SemiAnalysis Breaks Down the AI Compute Shortage: OpenAI and Anthropic Are Snapping Up All New Global Compute Capacity

Author: Wall Street View

On November 3, Dylan Patel, founder and CEO of semiconductor research firm SemiAnalysis, delivered a speech at Runtime 2026, focusing on compute, capital, and data centers.

He stated that OpenAI and Anthropic are heavily pre-book computing capacity. Taking 2027 as an example, the two companies will account for between 25% and 40% of new compute additions. Under demand squeeze, NVIDIA H100 rental prices have rebounded significantly from their lows, and future B300 rents could surge above $7 per hour.

The funding side is equally tight: U.S. Treasury yields have risen sharply over the past two months, with the 10-year yield climbing above 5% in early October, making it increasingly difficult for emerging cloud providers reliant on borrowed funds for data center construction to raise capital.

This speech came a week after Anthropic publicly filed its S-1 registration statement, bringing the compute accounting behind a top-tier AI lab into view. Patel noted that NVIDIA has stepped in to underwrite compute clusters totaling approximately 6.5 gigawatts, yet this still cannot alter the overall tight supply-demand dynamic.

Majority of New Compute Capacity Pre-Booked by Two Labs

The root cause of compute strain lies in buyer concentration. Global total compute capacity has grown rapidly: reaching about 60 gigawatts by the end of 2024 and exceeding 100 gigawatts by year-end this year, nearly doubling in two years. Patel said, "Global compute capacity stood at 60 gigawatts by the end of 2024 and has exceeded 100 gigawatts by the end of this year, nearly doubling in just two years."

The largest buyers of incremental capacity are OpenAI and Anthropic. Their combined compute holdings stood at roughly 4 gigawatts at the end of last year, about 10 gigawatts this year, and will exceed 20 gigawatts next year. Looking at increments alone, these two companies will absorb between 25% and 40% of new compute additions combined in 2027.

Patel estimated, "They are grabbing the 25%, 30%, and 40% shares of new compute capacity. At the end of last year, revenues for OpenAI and Anthropic began soaring with the launch of Claude Code, prompting them to frantically purchase compute, while Meta also started buying aggressively. For many now, acquiring compute capacity has become virtually impossible."

Another notable detail: this compute is primarily used for training rather than serving users (inference). Patel mentioned that Anthropic currently holds 4 gigawatts, while the entire Claude Code project only consumes tens of megawatts of it.

Figure: Forecast of global data center and top-tier lab compute demand from 2024 to 2028

Robust Inference Profits for Open-Source Models Drive Rental Price Surges

With more buyers, prices naturally rise. Patel believes another force driving up rental rates is the exceptionally high profit margin inherent in selling tokens (the metric unit for model outputs), meaning there are far more willing to rent GPUs at a premium than those willing to sell.

His calculations are per megawatt. Building and depreciating a B300 cluster along with five years of operations costs roughly $9 million annually; renting from a cloud provider might cost you around $14 million a year if you're lucky. Yet, leasing it out for open-source model inference generates profits far exceeding that figure. Even at a mere 30% utilization rate, returns remain substantial.

These estimates are based on SemiAnalysis's open-sourced InferenceX benchmark, utilizing containers downloadable by anyone and relying solely on two open-source inference frameworks, vLLM and SGLang, without any custom optimizations.

Price data aligns with this logic. Patel noted that last October marked a low point, with H100 rents around $1.70 and as low as $1.45; current transaction prices range between $2.40 and $3.20. Signing at these rates would have been considered unbelievable two years ago, but now people are rushing to lock in contracts within this bracket.

SemiAnalysis's own one-year contract index reached $2.35 in March this year, while listing prices from vendors like Nebius continued climbing in October. Patel's discussion primarily covers short-term and secondary market prices, excluding bulk deals from hyperscale cloud providers.

Looking ahead, he predicts, "I wouldn't be surprised if H100 climbs above $3 and B300 surpasses $7. If you can find B300 at $6 per hour, you should sign the deal and start selling tokens; that is the right move."

The obstacle lies in upfront payment requirements imposed by cloud providers, which poses even greater challenges for startups: securing one-year contracts is extremely difficult, and startups can hardly guarantee their own survival a year from now. Regarding closed-source models, he stated that Anthropic has already sacrificed API gross margins in pursuit of scale.

Figure: Quarterly compute load breakdown for Anthropic and OpenAI

Figure: One-year contract trend for NVIDIA H100 GPU rental rates

Delivery Bottlenecks and Divergence in Cloud Provider Service Reviews

Securing compute capacity is merely the first step; the true value hinges on cluster performance. SemiAnalysis conducted a review called ClusterMAX, tracking 323 cloud suppliers and evaluating 77 of them. Ultimately, only 19 awards were issued, with the vast majority rated as "unusable" or "underperforming."

In Patel's view, the top performers are CoreWeave and Nebius, but both suffer from poor availability: neither will sell compute for the next six months as they are sold out. Nebius requires upfront payment for either a full year or six months; despite the high price, it remains one of the few ways to secure capacity.

Oracle and Google also earned excellent ratings, with Google upgrading from a bronze award in the initial round. More downwards adjustments occurred: as evaluation criteria tightened, Azure was downgraded to silver, while Together and Lambda dropped to bronze. AWS consistently lagged behind, hampered primarily by overly burdensome delivery processes; obtaining a cluster requires navigating extensive multi-stakeholder coordination, leading to prolonged timelines and complex network configurations.

Security flaws were the most glaring aspect of the reviews. SemiAnalysis deployed AI models to test attacks against these cloud platforms, discovering vulnerabilities already publicized online that the models could directly exploit to access other clients' storage.

During one test, they used an open-source model to gain full Slurm control over an entire cluster and spotted a national defense department's name. Many cloud providers only discover security risks when prodded by clients, and some take days to respond even after being notified.

Figure: SemiAnalysis ClusterMAX rankings for compute cloud supplier delivery capability and security assessment

Rising Interest Rates and the Debt Dilemma of Compute Supply

Expanding compute supply demands massive capital, and the cost of that capital is rising significantly. Patel reiterated that global AI capital expenditure will reach approximately $11 trillion by the end of 2029, with roughly $7 trillion sourced from debt financing, mostly allocated to chips, networking, and data center construction.

Against this interest rate backdrop, lending flows show stark polarization. The 10-year U.S. Treasury yield climbed from 3.96% earlier this year to 5.31% in early October, causing AI-related debt issuance to cool considerably in September. Patel pointed out that if an emerging cloud provider aims to raise $2 billion to build a 1-gigawatt data center, securing financing in the current debt market is nearly impossible.

The only alternative is to pre-sell compute capacity to hyperscale cloud providers, who then turn to the bond market for financing. Although loans are secured, this compute capacity cannot enter the open market accessible to the broader industry.

Figure: Forecast trend of global AI capital expenditure and debt financing scale from 2024 to 2030

As a result, of the compute capacity contracted for next year, approximately 65% to 70% will flow to hyperscale cloud providers and top AI labs. For startups, renting a cluster requires a 50% deposit upfront, with deliveries taking three to nine months, making it extremely difficult to acquire compute resources.

NVIDIA's approach involves underwriting clusters, which Patel categorizes into three types: SoftBank Energy only underwrites land, power, and facilities for a 9-gigawatt site in Ohio expected to come online around 2030, exclusively reserved for OpenAI; Firmus involves NVIDIA committing to lease the cluster at a fixed price to ensure break-even operation, giving lenders the confidence to provide financing; the third category provides hardware residual value guarantees.

In terms of scale, NVIDIA has underwritten roughly 6.5 gigawatts, which will be sold to the open market rather than exclusively to hyperscalers. He sets an upper limit of 45 gigawatts, which, even then, constitutes only a fraction of new compute capacity added before 2030.

Concluding his speech, he emphasized the massive volume of token consumption: several top providers collectively process over 100 trillion tokens daily, with each company's annualized revenue climbing. He summarized that whether compute resources can achieve broader accessibility will determine the developmental vitality of the entire ecosystem and prevent the industry from becoming overly dependent on a handful of highly monopolistic labs.

Figure: Comparison of interest rates and rating spreads across different lending agreements among compute cloud providers