In February 2026, Anthropic raised $30 billion at a $380 billion valuation. By May, it had raised $65 billion more at $965 billion — closing in on a trillion dollars for a company that didn’t exist a decade ago — and filed confidentially for an IPO. Nobody in the industry called that the top of the market. They called it Tuesday.
The Number That Broke the Scale
Anthropic’s revenue run rate went from $9 billion at the end of 2025 to $14 billion in February, $19 billion in March, $30 billion in April, and $47 billion by May — five months, 5x growth. CEO Dario Amodei has described the pace as something the company is finding genuinely difficult to manage operationally. In a CNBC interview, he characterized the underlying capability gains as “reasonably exponential” year over year, and argued the next phase of the race won’t be won by whoever runs the single biggest training job — it’ll go to whoever delivers the most capability per dollar of compute.
The Hyperscaler Bill
Anthropic is the headline, but it’s not the whole story. Combined 2026 AI infrastructure spending across Amazon, Google, Meta, and Microsoft is on pace for roughly $725 billion — up 77% from about $410 billion in 2025. Google alone raised its 2026 capex forecast to as much as $205 billion. OpenAI, for its part, has raised its planned compute spending through 2030 to around $750 billion, up from a $600 billion figure it had set only months earlier, as it shifts from renting compute to building and owning data centers directly.
What a Training Run Actually Costs
The physical cost behind those balance sheets is electricity. Research firm Epoch AI estimates frontier training runs on clusters of 50,000 to 100,000 GPUs now consume somewhere between 100 and 300 GWh — up from roughly 50–60 GWh for GPT-4 and just 1.3 GWh for GPT-3. Epoch projects power demand for frontier training to keep growing 2.2x to 2.9x per year, with individual runs potentially requiring 4–16 gigawatts by 2030 — on the order of what a small country’s grid can supply.
The Grid Is Pushing Back
Regulators have noticed. In March 2026, Senator Bernie Sanders and Rep. Alexandria Ocasio-Cortez introduced federal legislation that would halt construction of new data centers drawing 20 megawatts or more until stronger safeguards are in place. Virginia — home to the densest concentration of data centers in the world — ended a major data-center tax exemption in February and restricted new construction to industrial-zoned land in January, shifting grid-upgrade costs onto the facilities themselves rather than ratepayers generally. Ireland, often cited as the cautionary tale, actually reopened its Dublin-area grid to new data centers in December 2025 after a multi-year freeze — but only for facilities that supply 100% of their own backup power and source at least 80% of electricity from new renewable projects. The moratorium era isn’t over so much as it’s being replaced by a much stricter set of conditions.
Winning Might Not Be a Strategy
Even the people building this admit the trajectory is strange. Amodei’s own framing — that the race will be decided by capability per dollar, not raw scale — is itself an implicit concession that spending alone doesn’t guarantee a moat. Every lab is now betting nine and ten figures on the idea that being a few months ahead is worth almost any price, while simultaneously arguing that efficiency, not size, is what will actually matter. Both things can’t be the main strategy forever. At some point the bills come due, the grid pushes back harder, or efficiency gains outrun the need for more raw compute. None of that has happened yet. The number keeps going up.
