Through the first half of 2026 the money and hardware behind AI reached a scale that reshapes who can compete. OpenAI closed a round it describes as $122 billion in committed capital, hyperscalers and labs signed multi-gigawatt data-center deals, and NVIDIA moved its next-generation Vera Rubin platform into production. The common thread is compute: frontier AI now runs on capital commitments that only a handful of companies can make.

What happened

On the capital side, OpenAI states it closed its latest round with $122 billion in committed capital at an $852 billion post-money valuation, co-led by SoftBank with a16z, D. E. Shaw, MGX, and TPG, and says it is generating around $2 billion in revenue per month. (OpenAI’s site blocks automated retrieval; these figures are quoted from OpenAI’s own announcement.) Anthropic raised $65 billion at a $965 billion valuation the same period, and Databricks completed roughly $5 billion of equity at a $134 billion valuation on 9 February 2026.

The compute deals were as large as the raises. CoreWeave and Meta expanded an AI-cloud agreement to about $21 billion through December 2032 (9 April 2026). NVIDIA took equity stakes in its own cloud customers, including $2 billion in Nebius (11 March) and a share right of up to $2.1 billion in IREN (7 May), a pattern that has drawn scrutiny as circular financing. On hardware, NVIDIA launched the seven-chip Vera Rubin platform at GTC on 16 March and reported it ramping into full production by 31 May; AMD and Meta agreed to deploy 6 gigawatts of AMD GPUs (24 February); and SK hynix shipped first 12-layer HBM4E memory samples on 18 June.

Why it matters for builders

You do not need a gigawatt data center to build on AI, but the buildout shapes what you build on. Three effects matter. First, capacity: the gigawatt deals are why frontier APIs can serve the traffic they do, and why new capacity keeps arriving. Second, concentration: when training a frontier model requires tens of billions in compute, the number of organizations that can do it stays small, which is why open-weight releases (like NVIDIA’s Nemotron 3 ) matter for everyone else. Third, cost trajectory: new memory (HBM4E) and new silicon (Vera Rubin) are aimed squarely at the long-context agentic workloads that are getting expensive to serve.

For where to actually rent this compute, see the GPU clouds and neoclouds comparison . The practical takeaway: the infrastructure is being built for agentic, long-context AI, so design for that direction.

Sources

Further reading