00:12:01,560 –> 00:12:04,560
So when someone claims impact, ask three questions.

335
00:12:04,560 –> 00:12:06,240
First, what changed in defaults?

336
00:12:06,240 –> 00:12:09,200
If nothing changed in defaults, behavior didn’t change at scale.

337
00:12:09,200 –> 00:12:10,640
Second, what changed in budgets?

338
00:12:10,640 –> 00:12:12,320
If nothing changed in budgets,

339
00:12:12,320 –> 00:12:13,960
the organization didn’t pay for the claim.

340
00:12:13,960 –> 00:12:16,120
Third, what changed in accountability?

341
00:12:16,120 –> 00:12:18,040
If nobody can be wrong, nobody can improve.

342
00:12:18,040 –> 00:12:20,840
That’s the fraud boundary, not vibes, not tone mechanisms.

343
00:12:20,840 –> 00:12:22,680
Now apply that boundary to Microsoft.

344
00:12:22,680 –> 00:12:24,480
Ignore speeches, ignore ads.

345
00:12:24,480 –> 00:12:26,800
Look at the artifacts, what Microsoft reports,

346
00:12:26,800 –> 00:12:29,280
what Microsoft charges itself internally

347
00:12:29,280 –> 00:12:32,160
and what Microsoft productizes as governance for others

348
00:12:32,160 –> 00:12:33,960
because those three things reveal

349
00:12:33,960 –> 00:12:36,880
what the company is actually doing when nobody is clapping.

350
00:12:36,880 –> 00:12:41,560
Microsoft’s environmental commitments, what they actually claim.

351
00:12:41,560 –> 00:12:43,760
Now Microsoft, not as a logo,

352
00:12:43,760 –> 00:12:45,840
as an operator of planetary scale compute

353
00:12:45,840 –> 00:12:48,040
with a sustainability story attached to it.

354
00:12:48,040 –> 00:12:49,760
Microsoft’s environmental commitments

355
00:12:49,760 –> 00:12:52,640
as they publicly state them in their sustainability reporting

356
00:12:52,640 –> 00:12:54,120
are straightforward on paper,

357
00:12:54,120 –> 00:12:56,120
become carbon negative by 203

358
00:12:56,120 –> 00:12:58,240
and by 205 remove from the atmosphere

359
00:12:58,240 –> 00:13:01,560
the carbon the company has emitted since its founding in 1975.

360
00:13:01,560 –> 00:13:03,360
That’s the headline claim and it matters

361
00:13:03,360 –> 00:13:05,280
because it’s not a promise to do better.

362
00:13:05,280 –> 00:13:07,440
It’s a promise to hit a measurable condition.

363
00:13:07,440 –> 00:13:10,240
Removal succeed emissions within a defined timeline

364
00:13:10,240 –> 00:13:12,880
and historical emissions get addressed, not just future ones.

365
00:13:12,880 –> 00:13:15,560
They also position parallel targets around water and waste,

366
00:13:15,560 –> 00:13:17,040
water positive and zero waste

367
00:13:17,040 –> 00:13:18,760
because carbon isn’t the only constraint

368
00:13:18,760 –> 00:13:20,880
that turns computing into a physical industry.

369
00:13:20,880 –> 00:13:23,680
That matters because it signals Microsoft understands this

370
00:13:23,680 –> 00:13:26,480
as a resource problem, not a public relations theme.

371
00:13:26,480 –> 00:13:28,880
Data centers don’t just burn electricity.

372
00:13:28,880 –> 00:13:30,760
They interact with water availability,

373
00:13:30,760 –> 00:13:32,680
local infrastructure and supply chains

374
00:13:32,680 –> 00:13:34,600
that produce actual material waste.

375
00:13:34,600 –> 00:13:37,280
But the big one, the one everyone fixates on, is carbon.

376
00:13:37,280 –> 00:13:39,520
And Microsoft’s framing is intentionally too layered.

377
00:13:39,520 –> 00:13:42,360
Layer one is reduction, lower the emissions associated

378
00:13:42,360 –> 00:13:44,080
with operations and the value chain.

379
00:13:44,080 –> 00:13:46,600
Layer two is removal, compensate for the part

380
00:13:46,600 –> 00:13:47,920
they can’t reduce fast enough

381
00:13:47,920 –> 00:13:50,160
by purchasing or enabling carbon removal.

382
00:13:50,160 –> 00:13:51,840
That split is not moral philosophy.

383
00:13:51,840 –> 00:13:53,840
It’s systems engineering under constraint.

384
00:13:53,840 –> 00:13:55,760
The reason this is the Microsoft case study

385
00:13:55,760 –> 00:13:57,960
is that they’re not just declaring goals.

386
00:13:57,960 –> 00:14:00,080
They’re building mechanisms around those goals,

387
00:14:00,080 –> 00:14:03,000
disclosures, internal economics and productized measurement,

388
00:14:03,000 –> 00:14:04,840
start with how they talk about renewable energy

389
00:14:04,840 –> 00:14:05,640
and infrastructure.

390
00:14:05,640 –> 00:14:07,200
In the sustainability narrative,

391
00:14:07,200 –> 00:14:10,000
renewables aren’t presented as personal virtue.

392
00:14:10,000 –> 00:14:12,280
They’re presented as infrastructure strategy,

393
00:14:12,280 –> 00:14:15,160
long-term power procurement, operational resiliency,

394
00:14:15,160 –> 00:14:18,360
and a way to bend the emissions curve of electricity consumption,

395
00:14:18,360 –> 00:14:20,360
again, not a sentiment, a control lever.

396
00:14:21,720 –> 00:14:23,400
And here’s the uncomfortable truth.

397
00:14:23,400 –> 00:14:25,640
Microsoft’s footprint matters because digital

398
00:14:25,640 –> 00:14:26,840
stopped being small.

399
00:14:26,840 –> 00:14:29,320
Cloud is an industrial consolidation of compute.

400
00:14:29,320 –> 00:14:32,480
It takes what used to be thousands of inefficient server rooms

401
00:14:32,480 –> 00:14:34,840
and concentrates it into hyper scale facilities

402
00:14:34,840 –> 00:14:38,160
that are designed, financed and operated like infrastructure

403
00:14:38,160 –> 00:14:41,160
that already changes the sustainability equation.

404
00:14:41,160 –> 00:14:43,440
Then AI arrives and turns the dial again,

405
00:14:43,440 –> 00:14:46,040
more accelerators, more dense racks, more cooling,

406
00:14:46,040 –> 00:14:48,200
more power delivery, more hardware churn.

407
00:14:48,200 –> 00:14:50,040
So Microsoft is now selling compute

408
00:14:50,040 –> 00:14:52,280
while claiming climate leadership in the same breath.

409
00:14:52,280 –> 00:14:53,840
That is not hypocrisy by default.

410
00:14:53,840 –> 00:14:55,960
It is simply what happens when a company’s product

411
00:14:55,960 –> 00:14:57,560
is the thing that creates the footprint.

412
00:14:57,560 –> 00:14:59,400
In other words, Microsoft can’t decouple

413
00:14:59,400 –> 00:15:01,760
the sustainability conversation from product strategy

414
00:15:01,760 –> 00:15:04,320
because the product strategy is the emissions strategy.

415
00:15:04,320 –> 00:15:06,440
This is also why their sustainability commitments

416
00:15:06,440 –> 00:15:08,840
can’t be judged like a consumer brand’s commitments.

417
00:15:08,840 –> 00:15:11,000
For a software company in 2005,

418
00:15:11,000 –> 00:15:13,640
going green could plausibly mean buying offsets,

419
00:15:13,640 –> 00:15:15,760
switching office lighting and calling it a day.

420
00:15:15,760 –> 00:15:17,680
For a hyper scale in 2026,

421
00:15:17,680 –> 00:15:19,800
the footprint is dominated by infrastructure growth

422
00:15:19,800 –> 00:15:21,000
and supply chains.

423
00:15:21,000 –> 00:15:22,560
The posture has to be different

424
00:15:22,560 –> 00:15:24,080
because the physics is different.

425
00:15:24,080 –> 00:15:26,560
Now, the audience should notice what Microsoft does

426
00:15:26,560 –> 00:15:27,760
and does not claim.

427
00:15:27,760 –> 00:15:30,520
They do claim specific timelines and outcomes,

428
00:15:30,520 –> 00:15:34,040
carbon negative by 203, historical removal by 205

429
00:15:34,040 –> 00:15:35,680
because those are auditable targets.

430
00:15:35,680 –> 00:15:37,720
They also disclose progress and context

431
00:15:37,720 –> 00:15:40,200
in annual reporting which forces continuity.

432
00:15:40,200 –> 00:15:42,920
If emissions rise, the story can’t hide it forever.

433
00:15:42,920 –> 00:15:44,600
The report becomes a record of drift.

434
00:15:44,600 –> 00:15:46,360
They do not claim that growth stops.

435
00:15:46,360 –> 00:15:49,720
They do not claim that compute demand politely levels off

436
00:15:49,720 –> 00:15:51,840
so sustainability becomes easy.

437
00:15:51,840 –> 00:15:54,280
Their commitments implicitly assume the opposite.

438
00:15:54,280 –> 00:15:57,000
Growth continues therefore governance has to get sharper

439
00:15:57,000 –> 00:15:58,440
and that’s where the tension shows up.

440
00:15:58,440 –> 00:16:00,400
Microsoft has a business model that scales

441
00:16:00,400 –> 00:16:02,760
by selling more cloud and more AI capability

442
00:16:02,760 –> 00:16:04,080
that’s the success condition.

443
00:16:04,080 –> 00:16:05,480
But the sustainability commitments

444
00:16:05,480 –> 00:16:07,400
set a parallel success condition.

445
00:16:07,400 –> 00:16:10,200
Reduce and remove fast enough that the net claim holds.

446
00:16:10,200 –> 00:16:11,920
These are competing optimization targets

447
00:16:11,920 –> 00:16:13,240
inside the same company.

448
00:16:13,240 –> 00:16:14,360
That distinction matters

449
00:16:14,360 –> 00:16:16,520
because it reframes the whole debate.

450
00:16:16,520 –> 00:16:18,640
The question isn’t, does Microsoft care?

451
00:16:18,640 –> 00:16:20,720
The question is, can Microsoft maintain

452
00:16:20,720 –> 00:16:22,600
a deterministic sustainability model

453
00:16:22,600 –> 00:16:25,360
while scaling an inherently physical platform?

454
00:16:25,360 –> 00:16:27,320
And the answer depends on whether commitments become

455
00:16:27,320 –> 00:16:28,160
enforcement.

456
00:16:28,160 –> 00:16:29,280
Targets are intent.

457
00:16:29,280 –> 00:16:30,480
Intent is cheap.

458
00:16:30,480 –> 00:16:32,120
Systems drift without control

459
00:16:32,120 –> 00:16:33,880
because exceptions accumulate.

460
00:16:33,880 –> 00:16:36,520
A new region here, a new workload class there,

461
00:16:36,520 –> 00:16:39,320
a supply chain constraint, a procurement shortcut,

462
00:16:39,320 –> 00:16:41,480
a customer demand that overrides the default.

463
00:16:41,480 –> 00:16:43,600
So the only way to treat these commitments seriously

464
00:16:43,600 –> 00:16:45,720
is to treat them as architecture requirements.

465
00:16:45,720 –> 00:16:47,080
They have to bind decisions.

466
00:16:47,080 –> 00:16:48,560
They have to constrain design.

467
00:16:48,560 –> 00:16:50,360
They have to show up in budgets and procurement

468
00:16:50,360 –> 00:16:51,720
and engineering trade-offs.

469
00:16:51,720 –> 00:16:54,440
Next, we need to clean up the foundational misunderstanding

470
00:16:54,440 –> 00:16:57,080
that makes most sustainability discussions useless.

471
00:16:57,080 –> 00:16:58,680
Carbon negative isn’t a feeling.

472
00:16:58,680 –> 00:17:00,000
It’s an accounting outcome.

473
00:17:00,000 –> 00:17:01,400
The foundational misunderstanding here

474
00:17:01,400 –> 00:17:03,120
and carbon negative isn’t a feeling.

475
00:17:03,120 –> 00:17:05,400
The foundational misunderstanding is that carbon negative

476
00:17:05,400 –> 00:17:08,640
is an attitude like eco-friendly but with better branding.

477
00:17:08,640 –> 00:17:09,320
It is not.

478
00:17:09,320 –> 00:17:11,680
It’s a balanced sheet claim that only exists

479
00:17:11,680 –> 00:17:14,280
if the math survives contact with system boundaries.

480
00:17:14,280 –> 00:17:16,680
Operationally, carbon negative means removal

481
00:17:16,680 –> 00:17:18,600
exceed emissions over a defined period

482
00:17:18,600 –> 00:17:21,080
using defined scopes, using defined accounting rules.

483
00:17:21,080 –> 00:17:21,600
That’s it.

484
00:17:21,600 –> 00:17:23,760
No inspirational wording can soften that.

485
00:17:23,760 –> 00:17:26,320
If the company emits X and removes Y,

486
00:17:26,320 –> 00:17:29,040
the claim is true only when Y is larger than X

487
00:17:29,040 –> 00:17:31,480
and only when both numbers refer to the same boundary.

488
00:17:31,480 –> 00:17:34,160
That boundary is where the argument always hides.

489
00:17:34,160 –> 00:17:35,720
A company can say carbon negative

490
00:17:35,720 –> 00:17:38,120
but means scopes one and two only.

491
00:17:38,120 –> 00:17:40,000
Another can mean scopes one, two and three.

492
00:17:40,000 –> 00:17:41,400
Those are not comparable statements.

493
00:17:41,400 –> 00:17:42,520
They are different systems

494
00:17:42,520 –> 00:17:45,520
and Microsoft explicitly talks in terms of scopes

495
00:17:45,520 –> 00:17:48,080
which is why the conversation gets messy fast.

496
00:17:48,080 –> 00:17:49,920
The closer you get to scope three,

497
00:17:49,920 –> 00:17:52,360
the less deterministic the model becomes.

498
00:17:52,360 –> 00:17:55,800
Now separate three actions people constantly blur together.

499
00:17:55,800 –> 00:17:57,640
Reduction, removal and offsetting.

500
00:17:57,640 –> 00:17:59,520
Reduction is cutting the emissions you create

501
00:17:59,520 –> 00:18:00,640
in the first place.

502
00:18:00,640 –> 00:18:04,320
Efficiency, renewable electricity, lower carbon materials,

503
00:18:04,320 –> 00:18:06,960
better logistics, less waste, fewer flights.

504
00:18:06,960 –> 00:18:09,080
It’s the part engineers instinctively respect

505
00:18:09,080 –> 00:18:10,720
because it maps to use less.

506
00:18:10,720 –> 00:18:12,800
Removal is pulling carbon out of the atmosphere

507
00:18:12,800 –> 00:18:14,840
and storing it in a way that counts as durable

508
00:18:14,840 –> 00:18:16,520
according to the rules being used.

509
00:18:16,520 –> 00:18:18,360
It’s not we did something nice.

510
00:18:18,360 –> 00:18:20,400
It’s we created a measured negative flow.

511
00:18:20,400 –> 00:18:23,280
In a carbon negative claim, removal is the mechanism

512
00:18:23,280 –> 00:18:25,560
that allows the net number to cross below zero

513
00:18:25,560 –> 00:18:28,000
when reductions can’t keep pace with growth.

514
00:18:28,000 –> 00:18:30,640
Offsetting is the broad category that people use to mean

515
00:18:30,640 –> 00:18:32,000
we paid someone else.

516
00:18:32,000 –> 00:18:34,120
Which includes a spectrum of credit types

517
00:18:34,120 –> 00:18:35,920
with wildly different properties.

518
00:18:35,920 –> 00:18:38,200
Some are reduction credits, some are avoidance claims,

519
00:18:38,200 –> 00:18:39,560
some are removal claims.

520
00:18:39,560 –> 00:18:41,240
The word offset tells you almost nothing

521
00:18:41,240 –> 00:18:42,640
about what happened physically.

522
00:18:42,640 –> 00:18:44,280
It tells you what happened financially.

523
00:18:44,280 –> 00:18:46,360
That distinction matters because net claims

524
00:18:46,360 –> 00:18:48,280
get built on category confusion.

525
00:18:48,280 –> 00:18:49,960
Reduction feels like engineering.

526
00:18:49,960 –> 00:18:52,640
Removal feels like engineering plus procurement.

527
00:18:52,640 –> 00:18:54,480
Offsetting often feels like paperwork

528
00:18:54,480 –> 00:18:56,400
but they all end up in the same spreadsheet

529
00:18:56,400 –> 00:18:58,880
therefore marketing treats them as interchangeable.

530
00:18:58,880 –> 00:18:59,720
They are not.

531
00:18:59,720 –> 00:19:01,320
And then scope three shows up

532
00:19:01,320 –> 00:19:03,040
and ruins the simplistic narrative

533
00:19:03,040 –> 00:19:05,240
because scope three is where most companies actually live.

534
00:19:05,240 –> 00:19:06,720
Scope one is your direct emissions.

535
00:19:06,720 –> 00:19:08,800
Scope two is your purchased electricity.

536
00:19:08,800 –> 00:19:10,280
Those are comparatively measurable

537
00:19:10,280 –> 00:19:12,520
because the company can map them to owned assets

538
00:19:12,520 –> 00:19:13,920
and utility bills.

539
00:19:13,920 –> 00:19:15,640
Scope three is everything else.

540
00:19:15,640 –> 00:19:18,160
Supply chain emissions, purchased goods, logistics,

541
00:19:18,160 –> 00:19:19,840
business travel, employee commuting

542
00:19:19,840 –> 00:19:22,600
and the downstream emissions associated with product use

543
00:19:22,600 –> 00:19:23,840
depending on the category.

544
00:19:23,840 –> 00:19:26,640
In other words, scope three is the emissions you cause

545
00:19:26,640 –> 00:19:28,560
without owning the thing that emits.

546
00:19:28,560 –> 00:19:31,480
That means your carbon accounting becomes an identity problem.

547
00:19:31,480 –> 00:19:33,680
You are no longer measuring exhaust pipes.

548
00:19:33,680 –> 00:19:36,440
You’re measuring relationships, suppliers, contractors,

549
00:19:36,440 –> 00:19:38,800
customers, grids, regions and time.

550
00:19:38,800 –> 00:19:41,200
You are building an attribution model for reality.

551
00:19:41,200 –> 00:19:42,920
And attribution models are always contested.

552
00:19:42,920 –> 00:19:45,600
This clicked for me when I started treating carbon accounting

553
00:19:45,600 –> 00:19:47,800
like authorization in a distributed system.

554
00:19:47,800 –> 00:19:49,880
You can make a clean claim about who owns

555
00:19:49,880 –> 00:19:51,520
an action inside one tenant.

556
00:19:51,520 –> 00:19:53,520
But scope three is cross-tenant by design.

557
00:19:53,520 –> 00:19:54,560
It’s a graph.

558
00:19:54,560 –> 00:19:56,960
It has missing data, inconsistent standards

559
00:19:56,960 –> 00:19:58,680
and incentives to under-report.

560
00:19:58,680 –> 00:20:00,480
So if you want a deterministic security model,

561
00:20:00,480 –> 00:20:02,280
you enforce it by design.

562
00:20:02,280 –> 00:20:04,160
If you want deterministic carbon governance,

563
00:20:04,160 –> 00:20:05,440
you need the same thing.

564
00:20:05,440 –> 00:20:07,760
Enforcement built into purchasing, budgeting

565
00:20:07,760 –> 00:20:10,040
and deployment decisions, not just reporting.

566
00:20:10,040 –> 00:20:12,360
So the unavoidable conclusion is this.

567
00:20:12,360 –> 00:20:14,560
You don’t audit intent, you audit flows,

568
00:20:14,560 –> 00:20:16,840
you audit energy consumption, hardware purchases,

569
00:20:16,840 –> 00:20:19,760
logistics, travel, supplier reporting and removal procurement.

570
00:20:19,760 –> 00:20:21,640
You audit the boundaries and the deltas,

571
00:20:21,640 –> 00:20:24,400
you audit whether a carbon negative outcome is being achieved

572
00:20:24,400 –> 00:20:26,520
through real reductions, credible removals

573
00:20:26,520 –> 00:20:27,880
or simply accounting choices

574
00:20:27,880 –> 00:20:30,280
that push emissions outside the declared boundary.

575
00:20:30,280 –> 00:20:33,000
And once you accept that, the whole tone changes.

576
00:20:33,000 –> 00:20:34,280
The question stops being,

577
00:20:34,280 –> 00:20:36,640
do we like their sustainability story?

578
00:20:36,640 –> 00:20:38,480
The question becomes, is there carbon control

579
00:20:38,480 –> 00:20:40,160
plain capable of maintaining integrity

580
00:20:40,160 –> 00:20:41,720
while the platform scales?

581
00:20:41,720 –> 00:20:44,720
Microsoft tried to answer that with an internal market mechanism.

582
00:20:44,720 –> 00:20:46,720
Make carbon behave like a budget line item

583
00:20:46,720 –> 00:20:48,960
so teams feel it without being asked to care.

584
00:20:48,960 –> 00:20:50,360
That’s where we go next.

585
00:20:50,360 –> 00:20:54,040
Artifact one, the internal carbon fee, carbon as a budget line.

586
00:20:54,040 –> 00:20:56,680
Microsoft’s internal carbon fee is the first artifact

587
00:20:56,680 –> 00:20:58,440
that matters because it’s not a slogan.

588
00:20:58,440 –> 00:21:01,160
It’s an accounting mechanism that creates consequences.

589
00:21:01,160 –> 00:21:02,760
And in systems terms, consequences

590
00:21:02,760 –> 00:21:05,000
are what separate governance from theater.

591
00:21:05,000 –> 00:21:06,960
The publicly described model is simple.

592
00:21:06,960 –> 00:21:10,360
Since 2012, Microsoft has charged internal business divisions

593
00:21:10,360 –> 00:21:13,040
a fee per metric ton of carbon dioxide equivalent

594
00:21:13,040 –> 00:21:17,000
or TCO2e tied to the emissions associated with their activity.

595
00:21:17,000 –> 00:21:19,080
Not a shadow price that lives in a spreadsheet

596
00:21:19,080 –> 00:21:20,240
for moral education.

597
00:21:20,240 –> 00:21:23,120
A real charge that hits budgets, that distinction matters

598
00:21:23,120 –> 00:21:25,000
because the moment a number affects a budget,

599
00:21:25,000 –> 00:21:26,760
it stops being a values conversation

600
00:21:26,760 –> 00:21:28,640
and becomes an optimization problem.

601
00:21:28,640 –> 00:21:30,880
The fee is designed around the emissions scopes

602
00:21:30,880 –> 00:21:32,480
used in greenhouse gas accounting.

603
00:21:32,480 –> 00:21:35,360
It covers scope one, scope two, and scope three categories

604
00:21:35,360 –> 00:21:37,680
as Microsoft expanded the program over time.

605
00:21:37,680 –> 00:21:40,360
In plain language, not just direct emissions and electricity,

606
00:21:40,360 –> 00:21:42,320
but also the broader value chain categories

607
00:21:42,320 –> 00:21:44,120
that normally let companies shrug and say,

608
00:21:44,120 –> 00:21:45,560
“That’s not ours.”

609
00:21:45,560 –> 00:21:47,320
Microsoft didn’t remove the ambiguity

610
00:21:47,320 –> 00:21:49,000
because scope three is still messy,

611
00:21:49,000 –> 00:21:51,280
but it did something most organizations avoid.

612
00:21:51,280 –> 00:21:54,240
It pulled the mess into the internal economic model anyway.

613
00:21:54,240 –> 00:21:55,760
There’s a reason they did that.

614
00:21:55,760 –> 00:21:57,920
Scope three is where the platform lives,

615
00:21:57,920 –> 00:22:00,920
hardware supply chains, logistics, employee travel

616
00:22:00,920 –> 00:22:03,880
and the emissions associated with operating at global scale.

617
00:22:03,880 –> 00:22:05,440
If you don’t charge for scope three,

618
00:22:05,440 –> 00:22:07,840
you incentivize the organization to push emissions

619
00:22:07,840 –> 00:22:09,440
into the category that doesn’t count.

620
00:22:09,440 –> 00:22:11,040
You build a perverse control plane,

621
00:22:11,040 –> 00:22:13,320
you reward the behavior you claim to hate.

622
00:22:13,320 –> 00:22:14,520
And the fee is not static.

623
00:22:14,520 –> 00:22:16,880
Microsoft has revised and increased it over time,

624
00:22:16,880 –> 00:22:19,760
including a redesigned described in sustainability disclosures

625
00:22:19,760 –> 00:22:22,680
around 2022 to accelerate progress.

626
00:22:22,680 –> 00:22:25,520
And they’ve talked about escalating specific components

627
00:22:25,520 –> 00:22:28,440
like business travel to higher-per-ton pricing.

628
00:22:28,440 –> 00:22:30,480
The exact rate isn’t the important part here

629
00:22:30,480 –> 00:22:33,600
because rates change, the architecture is the important part.

630
00:22:33,600 –> 00:22:36,160
Microsoft treats carbon as a unit cost

631
00:22:36,160 –> 00:22:37,560
that gets louder over time,

632
00:22:37,560 –> 00:22:39,720
specifically in areas where the organization

633
00:22:39,720 –> 00:22:42,880
has historically treated emissions as background noise.

634
00:22:42,880 –> 00:22:44,120
Now, what happens to the money?

635
00:22:44,120 –> 00:22:46,720
Microsoft’s disclosures describe the proceeds

636
00:22:46,720 –> 00:22:49,960
being reinvested into sustainability initiatives,

637
00:22:49,960 –> 00:22:53,320
energy efficiency work, renewable energy procurement,

638
00:22:53,320 –> 00:22:57,120
lower carbon operations, supplier decarbonization efforts

639
00:22:57,120 –> 00:22:58,880
and related investments.

640
00:22:58,880 –> 00:23:01,400
This is the second half of the mechanism that people ignore.

641
00:23:01,400 –> 00:23:03,920
If you charge a fee and then send the money into the void,

642
00:23:03,920 –> 00:23:05,160
you’ve built a tax.

643
00:23:05,160 –> 00:23:08,640
If you charge a fee and reinvest the proceeds into decarbonization,

644
00:23:08,640 –> 00:23:10,400
you’ve built a closed-loop system.

645
00:23:10,400 –> 00:23:12,120
Behavior gets pressured and the pressure

646
00:23:12,120 –> 00:23:14,440
funds the tools to reduce future pressure.

647
00:23:14,440 –> 00:23:16,880
It’s the same idea as a well-designed internal chargeback

648
00:23:16,880 –> 00:23:18,240
model for cloud spend.

649
00:23:18,240 –> 00:23:20,440
You allocate costs to the thing that caused it,

650
00:23:20,440 –> 00:23:22,800
therefore, teams have incentive to reduce it.

651
00:23:22,800 –> 00:23:24,560
Then you use some of the recovered money

652
00:23:24,560 –> 00:23:26,360
to fix the underlying inefficiencies.

653
00:23:26,360 –> 00:23:27,560
Nobody calls that virtue.

654
00:23:27,560 –> 00:23:28,560
They call it governance.

655
00:23:28,560 –> 00:23:30,520
This is the part where most organizations mess up

656
00:23:30,520 –> 00:23:32,880
because they want the branding of internal carbon pricing

657
00:23:32,880 –> 00:23:35,360
without the political pain of actually charging business

658
00:23:35,360 –> 00:23:36,360
units.

659
00:23:36,360 –> 00:23:37,080
They deploy shadow pricing.

660
00:23:37,080 –> 00:23:38,840
Shadow pricing is emotionally safe.

661
00:23:38,840 –> 00:23:40,600
It’s also operationally weak.

662
00:23:40,600 –> 00:23:42,760
Real chargeback changes behavior because it

663
00:23:42,760 –> 00:23:45,880
competes with other priorities in the only arena that matters.

664
00:23:45,880 –> 00:23:48,360
Budget planning, quarterly targets, and business unit

665
00:23:48,360 –> 00:23:48,960
performance.

666
00:23:48,960 –> 00:23:49,960
It forces a meeting.

667
00:23:49,960 –> 00:23:51,000
It forces trade-offs.

668
00:23:51,000 –> 00:23:52,520
It forces someone to say, out loud,

669
00:23:52,520 –> 00:23:55,240
we can afford this feature, but we can’t afford the emissions

670
00:23:55,240 –> 00:23:57,040
that come with it at this internal price.

671
00:23:57,040 –> 00:23:58,000
That sounds dramatic.

672
00:23:58,000 –> 00:23:58,840
It isn’t.

673
00:23:58,840 –> 00:24:00,680
It’s just how systems behave when you make externalities

674
00:24:00,680 –> 00:24:01,280
internal.

675
00:24:01,280 –> 00:24:03,680
And this is why the carbon fee is an operational artifact,

676
00:24:03,680 –> 00:24:05,000
not a marketing asset.

677
00:24:05,000 –> 00:24:06,680
It reveals Microsoft’s theory of change.

678
00:24:06,680 –> 00:24:07,600
Their theory is not.

679
00:24:07,600 –> 00:24:09,520
Convince every engineer to care.

680
00:24:09,520 –> 00:24:11,800
Their theory is make carbon behave like cost.

681
00:24:11,800 –> 00:24:13,440
Therefore, engineers will optimize it

682
00:24:13,440 –> 00:24:15,600
because engineers optimize constraints.

683
00:24:15,600 –> 00:24:17,560
But there’s a third piece baked into the fee

684
00:24:17,560 –> 00:24:19,920
that creates both power and risk.

685
00:24:19,920 –> 00:24:21,680
It depends on measurement integrity.

686
00:24:21,680 –> 00:24:24,120
The fee only works if the emissions attribution model

687
00:24:24,120 –> 00:24:25,200
is defensible.

688
00:24:25,200 –> 00:24:27,880
If the data is wrong, teams will optimize the wrong thing.

689
00:24:27,880 –> 00:24:30,040
If the boundaries are fuzzy, teams will argue

690
00:24:30,040 –> 00:24:32,320
the accounting instead of changing behavior.

691
00:24:32,320 –> 00:24:34,200
If exceptions multiply, the program

692
00:24:34,200 –> 00:24:37,120
becomes another example of conditional chaos.

693
00:24:37,120 –> 00:24:39,040
A control plane that exists on paper

694
00:24:39,040 –> 00:24:40,920
while the underlying system roots around it.

695
00:24:40,920 –> 00:24:42,560
So yes, the internal carbon fee

696
00:24:42,560 –> 00:24:44,880
is the most structurally serious thing Microsoft does

697
00:24:44,880 –> 00:24:46,280
in sustainability governance.

698
00:24:46,280 –> 00:24:49,160
It turns carbon from a moral narrative into a budget line.

699
00:24:49,160 –> 00:24:50,720
But incentives don’t create truth.

700
00:24:50,720 –> 00:24:53,280
They amplify whatever measurement model you build.

701
00:24:53,280 –> 00:24:55,240
And that’s why the next section matters more

702
00:24:55,240 –> 00:24:56,600
than people wanted to.

703
00:24:56,600 –> 00:24:58,760
Carbon accounting is not an emissions problem.

704
00:24:58,760 –> 00:25:00,440
It’s an identity problem.

705
00:25:00,440 –> 00:25:01,760
Measurement reality.

706
00:25:01,760 –> 00:25:03,920
Carbon accounting is an identity problem.

707
00:25:03,920 –> 00:25:06,480
If the internal carbon fee is the enforcement mechanism,

708
00:25:06,480 –> 00:25:09,240
carbon accounting is the identity layer underneath it.

709
00:25:09,240 –> 00:25:11,160
And identity layer is always disappointing people

710
00:25:11,160 –> 00:25:13,760
because everyone expects clean answers from messy systems.

711
00:25:13,760 –> 00:25:16,520
Carbon data is not the truth about emissions.

712
00:25:16,520 –> 00:25:18,480
It’s a set of claims about activity.

713
00:25:18,480 –> 00:25:21,240
What happened, where it happened, what was consumed,

714
00:25:21,240 –> 00:25:23,080
and who should be held responsible for it?

715
00:25:23,080 –> 00:25:24,480
That’s an attribution problem.

716
00:25:24,480 –> 00:25:27,240
When attribution problems don’t get solved by good intention,

717
00:25:27,240 –> 00:25:29,120
they get solved by boundaries, definitions,

718
00:25:29,120 –> 00:25:31,160
and relentless refusal to accept close enough

719
00:25:31,160 –> 00:25:34,120
when money and accountability depend on the number.

720
00:25:34,120 –> 00:25:36,520
Scope one and scope two are comparatively sane

721
00:25:36,520 –> 00:25:39,560
because they map to owned assets and purchased electricity.

722
00:25:39,560 –> 00:25:42,560
Vehicles, generators, buildings, utility bills.

723
00:25:42,560 –> 00:25:43,800
The accounting can still be wrong,

724
00:25:43,800 –> 00:25:46,160
but at least the object model is stable.

725
00:25:46,160 –> 00:25:47,760
The company can point to a thing it owns

726
00:25:47,760 –> 00:25:49,360
and a meter it pays for, therefore, the claim

727
00:25:49,360 –> 00:25:51,680
has a fighting chance of being deterministic.

728
00:25:51,680 –> 00:25:54,320
Scope three is where determinism goes to die.

729
00:25:54,320 –> 00:25:57,680
Scope three is dominated by supplier data, logistics assumptions,

730
00:25:57,680 –> 00:25:59,560
and category level emission factors.

731
00:25:59,560 –> 00:26:01,160
It is emissions by proxy.

732
00:26:01,160 –> 00:26:04,040
And that means the carbon control plane becomes probabilistic.

733
00:26:04,040 –> 00:26:06,800
The company estimates models and aggregates values

734
00:26:06,800 –> 00:26:09,120
that are partially observed and often delayed.

735
00:26:09,120 –> 00:26:11,680
It isn’t fake, but it is not the same kind of truth

736
00:26:11,680 –> 00:26:12,840
as a power meter.

737
00:26:12,840 –> 00:26:15,240
That distinction matters because once carbon numbers

738
00:26:15,240 –> 00:26:18,000
become budget-relevant, people will treat them like cost.

739
00:26:18,000 –> 00:26:19,680
They’ll ask for precision, they’ll argue,

740
00:26:19,680 –> 00:26:21,440
they’ll want to route around the charge,

741
00:26:21,440 –> 00:26:23,000
and the system will respond exactly

742
00:26:23,000 –> 00:26:24,840
like any distributed system under load.

743
00:26:24,840 –> 00:26:27,000
It will find the weakest boundary and exploit it.

744
00:26:27,000 –> 00:26:28,200
Here’s the uncomfortable truth.

745
00:26:28,200 –> 00:26:30,160
The moment you start charging teams for carbon,

746
00:26:30,160 –> 00:26:32,840
you create an incentive to fight identity, not behavior.

747
00:26:32,840 –> 00:26:35,080
Teams don’t wake up thinking, how do we emit less?

748
00:26:35,080 –> 00:26:37,520
They wake up thinking, why is our number so high

749
00:26:37,520 –> 00:26:38,920
and why are we paying for it?

750
00:26:38,920 –> 00:26:39,800
That’s rational.

751
00:26:39,800 –> 00:26:40,800
Budgets are finite.

752
00:26:40,800 –> 00:26:43,080
And if the accounting model can’t explain itself,

753
00:26:43,080 –> 00:26:45,560
the program turns into an internal tax dispute,

754
00:26:45,560 –> 00:26:47,520
not an emissions reduction mechanism.

755
00:26:47,520 –> 00:26:49,600
So measurement integrity isn’t a nice to have.

756
00:26:49,600 –> 00:26:50,640
It is the program.

757
00:26:50,640 –> 00:26:52,960
Carbon accounting must do three jobs at once.

758
00:26:52,960 –> 00:26:56,080
First, it has to map activity to an emitting reality,

759
00:26:56,080 –> 00:26:58,520
energy consumption, travel, hardware purchases,

760
00:26:58,520 –> 00:26:59,840
supply chain inputs.

761
00:26:59,840 –> 00:27:02,240
Second, it has to map that reality to a boundary,

762
00:27:02,240 –> 00:27:03,960
what counts inside the company’s claim.

763
00:27:03,960 –> 00:27:06,640
Third, it has to map the boundary to responsibility,

764
00:27:06,640 –> 00:27:09,480
which team owns the decision that caused the impact.

765
00:27:09,480 –> 00:27:11,240
That third one is the identity problem.

766
00:27:11,240 –> 00:27:13,720
In cloud terms, this is like trying to charge back network

767
00:27:13,720 –> 00:27:16,160
agress without knowing which service generated it,

768
00:27:16,160 –> 00:27:19,000
or trying to allocate security incidents to a team

769
00:27:19,000 –> 00:27:22,760
when the system spans shared services, managed identities,

770
00:27:22,760 –> 00:27:24,560
and inherited permissions.

771
00:27:24,560 –> 00:27:26,280
If you can’t tie the event to an owner,

772
00:27:26,280 –> 00:27:27,440
you can’t change behavior.

773
00:27:27,440 –> 00:27:30,600
You can only publish reports and reports don’t fix systems

774
00:27:30,600 –> 00:27:31,720
incentives do.

775
00:27:31,720 –> 00:27:34,640
Now, layer in the distortion effect, people optimize what gets

776
00:27:34,640 –> 00:27:36,440
measured and they ignore what doesn’t.

777
00:27:36,440 –> 00:27:38,560
If the program measures travel emissions well,

778
00:27:38,560 –> 00:27:39,960
travel gets optimized.

779
00:27:39,960 –> 00:27:41,960
If it measures supplier emissions poorly,

780
00:27:41,960 –> 00:27:44,680
supplier emissions become a negotiable story.

781
00:27:44,680 –> 00:27:47,800
And over time, that creates a governance asymmetry.

782
00:27:47,800 –> 00:27:50,200
Teams get punished for the measurable category

783
00:27:50,200 –> 00:27:53,160
and rewarded for moving impact into the unmeasurable one.

784
00:27:53,160 –> 00:27:55,280
That is how carbon programs become theater

785
00:27:55,280 –> 00:27:57,800
without anyone explicitly trying to commit fraud.

786
00:27:57,800 –> 00:28:00,760
The system drifts toward whatever measurement tolerates.

787
00:28:00,760 –> 00:28:02,760
The other distortion is boundary gaming.

788
00:28:02,760 –> 00:28:04,520
If the accounting is based on averages,

789
00:28:04,520 –> 00:28:05,800
teams will chase the average.

790
00:28:05,800 –> 00:28:07,680
If the accounting is based on categories,

791
00:28:07,680 –> 00:28:09,040
teams will debate categories.

792
00:28:09,040 –> 00:28:11,000
If the accounting is lagging by a year,

793
00:28:11,000 –> 00:28:13,360
teams will treat it as historical trivia.

794
00:28:13,360 –> 00:28:15,360
And if the accounting relies on suppliers,

795
00:28:15,360 –> 00:28:17,280
suppliers will learn what answers keep them

796
00:28:17,280 –> 00:28:18,760
in the preferred vendor list.

797
00:28:18,760 –> 00:28:20,480
None of this is cynical speculation.

798
00:28:20,480 –> 00:28:22,160
It is how incentives behave.

799
00:28:22,160 –> 00:28:24,120
So the governance truth is simple.

800
00:28:24,120 –> 00:28:27,000
Accuracy determines whether the carbon fee changes

801
00:28:27,000 –> 00:28:29,840
engineering decisions or becomes another compliance ritual

802
00:28:29,840 –> 00:28:31,200
with a finance wrapper.

803
00:28:31,200 –> 00:28:33,000
The fee amplifies the measurement model.

804
00:28:33,000 –> 00:28:34,480
If the model is defensible,

805
00:28:34,480 –> 00:28:36,600
the organization optimizes reality.

806
00:28:36,600 –> 00:28:38,960
If the model is weak, the organization optimizes

807
00:28:38,960 –> 00:28:39,640
the accounting.

808
00:28:39,640 –> 00:28:42,400
And this is where the carbon control plane starts colliding

809
00:28:42,400 –> 00:28:44,320
with the current era of growth.

810
00:28:44,320 –> 00:28:46,600
Because the workloads being added now are not small,

811
00:28:46,600 –> 00:28:47,800
steady and predictable.

812
00:28:47,800 –> 00:28:49,680
They’re spiky, they’re GPU dense,

813
00:28:49,680 –> 00:28:51,800
they’re tied to regions with grid variability,

814
00:28:51,800 –> 00:28:53,280
they’re supply chain heavy,

815
00:28:53,280 –> 00:28:56,360
they’re AI, AI growth stress tests, the whole model,

816
00:28:56,360 –> 00:29:00,440
attribution, scope three, procurement and regional power.

817
00:29:00,440 –> 00:29:02,520
It forces the company to answer a question

818
00:29:02,520 –> 00:29:03,680
it would rather avoid.

819
00:29:03,680 –> 00:29:05,480
When emissions rise is that drift,

820
00:29:05,480 –> 00:29:07,080
or is that the cost of scale?

821
00:29:07,080 –> 00:29:09,680
Microsoft’s reporting admits the uncomfortable part.

822
00:29:09,680 –> 00:29:11,520
Platform growth changes the baseline.

823
00:29:11,520 –> 00:29:14,040
The carbon control plane has to keep functioning anyway.

824
00:29:14,040 –> 00:29:16,200
That’s next, Microsoft’s emissions trend,

825
00:29:16,200 –> 00:29:18,120
the platform grows, the carbon grows.

826
00:29:18,120 –> 00:29:20,280
Microsoft’s public sustainability reporting

827
00:29:20,280 –> 00:29:23,400
creates an awkward moment that most companies try to avoid.

828
00:29:23,400 –> 00:29:26,000
It forces the organization to say what happened,

829
00:29:26,000 –> 00:29:27,760
not what it wished happened.

830
00:29:27,760 –> 00:29:29,520
And what happened in the disclosed totals

831
00:29:29,520 –> 00:29:32,200
is that emissions increased from 12 million TCO2e

832
00:29:32,200 –> 00:29:35,440
in 202 to 17 million TCO2e in 2023,

833
00:29:35,440 –> 00:29:38,400
driven by the buildout required for AI and data center growth.

834
00:29:38,400 –> 00:29:40,920
That number matters, not because it proves virtue or fraud,

835
00:29:40,920 –> 00:29:43,000
but because it exposes the system collision point.

836
00:29:43,000 –> 00:29:45,360
The carbon control plane is trying to impose constraints

837
00:29:45,360 –> 00:29:47,800
while the product strategy is trying to scale capacity.

838
00:29:47,800 –> 00:29:49,880
Those objectives can align over long horizons,

839
00:29:49,880 –> 00:29:52,680
but in the short run they fight like rival schedulers.

840
00:29:52,680 –> 00:29:54,680
Most listeners here emissions rose

841
00:29:54,680 –> 00:29:58,160
and immediately reach for the easiest story hypocrisy.

842
00:29:58,160 –> 00:29:59,440
It’s the comfortable narrative

843
00:29:59,440 –> 00:30:01,880
because it removes the need to understand the mechanics,

844
00:30:01,880 –> 00:30:03,560
but hypocrisy is a motive claim.

845
00:30:03,560 –> 00:30:06,000
The more accurate diagnosis is architectural.

846
00:30:06,000 –> 00:30:08,440
The platform grew, therefore the footprint grew

847
00:30:08,440 –> 00:30:10,000
because infrastructure is physical

848
00:30:10,000 –> 00:30:11,760
and AI is not a rounding error.

849
00:30:11,760 –> 00:30:14,000
If you want to understand why this happens,

850
00:30:14,000 –> 00:30:17,000
stop thinking in annual goals and start thinking in multipliers.

851
00:30:17,000 –> 00:30:18,560
AI doesn’t just add more workloads,

852
00:30:18,560 –> 00:30:20,400
it changes the shape of the infrastructure,

853
00:30:20,400 –> 00:30:22,920
higher power density, more specialized hardware

854
00:30:22,920 –> 00:30:24,800
and faster hardware replacement cycles

855
00:30:24,800 –> 00:30:26,480
driven by performance demands

856
00:30:26,480 –> 00:30:29,680
that pulls the emissions profile in two directions at once.

857
00:30:29,680 –> 00:30:30,960
Operational emissions rise

858
00:30:30,960 –> 00:30:33,440
because data centers consume more electricity,

859
00:30:33,440 –> 00:30:34,640
scope three emissions rise

860
00:30:34,640 –> 00:30:37,040
because the supply chain for hardware scales up

861
00:30:37,040 –> 00:30:39,920
and supply chains do not decarbonize on your timeline.

862
00:30:39,920 –> 00:30:42,800
So the platform doesn’t fail sustainability by growing.

863
00:30:42,800 –> 00:30:45,840
It reveals the constraint that sustainability claims

864
00:30:45,840 –> 00:30:47,920
were always trying to negotiate.

865
00:30:47,920 –> 00:30:50,960
Growth creates emissions now and reductions take time.

866
00:30:50,960 –> 00:30:52,440
That means the carbon negative target

867
00:30:52,440 –> 00:30:54,680
isn’t a simple reduced to zero narrative.

868
00:30:54,680 –> 00:30:57,600
It is a balancing act across reductions, renewables

869
00:30:57,600 –> 00:31:00,560
and removals under pressure from customer demand

870
00:31:00,560 –> 00:31:01,960
and competitive urgency.

871
00:31:01,960 –> 00:31:03,960
This is where the internal carbon fee

872
00:31:03,960 –> 00:31:05,920
stops looking like a moral program

873
00:31:05,920 –> 00:31:07,480
and starts looking like what it is,

874
00:31:07,480 –> 00:31:10,600
a throttling mechanism that tries to make growth more expensive

875
00:31:10,600 –> 00:31:12,400
in the dimensions that matter.

876
00:31:12,400 –> 00:31:15,040
But throttles don’t stop growth if the business can afford them.

877
00:31:15,040 –> 00:31:17,000
They just change how the growth is shaped.

878
00:31:17,000 –> 00:31:19,200
They move workloads, they change purchasing decisions,

879
00:31:19,200 –> 00:31:21,400
they create internal fights about efficiency

880
00:31:21,400 –> 00:31:22,560
that otherwise never happen

881
00:31:22,560 –> 00:31:24,160
and that’s the first uncomfortable lesson

882
00:31:24,160 –> 00:31:25,960
for any admin listening.

883
00:31:25,960 –> 00:31:28,000
A carbon control plane is not designed

884
00:31:28,000 –> 00:31:30,760
to make emissions always go down year over year.

885
00:31:30,760 –> 00:31:33,480
It is designed to ensure that when emissions go up,

886
00:31:33,480 –> 00:31:34,960
it happens with eyes open

887
00:31:34,960 –> 00:31:36,480
and with explicit trade-offs.

888
00:31:36,480 –> 00:31:38,240
That sounds like a low bar, it isn’t.

889
00:31:38,240 –> 00:31:39,800
Most organizations don’t even get to

890
00:31:39,800 –> 00:31:41,600
we understand why this happened.

891
00:31:41,600 –> 00:31:43,400
They get to the report is due.

892
00:31:43,400 –> 00:31:45,920
Now this is also where the narrative gets sharper.

893
00:31:45,920 –> 00:31:48,280
If Microsoft’s emissions rose into 2023,

894
00:31:48,280 –> 00:31:50,040
how does the company still claim a path

895
00:31:50,040 –> 00:31:51,800
to carbon negative by 203?

896
00:31:51,800 –> 00:31:53,800
The answer isn’t magic, it’s removals.

897
00:31:53,800 –> 00:31:58,360
Microsoft’s disclosed reporting has already conditioned the audience

898
00:31:58,360 –> 00:32:00,520
to the idea that the 203 target depends

899
00:32:00,520 –> 00:32:02,560
on scaling carbon removal procurement

900
00:32:02,560 –> 00:32:04,680
to handle what reductions can’t.

901
00:32:04,680 –> 00:32:06,440
In their own framing, they anticipate

902
00:32:06,440 –> 00:32:09,040
needing single digit millions of removal credits annually

903
00:32:09,040 –> 00:32:09,840
by 203.

904
00:32:09,840 –> 00:32:12,800
That detail matters because it exposes a dependency.

905
00:32:12,800 –> 00:32:14,840
The carbon negative claim is no longer only

906
00:32:14,840 –> 00:32:15,920
an engineering target.

907
00:32:15,920 –> 00:32:17,280
It is also a market dependency

908
00:32:17,280 –> 00:32:18,800
on the availability and quality

909
00:32:18,800 –> 00:32:20,560
of removal credits at scale.

910
00:32:20,560 –> 00:32:22,360
And that’s the second uncomfortable lesson.

911
00:32:22,360 –> 00:32:24,920
Sustainability targets don’t just depend on your behavior.

912
00:32:24,920 –> 00:32:26,760
They depend on markets you don’t control.

913
00:32:26,760 –> 00:32:28,480
You can build efficient data centers.

914
00:32:28,480 –> 00:32:30,480
You can negotiate renewable energy contracts.

915
00:32:30,480 –> 00:32:31,880
You can pressure suppliers.

916
00:32:31,880 –> 00:32:33,840
But if your net claim depends on removals,

917
00:32:33,840 –> 00:32:36,240
then you inherit the risks of a young, stressed,

918
00:32:36,240 –> 00:32:38,840
and politically contested carbon removal market.

919
00:32:38,840 –> 00:32:40,240
Not because Microsoft is unique,

920
00:32:40,240 –> 00:32:43,720
but because net claims at hyper scale require flows

921
00:32:43,720 –> 00:32:46,160
that the world hasn’t historically produced in bulk.

922
00:32:46,160 –> 00:32:47,920
So emissions rising isn’t the scandal.

923
00:32:47,920 –> 00:32:48,960
It’s the symptom.

924
00:32:48,960 –> 00:32:51,680
The real story is that the unit economics of compute

925
00:32:51,680 –> 00:32:54,000
are colliding with the unit economics of carbon.

926
00:32:54,000 –> 00:32:55,640
The platform wants to scale faster

927
00:32:55,640 –> 00:32:57,560
than the decarbonization stack can keep up.

928
00:32:57,560 –> 00:32:59,560
Therefore, the control plane has to extend

929
00:32:59,560 –> 00:33:02,680
into procurement, finance, and removal contracts

930
00:33:02,680 –> 00:33:04,560
just to keep the claim coherent.

931
00:33:04,560 –> 00:33:07,200
This is the open loop you should keep in mind as we move forward.

932
00:33:07,200 –> 00:33:08,600
When growth drives emissions up,

933
00:33:08,600 –> 00:33:10,520
the only way to hit a net target on schedule

934
00:33:10,520 –> 00:33:11,760
is to buy negative flow.

935
00:33:11,760 –> 00:33:12,920
That’s not a moral judgment.

936
00:33:12,920 –> 00:33:14,520
It’s an architectural consequence.

937
00:33:14,520 –> 00:33:16,840
And it puts the next battleground in the right place.

938
00:33:16,840 –> 00:33:19,360
Removal procurement becomes part of the control plane

939
00:33:19,360 –> 00:33:20,680
because it’s the only lever left

940
00:33:20,680 –> 00:33:22,520
that can move fast enough to compensate

941
00:33:22,520 –> 00:33:24,120
for growth that already happened.

942
00:33:24,120 –> 00:33:27,920
Carbon removal, procurement, buying time, buying risk.

943
00:33:27,920 –> 00:33:30,120
So now the carbon control plane extends

944
00:33:30,120 –> 00:33:32,880
into the least comfortable domain for engineers, procurement.

945
00:33:32,880 –> 00:33:34,680
When Microsoft buys carbon removal,

946
00:33:34,680 –> 00:33:36,160
it isn’t doing something symbolic.

947
00:33:36,160 –> 00:33:39,520
It’s acquiring a negative flow to counterbalance positive flows.

948
00:33:39,520 –> 00:33:41,200
It couldn’t eliminate fast enough.

949
00:33:41,200 –> 00:33:43,680
That’s what carbon negative becomes at scale,

950
00:33:43,680 –> 00:33:46,920
reduction where you can and removal where you can’t.

951
00:33:46,920 –> 00:33:49,320
The spreadsheet doesn’t care which one feels better.

952
00:33:49,320 –> 00:33:50,880
The spreadsheet just reconciles.

953
00:33:50,880 –> 00:33:53,680
This is where a lot of audiences get emotionally confused.

954
00:33:53,680 –> 00:33:55,040
They hear buying removals

955
00:33:55,040 –> 00:33:57,280
and assume the company is outsourcing guilt.

956
00:33:57,280 –> 00:33:59,360
Sometimes that’s true in the broader market.

957
00:33:59,360 –> 00:34:02,200
But in Microsoft’s case, their own sustainability disclosures

958
00:34:02,200 –> 00:34:03,880
make the dependency explicit.

959
00:34:03,880 –> 00:34:05,840
If emissions rise while the platform scales,

960
00:34:05,840 –> 00:34:08,360
the net claim requires removals at meaningful volume.

961
00:34:08,360 –> 00:34:09,480
That’s not a PR choice.

962
00:34:09,480 –> 00:34:10,920
That’s an arithmetic requirement.

963
00:34:10,920 –> 00:34:14,000
And Microsoft has been behaving like a buyer, not a commentator.

964
00:34:14,000 –> 00:34:18,120
Publicly reported deals show Microsoft securing large quantities

965
00:34:18,120 –> 00:34:19,480
of removal credits.

966
00:34:19,480 –> 00:34:23,120
In early 2006, Microsoft signed for nearly five million tons

967
00:34:23,120 –> 00:34:24,960
of removal credits across two structures,

968
00:34:24,960 –> 00:34:27,440
about two million tons of aforestation,

969
00:34:27,440 –> 00:34:30,720
reforestation and re-vegetation credits over nine years,

970
00:34:30,720 –> 00:34:34,520
via Rubicon Carbon, source through Kejani Forestry in Uganda,

971
00:34:34,520 –> 00:34:37,760
and about 2.85 million tons of soil-based credits

972
00:34:37,760 –> 00:34:39,600
over 12 years through Indigo Ag.

973
00:34:39,600 –> 00:34:41,960
Those numbers aren’t useful because they sound big.

974
00:34:41,960 –> 00:34:43,760
They’re useful because they tell you Microsoft

975
00:34:43,760 –> 00:34:46,200
is trying to lock supply years in advance.

976
00:34:46,200 –> 00:34:48,360
That detail matters.

977
00:34:48,360 –> 00:34:50,840
When a company signs multi-year removal contracts,

978
00:34:50,840 –> 00:34:53,280
it’s admitting a structural truth by two or three

979
00:34:53,280 –> 00:34:55,480
removals aren’t an emergency lever.

980
00:34:55,480 –> 00:34:56,760
They’re part of the operating model.

981
00:34:56,760 –> 00:34:59,360
You don’t buy that way unless you expect to need the flow

982
00:34:59,360 –> 00:35:01,320
predictably year after year.

983
00:35:01,320 –> 00:35:03,560
Microsoft has also described itself as purchasing

984
00:35:03,560 –> 00:35:07,440
more than 10 million tons of durable removals recently.

985
00:35:07,440 –> 00:35:09,360
Again, the important point is not the headline.

986
00:35:09,360 –> 00:35:10,880
The point is the architecture.

987
00:35:10,880 –> 00:35:13,520
They are building a portfolio, not placing one-time bets.

988
00:35:13,520 –> 00:35:15,280
A portfolio is a risk management tool.

989
00:35:15,280 –> 00:35:18,200
It exists because no single project type can carry

990
00:35:18,200 –> 00:35:20,880
the whole claim safely and safely is the keyword

991
00:35:20,880 –> 00:35:22,600
because removal procurement buys time,

992
00:35:22,600 –> 00:35:23,960
but it also buys risk.

993
00:35:23,960 –> 00:35:25,320
The first risk is quality.

994
00:35:25,320 –> 00:35:28,240
Carbon removal credits aren’t a single commodity.

995
00:35:28,240 –> 00:35:29,200
They have properties,

996
00:35:29,200 –> 00:35:32,600
additionality, permanence, verification, rigor, leakage,

997
00:35:32,600 –> 00:35:33,920
and reversal risk.

998
00:35:33,920 –> 00:35:35,840
If those words feel like ESG jargon

999
00:35:35,840 –> 00:35:37,920
translate them into system behavior,

1000
00:35:37,920 –> 00:35:41,040
additionality asks whether the negative flow exists

1001
00:35:41,040 –> 00:35:43,840
because you paid for it or whether it would have happened anyway.

1002
00:35:43,840 –> 00:35:46,080
Permanence asks whether the carbon stays stored

1003
00:35:46,080 –> 00:35:47,960
or whether it returns to the atmosphere later.

1004
00:35:47,960 –> 00:35:50,640
Verification asks whether the measurement is auditable

1005
00:35:50,640 –> 00:35:52,640
or whether it’s a narrative with spreadsheets.

1006
00:35:52,640 –> 00:35:54,480
This is the same problem class as identity

1007
00:35:54,480 –> 00:35:56,720
and carbon accounting, just in a different layer.

1008
00:35:56,720 –> 00:35:58,920
You’re still dealing with claims about reality.

1009
00:35:58,920 –> 00:36:01,080
You’re just dealing with them through counterparties,

1010
00:36:01,080 –> 00:36:03,880
methodologies and contracts instead of internal meters.

1011
00:36:03,880 –> 00:36:05,560
The second risk is dependency.

1012
00:36:05,560 –> 00:36:08,080
Once removals become material to your net claim,

1013
00:36:08,080 –> 00:36:11,960
you inherit market risk, price volatility, supply constraints,

1014
00:36:11,960 –> 00:36:15,120
methodology changes, and reputational exposure

1015
00:36:15,120 –> 00:36:17,360
when the public learns that net depends

1016
00:36:17,360 –> 00:36:19,560
on purchased negative flow.

1017
00:36:19,560 –> 00:36:21,520
Even if your procurement is defensible,

1018
00:36:21,520 –> 00:36:24,800
your story becomes hostage to how the broader market behaves

1019
00:36:24,800 –> 00:36:27,080
and the voluntary carbon market has never been famous

1020
00:36:27,080 –> 00:36:28,120
for boring stability.

1021
00:36:28,120 –> 00:36:30,880
The third risk is moral hazard inside the organization.

1022
00:36:30,880 –> 00:36:33,240
If teams believe removals will always be available,

1023
00:36:33,240 –> 00:36:37,000
cheap and politically acceptable, reductions lose urgency.

1024
00:36:37,000 –> 00:36:38,960
The control plane can drift into a pattern

1025
00:36:38,960 –> 00:36:41,520
where growth happens first and cleanup gets purchased later.

1026
00:36:41,520 –> 00:36:42,800
That’s not because people are evil.

1027
00:36:42,800 –> 00:36:45,240
It’s because incentives always pick the fastest path

1028
00:36:45,240 –> 00:36:46,960
unless you enforce a slower, better one.

1029
00:36:46,960 –> 00:36:49,000
So removals have to be framed correctly.

1030
00:36:49,000 –> 00:36:51,000
According to many sustainability experts,

1031
00:36:51,000 –> 00:36:54,520
removals are a last mile mechanism, not a primary strategy.

1032
00:36:54,520 –> 00:36:56,680
That framing exists because reduction changes

1033
00:36:56,680 –> 00:36:59,280
the underlying system while removals compensate

1034
00:36:59,280 –> 00:37:00,080
for what remains.

1035
00:37:00,080 –> 00:37:02,240
If you reverse that order, you build a company

1036
00:37:02,240 –> 00:37:05,120
that can only stay net by continually buying its way

1037
00:37:05,120 –> 00:37:05,880
out of physics.

1038
00:37:05,880 –> 00:37:07,200
That works until it doesn’t.

1039
00:37:07,200 –> 00:37:09,760
And Microsoft is operating right on that boundary.

1040
00:37:09,760 –> 00:37:11,760
They’re scaling AI and cloud capacity,

1041
00:37:11,760 –> 00:37:13,800
therefore emissions pressure rises.

1042
00:37:13,800 –> 00:37:15,440
They’re also building internal incentives

1043
00:37:15,440 –> 00:37:18,280
like the carbon fee and publishing annual disclosures

1044
00:37:18,280 –> 00:37:19,960
then they bridge the gap with removals.

1045
00:37:19,960 –> 00:37:22,560
That’s the carbon control plane in its full form.

1046
00:37:22,560 –> 00:37:24,480
Measurement, enforcement, and procurement

1047
00:37:24,480 –> 00:37:27,960
stitched into one coherent story that can survive and audit.

1048
00:37:27,960 –> 00:37:30,000
But coherence isn’t the same as victory.

1049
00:37:30,000 –> 00:37:31,840
Removal procurement is the part of the system

1050
00:37:31,840 –> 00:37:34,520
that can keep the promise alive while the platform grows.

1051
00:37:34,520 –> 00:37:35,920
It’s also the part that introduces

1052
00:37:35,920 –> 00:37:38,880
the most external dependency and the most fragile trust surface.

1053
00:37:38,880 –> 00:37:41,120
And now we can move from corporate strategy

1054
00:37:41,120 –> 00:37:43,040
to what actually matters for admins,

1055
00:37:43,040 –> 00:37:45,640
how those incentives change infrastructure decisions

1056
00:37:45,640 –> 00:37:47,200
inside the machine.

1057
00:37:47,200 –> 00:37:48,760
Case study, one.

1058
00:37:48,760 –> 00:37:52,160
Internal carbon fee applied to cloud consumption behavior.

1059
00:37:52,160 –> 00:37:53,600
Take the internal carbon fee

1060
00:37:53,600 –> 00:37:55,840
and stop treating it like a sustainability program.

1061
00:37:55,840 –> 00:37:57,480
Treat it like a behavioral instrument

1062
00:37:57,480 –> 00:37:59,320
applied to cloud operations.

1063
00:37:59,320 –> 00:38:01,760
Inside Microsoft teams don’t just use Azure.

1064
00:38:01,760 –> 00:38:04,280
They build services on top of shared infrastructure.

1065
00:38:04,280 –> 00:38:06,560
They choose regions, they choose redundancy patterns.

1066
00:38:06,560 –> 00:38:09,280
They choose whether something runs 24/7 or only when

1067
00:38:09,280 –> 00:38:10,120
there’s demand.

1068
00:38:10,120 –> 00:38:13,240
And when a carbon fee exists, those choices stop being silent.

1069
00:38:13,240 –> 00:38:14,720
They become financially loud.

1070
00:38:14,720 –> 00:38:15,840
This is the core mechanic.

1071
00:38:15,840 –> 00:38:18,640
You attach an internal cost to emissions tied to operations

1072
00:38:18,640 –> 00:38:20,640
and services, therefore engineers and product owners

1073
00:38:20,640 –> 00:38:23,360
feel carbon the same way they feel cloud spend.

1074
00:38:23,360 –> 00:38:25,560
Not because they suddenly became climate philosophers

1075
00:38:25,560 –> 00:38:28,600
because their budget now has a line item that punishes waste.

1076
00:38:28,600 –> 00:38:30,000
Now here’s where most people mess up

1077
00:38:30,000 –> 00:38:31,560
when they try to learn from this.

1078
00:38:31,560 –> 00:38:34,400
They assume the carbon fee makes people turn things off.

1079
00:38:34,400 –> 00:38:37,120
Sometimes it does, but the bigger impact is that it changes

1080
00:38:37,120 –> 00:38:38,360
how teams argue.

1081
00:38:38,360 –> 00:38:40,880
Without a carbon price, cloud conversations default

1082
00:38:40,880 –> 00:38:43,520
to performance, reliability, and cost.

1083
00:38:43,520 –> 00:38:45,360
Carbon sits outside the trade space, which

1084
00:38:45,360 –> 00:38:46,920
means it loses every time.

1085
00:38:46,920 –> 00:38:49,240
With a carbon price, carbon enters the trade space

1086
00:38:49,240 –> 00:38:51,160
as a constraint with consequences.

1087
00:38:51,160 –> 00:38:52,200
The meeting changes.

1088
00:38:52,200 –> 00:38:53,200
The backlog changes.

1089
00:38:53,200 –> 00:38:54,760
The architecture review changes.

1090
00:38:54,760 –> 00:38:56,880
Start with region choice because it’s the simplest place

1091
00:38:56,880 –> 00:38:57,800
to see the mechanism.

1092
00:38:57,800 –> 00:39:00,880
Engineers pick regions for latency, data residency,

1093
00:39:00,880 –> 00:39:03,000
resiliency, and sometimes pure habit.

1094
00:39:03,000 –> 00:39:05,400
But carbon exposure differs by where compute runs

1095
00:39:05,400 –> 00:39:06,840
because the grid behind that region

1096
00:39:06,840 –> 00:39:09,040
has a different carbon intensity profile

1097
00:39:09,040 –> 00:39:11,560
and because the supporting infrastructure differs.

1098
00:39:11,560 –> 00:39:14,840
If the internal accounting model attributes emissions

1099
00:39:14,840 –> 00:39:18,080
by region and workload, then region becomes more

1100
00:39:18,080 –> 00:39:18,960
than a map pin.

1101
00:39:18,960 –> 00:39:20,160
It becomes a cost decision.

1102
00:39:20,160 –> 00:39:21,840
And that’s the first uncomfortable question

1103
00:39:21,840 –> 00:39:23,440
in a carbon control plane forces.

1104
00:39:23,440 –> 00:39:25,960
If two regions satisfy the technical requirements,

1105
00:39:25,960 –> 00:39:28,840
why are you deploying to the higher emissions one?

1106
00:39:28,840 –> 00:39:30,960
In most organizations, nobody can answer that

1107
00:39:30,960 –> 00:39:32,560
because nobody even asked it.

1108
00:39:32,560 –> 00:39:33,960
The system never made them.

1109
00:39:33,960 –> 00:39:35,320
Next, workload shape.

1110
00:39:35,320 –> 00:39:37,680
Most teams still deploy like it’s 2012,

1111
00:39:37,680 –> 00:39:40,120
always on by default, oversized, just in case,

1112
00:39:40,120 –> 00:39:43,000
and scaled for peak even when peak happens twice a week.

1113
00:39:43,000 –> 00:39:43,960
It feels safe.

1114
00:39:43,960 –> 00:39:46,000
It also burns electricity continuously.

1115
00:39:46,000 –> 00:39:48,160
When carbon behaves like cost, always on becomes

1116
00:39:48,160 –> 00:39:50,280
an explicit tax you pay every month.

1117
00:39:50,280 –> 00:39:52,840
Elastic patterns, auto scaling, Q-based designs,

1118
00:39:52,840 –> 00:39:54,920
turning off non-production resources,

1119
00:39:54,920 –> 00:39:58,800
scheduling batch jobs, stop being nice to have optimizations.

1120
00:39:58,800 –> 00:40:00,000
They become budget defense.

1121
00:40:00,000 –> 00:40:02,520
This is where carbon governance quietly aligns with fin ops.

1122
00:40:02,520 –> 00:40:03,960
The goal is not morality.

1123
00:40:03,960 –> 00:40:05,520
The goal is efficiency under constraint.

1124
00:40:05,520 –> 00:40:07,840
Now add compute intensity and service selection.

1125
00:40:07,840 –> 00:40:10,680
Different workload architectures have different energy profiles

1126
00:40:10,680 –> 00:40:13,320
because they drive different utilization patterns.

1127
00:40:13,320 –> 00:40:16,720
A system that turns through constant retries, oversized instances,

1128
00:40:16,720 –> 00:40:19,440
and inefficient data movement doesn’t just cost more money.

1129
00:40:19,440 –> 00:40:22,840
It also costs more carbon because wasted compute is still compute.

1130
00:40:22,840 –> 00:40:25,280
The internal fee makes that visible in a way

1131
00:40:25,280 –> 00:40:26,680
performance metrics never will.

1132
00:40:26,680 –> 00:40:29,240
The thing most people miss is that carbon exposure

1133
00:40:29,240 –> 00:40:31,520
isn’t just how much compute you use.

1134
00:40:31,520 –> 00:40:33,600
It’s when, where, and how you use it.

1135
00:40:33,600 –> 00:40:37,400
A workload running at 10% utilization all month is a carbon leak.

1136
00:40:37,400 –> 00:40:40,720
A workload running at 80% utilization for a shorter window

1137
00:40:40,720 –> 00:40:42,240
can be materially different.

1138
00:40:42,240 –> 00:40:44,280
Even if the total capacity provisioned

1139
00:40:44,280 –> 00:40:45,800
looks similar in a diagram.

1140
00:40:45,800 –> 00:40:47,600
The fee pushes teams toward designs

1141
00:40:47,600 –> 00:40:50,480
that match consumption to demand because demand mismatch

1142
00:40:50,480 –> 00:40:52,120
becomes billable emissions.

1143
00:40:52,120 –> 00:40:54,840
And then there’s availability architecture, the sacred cow.

1144
00:40:54,840 –> 00:40:56,800
Multi-region deployments improve resilience.

1145
00:40:56,800 –> 00:40:58,680
They also duplicate infrastructure.

1146
00:40:58,680 –> 00:41:01,640
A carbon price doesn’t tell you to abandon resiliency.

1147
00:41:01,640 –> 00:41:04,000
It forces you to stop treating resiliency as free.

1148
00:41:04,000 –> 00:41:06,200
If a service runs active active across regions

1149
00:41:06,200 –> 00:41:09,040
that redundancy has a footprint, if it runs active passive

1150
00:41:09,040 –> 00:41:11,400
with cold standby, that footprint changes.

1151
00:41:11,400 –> 00:41:13,160
If it uses data replication patterns

1152
00:41:13,160 –> 00:41:15,960
that increase constant data movement that has a footprint too,

1153
00:41:15,960 –> 00:41:18,760
the carbon control plane turns those into explicit trade-offs

1154
00:41:18,760 –> 00:41:20,320
instead of invisible defaults.

1155
00:41:20,320 –> 00:41:22,000
You still choose the resiliency level.

1156
00:41:22,000 –> 00:41:25,200
You just stop pretending it’s costless in every dimension.

1157
00:41:25,200 –> 00:41:28,160
This clicked for a lot of teams only after they saw carbon

1158
00:41:28,160 –> 00:41:30,760
show up next to cost in the same financial conversation

1159
00:41:30,760 –> 00:41:32,920
because that’s what makes the decision uncomfortable

1160
00:41:32,920 –> 00:41:33,760
in the right places.

1161
00:41:33,760 –> 00:41:35,200
It doesn’t ask the engineer to care.

1162
00:41:35,200 –> 00:41:36,920
It asks the engineer to justify.

1163
00:41:36,920 –> 00:41:39,280
So the practical lens for admins listening is simple.

1164
00:41:39,280 –> 00:41:41,680
Treat carbon as a non-functional requirement,

1165
00:41:41,680 –> 00:41:44,360
alongside latency, security, and resiliency.

1166
00:41:44,360 –> 00:41:47,440
Not a report, a requirement.

1167
00:41:47,440 –> 00:41:50,600
And once you do that, you get a transferable method.

1168
00:41:50,600 –> 00:41:53,320
Any workload decision you can already model financially,

1169
00:41:53,320 –> 00:41:55,600
you can also model for emissions.

1170
00:41:55,600 –> 00:41:57,600
If your accounting boundary is good enough.

1171
00:41:57,600 –> 00:42:00,400
Now, before we go further, remember the open loop.

1172
00:42:00,400 –> 00:42:03,120
If traditional cloud workloads make these trade-offs visible,

1173
00:42:03,120 –> 00:42:05,680
AI workloads amplify the whole problem.

1174
00:42:05,680 –> 00:42:07,880
They increase power density, they increase scope

1175
00:42:07,880 –> 00:42:09,440
three pressure through hardware,

1176
00:42:09,440 –> 00:42:12,160
and they make efficient by default harder.

1177
00:42:12,160 –> 00:42:13,800
That’s why the next case study matters.

1178
00:42:13,800 –> 00:42:15,680
AI doesn’t just consume more.

1179
00:42:15,680 –> 00:42:17,200
It changes what the platform is.

1180
00:42:17,200 –> 00:42:20,480
Case study two, AI workload energy budgeting,

1181
00:42:20,480 –> 00:42:22,880
training versus inference reality.

1182
00:42:22,880 –> 00:42:26,960
AI creates a special kind of organizational self-deception.

1183
00:42:26,960 –> 00:42:29,480
Because it lives behind APIs and notebooks,

1184
00:42:29,480 –> 00:42:32,280
people treat it like it’s just software.

1185
00:42:32,280 –> 00:42:33,000
It isn’t.

1186
00:42:33,000 –> 00:42:34,920
It’s industrial compute with a user interface.

1187
00:42:34,920 –> 00:42:37,320
And if you want to understand why Microsoft’s carbon control

1188
00:42:37,320 –> 00:42:40,800
plane gets so stressed by AI, you need one clean distinction

1189
00:42:40,800 –> 00:42:44,240
that most teams refuse to make until finance forces them to.

1190
00:42:44,240 –> 00:42:46,440
Training is not inference.

1191
00:42:46,440 –> 00:42:49,360
Training is the one time or occasional capital expense.

1192
00:42:49,360 –> 00:42:52,320
Weeks of dense compute to produce a model artifact.

1193
00:42:52,320 –> 00:42:53,920
Infrains is the operating expense.

1194
00:42:53,920 –> 00:42:56,800
Every prompt, every completion, every embedding,

1195
00:42:56,800 –> 00:43:00,440
every classification repeated forever at production scale.

1196
00:43:00,440 –> 00:43:03,040
Teams obsess over training because it’s dramatic.

1197
00:43:03,040 –> 00:43:05,120
But the business footprint often lives in inference

1198
00:43:05,120 –> 00:43:06,920
because that’s where usage compounds.

1199
00:43:06,920 –> 00:43:09,120
This separation matters because the optimization

1200
00:43:09,120 –> 00:43:10,280
leaves us differ.

1201
00:43:10,280 –> 00:43:13,360
Training is where you can choose architecture, data set strategy,

1202
00:43:13,360 –> 00:43:15,920
precision, check pointing, and stopping criteria.

1203
00:43:15,920 –> 00:43:17,280
You can decide whether you actually

1204
00:43:17,280 –> 00:43:19,440
need a frontier scale run or whether you’re

1205
00:43:19,440 –> 00:43:22,520
doing ego-driven experimentation that produces heat.

1206
00:43:22,520 –> 00:43:24,480
Infrains is where you can choose caching, batching,

1207
00:43:24,480 –> 00:43:27,120
routing, context length discipline, and model selection.

1208
00:43:27,120 –> 00:43:30,200
Infrains is where good enough becomes an environmental strategy

1209
00:43:30,200 –> 00:43:32,640
because every unnecessary token becomes a tiny text

1210
00:43:32,640 –> 00:43:34,440
that repeats millions of times.

1211
00:43:34,440 –> 00:43:36,960
So the case study is not AI is energy hungry.

1212
00:43:36,960 –> 00:43:38,160
Everyone knows that vaguely.

1213
00:43:38,160 –> 00:43:39,600
The case study is this.

1214
00:43:39,600 –> 00:43:42,480
Microsoft’s sustainability posture forces the organization

1215
00:43:42,480 –> 00:43:44,840
to treat AI like any other constrained resource,

1216
00:43:44,840 –> 00:43:48,320
not free because it’s cloud, constrained because it’s physical.

1217
00:43:48,320 –> 00:43:51,680
That’s why carbon budgeting per AI workload becomes inevitable,

1218
00:43:51,680 –> 00:43:55,440
not as a moral gesture, as capacity governance.

1219
00:43:55,440 –> 00:43:58,040
Most organizations already do cost budgeting.

1220
00:43:58,040 –> 00:44:00,440
This model can’t exceed $x per month.

1221
00:44:00,440 –> 00:44:01,040
That’s familiar.

1222
00:44:01,040 –> 00:44:03,680
It’s also incomplete because cost is a market signal

1223
00:44:03,680 –> 00:44:05,200
and carbon is a physical signal.

1224
00:44:05,200 –> 00:44:07,200
They can correlate, but they’re not identical.

1225
00:44:07,200 –> 00:44:09,720
Power prices, contractual discounts, region choice,

1226
00:44:09,720 –> 00:44:12,240
and utilization patterns can make cost look acceptable

1227
00:44:12,240 –> 00:44:13,720
while carbon stays ugly.

1228
00:44:13,720 –> 00:44:15,960
So carbon budgeting is simply the same discipline

1229
00:44:15,960 –> 00:44:17,800
applied to a different constraint.

1230
00:44:17,800 –> 00:44:20,840
Set an emissions budget per model or per workload class,

1231
00:44:20,840 –> 00:44:23,520
then enforce it with the same seriousness as spend.

1232
00:44:23,520 –> 00:44:25,880
Here’s the practical baseline formula teams use

1233
00:44:25,880 –> 00:44:27,960
when they want to stop hand waving.

1234
00:44:27,960 –> 00:44:30,000
Carbon emissions equal compute hours

1235
00:44:30,000 –> 00:44:32,600
times power draw times data center efficiency

1236
00:44:32,600 –> 00:44:35,000
multiplied by grid carbon intensity.

1237
00:44:35,000 –> 00:44:37,640
Compute hours times power draw gives you energy.

1238
00:44:37,640 –> 00:44:40,160
Data center efficiency gets represented by PUE

1239
00:44:40,160 –> 00:44:42,400
because the data center consumes extra power

1240
00:44:42,400 –> 00:44:43,640
for cooling and overhead.

1241
00:44:43,640 –> 00:44:45,680
Grid carbon intensity converts electricity

1242
00:44:45,680 –> 00:44:48,160
into emissions based on where that electricity came from.

1243
00:44:48,160 –> 00:44:49,240
That’s not a perfect model.

1244
00:44:49,240 –> 00:44:51,680
But it’s a useful one because it turns AI is bad

1245
00:44:51,680 –> 00:44:54,160
for the climate into numbers you can argue with

1246
00:44:54,160 –> 00:44:55,920
and numbers are where governance begins.

1247
00:44:55,920 –> 00:44:58,520
This is also where training versus inference

1248
00:44:58,520 –> 00:45:01,560
becomes a budget conversation instead of a technical footnote.

1249
00:45:01,560 –> 00:45:03,560
Training budgets can be managed as projects.

1250
00:45:03,560 –> 00:45:04,960
The run starts, it ends,

1251
00:45:04,960 –> 00:45:06,320
you review the footprint.

1252
00:45:06,320 –> 00:45:08,800
Infrains budgets must be managed as product behavior,

1253
00:45:08,800 –> 00:45:11,800
usage patterns, user experience choices, defaults,

1254
00:45:11,800 –> 00:45:12,480
and throttles.

1255
00:45:12,480 –> 00:45:14,800
If you let product design treat inference

1256
00:45:14,800 –> 00:45:16,080
like an infinite sink,

1257
00:45:16,080 –> 00:45:18,640
the carbon control plane becomes a spectator sport.

1258
00:45:18,640 –> 00:45:20,520
This is where the industry’s measurement standards

1259
00:45:20,520 –> 00:45:23,440
start to matter and not in a nerdy way.

1260
00:45:23,440 –> 00:45:25,120
The software carbon intensity approach

1261
00:45:25,120 –> 00:45:27,520
exists to standardize how software workloads

1262
00:45:27,520 –> 00:45:28,800
quantify emissions.

1263
00:45:28,800 –> 00:45:32,120
And in December 2025, the SCI for AI specification

1264
00:45:32,120 –> 00:45:35,520
was ratified to extend that thinking to AI’s life cycle

1265
00:45:35,520 –> 00:45:37,840
specifically, including the heavy training

1266
00:45:37,840 –> 00:45:39,960
and data preparation stages that people love

1267
00:45:39,960 –> 00:45:42,360
to ignore when they only measure inference.

1268
00:45:42,360 –> 00:45:45,200
That distinction matters because AI systems have two audiences

1269
00:45:45,200 –> 00:45:46,360
who can act.

1270
00:45:46,360 –> 00:45:49,000
Providers who build and train and consumers who run inference.

1271
00:45:49,000 –> 00:45:50,440
If you don’t separate those boundaries,

1272
00:45:50,440 –> 00:45:52,720
everyone points at each other and nobody changes anything,

1273
00:45:52,720 –> 00:45:54,480
providers say users drove demand.

1274
00:45:54,480 –> 00:45:56,680
Users say providers chose the architecture.

1275
00:45:56,680 –> 00:45:57,880
The carbon stays.

1276
00:45:57,880 –> 00:46:01,840
So an AI carbon budget needs two explicit owners,

1277
00:46:01,840 –> 00:46:04,360
one for the model and one for the workload using it.

1278
00:46:04,360 –> 00:46:06,960
Otherwise, you’re budgeting AI like it’s a single thing

1279
00:46:06,960 –> 00:46:09,040
and that’s how you get conditional chaos disguised

1280
00:46:09,040 –> 00:46:09,960
as responsibility.

1281
00:46:09,960 –> 00:46:12,080
Now here’s the part admins will recognize

1282
00:46:12,080 –> 00:46:15,320
AI workloads create the worst possible utilization patterns,

1283
00:46:15,320 –> 00:46:17,640
spiky demand, high peak requirements,

1284
00:46:17,640 –> 00:46:19,400
and expensive accelerators that sit idle

1285
00:46:19,400 –> 00:46:20,600
if you don’t shape the traffic.

1286
00:46:20,600 –> 00:46:23,240
Carbon budgeting forces the same uncomfortable question.

1287
00:46:23,240 –> 00:46:24,160
FinOps forces.

1288
00:46:24,160 –> 00:46:26,720
Are you paying for capacity or are you paying for waste?

1289
00:46:26,720 –> 00:46:28,400
And once you start asking that,

1290
00:46:28,400 –> 00:46:31,160
the optimization options get brutally practical.

1291
00:46:31,160 –> 00:46:33,480
Reduce compute time, smaller models where they work,

1292
00:46:33,480 –> 00:46:35,200
better prompts, better rooting.

1293
00:46:35,200 –> 00:46:37,400
Reduce power draw, hardware choice and right sizing.

1294
00:46:37,400 –> 00:46:38,680
Reduce overhead.

1295
00:46:38,680 –> 00:46:41,040
Use efficient infrastructure instead of pretending

1296
00:46:41,040 –> 00:46:42,640
PE is someone else’s job.

1297
00:46:42,640 –> 00:46:45,000
Reduce grid intensity,

1298
00:46:45,000 –> 00:46:47,800
region selection and scheduling, where possible.

1299
00:46:47,800 –> 00:46:49,120
None of this is ideological.

1300
00:46:49,120 –> 00:46:51,560
It’s the mechanics of running a physical platform

1301
00:46:51,560 –> 00:46:54,280
under constraint and the payoff is that AI stops

1302
00:46:54,280 –> 00:46:55,680
being a mystical line item.

1303
00:46:55,680 –> 00:46:57,560
It becomes governable, which is the whole point

1304
00:46:57,560 –> 00:46:58,720
of a carbon control plane.

1305
00:46:58,720 –> 00:46:59,600
But to be clear,

1306
00:46:59,600 –> 00:47:03,040
this budgeting discipline crashes into the next hard limit fast.

1307
00:47:03,040 –> 00:47:04,760
Because even if you optimize models,

1308
00:47:04,760 –> 00:47:07,480
the data center still has to deliver the power.

1309
00:47:07,480 –> 00:47:10,400
The data center constraint power density ends the fantasy.

1310
00:47:10,400 –> 00:47:12,920
And this is where the conversation stops being about dashboards

1311
00:47:12,920 –> 00:47:14,560
and starts being about thermodynamics.

1312
00:47:14,560 –> 00:47:17,960
The cloud industry spend years acting like compute is elastic

1313
00:47:17,960 –> 00:47:18,920
in the purest sense.

1314
00:47:18,920 –> 00:47:20,840
You click, capacity appears,

1315
00:47:20,840 –> 00:47:23,240
and the only real constraint is your budget.

1316
00:47:23,240 –> 00:47:26,360
That story worked when workloads were mostly CPU,

1317
00:47:26,360 –> 00:47:28,360
mostly predictable and mostly tolerant

1318
00:47:28,360 –> 00:47:30,520
of being spread across commodity infrastructure.

1319
00:47:30,520 –> 00:47:32,880
AI breaks that story because it changes the density

1320
00:47:32,880 –> 00:47:35,640
of the problem, not more servers.

1321
00:47:35,640 –> 00:47:37,880
Denser servers, hotter servers,

1322
00:47:37,880 –> 00:47:40,000
more concentrated demand for power,

1323
00:47:40,000 –> 00:47:41,800
cooling and physical space.

1324
00:47:41,800 –> 00:47:43,200
And once density goes up,

1325
00:47:43,200 –> 00:47:45,080
the control plane you thought was abstract

1326
00:47:45,080 –> 00:47:47,440
gets pinned to a building with wires, pipes,

1327
00:47:47,440 –> 00:47:50,280
and a local utility that can say no.

1328
00:47:50,280 –> 00:47:51,760
Start with the basic fact pattern,

1329
00:47:51,760 –> 00:47:53,920
the industry can’t market its way around.

1330
00:47:53,920 –> 00:47:56,080
Modern accelerators pull serious power.

1331
00:47:56,080 –> 00:47:59,960
Current GPU class chips land roughly in the 700 to 120 watt range

1332
00:47:59,960 –> 00:48:00,480
per chip.

1333
00:48:00,480 –> 00:48:01,720
That’s per component.

1334
00:48:01,720 –> 00:48:04,080
Once you aggregate them into high density servers,

1335
00:48:04,080 –> 00:48:06,520
you’re talking about multi kilowatt machines

1336
00:48:06,520 –> 00:48:08,920
and then racks that trend toward extreme loads.

1337
00:48:08,920 –> 00:48:11,840
That one shift rewrites the data center design constraints.

1338
00:48:11,840 –> 00:48:13,400
Air cooling used to be the default

1339
00:48:13,400 –> 00:48:15,200
because it was operationally simple.

1340
00:48:15,200 –> 00:48:17,640
Push cold air in, pull hot air out

1341
00:48:17,640 –> 00:48:20,440
and pretend the laws of heat transfer are negotiable.

1342
00:48:20,440 –> 00:48:21,320
They aren’t.

1343
00:48:21,320 –> 00:48:24,880
As rack density’s climb, air cooling stops being inefficient

1344
00:48:24,880 –> 00:48:26,360
and becomes insufficient.

1345
00:48:26,360 –> 00:48:27,920
There isn’t enough airflow in the world

1346
00:48:27,920 –> 00:48:30,440
to move that much heat out of that much space,

1347
00:48:30,440 –> 00:48:33,520
reliably without burning absurd energy on fans and chilling.

1348
00:48:33,520 –> 00:48:35,520
So the industry moves to liquid direct to chip

1349
00:48:35,520 –> 00:48:37,000
liquid cooling at minimum,

1350
00:48:37,000 –> 00:48:39,520
sometimes immersion or other liquid based approaches.

1351
00:48:39,520 –> 00:48:41,080
Not because it looks futuristic,

1352
00:48:41,080 –> 00:48:44,240
but because it’s the only path that keeps hardware inside

1353
00:48:44,240 –> 00:48:46,200
operating temperatures while maintaining

1354
00:48:46,200 –> 00:48:48,080
reasonable facility efficiency.

1355
00:48:48,080 –> 00:48:49,320
The moment you move to liquid,

1356
00:48:49,320 –> 00:48:52,720
the fantasy of data centers are just big server rooms, dies.

1357
00:48:52,720 –> 00:48:54,280
You’ve built an industrial plant.

1358
00:48:54,280 –> 00:48:56,360
You need plumbing, monitoring, leak detection,

1359
00:48:56,360 –> 00:48:58,520
maintenance protocols and people who understand

1360
00:48:58,520 –> 00:49:00,040
both facilities and IT.

1361
00:49:00,040 –> 00:49:02,040
That distinction matters for sustainability

1362
00:49:02,040 –> 00:49:03,720
because cooling is not an accessory.

1363
00:49:03,720 –> 00:49:04,840
It’s part of the energy model.

1364
00:49:04,840 –> 00:49:07,800
It drives overhead and overhead drives emissions.

1365
00:49:07,800 –> 00:49:10,360
When you hear PUE, you’re hearing the data center

1366
00:49:10,360 –> 00:49:13,320
confess that compute isn’t the only thing drawing power.

1367
00:49:13,320 –> 00:49:16,160
Cooling is an entire second system running in parallel

1368
00:49:16,160 –> 00:49:18,280
and AI makes that system louder.

1369
00:49:18,280 –> 00:49:21,160
Now add the grid constraint, which is where reality starts

1370
00:49:21,160 –> 00:49:21,960
pushing back.

1371
00:49:21,960 –> 00:49:23,440
Power isn’t just a cost.

1372
00:49:23,440 –> 00:49:24,280
It’s capacity.

1373
00:49:24,280 –> 00:49:26,800
You can’t buy what the grid can’t deliver to your site

1374
00:49:26,800 –> 00:49:27,880
on your timeline.

1375
00:49:27,880 –> 00:49:29,520
And even if the grid can deliver it,

1376
00:49:29,520 –> 00:49:31,800
the local community might not tolerate the outcome

1377
00:49:31,800 –> 00:49:33,000
when demand explodes.

1378
00:49:33,000 –> 00:49:34,920
A widely cited example is Virginia

1379
00:49:34,920 –> 00:49:36,680
where data centers already consume

1380
00:49:36,680 –> 00:49:38,600
almost a quarter of the state’s electricity.

1381
00:49:38,600 –> 00:49:39,840
That’s not a climate debate.

1382
00:49:39,840 –> 00:49:41,640
That’s a political and economic constraint.

1383
00:49:41,640 –> 00:49:43,240
When residents face rising costs

1384
00:49:43,240 –> 00:49:45,280
and utilities face capacity stress,

1385
00:49:45,280 –> 00:49:47,720
data center growth stops being innovation.

1386
00:49:47,720 –> 00:49:49,160
It becomes a governance problem.

1387
00:49:49,160 –> 00:49:50,640
This is the uncomfortable truth.

1388
00:49:50,640 –> 00:49:53,560
Cloud growth is now a local infrastructure event.

1389
00:49:53,560 –> 00:49:56,160
It competes with housing growth, industrial demand,

1390
00:49:56,160 –> 00:49:59,600
transportation electrification, and basic grid stability.

1391
00:49:59,600 –> 00:50:01,800
It also inherits the pace of permitting,

1392
00:50:01,800 –> 00:50:04,200
transmission build out, and generation projects,

1393
00:50:04,200 –> 00:50:06,360
which is slow, slower than product roadmaps,

1394
00:50:06,360 –> 00:50:09,360
slower than quarterly revenue pressure, slower than ship the model.

1395
00:50:09,360 –> 00:50:11,720
So when Microsoft talks about being carbon negative

1396
00:50:11,720 –> 00:50:13,600
while scaling AI, the hidden requirement

1397
00:50:13,600 –> 00:50:15,920
isn’t just better accounting, it’s power planning,

1398
00:50:15,920 –> 00:50:19,080
it’s site selection, it’s long term energy procurement,

1399
00:50:19,080 –> 00:50:20,800
it’s cooling modernization,

1400
00:50:20,800 –> 00:50:22,880
it’s supply chain coordination for hardware

1401
00:50:22,880 –> 00:50:24,000
and infrastructure components

1402
00:50:24,000 –> 00:50:25,920
that are not infinitely available.

1403
00:50:25,920 –> 00:50:28,480
And this is why sustainability becomes capacity planning,

1404
00:50:28,480 –> 00:50:29,480
not branding.

1405
00:50:29,480 –> 00:50:31,440
The carbon control plane can tell you what happened,

1406
00:50:31,440 –> 00:50:33,920
the internal carbon fee can make waste painful,

1407
00:50:33,920 –> 00:50:36,600
cloud for sustainability can make emissions visible.

1408
00:50:36,600 –> 00:50:38,240
But none of those things conjure electrons

1409
00:50:38,240 –> 00:50:40,400
out of a constrained grid or make heat disappear

1410
00:50:40,400 –> 00:50:41,240
from a sealed room.

1411
00:50:41,240 –> 00:50:42,920
The system is pinned to physics

1412
00:50:42,920 –> 00:50:45,840
and physics is the only auditor that never negotiates.

1413
00:50:45,840 –> 00:50:47,880
Now this is where the episode earns its title,

1414
00:50:47,880 –> 00:50:50,600
impossible audit, because the organization isn’t trying

1415
00:50:50,600 –> 00:50:52,000
to audit a static footprint.

1416
00:50:52,000 –> 00:50:53,880
It’s trying to audit a moving target

1417
00:50:53,880 –> 00:50:56,120
whose underlying infrastructure expands faster

1418
00:50:56,120 –> 00:50:58,880
than decarbonization can propagate through supply chains

1419
00:50:58,880 –> 00:50:59,480
and grids.

1420
00:50:59,480 –> 00:51:02,080
So the control plane has to evolve from report and explain

1421
00:51:02,080 –> 00:51:03,840
into design and constraint.

1422
00:51:03,840 –> 00:51:06,240
It has to move upstream into architecture patterns,

1423
00:51:06,240 –> 00:51:08,120
region strategy, workload defaults,

1424
00:51:08,120 –> 00:51:09,480
and procurement commitments that

1425
00:51:09,480 –> 00:51:12,920
make sustainability into the growth engine itself.

1426
00:51:12,920 –> 00:51:14,920
Otherwise, the pattern is predictable.

1427
00:51:14,920 –> 00:51:16,520
Emissions rise with capacity,

1428
00:51:16,520 –> 00:51:18,280
removals expand to compensate

1429
00:51:18,280 –> 00:51:20,880
and trust becomes dependent on markets and methodologies

1430
00:51:20,880 –> 00:51:22,600
that engineers don’t control.

1431
00:51:22,600 –> 00:51:24,400
Once you accept the data center constraint,

1432
00:51:24,400 –> 00:51:25,880
the next move becomes obvious.

1433
00:51:25,880 –> 00:51:27,640
Microsoft can’t just run sustainability

1434
00:51:27,640 –> 00:51:28,600
as a corporate program.

1435
00:51:28,600 –> 00:51:31,040
It has to productize the measurement and governance stack

1436
00:51:31,040 –> 00:51:32,600
because at high-per-scale visibility

1437
00:51:32,600 –> 00:51:35,560
is the only way to get friction into the right decisions.

1438
00:51:35,560 –> 00:51:37,840
Artifact 2, cloud for sustainability,

1439
00:51:37,840 –> 00:51:40,360
emissions data as an operating system.

1440
00:51:40,360 –> 00:51:43,800
So the next artifact is Microsoft cloud for sustainability.

1441
00:51:43,800 –> 00:51:47,120
And it’s easy to dismiss it as a dashboard for ESG people.

1442
00:51:47,120 –> 00:51:49,440
That’s the comfortable interpretation.

1443
00:51:49,440 –> 00:51:51,600
In reality, it’s an attempt to turn emissions

1444
00:51:51,600 –> 00:51:55,520
into a managed data domain, collected, modeled, reconciled,

1445
00:51:55,520 –> 00:51:57,000
and reported with enough structure

1446
00:51:57,000 –> 00:51:59,080
that it can influence decisions, which

1447
00:51:59,080 –> 00:52:01,440
is the only point of having a carbon control plane

1448
00:52:01,440 –> 00:52:02,480
in the first place.

1449
00:52:02,480 –> 00:52:04,520
Because here’s what Microsoft learned the hard way

1450
00:52:04,520 –> 00:52:06,760
and what every large enterprise learns eventually.

1451
00:52:06,760 –> 00:52:09,800
You can’t govern what you can’t compile.

1452
00:52:09,800 –> 00:52:11,640
Cloud for sustainability is essentially

1453
00:52:11,640 –> 00:52:12,920
an emissions management system.

1454
00:52:12,920 –> 00:52:16,080
It ingests activity data, maps that activity to emissions

1455
00:52:16,080 –> 00:52:19,280
factors, and produces outputs that can be used for reporting

1456
00:52:19,280 –> 00:52:20,720
and operational tracking.

1457
00:52:20,720 –> 00:52:22,280
That sounds simple until you remember

1458
00:52:22,280 –> 00:52:25,400
what activity data means in a modern company.

1459
00:52:25,400 –> 00:52:28,440
Energy usage, travel records, procurement data,

1460
00:52:28,440 –> 00:52:31,640
supplier disclosures, logistics, and the endless gray area

1461
00:52:31,640 –> 00:52:32,840
of scope three.

1462
00:52:32,840 –> 00:52:35,320
So this product exists because spreadsheets don’t scale

1463
00:52:35,320 –> 00:52:37,600
and manual reconciliation turns sustainability

1464
00:52:37,600 –> 00:52:39,480
into a quarterly panic ritual.

1465
00:52:39,480 –> 00:52:41,880
It’s the same failure mode you see in identity governance.

1466
00:52:41,880 –> 00:52:44,120
A company can survive on tribal knowledge and exports

1467
00:52:44,120 –> 00:52:47,040
for a while, and then the system size crosses a threshold.

1468
00:52:47,040 –> 00:52:49,800
And the only thing left is automation or collapse.

1469
00:52:49,800 –> 00:52:52,480
Cloud for sustainability is Microsoft trying to automate,

1470
00:52:52,480 –> 00:52:54,360
not being sustainable.

1471
00:52:54,360 –> 00:52:57,560
Automating the ability to make sustainability measurable.

1472
00:52:57,560 –> 00:52:59,200
That distinction matters.

1473
00:52:59,200 –> 00:53:01,600
The foundational move is to treat emissions

1474
00:53:01,600 –> 00:53:04,480
as a first class data set, not a slide deck.

1475
00:53:04,480 –> 00:53:06,560
Data comes in from multiple sources,

1476
00:53:06,560 –> 00:53:09,160
gets normalized, and then gets tied to an accounting model

1477
00:53:09,160 –> 00:53:12,240
that fits recognized categories like the emissions scopes.

1478
00:53:12,240 –> 00:53:15,400
The output becomes something you can query, trend, and audit.

1479
00:53:15,400 –> 00:53:17,600
And if that sounds like an ERP pitch, good.

1480
00:53:17,600 –> 00:53:18,440
That’s what this is.

1481
00:53:18,440 –> 00:53:21,400
Emissions as a management system with the implied promise

1482
00:53:21,400 –> 00:53:23,960
that you can run governance workflows on top of it.

1483
00:53:23,960 –> 00:53:25,920
You can’t enforce what you can’t see,

1484
00:53:25,920 –> 00:53:27,920
and you can’t see what you can’t model.

1485
00:53:27,920 –> 00:53:30,680
Now, the key reframing for admins is this.

1486
00:53:30,680 –> 00:53:33,880
Cloud for sustainability isn’t there to make Microsoft look good.

1487
00:53:33,880 –> 00:53:35,880
It’s there because Microsoft needs an internal

1488
00:53:35,880 –> 00:53:37,720
and external way to turn carbon into something

1489
00:53:37,720 –> 00:53:40,440
that the organization can root through its existing control

1490
00:53:40,440 –> 00:53:44,400
mechanisms, budgeting, procurement, and operational review.

1491
00:53:44,400 –> 00:53:46,280
In other words, it’s an attempt to give carbon

1492
00:53:46,280 –> 00:53:47,880
the same treatment cost already gets.

1493
00:53:47,880 –> 00:53:50,240
Cost has telemetry, cost has allocation,

1494
00:53:50,240 –> 00:53:54,040
cost has chargeback, cost has alerts, cost has governance,

1495
00:53:54,040 –> 00:53:56,120
cost has people who get fired for ignoring it.

1496
00:53:56,120 –> 00:53:58,040
Carbon historically has vibes.

1497
00:53:58,040 –> 00:54:00,280
So Cloud for sustainability is Microsoft saying,

1498
00:54:00,280 –> 00:54:03,280
enough with vibes, build the substrate.

1499
00:54:03,280 –> 00:54:05,240
And this is where the operating system analogy

1500
00:54:05,240 –> 00:54:06,440
actually earns its place.

1501
00:54:06,440 –> 00:54:08,400
An operating system isn’t the app you launch.

1502
00:54:08,400 –> 00:54:11,320
It’s the layer that coordinates resources, mediates access,

1503
00:54:11,320 –> 00:54:14,000
and creates a consistent model for higher level policy.

1504
00:54:14,000 –> 00:54:16,520
That’s what this product tries to be for emissions data.

1505
00:54:16,520 –> 00:54:19,240
A layer that standardizes how emissions get represented

1506
00:54:19,240 –> 00:54:22,520
so the organization can build reporting and action on top

1507
00:54:22,520 –> 00:54:24,800
because annual reporting alone isn’t control.

1508
00:54:24,800 –> 00:54:25,760
It’s hindsight.

1509
00:54:25,760 –> 00:54:28,400
The difference between reporting and governance is time.

1510
00:54:28,400 –> 00:54:29,800
Reporting tells you what happened.

1511
00:54:29,800 –> 00:54:31,280
Governance changes what happens next

1512
00:54:31,280 –> 00:54:33,360
by altering defaults and adding friction.

1513
00:54:33,360 –> 00:54:36,080
Cloud for sustainability sits in that middle zone.

1514
00:54:36,080 –> 00:54:39,440
If you wire it correctly, it can produce carbon per workload views.

1515
00:54:39,440 –> 00:54:41,360
It can create a consistent methodology

1516
00:54:41,360 –> 00:54:43,480
for translating activity into emissions.

1517
00:54:43,480 –> 00:54:45,800
It can support the kind of accounting, integrity,

1518
00:54:45,800 –> 00:54:47,680
and internal carbon feed depends on.

1519
00:54:47,680 –> 00:54:50,720
It can also create the boring but critical audit trail

1520
00:54:50,720 –> 00:54:53,040
where a number came from, which source fatted,

1521
00:54:53,040 –> 00:54:54,120
and what changed when.

1522
00:54:54,120 –> 00:54:56,160
This is the part nobody sells in a keynote

1523
00:54:56,160 –> 00:54:57,360
because it’s not inspiring.

1524
00:54:57,360 –> 00:54:59,720
It is, however, the only thing that survives scrutiny.

1525
00:54:59,720 –> 00:55:01,800
Now the limitation, and Microsoft knows this,

1526
00:55:01,800 –> 00:55:04,640
is that tooling doesn’t enforce anything by itself.

1527
00:55:04,640 –> 00:55:06,400
A management system without enforcement

1528
00:55:06,400 –> 00:55:09,960
becomes a higher resolution version of the same old failure mode.

1529
00:55:09,960 –> 00:55:12,480
Disclosure theater, just with better charts.

1530
00:55:12,480 –> 00:55:15,760
If cloud for sustainability is deployed as a reporting overlay,

1531
00:55:15,760 –> 00:55:17,840
it will do exactly what overlays always do.

1532
00:55:17,840 –> 00:55:20,680
Produce beautiful artifacts that don’t change decisions.

1533
00:55:20,680 –> 00:55:22,720
The control plane only becomes real

1534
00:55:22,720 –> 00:55:25,080
when the organization couples the data domain

1535
00:55:25,080 –> 00:55:28,240
to enforcement points, procurement requirements,

1536
00:55:28,240 –> 00:55:31,520
design review gates, internal chargeback models,

1537
00:55:31,520 –> 00:55:35,280
and policy defaults like region justification required,

1538
00:55:35,280 –> 00:55:37,920
or always on must be defended.

1539
00:55:37,920 –> 00:55:39,680
And so the product is not the control plane,

1540
00:55:39,680 –> 00:55:42,400
it is the instrumentation layer for the control plane.

1541
00:55:42,400 –> 00:55:43,960
And this is the uncomfortable truth

1542
00:55:43,960 –> 00:55:45,680
of enterprise sustainability.

1543
00:55:45,680 –> 00:55:48,080
The hardest part isn’t calculating emissions.

1544
00:55:48,080 –> 00:55:50,800
The hardest part is making the people who own the decisions

1545
00:55:50,800 –> 00:55:52,760
feel the consequence of those numbers.

1546
00:55:52,760 –> 00:55:55,040
Cloud for sustainability can make carbon visible.

1547
00:55:55,040 –> 00:55:56,400
It can’t make you care.

1548
00:55:56,400 –> 00:55:58,040
That’s still your job as the architect

1549
00:55:58,040 –> 00:55:59,920
to wire visibility into control.

1550
00:55:59,920 –> 00:56:02,400
And once you do, you discover the next problem nobody

1551
00:56:02,400 –> 00:56:03,320
wants to talk about.

1552
00:56:03,320 –> 00:56:05,120
The emissions you didn’t even know you had

1553
00:56:05,120 –> 00:56:08,400
because they live in small automation and low-code sprawl

1554
00:56:08,400 –> 00:56:11,600
that never went through a real architecture review.

1555
00:56:11,600 –> 00:56:14,160
Case study three, power platform governance.

1556
00:56:14,160 –> 00:56:16,080
sprawl becomes environmental risk.

1557
00:56:16,080 –> 00:56:17,640
Now the carbon you don’t see.

1558
00:56:17,640 –> 00:56:19,640
Not the data center, not the AI cluster.

1559
00:56:19,640 –> 00:56:23,160
The quiet mess that grows inside every Microsoft 365 tenant,

1560
00:56:23,160 –> 00:56:25,520
the moment the business discovers it can ship apps

1561
00:56:25,520 –> 00:56:26,440
without asking you.

1562
00:56:26,440 –> 00:56:28,280
Power platform is productivity acceleration.

1563
00:56:28,280 –> 00:56:29,640
It’s also a sprawl engine.

1564
00:56:29,640 –> 00:56:32,360
And once you accept that carbon behaves like cost,

1565
00:56:32,360 –> 00:56:34,960
sprawl becomes an environmental risk for the same reason

1566
00:56:34,960 –> 00:56:36,600
it becomes a security risk.

1567
00:56:36,600 –> 00:56:40,360
It creates unmanaged, always on activity that nobody owns.

1568
00:56:40,360 –> 00:56:42,600
Most people still frame low-code governance

1569
00:56:42,600 –> 00:56:43,880
as a security story.

1570
00:56:43,880 –> 00:56:46,360
Data loss prevention policies, environment strategy,

1571
00:56:46,360 –> 00:56:48,160
connector control, audit logs.

1572
00:56:48,160 –> 00:56:48,960
Fine.

1573
00:56:48,960 –> 00:56:51,040
But the foundational mistake is treating sprawl

1574
00:56:51,040 –> 00:56:53,280
as a compliance nuisance instead of a resource leak.

1575
00:56:53,280 –> 00:56:55,560
Because every duplicated flow, every zombie app,

1576
00:56:55,560 –> 00:56:58,400
every temporary automation that runs forever is compute.

1577
00:56:58,400 –> 00:56:59,400
It schedules jobs.

1578
00:56:59,400 –> 00:57:00,680
It triggers API calls.

1579
00:57:00,680 –> 00:57:01,440
It writes data.

1580
00:57:01,440 –> 00:57:02,440
It wakes up systems.

1581
00:57:02,440 –> 00:57:06,000
And at scale, small inefficiencies become a standing order

1582
00:57:06,000 –> 00:57:07,080
against the grid.

1583
00:57:07,080 –> 00:57:08,760
Here’s how it happens.

1584
00:57:08,760 –> 00:57:11,520
A team builds a power automate flow to sync files,

1585
00:57:11,520 –> 00:57:13,800
push notifications or move approvals.

1586
00:57:13,800 –> 00:57:14,320
It works.

1587
00:57:14,320 –> 00:57:15,960
Then another team builds the same thing,

1588
00:57:15,960 –> 00:57:17,160
slightly differently, because they

1589
00:57:17,160 –> 00:57:18,400
couldn’t find the first one.

1590
00:57:18,400 –> 00:57:20,560
A month later, someone builds a third version

1591
00:57:20,560 –> 00:57:23,640
because the second version broke after a connector permission

1592
00:57:23,640 –> 00:57:24,040
changed.

1593
00:57:24,040 –> 00:57:25,160
None of these are malicious.

1594
00:57:25,160 –> 00:57:27,120
They are rational responses to entropy.

1595
00:57:27,120 –> 00:57:28,640
But they have one common trait.

1596
00:57:28,640 –> 00:57:29,320
They run.

1597
00:57:29,320 –> 00:57:30,200
They run in the background.

1598
00:57:30,200 –> 00:57:31,360
They run on schedules.

1599
00:57:31,360 –> 00:57:33,280
They run on triggers that nobody revisits.

1600
00:57:33,280 –> 00:57:34,800
And because they’re just automation,

1601
00:57:34,800 –> 00:57:37,200
they rarely get cost scrutiny, architecture, review,

1602
00:57:37,200 –> 00:57:38,440
or life cycle ownership.

1603
00:57:38,440 –> 00:57:41,080
So the organization accumulates invisible workload.

1604
00:57:41,080 –> 00:57:43,920
In financial terms, it’s like handing out corporate credit cards

1605
00:57:43,920 –> 00:57:46,840
with no expense review because each charge is small.

1606
00:57:46,840 –> 00:57:49,000
Then acting surprised when the total is ugly.

1607
00:57:49,000 –> 00:57:52,040
In carbon terms, it’s the same pattern with a different unit.

1608
00:57:52,040 –> 00:57:53,600
You don’t get one big emissions event.

1609
00:57:53,600 –> 00:57:55,760
You get thousands of tiny ones that never stop.

1610
00:57:55,760 –> 00:57:57,640
This is where the carbon control plane idea

1611
00:57:57,640 –> 00:58:00,120
becomes useful to admins because it gives you

1612
00:58:00,120 –> 00:58:03,440
a forcing function that security governance alone doesn’t.

1613
00:58:03,440 –> 00:58:06,400
Security governance says don’t exfiltrate data.

1614
00:58:06,400 –> 00:58:08,880
Carbon aware governance says, stop creating always

1615
00:58:08,880 –> 00:58:10,560
on behavior without ownership.

1616
00:58:10,560 –> 00:58:12,240
That is a different control objective.

1617
00:58:12,240 –> 00:58:13,600
And it changes what you measure.

1618
00:58:13,600 –> 00:58:16,040
The thing most people miss is that low-code sprawl

1619
00:58:16,040 –> 00:58:18,280
creates two kinds of hidden compute–

1620
00:58:18,280 –> 00:58:19,920
first, direct execution.

1621
00:58:19,920 –> 00:58:23,040
Flows, triggers, scheduled jobs, connectors,

1622
00:58:23,040 –> 00:58:24,920
retries, and polling patterns.

1623
00:58:24,920 –> 00:58:26,960
Badly designed flows don’t just fail.

1624
00:58:26,960 –> 00:58:27,840
They retry.

1625
00:58:27,840 –> 00:58:28,720
They fan out.

1626
00:58:28,720 –> 00:58:30,240
They hammer APIs.

1627
00:58:30,240 –> 00:58:32,200
They increase execution volume, which

1628
00:58:32,200 –> 00:58:33,760
increases back-end workload.

1629
00:58:33,760 –> 00:58:35,400
Second, indirect compute.

1630
00:58:35,400 –> 00:58:37,360
All the downstream systems you wake up.

1631
00:58:37,360 –> 00:58:40,040
Every automation that touches SharePoint, Dataverse,

1632
00:58:40,040 –> 00:58:42,440
Exchange Teams, or third-party connectors

1633
00:58:42,440 –> 00:58:44,520
creates activity across multiple services.

1634
00:58:44,520 –> 00:58:45,920
You don’t just pay for the flow.

1635
00:58:45,920 –> 00:58:47,480
You pay for the chain reaction.

1636
00:58:47,480 –> 00:58:48,960
This clicked for a lot of architects

1637
00:58:48,960 –> 00:58:51,560
when they saw a tenant with hundreds of flows still firing

1638
00:58:51,560 –> 00:58:54,200
every day long after the business process changed.

1639
00:58:54,200 –> 00:58:56,040
Nobody removed them because nobody owned them.

1640
00:58:56,040 –> 00:58:57,640
The flow became infrastructure.

1641
00:58:57,640 –> 00:58:58,960
Accidental infrastructure.

1642
00:58:58,960 –> 00:59:00,920
And accidental infrastructure is always

1643
00:59:00,920 –> 00:59:04,000
expensive in cost, in risk, in carbon.

1644
00:59:04,000 –> 00:59:06,560
So what does a serious power platform governance model

1645
00:59:06,560 –> 00:59:09,280
look like when you include environmental accountability?

1646
00:59:09,280 –> 00:59:12,040
It starts with ownership as a non-negotiable control.

1647
00:59:12,040 –> 00:59:13,760
Every environment needs a named owner.

1648
00:59:13,760 –> 00:59:15,480
Every production flow needs a named owner.

1649
00:59:15,480 –> 00:59:18,200
Every always-on automation needs a justification

1650
00:59:18,200 –> 00:59:19,400
and a review cadence.

1651
00:59:19,400 –> 00:59:21,960
If you can’t identify who gets paid when it breaks,

1652
00:59:21,960 –> 00:59:24,320
you also can’t justify why it exists.

1653
00:59:24,320 –> 00:59:26,840
Then you enforce life cycle, not just access.

1654
00:59:26,840 –> 00:59:29,560
App and flow inventories aren’t documentation projects.

1655
00:59:29,560 –> 00:59:30,880
They’re deletion pipelines.

1656
00:59:30,880 –> 00:59:33,040
You need a process that finds unused assets,

1657
00:59:33,040 –> 00:59:34,840
flags them, and decommissions them

1658
00:59:34,840 –> 00:59:37,400
because the default state of low code is not maintained.

1659
00:59:37,400 –> 00:59:38,960
The default state is abandoned.

1660
00:59:38,960 –> 00:59:41,560
Now add the part that makes people uncomfortable.

1661
00:59:41,560 –> 00:59:44,880
You treat background automation as a consumption class,

1662
00:59:44,880 –> 00:59:47,920
not as innovation as consumption.

1663
00:59:47,920 –> 00:59:50,400
You measure it like you measure cloud spend.

1664
00:59:50,400 –> 00:59:53,920
Number of flows, execution counts, connector usage,

1665
00:59:53,920 –> 00:59:56,320
error rates and retries, and environments

1666
00:59:56,320 –> 00:59:58,840
that exist with no active business owner.

1667
00:59:58,840 –> 01:00:00,160
Those are your leak indicators.

1668
01:00:00,160 –> 01:00:01,560
You don’t need perfect carbon numbers

1669
01:00:01,560 –> 01:00:03,360
at this layer to improve behavior.

1670
01:00:03,360 –> 01:00:06,240
You need enough telemetry to identify waste patterns

1671
01:00:06,240 –> 01:00:08,800
that correlate strongly with unnecessary activity.

1672
01:00:08,800 –> 01:00:11,640
And yes, this is where cloud for sustainability fits.

1673
01:00:11,640 –> 01:00:14,080
Conceptually, even if the product doesn’t magically

1674
01:00:14,080 –> 01:00:16,000
give you a carbon readout per flow.

1675
01:00:16,000 –> 01:00:17,600
The value is the mental model.

1676
01:00:17,600 –> 01:00:19,920
Emissions are downstream of activity.

1677
01:00:19,920 –> 01:00:22,440
If you reduce unnecessary activity, you reduce impact.

1678
01:00:22,440 –> 01:00:24,720
That’s deterministic enough to be operationally useful.

1679
01:00:24,720 –> 01:00:27,320
Finally, you align incentives.

1680
01:00:27,320 –> 01:00:29,800
If teams can create environments and automations

1681
01:00:29,800 –> 01:00:31,720
at zero perceived cost, they will.

1682
01:00:31,720 –> 01:00:33,720
If you introduce friction, approval gates

1683
01:00:33,720 –> 01:00:36,400
for production environments, default retention policies,

1684
01:00:36,400 –> 01:00:39,600
periodic iterations, and chargeback signals where possible.

1685
01:00:39,600 –> 01:00:40,840
Behavior changes.

1686
01:00:40,840 –> 01:00:42,280
Not because people became better,

1687
01:00:42,280 –> 01:00:44,320
because the system stopped rewarding entropy.

1688
01:00:44,320 –> 01:00:46,200
So the case study takeaway is blunt.

1689
01:00:46,200 –> 01:00:48,240
Power platform sprawl isn’t just security dead.

1690
01:00:48,240 –> 01:00:49,000
It’s carbon dead.

1691
01:00:49,000 –> 01:00:51,160
Not because each flow emits a lot.

1692
01:00:51,160 –> 01:00:53,040
Because the platform makes it easy to create

1693
01:00:53,040 –> 01:00:55,600
perpetual background activity without ownership

1694
01:00:55,600 –> 01:00:58,560
and perpetual background activity becomes infrastructure load.

1695
01:00:58,560 –> 01:01:01,800
And once you admit that you can’t hide behind its only low code,

1696
01:01:01,800 –> 01:01:03,200
low code is not low impact.

1697
01:01:03,200 –> 01:01:05,600
It is high scale.

1698
01:01:05,600 –> 01:01:07,200
Now, after all this governance talk,

1699
01:01:07,200 –> 01:01:10,040
the next question is the one people always want to jump to early.

1700
01:01:10,040 –> 01:01:11,640
Are we reducing emissions?

1701
01:01:11,640 –> 01:01:15,080
Or are we just buying the right kind of story?

1702
01:01:15,080 –> 01:01:18,360
The critical question, reduction, or outsourcing guilt?

1703
01:01:18,360 –> 01:01:20,320
So now we hit the question, everyone asks,

1704
01:01:20,320 –> 01:01:22,320
the moment removals show up in the story.

1705
01:01:22,320 –> 01:01:24,800
Is this reduction or is this outsourcing guilt?

1706
01:01:24,800 –> 01:01:26,440
And the annoying answer is that the spreadsheet

1707
01:01:26,440 –> 01:01:28,240
doesn’t care about your moral preference.

1708
01:01:28,240 –> 01:01:29,480
It cares about levers.

1709
01:01:29,480 –> 01:01:31,600
A serious carbon control plane needs three.

1710
01:01:31,600 –> 01:01:34,000
And it needs them separated, because each lever

1711
01:01:34,000 –> 01:01:37,280
behaves differently under pressure, reduce, replace, remove.

1712
01:01:37,280 –> 01:01:39,800
Reduce means you emit less for the same outcome.

1713
01:01:39,800 –> 01:01:41,560
Less energy per transaction.

1714
01:01:41,560 –> 01:01:45,240
Fewer wasted cycles, fewer flights, fewer trucks, fewer retreats,

1715
01:01:45,240 –> 01:01:47,040
fewer always on because nobody wanted

1716
01:01:47,040 –> 01:01:48,720
to design auto scaling.

1717
01:01:48,720 –> 01:01:51,720
Reduction is the engineering lever because it attacks the source term.

1718
01:01:51,720 –> 01:01:53,760
If you reduce, you don’t need to compensate later.

1719
01:01:53,760 –> 01:01:55,200
The system stays simpler.

1720
01:01:55,200 –> 01:01:57,160
Replace means you change the energy input,

1721
01:01:57,160 –> 01:01:59,680
so the same activity has a lower emissions factor.

1722
01:01:59,680 –> 01:02:02,720
This is where renewable electricity procurement and grid strategy live.

1723
01:02:02,720 –> 01:02:04,640
Replace doesn’t necessarily reduce consumption.

1724
01:02:04,640 –> 01:02:07,040
It changes what powers that consumption in cloud terms.

1725
01:02:07,040 –> 01:02:08,840
It’s not about fewer CPU cycles.

1726
01:02:08,840 –> 01:02:10,000
It’s about different electrons.

1727
01:02:10,000 –> 01:02:11,880
Remove means you create a negative flow

1728
01:02:11,880 –> 01:02:13,560
to counterbalance what remains.

1729
01:02:13,560 –> 01:02:16,600
This is where carbon removal credits enter.

1730
01:02:16,600 –> 01:02:18,840
Afforestation and reforestation structures,

1731
01:02:18,840 –> 01:02:20,720
soil-based sequestration structures,

1732
01:02:20,720 –> 01:02:22,920
and other durable removal categories Microsoft

1733
01:02:22,920 –> 01:02:25,240
has discussed in its public reporting and deals.

1734
01:02:25,240 –> 01:02:26,600
Remove is not a field good gesture.

1735
01:02:26,600 –> 01:02:28,200
Remove is the reconciliation mechanism

1736
01:02:28,200 –> 01:02:29,880
when the platform’s residual emissions

1737
01:02:29,880 –> 01:02:32,400
refuse to hit zero on your timeline.

1738
01:02:32,400 –> 01:02:33,920
Now here’s where the audience gets trapped.

1739
01:02:33,920 –> 01:02:37,080
They collapse those three levers into one word, offsets.

1740
01:02:37,080 –> 01:02:38,280
And then argue about the word.

1741
01:02:38,280 –> 01:02:39,320
That’s not analysis.

1742
01:02:39,320 –> 01:02:40,720
That’s branding warfare.

1743
01:02:40,720 –> 01:02:42,200
A reduction is not a removal.

1744
01:02:42,200 –> 01:02:45,280
A removal is not the same as avoiding someone else’s emissions.

1745
01:02:45,280 –> 01:02:48,400
And offset often means we bought something.

1746
01:02:48,400 –> 01:02:50,560
Without telling you what physical change occurred,

1747
01:02:50,560 –> 01:02:52,840
how it was measured or how long it persists,

1748
01:02:52,840 –> 01:02:55,640
if you remember nothing else from this section, remember this.

1749
01:02:55,640 –> 01:02:57,480
A net claim is a systems claim.

1750
01:02:57,480 –> 01:02:59,560
It lives or dies on boundaries and mechanisms,

1751
01:02:59,560 –> 01:03:00,480
not on sincerity.

1752
01:03:00,480 –> 01:03:03,160
So when people say Microsoft is just buying credits,

1753
01:03:03,160 –> 01:03:05,920
the only useful response is, which lever are they buying

1754
01:03:05,920 –> 01:03:07,960
and what does the rest of the system do?

1755
01:03:07,960 –> 01:03:10,760
Because removals can be either the last mile tool

1756
01:03:10,760 –> 01:03:12,120
in a reduction first strategy

1757
01:03:12,120 –> 01:03:13,920
or they can become the primary mechanism

1758
01:03:13,920 –> 01:03:16,280
that enables growth without structural change.

1759
01:03:16,280 –> 01:03:19,520
Same accounting vocabulary, completely different system behavior.

1760
01:03:19,520 –> 01:03:21,760
This is where Microsoft’s internal carbon fee

1761
01:03:21,760 –> 01:03:23,600
and its measurement tooling matter,

1762
01:03:23,600 –> 01:03:25,000
not because they guarantee purity,

1763
01:03:25,000 –> 01:03:26,680
but because they create internal pressure

1764
01:03:26,680 –> 01:03:28,360
to reduce before you remove.

1765
01:03:28,360 –> 01:03:31,360
The fee makes emissions expensive inside the organization.

1766
01:03:31,360 –> 01:03:35,160
That pushes teams toward efficiency, region discipline

1767
01:03:35,160 –> 01:03:37,800
and life cycle cleanup, because the organization feels

1768
01:03:37,800 –> 01:03:39,560
the cost now, not just in 203,

1769
01:03:39,560 –> 01:03:42,360
but here’s the uncomfortable truth that makes people angry.

1770
01:03:42,360 –> 01:03:45,160
Removals are structurally necessary in any world

1771
01:03:45,160 –> 01:03:46,960
where growth outpaces reduction.

1772
01:03:46,960 –> 01:03:49,920
You can improve efficiency and still increase total consumption

1773
01:03:49,920 –> 01:03:52,880
if demand scales faster than efficiency gains.

1774
01:03:52,880 –> 01:03:55,600
Cloud and AI are textbook examples of that dynamic.

1775
01:03:55,600 –> 01:03:58,080
The platform gets more efficient per unit of compute

1776
01:03:58,080 –> 01:04:00,840
while total compute demand explodes, both can be true.

1777
01:04:00,840 –> 01:04:03,120
And if both are true, then the residual emissions

1778
01:04:03,120 –> 01:04:05,120
do not vanish on your preferred schedule.

1779
01:04:05,120 –> 01:04:07,360
Therefore, the organization either accepts missing

1780
01:04:07,360 –> 01:04:09,880
the net target or it adds removals.

1781
01:04:09,880 –> 01:04:12,440
That’s not ideology, that’s arithmetic under scale.

1782
01:04:12,440 –> 01:04:15,320
So the real ethical question isn’t, did they use removals?

1783
01:04:15,320 –> 01:04:17,920
The real question is whether removals are being used

1784
01:04:17,920 –> 01:04:20,600
to cover residual hard to eliminate emissions

1785
01:04:20,600 –> 01:04:24,080
while the system aggressively reduces and replaces upstream.

1786
01:04:24,080 –> 01:04:26,240
And that question drags you back to system boundaries

1787
01:04:26,240 –> 01:04:28,840
because any net claim is only as good as what it includes.

1788
01:04:28,840 –> 01:04:30,640
If a company draws the boundary tightly,

1789
01:04:30,640 –> 01:04:32,840
it can look clean by excluding the ugly parts.

1790
01:04:32,840 –> 01:04:36,040
If it draws the boundary broadly, especially into scope three,

1791
01:04:36,040 –> 01:04:38,640
it inherits uncertainty, supplier dependency

1792
01:04:38,640 –> 01:04:39,880
and delayed data.

1793
01:04:39,880 –> 01:04:43,000
Microsoft talks in scopes and explicitly includes scope three

1794
01:04:43,000 –> 01:04:46,080
in its framing, which makes the claim harder, noisier,

1795
01:04:46,080 –> 01:04:47,120
and more attackable.

1796
01:04:47,120 –> 01:04:48,880
That doesn’t make it automatically more true.

1797
01:04:48,880 –> 01:04:50,720
But it does make the audit more real

1798
01:04:50,720 –> 01:04:52,680
because the boundary includes the categories

1799
01:04:52,680 –> 01:04:54,120
where the platform actually lives.

1800
01:04:54,120 –> 01:04:56,400
So don’t ask whether the story feels good.

1801
01:04:56,400 –> 01:04:59,000
Ask whether the control plane enforces the hierarchy,

1802
01:04:59,000 –> 01:05:02,080
reduce first, replace where possible, remove what remains,

1803
01:05:02,080 –> 01:05:03,520
and then ask the follow-up that makes

1804
01:05:03,520 –> 01:05:05,080
the whole episode uncomfortable.

1805
01:05:05,080 –> 01:05:06,960
What happens when the platform grows faster

1806
01:05:06,960 –> 01:05:10,080
than the removal market can supply credible negative flow

1807
01:05:10,080 –> 01:05:12,200
at prices the business will tolerate?

1808
01:05:12,200 –> 01:05:15,480
That’s the real stress test, not whether the slogan sounds sincere.

1809
01:05:15,480 –> 01:05:17,520
And it’s exactly why criticism isn’t optional.

1810
01:05:17,520 –> 01:05:19,160
It’s part of the audit.

1811
01:05:19,160 –> 01:05:22,160
Criticism and trade-offs, what a serious listener should challenge.

1812
01:05:22,160 –> 01:05:24,400
At this point, the worst thing a listener can do

1813
01:05:24,400 –> 01:05:26,160
is treat this like a morality play.

1814
01:05:26,160 –> 01:05:27,880
Microsoft good, Microsoft bad.

1815
01:05:27,880 –> 01:05:28,840
That’s not an audit.

1816
01:05:28,840 –> 01:05:31,240
That’s sports fandom with climate vocabulary.

1817
01:05:31,240 –> 01:05:33,840
A serious listener challenges the system on trade-offs

1818
01:05:33,840 –> 01:05:36,000
where the control plane is structurally strong

1819
01:05:36,000 –> 01:05:38,240
where it’s structurally weak and where the incentives

1820
01:05:38,240 –> 01:05:40,720
can drift into theater even when everyone involved

1821
01:05:40,720 –> 01:05:42,440
thinks they’re being responsible.

1822
01:05:42,440 –> 01:05:44,920
Start with the obvious tension, data center,

1823
01:05:44,920 –> 01:05:48,280
and AI footprint growth versus public commitments.

1824
01:05:48,280 –> 01:05:50,680
Microsoft’s own reporting discloses emissions rising

1825
01:05:50,680 –> 01:05:54,240
into 2023 driven by AI infrastructure expansion.

1826
01:05:54,240 –> 01:05:57,080
So the challenge isn’t why didn’t emissions instantly drop.

1827
01:05:57,080 –> 01:06:00,560
The challenge is, what is the credible path for reductions

1828
01:06:00,560 –> 01:06:04,040
to outpace growth inside the time window implied by 203?

1829
01:06:04,040 –> 01:06:06,680
Because a carbon control plane can force better behavior

1830
01:06:06,680 –> 01:06:09,000
at the margin, but it cannot wish away demand.

1831
01:06:09,000 –> 01:06:11,920
If product strategy continues to scale compute aggressively,

1832
01:06:11,920 –> 01:06:14,680
then the carbon negative claim becomes increasingly dependent

1833
01:06:14,680 –> 01:06:18,240
on external levers, grid decarbonization and removals.

1834
01:06:18,240 –> 01:06:19,840
That’s not automatically wrong.

1835
01:06:19,840 –> 01:06:22,880
But it changes what kind of claim carbon negative really is.

1836
01:06:22,880 –> 01:06:24,640
It stops being an engineering milestone

1837
01:06:24,640 –> 01:06:26,760
and becomes a portfolio management outcome.

1838
01:06:26,760 –> 01:06:28,400
Second challenge, removal dependence.

1839
01:06:28,400 –> 01:06:30,160
Microsoft’s large removal procurement

1840
01:06:30,160 –> 01:06:32,400
makes the 203 pathway legible, but it also

1841
01:06:32,400 –> 01:06:36,280
makes the claim inherit the fragility of the removal ecosystem,

1842
01:06:36,280 –> 01:06:39,600
quality debates, permanence risk, verification limits,

1843
01:06:39,600 –> 01:06:41,720
and changing standards over time.

1844
01:06:41,720 –> 01:06:44,480
The listener should not argue about removals in the abstract.

1845
01:06:44,480 –> 01:06:46,800
They should ask, what fraction of the net claim

1846
01:06:46,800 –> 01:06:48,920
is coming from reductions and replacements

1847
01:06:48,920 –> 01:06:50,320
versus purchase negative flow?

1848
01:06:50,320 –> 01:06:53,560
And how does that ratio change as AI capacity scales?

1849
01:06:53,560 –> 01:06:55,440
If the ratio trends toward buy more,

1850
01:06:55,440 –> 01:06:57,800
then the control plane is drifting from system change

1851
01:06:57,800 –> 01:06:59,400
to financial reconciliation.

1852
01:06:59,400 –> 01:07:01,480
And because removals are claims about the world,

1853
01:07:01,480 –> 01:07:03,440
not meters on a wall, the listener should

1854
01:07:03,440 –> 01:07:07,040
ask about auditability, what gets independently verified,

1855
01:07:07,040 –> 01:07:08,800
what relies on modeled estimates,

1856
01:07:08,800 –> 01:07:11,160
and what changes when methodologies evolve.

1857
01:07:11,160 –> 01:07:13,680
If a net target depends on accounting conventions

1858
01:07:13,680 –> 01:07:16,400
staying stable, that isn’t sustainability.

1859
01:07:16,400 –> 01:07:19,280
That’s a dependency on bureaucracy behaving deterministically.

1860
01:07:19,280 –> 01:07:21,840
Third challenge, scope three measurement integrity.

1861
01:07:21,840 –> 01:07:24,320
Microsoft includes scope three in the carbon fee model

1862
01:07:24,320 –> 01:07:27,040
and in its reporting, which is more serious than many companies.

1863
01:07:27,040 –> 01:07:28,840
But seriousness creates exposure.

1864
01:07:28,840 –> 01:07:31,960
Scope three is where proxies live, supply yourself reporting,

1865
01:07:31,960 –> 01:07:34,080
category level factors, and delayed data.

1866
01:07:34,080 –> 01:07:36,400
So the question becomes, does the system create supply

1867
01:07:36,400 –> 01:07:38,000
a behavior change or does it just create

1868
01:07:38,000 –> 01:07:39,320
supply a reporting improvement?

1869
01:07:39,320 –> 01:07:40,440
Those are not the same thing.

1870
01:07:40,440 –> 01:07:42,520
Supply a decarbonization is physical change,

1871
01:07:42,520 –> 01:07:43,960
supply reporting is legibility.

1872
01:07:43,960 –> 01:07:46,520
You need both, but you can accidentally reward

1873
01:07:46,520 –> 01:07:48,360
the second while hoping for the first.

1874
01:07:48,360 –> 01:07:49,800
That’s the classic governance trap.

1875
01:07:49,800 –> 01:07:51,760
You measure the easiest signal and assume

1876
01:07:51,760 –> 01:07:53,400
it represents the hardest reality.

1877
01:07:53,400 –> 01:07:56,640
Fourth challenge, global inequality and uneven capacity.

1878
01:07:56,640 –> 01:07:59,080
A lot of sustainability narratives assume every supplier

1879
01:07:59,080 –> 01:08:01,120
and region can decarbonize at the same pace.

1880
01:08:01,120 –> 01:08:01,720
They can’t.

1881
01:08:01,720 –> 01:08:04,360
Emerging markets face different grid realities,

1882
01:08:04,360 –> 01:08:06,600
different capital access and different regulatory

1883
01:08:06,600 –> 01:08:07,280
enforcement.

1884
01:08:07,280 –> 01:08:09,520
So when a hyper-scale buyer pushes requirements

1885
01:08:09,520 –> 01:08:12,280
down a supply chain, the question isn’t should they?

1886
01:08:12,280 –> 01:08:14,960
The question is, how the pressure redistributes risk?

1887
01:08:14,960 –> 01:08:16,760
Does it actually finance decarbonization

1888
01:08:16,760 –> 01:08:19,600
or does it simply shift emissions into less visible tears

1889
01:08:19,600 –> 01:08:22,080
where audits are thinner and margins are smaller?

1890
01:08:22,080 –> 01:08:25,720
If the system pushes cost onto the least resilient suppliers,

1891
01:08:25,720 –> 01:08:28,080
it may create compliance, theater and operational

1892
01:08:28,080 –> 01:08:29,800
fragility at the same time.

1893
01:08:29,800 –> 01:08:32,560
Fifth challenge, governance versus productization.

1894
01:08:32,560 –> 01:08:34,280
Microsoft’s cloud for sustainability

1895
01:08:34,280 –> 01:08:36,880
exists because measurement has to scale.

1896
01:08:36,880 –> 01:08:38,960
But tools create a new failure mode.

1897
01:08:38,960 –> 01:08:41,200
The organization can confuse, we have the platform

1898
01:08:41,200 –> 01:08:43,120
with we change the defaults.

1899
01:08:43,120 –> 01:08:45,920
So the listener should ask, where enforcement lives?

1900
01:08:45,920 –> 01:08:47,960
Which policies force region justification?

1901
01:08:47,960 –> 01:08:49,840
Which processes block always on waste?

1902
01:08:49,840 –> 01:08:51,960
Which procurement gates reject high emissions,

1903
01:08:51,960 –> 01:08:53,960
even when they’re cheaper or faster?

1904
01:08:53,960 –> 01:08:56,480
If the answer is we report it, then the control plane

1905
01:08:56,480 –> 01:08:58,040
is observational, not controlling.

1906
01:08:58,040 –> 01:09:01,240
And observational control planes are just telemetry.

1907
01:09:01,240 –> 01:09:04,080
Finally, the regulatory question, should governments regulate

1908
01:09:04,080 –> 01:09:06,880
big-tech sustainability claims more strictly?

1909
01:09:06,880 –> 01:09:09,640
The listener shouldn’t default to ideology here either.

1910
01:09:09,640 –> 01:09:12,320
Regulation is another control plane with the same failure

1911
01:09:12,320 –> 01:09:16,400
modes, weak definitions, easy loopholes, slow enforcement,

1912
01:09:16,400 –> 01:09:17,720
and metric gaming.

1913
01:09:17,720 –> 01:09:19,800
But without external constraint, net claims

1914
01:09:19,800 –> 01:09:21,200
become a private language.

1915
01:09:21,200 –> 01:09:23,520
Companies define boundaries, choose methodologies

1916
01:09:23,520 –> 01:09:24,880
and grade their own homework.

1917
01:09:24,880 –> 01:09:26,360
That’s not necessarily malicious.

1918
01:09:26,360 –> 01:09:29,200
It’s simply what happens when the same entity sets the rules

1919
01:09:29,200 –> 01:09:30,600
and benefits from the outcome.

1920
01:09:30,600 –> 01:09:32,360
So challenge the mechanics, not the marketing.

1921
01:09:32,360 –> 01:09:33,880
If the carbon control plane is real,

1922
01:09:33,880 –> 01:09:36,160
it should show up in changed architecture defaults,

1923
01:09:36,160 –> 01:09:38,120
constrained growth patterns, and budget decisions

1924
01:09:38,120 –> 01:09:39,120
that feel inconvenient.

1925
01:09:39,120 –> 01:09:41,600
If it doesn’t, it’s just a report with a better font.

1926
01:09:41,600 –> 01:09:43,000
Lessons for other businesses.

1927
01:09:43,000 –> 01:09:45,680
Stealing Microsoft’s mechanics, not its slogans.

1928
01:09:45,680 –> 01:09:47,720
So what does the listener do with all of this?

1929
01:09:47,720 –> 01:09:49,960
Besides, not grimly and forward the episode

1930
01:09:49,960 –> 01:09:53,160
to the one person who still thinks net zero is a logo.

1931
01:09:53,160 –> 01:09:55,880
They steal the mechanics, not Microsoft’s slogans,

1932
01:09:55,880 –> 01:09:58,720
not the aspirational language, not the 203 headline

1933
01:09:58,720 –> 01:10:00,200
that gets copied into a quarterly deck

1934
01:10:00,200 –> 01:10:01,520
and then quietly forgotten.

1935
01:10:01,520 –> 01:10:04,400
Mechanics are the only part that survives contact with budgets.

1936
01:10:04,400 –> 01:10:06,400
The first mechanic is measurable goals

1937
01:10:06,400 –> 01:10:08,600
that are defined like engineering requirements,

1938
01:10:08,600 –> 01:10:09,960
not like mission statements.

1939
01:10:09,960 –> 01:10:11,800
We care about the environment, isn’t a goal.

1940
01:10:11,800 –> 01:10:12,600
It’s a mood.

1941
01:10:12,600 –> 01:10:15,480
A goal has a boundary, a unit, and a review cadence.

1942
01:10:15,480 –> 01:10:17,840
If your sustainability target can’t be expressed

1943
01:10:17,840 –> 01:10:20,200
in the same language as uptime, cost, and risk,

1944
01:10:20,200 –> 01:10:22,800
it will never compete with uptime, cost, and risk.

1945
01:10:22,800 –> 01:10:24,840
The system will root around it automatically.

1946
01:10:24,840 –> 01:10:27,480
That distinction matters because organizations don’t fail

1947
01:10:27,480 –> 01:10:29,280
at sustainability from malice.

1948
01:10:29,280 –> 01:10:30,640
They fail from ambiguity.

1949
01:10:30,640 –> 01:10:32,520
Ambiguity is an entropy generator.

1950
01:10:32,520 –> 01:10:34,160
Every team interprets it differently

1951
01:10:34,160 –> 01:10:36,520
and drift becomes the default state.

1952
01:10:36,520 –> 01:10:38,920
The second mechanic is making carbon budget relevant.

1953
01:10:38,920 –> 01:10:40,960
You don’t need a carbon negative pledge to do this.

1954
01:10:40,960 –> 01:10:42,680
You need a way to make emissions show up

1955
01:10:42,680 –> 01:10:45,600
where decisions get made, cost management, procurement,

1956
01:10:45,600 –> 01:10:48,000
architecture review, and service ownership.

1957
01:10:48,000 –> 01:10:50,200
Microsoft uses an internal carbon fee.

1958
01:10:50,200 –> 01:10:52,920
Smaller organizations don’t need to copy the scale.

1959
01:10:52,920 –> 01:10:54,480
They need to copy the idea.

1960
01:10:54,480 –> 01:10:56,000
Attach consequence to a metric

1961
01:10:56,000 –> 01:10:57,800
therefore the metric stops being decorative.

1962
01:10:57,800 –> 01:10:59,800
If you can’t implement a real internal fee,

1963
01:10:59,800 –> 01:11:02,280
implement a shadow one, charge back as ideal.

1964
01:11:02,280 –> 01:11:04,640
Showback still works if leadership treats it as real.

1965
01:11:04,640 –> 01:11:05,920
The goal is not to punish.

1966
01:11:05,920 –> 01:11:08,160
The goal is to stop waste being invisible.

1967
01:11:08,160 –> 01:11:10,160
The third mechanic is accountability boundaries.

1968
01:11:10,160 –> 01:11:11,680
This is where most programs collapse.

1969
01:11:11,680 –> 01:11:13,200
You need to decide what you’re measuring

1970
01:11:13,200 –> 01:11:15,560
and then decide who owns the decision that drives it.

1971
01:11:15,560 –> 01:11:17,960
If you can’t name an owner, you don’t have governance.

1972
01:11:17,960 –> 01:11:19,200
You have reporting.

1973
01:11:19,200 –> 01:11:20,840
This is why scope three is such a trap.

1974
01:11:20,840 –> 01:11:22,920
It’s real, it’s huge, and it’s shared.

1975
01:11:22,920 –> 01:11:25,040
So you separate responsibility by control,

1976
01:11:25,040 –> 01:11:27,320
procurement owns supplier requirements,

1977
01:11:27,320 –> 01:11:30,040
engineering owns workload shape and runtime patterns,

1978
01:11:30,040 –> 01:11:33,000
facilities owns, PUE and cooling strategy,

1979
01:11:33,000 –> 01:11:36,440
finance owns incentives, sustainability owns methodology.

1980
01:11:36,440 –> 01:11:39,080
If you merge those into a single ESG team,

1981
01:11:39,080 –> 01:11:40,440
you didn’t centralize control.

1982
01:11:40,440 –> 01:11:41,640
You centralized blame.

1983
01:11:41,640 –> 01:11:43,960
The fourth mechanic is transparency cadence,

1984
01:11:43,960 –> 01:11:46,160
publish progress on a schedule and stick to it,

1985
01:11:46,160 –> 01:11:47,760
not because it’s virtuous,

1986
01:11:47,760 –> 01:11:51,000
because it’s a forcing function against internal story editing.

1987
01:11:51,000 –> 01:11:54,240
When you know you have to reconcile the numbers every year,

1988
01:11:54,240 –> 01:11:57,120
you build systems that can actually produce those numbers.

1989
01:11:57,120 –> 01:12:00,200
Transparency is a control mechanism for your own organization

1990
01:12:00,200 –> 01:12:01,800
and yes, it will be uncomfortable.

1991
01:12:01,800 –> 01:12:03,840
Good, comfort is the enemy of governance.

1992
01:12:03,840 –> 01:12:05,840
The fifth mechanic is treating sustainability

1993
01:12:05,840 –> 01:12:07,960
as risk management, not charity.

1994
01:12:07,960 –> 01:12:09,520
This is where the mental model shifts

1995
01:12:09,520 –> 01:12:11,240
for architects and executives.

1996
01:12:11,240 –> 01:12:14,120
If your workloads depend on regions with constrained power,

1997
01:12:14,120 –> 01:12:17,200
sustainability stops being ethics and becomes availability.

1998
01:12:17,200 –> 01:12:19,080
If your customers face mandatory reporting,

1999
01:12:19,080 –> 01:12:21,040
sustainability becomes sales friction.

2000
01:12:21,040 –> 01:12:24,520
If your supply chain can’t meet emissions disclosure requirements,

2001
01:12:24,520 –> 01:12:27,440
sustainability becomes procurement risk.

2002
01:12:27,440 –> 01:12:28,880
You are not doing good.

2003
01:12:28,880 –> 01:12:30,680
You are managing constraints that will hit you

2004
01:12:30,680 –> 01:12:32,920
whether you like the vocabulary or not.

2005
01:12:32,920 –> 01:12:35,280
Now, there’s also an uncomfortable architectural truth

2006
01:12:35,280 –> 01:12:37,120
that smaller organizations need to hear.

2007
01:12:37,120 –> 01:12:39,800
You can’t solve sustainability with a tool purchase.

2008
01:12:39,800 –> 01:12:42,400
You can buy platforms that help you model and report.

2009
01:12:42,400 –> 01:12:44,720
Fine, but tools don’t create discipline.

2010
01:12:44,720 –> 01:12:47,800
Discipline comes from default settings and enforcement points.

2011
01:12:47,800 –> 01:12:50,360
What gets approved, what gets funded, what gets reviewed,

2012
01:12:50,360 –> 01:12:51,280
what gets shut down.

2013
01:12:51,280 –> 01:12:53,360
So instead of shopping for sustainability software,

2014
01:12:53,360 –> 01:12:54,640
ask a better question.

2015
01:12:54,640 –> 01:12:57,920
Where in your delivery pipeline can you introduce one unit of friction

2016
01:12:57,920 –> 01:12:59,560
that prevents 10 units of waste?

2017
01:12:59,560 –> 01:13:01,040
That’s the control plane mindset.

2018
01:13:01,040 –> 01:13:03,680
One gate, one default, one budgeting rule,

2019
01:13:03,680 –> 01:13:05,120
one ownership requirement.

2020
01:13:05,120 –> 01:13:07,120
Small constraints applied consistently.

2021
01:13:07,120 –> 01:13:10,040
And if you want a place to start that doesn’t require a reog,

2022
01:13:10,040 –> 01:13:11,960
start where Microsoft started years ago.

2023
01:13:11,960 –> 01:13:14,760
Ty sustainability to procurement and consumption.

2024
01:13:14,760 –> 01:13:16,000
Make it part of the buying decision

2025
01:13:16,000 –> 01:13:17,080
and part of the running decision.

2026
01:13:17,080 –> 01:13:18,880
If it only lives in the annual report,

2027
01:13:18,880 –> 01:13:20,600
it will never touch the infrastructure.

2028
01:13:20,600 –> 01:13:23,240
Finally, don’t pretend you’re auditing morality,

2029
01:13:23,240 –> 01:13:24,560
you’re auditing flows.

2030
01:13:24,560 –> 01:13:26,240
Energy in, work out.

2031
01:13:26,240 –> 01:13:28,400
Emissions attached, responsibility assigned,

2032
01:13:28,400 –> 01:13:31,040
incentives applied, drift observed, corrections made.

2033
01:13:31,040 –> 01:13:32,400
That’s the whole game.

2034
01:13:32,400 –> 01:13:34,680
The organizations that win don’t care more,

2035
01:13:34,680 –> 01:13:36,800
they build systems that make the right behavior,

2036
01:13:36,800 –> 01:13:38,760
the easy behavior and the wasteful behavior,

2037
01:13:38,760 –> 01:13:39,960
the expensive behavior.

2038
01:13:39,960 –> 01:13:43,080
That’s what Microsoft’s carbon control plane is really trying to do.

2039
01:13:43,080 –> 01:13:44,560
And that’s the part you can copy

2040
01:13:44,560 –> 01:13:46,560
without having Microsoft’s budget,

2041
01:13:46,560 –> 01:13:49,920
Microsoft’s brand or Microsoft’s removal contracts.

2042
01:13:49,920 –> 01:13:51,400
One week implementation.

2043
01:13:51,400 –> 01:13:53,640
Adopt carbon per workload review.

2044
01:13:53,640 –> 01:13:55,600
If this episode stays theoretical,

2045
01:13:55,600 –> 01:13:58,680
it becomes another well-produced artifact that changes nothing.

2046
01:13:58,680 –> 01:14:01,760
The system will happily absorb it and keep behaving the same way.

2047
01:14:01,760 –> 01:14:03,480
So here’s the one week implementation.

2048
01:14:03,480 –> 01:14:07,600
One workload, one owner, one forcing question, no committee.

2049
01:14:07,600 –> 01:14:10,560
Pick a single production workload you actually care about.

2050
01:14:10,560 –> 01:14:11,960
Not the clean demo environment,

2051
01:14:11,960 –> 01:14:14,360
not the will-migrated next quarter zombie.

2052
01:14:14,360 –> 01:14:16,240
A real thing, customer-facing API,

2053
01:14:16,240 –> 01:14:17,480
internal line of business app,

2054
01:14:17,480 –> 01:14:20,600
a data pipeline, a team spot, whatever your organization depends on.

2055
01:14:20,600 –> 01:14:22,560
Now assign one accountable owner,

2056
01:14:22,560 –> 01:14:23,920
not the platform team.

2057
01:14:23,920 –> 01:14:26,360
One person who can answer questions and make changes.

2058
01:14:26,360 –> 01:14:27,560
If you can’t name an owner,

2059
01:14:27,560 –> 01:14:28,880
you’ve already found the problem.

2060
01:14:28,880 –> 01:14:31,840
The workload is unmanaged and everything else is theater.

2061
01:14:31,840 –> 01:14:33,440
Next, define the boundary.

2062
01:14:33,440 –> 01:14:35,800
You’re not building a perfect lifecycle assessment,

2063
01:14:35,800 –> 01:14:37,280
you’re drawing a box you can defend.

2064
01:14:37,280 –> 01:14:38,520
What services are in scope?

2065
01:14:38,520 –> 01:14:39,560
What region or regions?

2066
01:14:39,560 –> 01:14:40,720
What’s the runtime pattern?

2067
01:14:40,720 –> 01:14:41,760
What’s the traffic pattern?

2068
01:14:41,760 –> 01:14:43,600
You need a boundary because without one,

2069
01:14:43,600 –> 01:14:46,520
every question turns into an argument about what you didn’t include.

2070
01:14:46,520 –> 01:14:48,760
Then you collect four inputs that you already track

2071
01:14:48,760 –> 01:14:50,160
because this is the trick.

2072
01:14:50,160 –> 01:14:52,080
The goal is to piggyback on telemetry you have,

2073
01:14:52,080 –> 01:14:53,920
not invent a new bureaucracy.

2074
01:14:53,920 –> 01:14:55,520
First, cost per month.

2075
01:14:55,520 –> 01:14:57,080
Not because cost equals carbon,

2076
01:14:57,080 –> 01:15:00,040
but because cost is already the organization’s attention magnet.

2077
01:15:00,040 –> 01:15:01,640
You need it as the control variable.

2078
01:15:01,640 –> 01:15:03,240
Second, uptime and performance.

2079
01:15:03,240 –> 01:15:05,920
The carbon control plane does not get to break the service.

2080
01:15:05,920 –> 01:15:09,440
If your sustainability plan requires outages, it won’t ship.

2081
01:15:09,440 –> 01:15:12,240
Write down the S-loads that matter and keep them in frame.

2082
01:15:12,240 –> 01:15:14,080
So nobody plays games.

2083
01:15:14,080 –> 01:15:17,080
Third, deployment location, region availability zones,

2084
01:15:17,080 –> 01:15:19,160
multi-region topology if applicable.

2085
01:15:19,160 –> 01:15:21,520
This is where the conversation gets uncomfortable fast

2086
01:15:21,520 –> 01:15:25,600
because we always deploy there is not an architectural reason.

2087
01:15:25,600 –> 01:15:27,040
Fourth, runtime shape.

2088
01:15:27,040 –> 01:15:27,880
Is it always on?

2089
01:15:27,880 –> 01:15:28,720
Does it auto-scale?

2090
01:15:28,720 –> 01:15:30,280
Do you shut down non-prod at night?

2091
01:15:30,280 –> 01:15:31,920
Is batch scheduled intelligently?

2092
01:15:31,920 –> 01:15:35,600
Or does it run whenever someone remembered to click run?

2093
01:15:35,600 –> 01:15:38,320
Run time shape is where waste hides in plain sight.

2094
01:15:38,320 –> 01:15:39,920
Now you add the emissions estimate.

2095
01:15:39,920 –> 01:15:41,400
And yes, it will be imperfect.

2096
01:15:41,400 –> 01:15:42,280
That’s fine.

2097
01:15:42,280 –> 01:15:45,320
Imperfect numbers that change decisions beat perfect numbers

2098
01:15:45,320 –> 01:15:47,200
that arrive after the budget is spent.

2099
01:15:47,200 –> 01:15:48,720
Use a consistent method.

2100
01:15:48,720 –> 01:15:51,440
Energy as a function of compute time and power draw

2101
01:15:51,440 –> 01:15:53,160
adjusted by data center overhead,

2102
01:15:53,160 –> 01:15:55,280
adjusted by grid intensity for that region.

2103
01:15:55,280 –> 01:15:57,440
If you have vendor-provided estimates, use them.

2104
01:15:57,440 –> 01:15:59,200
If you have internal factors, use them.

2105
01:15:59,200 –> 01:16:01,000
The key is consistency over precision.

2106
01:16:01,000 –> 01:16:03,320
You’re trying to see direction and magnitude

2107
01:16:03,320 –> 01:16:04,960
not litigate decimals.

2108
01:16:04,960 –> 01:16:07,400
Once you have those inputs, ask the forcing question.

2109
01:16:07,400 –> 01:16:08,840
If carbon were priced like cost,

2110
01:16:08,840 –> 01:16:10,480
would you still deploy it this way?

2111
01:16:10,480 –> 01:16:13,280
That question is the entire control plane in one line.

2112
01:16:13,280 –> 01:16:16,360
It turns carbon from a value into a constraint

2113
01:16:16,360 –> 01:16:18,200
and constraints create architecture.

2114
01:16:18,200 –> 01:16:20,520
Now don’t let the conversation die in hand-waving.

2115
01:16:20,520 –> 01:16:22,840
You need one change you can implement in a week.

2116
01:16:22,840 –> 01:16:25,440
Pick from this list based on what your review revealed.

2117
01:16:25,440 –> 01:16:27,840
If the workload is always on for no good reason,

2118
01:16:27,840 –> 01:16:29,520
implement scheduling or auto-scaling

2119
01:16:29,520 –> 01:16:31,040
so idle time collapses,

2120
01:16:31,040 –> 01:16:34,040
the fastest carbon reduction is eliminating dead run time.

2121
01:16:34,040 –> 01:16:35,800
If the region choices habit,

2122
01:16:35,800 –> 01:16:37,560
do a region justification.

2123
01:16:37,560 –> 01:16:39,920
Write down the latency and data residency constraints,

2124
01:16:39,920 –> 01:16:42,400
then identify whether an alternate region meets them.

2125
01:16:42,400 –> 01:16:44,480
If two region satisfy requirements default

2126
01:16:44,480 –> 01:16:45,800
to the lower emissions option,

2127
01:16:45,800 –> 01:16:47,200
habit is not a requirement.

2128
01:16:47,200 –> 01:16:49,520
If the workload is oversized, write size it,

2129
01:16:49,520 –> 01:16:51,600
not by guessing, by measuring utilization

2130
01:16:51,600 –> 01:16:54,360
and choosing the smallest instance class that meets the SLO.

2131
01:16:54,360 –> 01:16:55,880
Over-provisioning is just waste

2132
01:16:55,880 –> 01:16:57,920
with a risk management story taped to it.

2133
01:16:57,920 –> 01:17:00,680
If batch jobs run at random times, schedule them.

2134
01:17:00,680 –> 01:17:02,280
This is where you can align runtime

2135
01:17:02,280 –> 01:17:03,960
with operational realities.

2136
01:17:03,960 –> 01:17:05,920
Maintenance windows, capacity constraints

2137
01:17:05,920 –> 01:17:07,920
and when possible cleaner grid windows.

2138
01:17:07,920 –> 01:17:09,440
Again, you’re not chasing perfection,

2139
01:17:09,440 –> 01:17:10,920
you’re shaping demand.

2140
01:17:10,920 –> 01:17:12,280
If the workload is chatty,

2141
01:17:12,280 –> 01:17:15,160
retries, polling, unnecessary data movement,

2142
01:17:15,160 –> 01:17:16,360
fix the behavior.

2143
01:17:16,360 –> 01:17:19,560
Network and compute inefficiency are carbon inefficiency.

2144
01:17:19,560 –> 01:17:22,000
They are the same bug just built in different units.

2145
01:17:22,000 –> 01:17:23,880
And before you stop, write a default policy

2146
01:17:23,880 –> 01:17:25,280
you’ll apply to new workloads.

2147
01:17:25,280 –> 01:17:27,400
Keep it simple so it survives adoption.

2148
01:17:27,400 –> 01:17:30,160
New production workloads require region justification

2149
01:17:30,160 –> 01:17:32,200
and the default is the lowest emissions region

2150
01:17:32,200 –> 01:17:35,280
that meets latency, availability and compliance needs.

2151
01:17:35,280 –> 01:17:38,080
That’s it, one line, one gate, a tiny amount of friction,

2152
01:17:38,080 –> 01:17:39,480
applied consistently.

2153
01:17:39,480 –> 01:17:42,200
Because the real enemy isn’t bad in 10, it’s drift.

2154
01:17:42,200 –> 01:17:44,360
Drift happens when defaults are convenient

2155
01:17:44,360 –> 01:17:46,200
and consequences are invisible.

2156
01:17:46,200 –> 01:17:48,280
This one week review forces visibility,

2157
01:17:48,280 –> 01:17:51,040
then forces one decision that makes the next decision easier

2158
01:17:51,040 –> 01:17:53,600
to do this once and you’ll learn something annoying.

2159
01:17:53,600 –> 01:17:56,800
Most sustainability strategy is just backlog work.

2160
01:17:56,800 –> 01:17:58,280
Nobody was prioritized to do.

2161
01:17:58,280 –> 01:18:01,000
Put carbon next to cost and suddenly it becomes real work

2162
01:18:01,000 –> 01:18:01,760
with real owners.

2163
01:18:01,760 –> 01:18:02,600
That’s the point.

2164
01:18:02,600 –> 01:18:05,000
The carbon control plane is behavioral design.

2165
01:18:05,000 –> 01:18:07,600
The carbon control plane works only when it changes defaults

2166
01:18:07,600 –> 01:18:09,520
and budgets, not when it decorates a report

2167
01:18:09,520 –> 01:18:10,840
with better vocabulary.

2168
01:18:10,840 –> 01:18:14,600
If you want to make this real, run the one workload carbon review

2169
01:18:14,600 –> 01:18:16,520
this week and see what decision you’d change

2170
01:18:16,520 –> 01:18:18,080
if carbon behave like cost.

2171
01:18:18,080 –> 01:18:19,240
If you want the next layer,

2172
01:18:19,240 –> 01:18:21,400
how to spot greenwashing versus real governance,

2173
01:18:21,400 –> 01:18:23,800
watch the next episode in this series and subscribe,

2174
01:18:23,800 –> 01:18:24,920
so you don’t miss it.