The old assumption for technology procurement: sign a contract, and what you signed for arrives within a quarter or two. That assumption no longer holds for AI computing capacity, called compute, the processing power that runs AI workloads. One of the largest AI companies in the world signed five major compute deals this year. Most of that capacity does not exist yet. Some of it requires infrastructure that has not been built, and some requires government approval to proceed at all. The buyer behind those deals is not a startup. It is a company whose growth led it to tell paying customers directly, in April, that demand had degraded service during peak hours. If a buyer with that market presence waits years for capacity, a mid-size business cannot plan as if it will move faster. The changed decision: treat a signed compute contract as a planning assumption, not a confirmed resource, unless the capacity behind it is already online.

Money Is Committed. Power Arrives Later.

Five deals, announced across this year, illustrate a single pattern: capital committed now, capacity delivered later. For any business whose vendor's service roadmap depends on infrastructure being built, that distinction determines what is actually available to run your AI services today.

Anthropic has committed more than one hundred billion dollars over ten years to AWS, Amazon's cloud, securing up to five gigawatts of new capacity, power measured in billions of watts. [1] It said in April that nearly one gigawatt of that capacity would be running by the end of this year. [1] The remainder arrives over the life of the agreement.

A separate agreement with Google and Broadcom brings multiple gigawatts of TPU capacity, Google's own AI chip, online starting in twenty twenty-seven. [2]

Reuters reported that Anthropic will spend forty-five billion dollars to rent computing power from Nscale's West Virginia campus, according to a person familiar with the matter. [3] Reuters also reported a thirty-five billion dollar agreement with the cloud firm Lambda for a Texas site of roughly three hundred fifty megawatts, citing a person familiar. [4]

Reuters reports the planned campus near Brisbane would draw two point one six gigawatts. [5] Reuters reports that facility is planned to start coming online in twenty twenty-seven, subject to Australian government approval. [5]

Across all five deals, the capital committed is substantial. The capacity running today is a fraction of what has been announced. Announcement and operation are different events with different timelines.

Announced Capacity and Online Capacity Are Not the Same

Two of the five deals sit closer to usable capacity than the others. The Nscale agreement rents power from an existing West Virginia campus. Contracting with running infrastructure is different from contracting with a facility that will be built.

The Brisbane campus and the Google TPU expansion are not in that category. Both come online later, and Brisbane remains subject to Australian government approval. Both are real commitments backed by real capital. Neither is available to buyers who need capacity today. The Lambda site in Texas is still being developed, Reuters reports, which places it between those two groups. [4] A vendor who prices services on announced capacity is pricing on a promise. A vendor who prices on confirmed, operational capacity is pricing on a fact.

The company making all five commitments is not a marginal buyer. Anthropic says its run-rate revenue, meaning the annualized pace of current sales, has passed thirty billion dollars, up from roughly nine billion at the end of the prior year. [2] It reports that more than one thousand business customers now each spend over one million dollars a year. [2] A company at that revenue level and with that customer base carries enormous negotiating weight. It still told those paying customers in April that growth had degraded reliability at peak hours. The wait is not a failure of negotiating skill. It is the structure of the market, and it applies regardless of how much leverage a buyer brings.

The Default That No Longer Works

Technology procurement worked on a predictable cycle for decades. Sign a contract, and delivery arrived within a quarter or two. Server provisioning, software licensing, cloud platform onboarding: delivery happened close to signing, not years after it.

AI computing capacity operates on a different schedule because it requires physical infrastructure. Power grid connections, cooling systems, data center construction, and specialized chips that take months to manufacture are all rate-limiting steps. A new facility committed today takes years to open. That gap between announcement and operation is not an exception. It is the normal construction timeline for data center infrastructure.

A reasonable counterargument: capital eventually becomes capacity. Every one of these deals funds real construction. By the time these facilities open, the constraint may have eased and today's bottleneck may look like a short-term episode. That is possible. Next quarter's AI workloads cannot run on supply that has not arrived.

One question a reasonable manager asks next: if frontier labs are absorbing all available compute, why would any cloud provider sell capacity to a mid-size business at all? Mid-size businesses buy from a different pool. The frontier training pipeline, which powers the largest experimental workloads, is separate from the commercial cloud pools that businesses rent for everyday AI services. The squeeze reaches a business as higher prices, tighter usage quotas, slower rollouts of new model features, and longer waits for upgrades, not as a refusal. Capacity exists. The conditions around it have tightened.

Adjustments for This Quarter

The response is a budget and planning problem, not a technology problem.

  • Separate committed capacity from usable capacity in your budget. Agreements that draw on existing, operating infrastructure belong in one line. Agreements that depend on facilities not yet built or not yet online belong in a separate line, flagged as a planning assumption rather than a confirmed resource.
  • Model next year's AI usage against capacity your vendors hold today. Ask directly: which service commitments draw on data centers already running, and which depend on facilities not yet online? A vendor who cannot answer that question is giving you a roadmap, not a delivery guarantee.
  • Assign a named owner to any vendor commitment that depends on capacity not yet available. That owner tracks construction milestones, monitors announcements for slippage, and maintains a fallback plan built on confirmed supply.
  • Keep current AI rollouts sized to run on existing capacity. Scale the plan to match confirmed supply, not announced supply.

The Operating Consequence

A signed contract tells you a commitment was made. It does not tell you the power is on. Build your AI plans around what your vendors can deliver from infrastructure that exists today. Treat capacity from future facilities as a potential addition to a working plan, not a precondition for starting one. The AI work that earns business value this year runs on power that is on today. That is the constraint to plan around.