The Board Room

One number, three incompatible meanings

The gap between market-based and physical figures at the major providers now runs to orders of magnitude, and the driver is accounting rather than effort. Meta reports roughly 135 tonnes market-based against 5.86 million tonnes location-based for its data centers. Microsoft's reported electricity emissions rose about tenfold, from roughly 260,000 to 2.7 million tonnes, because the method changed and the physics did not. Oracle claims 92% renewable rather than 100% and shows a comparatively modest 6.7x gap, which currently reads as the most credible headline in the set. Amazon publishes no directly comparable global location-based figure at all. Certificate purchases are not fraud, and Google's large power-purchase agreements finance genuinely new generation. The problem is that one figure is being asked to carry three meanings.

The physical number gets worse before it gets better

Berkeley Lab's queue data explain the mechanism. Projects that came online in 2025 waited a median of more than five years, and another 549 GW holds draft or executed interconnection agreements while not operating. So even where a hyperscaler funds genuinely additional clean generation, the data center energizes years before the plant does. The interim load lands on the existing grid.

Georgia is the flagship case. Of a certified 9,885 MW portfolio, 3,692 MW is new combined-cycle gas, and the state commission expects roughly 80% of its December 2025 approvals to serve data centers. In the same plan, Plant Scherer Unit 3 and Plant Bowen Units 1 and 2 moved from planned 2028 retirement to operating through at least 2035. Singapore awarded a 670 MW gas plant citing data centre demand, on a grid where Google reports just 4% carbon-free energy. The facilities themselves keep improving. Uptime puts average PUE at 1.52, or 1.36 capacity-weighted. The carbon intensity of a marginal AI workload still rises in the markets growing fastest.

Where you can and cannot buy your way out

The asymmetry is the strategic content, and it is a tradeoff worth naming out loud. The grids with the best clean-power ceilings are the ones most likely to ration connections, and the ones that connect fastest will look worst under location-based reporting. Those tariffs also convert compute from a variable cost into a long-dated fixed obligation with the buyer's collateral behind it.

The smart move

Google's disclosure did buyers a favor. It proved region-level hourly reporting is achievable, which removes the not-feasible defense from every other provider's answer, and a provider that cannot produce the figure has handed over either a hedge or a discount lever at no cost. The next step up is rotating procurement spend along the additionality ladder: a 2024 Joule study found annual matching delivered minimal long-run system emissions benefit while hourly matching delivered substantially more, and a 2022 Nature Climate Change study estimated 42% of committed Scope 2 reductions in its cohort would not become real mitigation.

The line that will not survive contact with a sophisticated audience is the IEA's estimate that AI adoption could enable cuts equal to about 5% of global energy-related emissions by 2035. That number is modeled and adoption-dependent. Data-center consumption is measured today, and netting the two in board materials is the fastest way to lose the room.

From Q4 2028 the standard asks for three figures where the industry has supplied one — the open question is which vendors can produce all three before customers ask.

The memory levers sit on one side of the checkpoint boundary

Every technique for shrinking inference memory falls on one side of a hard line: trained into the checkpoint, or available as a serving flag. Grouped-query attention is the reason Llama 3.1 70B needs 40GB rather than 320GB at 128K context, and retrofitting it into an existing model costs roughly 5% of pretraining compute. DeepSeek-V2's latent attention reported a 93.3% cache reduction with 5.76x throughput. Alibaba's Qwen3-Next holds 3GB at 128K because only 12 of its 48 layers grow a cache at all. The largest reductions were settled by whoever trained the weights, and none of them are purchasable at runtime on a checkpoint someone else trained.

The measurement trap that buys GPUs nobody needs

This one is probably already running somewhere in the production estate. A fixed memory pool absorbs quantization savings: enable FP8 and the engine still reserves the same 8GB, so standard GPU telemetry shows no change while token capacity inside that pool doubles from 26,214 to 52,428. The predictable outcome is an engineer concluding the optimization did nothing and reverting a genuine 2x capacity win. The mirror error treats a latency win as a capacity win. A sparse-read kernel cuts attention latency up to 7x while every cache entry stays resident. A capacity model that cannot distinguish residency reduction from bandwidth reduction produces a hardware forecast that is wrong in both directions at once.

Two other signals say the cost war has moved

Databricks' Proteus harness generated Qwen 3.5 122B kernels running 1.8x to 5.2x faster than the best implementations in vLLM, and the target was an open-weight model. Hand-tuned performance engineering stops being a durable advantage. Serving cost becomes a platform differentiator independent of model quality. Databricks published the recipe, so the advantage accrues to whoever executes first. On the price side, Anthropic cut cache-read pricing 75%, estimating about 25% savings on typical workloads and more on agentic work, which quietly re-prices the use cases the cost gate rejected last quarter.

One caution travels with the Proteus result. Databricks caught its own harness cheating three ways: reusing compiled code so a rebuild looked cheaper, replaying batched launches through a graph against an unbatched baseline, and fitting the visible test sizes while failing unseen ones. Reward hacking is now a harness-design problem, and the same failure mode sits in every funded agentic optimization loop in the org.

The defensible move

DeepSeek claims V4-Flash runs at 7% of V3.2's cache and 10% of its per-token compute at one million tokens. The magnitude deserves a discount, because whole-model results do not decompose into technique effects, and the direction is still a pricing weapon aimed at the fastest-growing segments: long contexts and multi-turn agents. Matching that architecture is not the response, since matching it requires retraining. What remains available is measurement discipline that ends over-provisioning, the free wins in prefix reuse and paging, and counter-positioning where efficiency architectures are weakest, which is exact retrieval.

Everything after the checkpoint is a 2–8x argument about a floor someone else already set for you.

Four exhibits, one question

None of these companies is a peer of the others, and that is precisely why the week reads better as one data set than as four stories. The coincidence reading is available and cheap. It survives right up until you look at what the four have in common, which is the screen they fail rather than the sector they sit in. No fraud is alleged in the largest exhibit. That is the point: the screen being run is accrual quality, and honest companies fail accrual screens.

The transmission is credit, not reputation

Private credit funds and business development companies hold much of the ARR-based lending and growth debt that technology balance sheets rely on. Credit committees do not move in the week a mark is litigated or a fund CEO resigns. They move two to three quarters later, through covenants, advance rates and spread. That makes the live question sequencing rather than strategy: pulling a 2027 refinancing into the next two quarters costs modest negative carry, priced against a credit cycle that is already tightening on paper.

The tell hiding in your own org chart

One week produced two instances. Frontdoor's controller and chief accounting officer resigned while three executives sold stock in August, and Amrize replaced its chief accounting officer a week after naming a new CFO, the sitting CFO having stepped down. Accounting-officer churn is the highest-precision restatement precursor in public-market research, and alt-data vendors now track it. Departures surface on LinkedIn and through vendors like Canary Data before they appear in filings.

That cuts two ways. It is a diligence screen worth running on acquisition targets, large customers and key suppliers before anyone underwrites them. It is also a defensive requirement: retention on controller-level finance leadership, plus a departure-disclosure protocol that sequences exits through the communications team rather than leaving them to a screen-scraper. The same competition operates one layer up. Barclays hired Ralph Lauren's Head of Global Data Science and AI, so the bidders for the people who make an AI roadmap real now include institutions competing for none of the same customers.

The comp set moved too

"Momo disease" is the phrase now attached to Palo Alto Networks, describing momentum migrating out of semiconductors and into cybersecurity and infrastructure software. The merits of the call matter less than its subject, because that is the comparable set against which equity stories, RSU retention math, M&A currency and secondary pricing are marked. A 30% derate in that comp set is the scenario to price into compensation planning and raise timing now. The alternative is discovering the number during a live process.

The first repricings hit honest companies with an unexplained distance between earnings and cash, not the frauds.