Anthropic has assembled a huge compute pipeline. The expensive part is the capacity it needs before that pipeline arrives.
The most expensive compute in AI may be the compute you realize you need two years too late.
In May, Anthropic announced that it had secured all of the compute capacity at SpaceX’s Colossus 1 data center, giving it more than 300 megawatts of capacity and more than 220,000 Nvidia GPUs within a month. Later reporting put the price at $1.25 billion per month through May 2029.
Annualized, the math is brutal:
$1.25 billion × 12 months ÷ 0.3 gigawatts = approximately $50 billion per gigawatt-year.
That annualized figure is the shorthand behind this headline. It does not mean Anthropic pays $50 billion per gigawatt across its entire infrastructure portfolio. The SpaceX agreement is short-notice bridge capacity, and bridge capacity is the point. Anthropic has plenty of large, potentially more economical compute deals coming. Much of that capacity arrives later.
OpenAI faced the same fundamental problem earlier and made a different bet. It committed aggressively before the shortage became obvious, accepted the risk that it might buy too much, and gained years of capacity certainty. Anthropic managed the overbuilding risk more cautiously. Demand then accelerated faster than its infrastructure could arrive.
OpenAI bought time before compute became scarce
In January 2023, OpenAI and Microsoft expanded their partnership. Microsoft increased its investment in specialized Azure supercomputing systems, and Azure remained the exclusive cloud provider for OpenAI’s research, API and products.
The most valuable part of that agreement may have been priority. Microsoft had a financial incentive to design infrastructure around OpenAI’s workloads, finance the buildout and make the capacity available as ChatGPT demand grew.
OpenAI later reported that its available compute increased from 0.2 GW in 2023 to 0.6 GW in 2024 and roughly 1.9 GW in 2025. Over the same period, it said annualized revenue grew from $2 billion to more than $20 billion. The figures are company-reported, but they show the relationship clearly: more available compute translated into more products served, more users supported and more revenue captured.
By the time the rest of the market fully understood the scale of inference demand, OpenAI already had an operating base. It could diversify from that base rather than shopping for every new megawatt under emergency conditions.
Figure 1. Announced capacity and delivery windows. The largest distinction is often when the compute arrives.
Anthropic made the rational opposite bet
Anthropic was hardly asleep at the wheel. In September 2023, Amazon agreed to invest up to $4 billion, made AWS Anthropic’s primary cloud provider for mission-critical workloads and gave the company access to Trainium and Inferentia chips. That was a serious early infrastructure partnership. It simply came without a publicly disclosed gigawatt ramp comparable to the capacity OpenAI later reported.
Anthropic also had a coherent reason to avoid enormous speculative commitments. Data centers take years to build, while AI revenue is extraordinarily difficult to forecast. Buying too much can turn a one-year forecasting error into a balance-sheet disaster.
Dario Amodei made that argument publicly in December 2025. He warned that some AI companies were “YOLOing” their compute spending and said Anthropic was taking a more controlled path. His warning was financially sensible. The catch is that underbuying carries its own penalty.
When demand outruns committed infrastructure, the company has three choices: throttle customers, delay products or buy whatever capacity can arrive immediately. Anthropic appears to have experienced all the pressure points. Its April 2026 Amazon announcement said rapid enterprise, developer and consumer growth had strained reliability and peak-hour performance.
The long-term pipeline is enormous
Anthropic is now building a diversified fleet across Amazon Trainium, Google TPUs and Nvidia GPUs. The announced pipeline includes:
More than $100 billion committed to AWS technologies over ten years, securing up to 5 GW. Anthropic says significant capacity is arriving during 2026 and nearly 1 GW should be online by year-end.
A Google Cloud expansion worth tens of billions, initially expected to bring well over 1 GW online during 2026.
A later Google and Broadcom agreement summarized by Anthropic as 5 GW beginning to come online in 2027.
$30 billion in Azure compute and up to 1 GW of Nvidia-based capacity through Microsoft.
$50 billion in custom U.S. data-center infrastructure with Fluidstack, with sites coming online through 2026.
A multi-year CoreWeave agreement that begins bringing additional production capacity online later in 2026.
Anthropic’s official announcements describe up to 5 GW from Amazon, well over 1 GW from its initial Google expansion, 5 GW from its later Google-Broadcom program, up to 1 GW through Microsoft and Nvidia, and additional Fluidstack and CoreWeave capacity.
The portfolio shows a serious compute strategy whose demand curve arrived before several major pieces of the plan.
Figure 2. Selected commitments. Conditional and rumored arrangements are visually separated from definitive deals.
Only a few public agreements disclose enough information to attempt a rough cost-per-GW comparison. Even then, the products are different. A cloud-services contract can include networking, storage, CPUs, software and support. A data-center investment includes construction. A custom-silicon commitment ramps over time. Any single number should be treated as a directional indicator, rather than a clean commodity price.
With that caveat, the available math is still striking.
The AWS result is only a lower-bound normalization because the agreement ramps toward “up to” 5 GW and includes more than raw accelerator capacity. The Oracle figure combines OpenAI’s official 4.5 GW Stargate announcement with the reported $300 billion, five-year purchase commitment. The SpaceX result reflects immediate access to an already operating facility.
This makes the SpaceX lease especially revealing. Anthropic is paying nearly four times the annualized Oracle figure for capacity that can improve Claude limits immediately. Availability has become a product feature, and the premium is the cost of getting that feature now.
Figure 3. Rough annualized comparison for the few deals with disclosed price, capacity and duration.
The rumored Meta deal fits the same pattern
On July 17, Reuters reported that Meta and Anthropic were in early discussions over a potential $10 billion compute lease lasting up to two years. Anthropic reportedly proposed the arrangement, payments would be monthly, either party could exit early, and the talks may never produce a final agreement. No capacity number has been disclosed.
Without megawatts, no honest per-GW price can be calculated. The sensitivity range shows why the missing denominator matters:
At 100 MW, $10 billion over two years equals $50 billion per GW-year.
At 300 MW, it equals about $16.7 billion per GW-year.
At 1 GW, it equals $5 billion per GW-year.
The structure matters even before the final price is known. Anthropic is reportedly considering a short-duration lease from a company that has never operated as a conventional public cloud provider. That looks less like ordinary capacity planning and more like a market in which every available cluster has become negotiable.
How OpenAI avoided the same gap
OpenAI did not avoid expensive compute. Its reported Oracle commitment is colossal, and its later infrastructure portfolio includes a $38 billion AWS agreement, up to $22.4 billion of CoreWeave contracts, a 6 GW AMD deployment and a 10 GW Broadcom custom-accelerator program. Some of those arrangements are still years from full deployment.
If this helped you see the signal more clearly, subscribe to Staticbreaker for defensible AI interpretation for busy, intelligent non-specialists: https://thestaticbreaker.com/subscribe
Its advantage came from sequencing. OpenAI had meaningful capacity operating before it began stacking the next wave of suppliers. Microsoft carried much of the early infrastructure burden. Stargate, Oracle, AWS, AMD, Broadcom and CoreWeave then expanded a platform that already existed.
Anthropic’s sequence was tighter. Its largest disclosed capacity additions are arriving during 2026 and 2027, while Claude demand is already straining the system. The company is therefore combining economical long-term custom silicon with very expensive short-term GPU access.
OpenAI paid for the possibility of overcapacity. Anthropic is paying for the certainty of undercapacity.
One accounting trap: the giant GW numbers overlap
The industry’s headline numbers cannot simply be added. OpenAI’s 10 GW Stargate target, 4.5 GW Oracle buildout, 6 GW AMD agreement and 10 GW Broadcom program describe different layers of the stack. AMD and Broadcom systems can be installed inside Stargate, Oracle, Microsoft, AWS or partner facilities. Adding every announcement would count the same physical power more than once.
The same caution applies to Anthropic. Its 2025 Google TPU expansion and later Google-Broadcom agreement may represent overlapping phases of a larger program. “Up to” capacity may never be consumed in full. Investment amounts can include equity, construction, networking, software and services beyond GPU or TPU time.
That uncertainty prevents a definitive company-wide average price per gigawatt. It does not erase the immediate-capacity signal. The few bridge deals with visible terms are substantially more expensive than long-horizon commitments.
The compute moat is measured in calendars
Anthropic may close this gap. Nearly 1 GW of new Amazon capacity is expected by the end of 2026, Google capacity is ramping, and the broader 5 GW Google-Broadcom program begins in 2027. If those systems arrive on schedule and deliver the expected economics, today’s emergency leases may become a temporary footnote.
But temporary constraints can have permanent competitive consequences. Rate limits push developers toward other APIs. Reliability issues weaken enterprise confidence. Delayed model releases surrender attention. In a market growing this quickly, a year of constrained supply can shape customer habits for much longer than a year.
OpenAI’s early compute commitments became more than an infrastructure decision. They became permission to ship, permission to scale and permission to capture demand while competitors waited for substations, cooling systems and accelerators to arrive.
In the AI race, delivery dates may matter as much as the chips themselves.
Methodology note: This analysis uses public company announcements and reported contract terms available through July 17, 2026. “GW” refers to stated power or system capacity, rather than a standardized quantity of useful model output. Annualized comparisons are rough headline normalizations and should not be read as audited unit economics.
Image captions: Figure 1. Announced capacity and delivery windows. Figure 2. Selected commitments. Figure 3. Rough annualized comparison.