GPU Prices Surge 48% as Compute Scarcity Bites — Anthropic Gating Top Models to ~40 Orgs
Nvidia Blackwell GPU rental rates hit $4.08/hr this week — up 48% from $2.75 in February — as demand outstrips supply across cloud providers. CoreWeave has raised prices 20% and extended minimum contract lengths from 1 to 3 years. Anthropic has reportedly limited access to its newest model to approximately 40 organizations.
Original sourceFor the first time since the dot-com era, AI infrastructure is hitting genuine supply-chain limits. A new analysis from Tom Tunguz at Theory Ventures documents what industry insiders have been whispering about for weeks: GPU compute is becoming a seller's market with structural scarcity that won't resolve quickly.
**The price data.** Nvidia Blackwell H100-equivalent GPU rental on CoreWeave has risen from $2.75/hr in February 2026 to $4.08/hr this week — a 48% increase in under two months. CoreWeave simultaneously raised its list prices 20% and extended minimum contract durations from 1 year to 3 years, locking customers in for longer at higher rates. AWS and Azure have both seen similar upticks on spot instance pricing for A100/H100 inventory.
**The capacity crunch.** OpenAI CFO Sarah Friar acknowledged in an earnings call this week that the company is making "very tough trades" on compute allocation. OpenAI is not alone. Anthropic has reportedly limited access to its highest-capability new model to approximately 40 partner organizations, with a waitlist for everyone else. For most developers, this means the frontier is functionally inaccessible regardless of ability to pay.
**The energy bottleneck.** Underneath the GPU shortage is a harder constraint: power. New data center construction is stalling on grid interconnection queues that run 4-7 years in most U.S. markets. Microsoft has already cancelled some data center leases. Google has signed power purchase agreements with nuclear operators to secure dedicated generation. The compute scarcity of 2026 may look tame compared to 2028-2029 if energy permitting doesn't accelerate.
**What this means for builders.** Teams building on frontier APIs need contingency plans. Local model hosting — once a cost optimization — is increasingly a reliability hedge. The Ternary Bonsai and similar ultra-compressed model releases are arriving exactly when the cost of cloud inference is making developers receptive to them. Compute scarcity may accelerate the edge inference trend more than any technical breakthrough could.
Panel Takes
The Builder
Developer Perspective
“The 3-year minimum contract requirement from CoreWeave is the data point that changes my infrastructure planning. That's not a temporary squeeze — that's a structural shift. Teams that don't lock in compute now may find themselves priced out of frontier training runs for years.”
The Skeptic
Reality Check
“GPU price spikes have happened before and corrected quickly when supply caught up. Nvidia's fab capacity is ramping and TSMC has dedicated CoWoS packaging lines to AI chips. The 48% spike may reverse by Q4 2026 as Blackwell supply hits the market. Don't panic-buy 3-year contracts based on a 10-week price trend.”
The Futurist
Big Picture
“Compute scarcity is the best thing that could happen for the open-source model ecosystem. When frontier APIs are rationed and expensive, the economic incentive to run capable local models surges. This is the forcing function that will make edge inference mainstream — and it arrived faster than most predicted.”