- GPU rental prices have surged 48% in two months — Nvidia’s top Blackwell chip now costs $4.08/hour on spot markets, up from $2.75 — as AI demand far outpaces infrastructure buildout
- Anthropic hit a 98.95% uptime rate over the last 90 days — well below the “four nines” standard — as frequent outages push enterprise clients to switch to rivals
- Anthropic’s revenue nearly tripled in three months: from $9B annualized run rate at year-end 2025, to $14B by February, to $30B by April — the demand surge is causing the supply crisis
- OpenAI scrapped its Sora video app to free computing capacity for coding and enterprise products, while token usage on its API more than doubled — from 6B to 15B per minute — since October
What Happened?
The AI industry is running headlong into a compute wall. Explosive demand for “agentic” AI — autonomous tools that independently execute complex tasks like writing code, scheduling appointments, or managing workflows — has consumed computing resources faster than any company can provision them. Anthropic, whose revenue surged from a $9 billion annualized run rate at year-end 2025 to $30 billion by April, has responded by rationing token usage during peak hours (5 a.m. to 11 p.m. PT on weekdays), triggering user backlash. Meanwhile, GPU hourly rental prices on spot markets jumped 48% in two months, data-center build times stretch years into the future, and all available power capacity through 2026 is already spoken for, according to cloud infrastructure CEO J.J. Kardwell of Vultr.
Why It Matters?
This crunch is a structural threat to the AI adoption flywheel. The entire premise of frontier AI’s business model depends on rapidly converting users into paying customers and then expanding their usage. But if the product goes down constantly — Anthropic’s API achieved only 98.32% uptime in March, well below the 99.99% “four nines” standard that enterprise software customers expect — those customers will switch. Several already have: enterprise clients have begun migrating from Anthropic to OpenAI specifically because of reliability failures. The deeper problem is that the bottleneck cannot be solved quickly. Data center construction takes years, power grid connections require regulatory approvals, and GPU lead times from Nvidia extend months out. Higher prices are the classical response to supply crunches, but in a ferociously competitive market where OpenAI, Google, and Anthropic are all fighting for the same users, price hikes risk accelerating churn.
What’s Next?
CoreWeave has already raised prices more than 20% and is locking smaller customers into three-year contracts — a sign that cloud providers know the crunch is structural, not temporary. Bank of America analysts project demand will outstrip AI compute supply through at least 2029. For frontier AI labs, the race is now as much about securing compute as it is about building better models. Companies that lock in capacity — through owned data centers, long-term GPU contracts, or exclusive chip partnerships — will have a durable competitive advantage over those that rely on spot markets. For enterprise customers, the lesson is clearer still: AI reliability cannot yet be taken for granted, and redundancy planning is no longer optional.
Source: The Wall Street Journal
