Fluidstack, which designs and operates data centers for AI workloads, has raised 1.5 billion USD (about 231 billion yen) in a round led by Jane Street Capital. The company is now valued at 18 billion USD (about 2.77 trillion yen). What stands out is that it has been entrusted with building Anthropic's large-scale compute footprint without owning a single GPU itself.
※1 USD = 154 JPY (as of September 8, 2026)
A Company That Buys No GPUs Reaches an 18 Billion Valuation
Neoclouds, the newer breed of AI compute providers, have generally been valued on the hardware they own. CoreWeave and Nebius built their positions by purchasing NVIDIA accelerators in volume and stacking valuation on those assets and their utilization.
Fluidstack removes that premise. It secures power, designs and builds the facilities, and writes and runs the software inside them, while the silicon itself belongs to the customer. Because it carries no GPU inventory, its balance sheet does not suffer when a chip generation ages out. The new round takes total funding past 2.6 billion USD (about 400 billion yen). Given that the company was valued at 7.5 billion USD (about 1.16 trillion yen) only months earlier, the pace of the climb is steep.
Within the past week, it ranks as the second-largest AI deal after rival Crusoe's 3 billion USD (about 460 billion yen) Series F. Two consecutive rounds have gone not to compute itself, but to the sites that house it and the software that runs it.
Carrying Anthropic's 50 Billion Dollar Plan
What defined Fluidstack's position was the 50 billion USD (about 7.7 trillion yen) commitment to American AI infrastructure that Anthropic announced in November 2025. Fluidstack is the party building the Anthropic-dedicated data centers in Texas and New York under that plan, with sites scheduled to come online through 2026.
For Anthropic, it is the largest independent infrastructure contract signed outside its relationships with Amazon Web Services and Google Cloud. General-purpose clouds are optimized for breadth, which leaves little room to tune for the single demand of frontier-scale model training. That is the reasoning behind commissioning purpose-built facilities separately.
Following the contract, Fluidstack moved its global headquarters from London to New York in December 2025. In March 2026 it withdrew entirely from the French government's 10 billion euro data center project, making its focus on the US market explicit.
Six-Month Builds and Software That Does Not Pick a Chip
The target the company repeats is delivering gigawatt-scale compute in 6 months. New data centers normally take 18 to 24 months, and Fluidstack compresses that by assembling factory-made, standardized modules on site. For labs whose compute needs can double within a year, that compression is itself the product.
The software inside the buildings is also in-house. Atlas OS, a bare-metal platform, automates server imaging and provisions Kubernetes or Slurm environments in days rather than the weeks typical of standard cloud onboarding. Lighthouse, the monitoring side, tracks performance in real time, restarts training jobs that fail, and swaps broken hardware for spares. On clusters of thousands of accelerators, a single faulty unit can halt everything, so this automated recovery determines effective utilization.
Neither Atlas OS nor Lighthouse depends on a particular chip. Alongside NVIDIA H100, H200, B200 and GB200, the platform supports AMD MI300X and MI355X, all validated to deliver at least 95 percent of theoretical performance. Clusters are single-tenant by default, fully isolated per customer, with InfiniBand between nodes. There is no contention for compute with a neighboring workload.
Taking no side among chip vendors matters at a moment when Google is pushing its TPUs out to external cloud operators. Anthropic disclosed in October 2025 that it plans to use as many as 1 million Google TPUs, and the scarcity of operators able to run something other than NVIDIA appears to be priced directly into the valuation.
Nine Years From an Oxford Student Project
Fluidstack was founded in 2017 by Gary Wu and James Cox while at Oxford University. The original idea was to pool idle GPUs sitting in data centers and research labs and rent them to developers, an approach described at the time as an Airbnb for GPUs.
The arrival of ChatGPT in late 2022 changed the picture. Frontier labs wanted purpose-built clusters at scale, not spare capacity gathered from a marketplace. The company shifted from brokering to designing, building and operating. UK corporate filings show revenue rising from 1.8 million USD (about 280 million yen) in 2022 to 66.2 million USD (about 10.2 billion yen) in 2024.
An investor memo projects 2026 revenue of 660 million USD (about 102 billion yen), more than tripling from 200 million USD (about 31 billion yen) in 2025. The same memo assumes more than 17 gigawatts of capacity under management by 2030, a figure that exceeds the combined current US data center capacity of Amazon and Microsoft. These are investor-facing projections, and whether power procurement and construction keep to schedule can only be confirmed by results. Customers named alongside Anthropic include Meta, Mistral and Poolside.
Summary
Fluidstack raised 1.5 billion USD, reaching an 18 billion valuation and total funding above 2.6 billion USD. A structure built on power, buildings and software rather than owned GPUs has been valued as a business insulated from chip generation turnover. Its role in Anthropic's 50 billion dollar buildout and a platform that favors neither NVIDIA nor AMD are what support that view. If demand keeps shifting from training toward inference while power constraints cap how fast compute can grow, operators that own no chips look set to matter more, not less.
