Microsoft Signs 10GW in Binding Datacenter Contracts for AI Inference

Full analysis

Microsoft has signed over 10 gigawatts of datacenter capacity in binding contracts this year, north of $300B in commitments, and SemiAnalysis says the driver is $100M per megawatt per year from serving OpenAI inference through Azure. The question for anyone building on this stack: is this real demand you can lean on, or option-value hoarding that reprices the moment utilization disappoints?

Reversibility. Mostly Type 1 for Microsoft (you don't unwind $300B of steel and power quickly), but the SpaceX slice with 90-day cancellation clauses is deliberately Type 2. That split is the whole story.

What's actually being decided. Not "will Microsoft build datacenters." It's whether inference demand curves are steep enough to justify treating capacity as base load instead of a bet. And, underneath, whether the April 2026 OpenAI renegotiation, which killed the 20% revenue share, turned Microsoft from a cautious lessor into a company whose only incentive is maximum throughput.

Forcing function. No clean one. The 2027 capacity ramp and Microsoft's next few earnings prints are where the projection gets graded.

The Skeptic. "Binding" in hyperscaler land means binding-with-exit-ramps, and the 90-day SpaceX cancellation clause is the tell. SemiAnalysis projects Azure growth from 42% to over 100% annually. Extraordinary, and the demand evidence to back it isn't in the piece. Do the arithmetic: $300B against $100M/MW/year implies roughly 3,000 MW of fully utilized capacity. Show me those customers. The December 2024 leasing pause was real, so a "dramatic reversal" narrative papers over whether demand actually arrived or whether the internal politics just flipped. For a PM: signing a lease is not the same as filling it, and nobody here has shown the seats are sold.

The Compute Pragmatist. 10GW is a geological feature. The entire US datacenter industry drew roughly 17GW in 2023. Committing 10GW in one year says inference, not training, now drives the capex, and that the utilization math assumes sustained load. But signed megawatts and energized megawatts live on different timelines. The gap between a contract and a powered rack, cooling, substations, interconnect queues, is exactly where the revenue ramp gets stress-tested. The SpaceX structure fits variable, location-dependent throughput: fine for batch inference, bad for real-time bidding-style latency. For a PM: they've promised the power. Whether the grid delivers it on schedule is a separate, harder problem.

The Researcher. The figure worth interrogating is $100M/MW/year, because if it holds it implies demand steeper than anyone modeled two years ago, and dropping the 20% revenue share changes the marginal economics of every additional megawatt. What's missing is which state that number lives in: realized ARR, contracted, or projected. Those are three different theses wearing one number. Copilot and Foundry API revenue are real but not obviously 3,000-MW real yet. For a PM: the same dollar figure can mean "money in the bank," "money we're owed," or "money we hope for," and this piece doesn't say which.

The Safety Lens. One company holding 10GW of inference and acting as the exclusive commercial pipe for the dominant frontier model is a concentration story regulators haven't priced. Incident response, model update rollout, abuse mitigation, all flow through one chokepoint. And the renegotiated deal removed the friction that a revenue share quietly imposed: Microsoft's incentive is now throughput, not caution. For a PM: when your model, your capacity, and your safety review all sit behind one vendor, an outage or a bad rollout there is your outage too.

The Enterprise Buyer. Full economics on inference cuts two ways for a buyer. Microsoft now has real skin in the game, which should mean faster capacity unlocks and less throttling during spikes. But it also means your inference bill funds a $300B build, and the SpaceX overflow could route your latency-sensitive workload through nodes with reliability characteristics that aren't in your SLA. Before signing a multi-year Foundry commit, I want geographic routing guarantees in writing and a clause that keeps my traffic out of surge capacity unless I opt in.

Where they split. The Skeptic and the Researcher agree the $100M/MW figure is unverified, but they diverge on what that means: the Researcher wants to know its state (contracted vs. projected) before judging, while the Skeptic treats the missing customer list as reason enough to discount the 100% growth claim outright. The Compute Pragmatist and the Enterprise Buyer clash on the SpaceX capacity: the Pragmatist sees a clever, no-balance-sheet-risk surge buffer, the Buyer sees an SLA landmine routed through his production traffic. And the Safety Lens sees a concentration risk exactly where everyone else sees efficiency, the removal of the 20% revenue share.

What it hinges on. Two beliefs. First, whether $100M/MW/year is realized or projected, because that single distinction separates "once-in-a-generation" from "expensive option on demand that may not show." Second, whether powered capacity actually lands on the 2027 timeline, since signed contracts with cancellation clauses are cheap and energized gigawatts are not. The council leans skeptical on the growth projection and pragmatic on the build: Microsoft is real, the $300B is real, the 100%+ Azure growth is a projection dressed in a lot of confidence.

If you're building on this stack, verify one thing before you plan around abundant capacity: get Microsoft to commit in writing that your inference stays in core Azure regions, not routed to cancellable surge nodes, and load-test what happens if it isn't.

Prediction: Azure's reported cloud revenue growth will not reach 100% year-over-year in any quarter through Microsoft's June 2026 fiscal-year-end earnings report; it will stay under 55%.

Confidence: High. 100%+ growth off Azure's base has no precedent and requires energized capacity that doesn't exist yet.

Why: SemiAnalysis projects Azure growth jumping from ~42% to over 100% annually, but that leap depends on 3,000+ MW of fully utilized capacity that is contracted, not yet powered, and the interconnect-and-cooling timeline for new gigawatts runs years, not quarters. Azure grew in the low-to-mid 40s recently, and even a demand surge can only be served by megawatts that are actually energized, so the revenue can't outrun the substations. The opposite outcome, a doubling of Azure growth within a few quarters, would require both instant capacity delivery and instant customer uptake at a scale no hyperscaler has shown, which is why it's the far less likely path.

Revisit by 2026-11-15: We're right if no reported Azure growth quarter through the next two earnings prints hits 100% and it stays under 55%. We're wrong if Microsoft reports Azure or its cloud segment growing at or above 100% year-over-year in any quarter in that window.