Fluidstack

Fluidstack started in 2017 as an Airbnb for GPUs, matching startups that needed cheap compute with gamers and data centers sitting on idle Nvidia cards. That bootstrapped business is barely recognizable now. Today Fluidstack acquires power, designs and builds data centers, and operates them, and it is the company Anthropic picked to physically build the $50 billion of Texas and New York compute behind Claude.

The pitch is speed. Fluidstack says it can stand up gigawatts of compute in six months instead of the industry's usual eighteen to twenty-four, and that time advantage is what turned a tiny UK startup into a business reportedly valued at $18 billion within a few years of doing real revenue.

Fluidstack does two things that usually belong to different companies. It builds custom data centers from raw land up, handling power acquisition, site selection, and the civil, electrical, and mechanical engineering, and it runs fully managed GPU clusters that customers never have to touch. There are no Kubernetes configs to write and no bare metal to babysit; Fluidstack owns everything from the substation to the GPU scheduler.

For the AI labs it sells to, that full-stack control is the point. When the constraint on training a frontier model is how fast you can energize megawatts, a vendor that owns the power and the buildout gets to move at a different speed than one renting rack space.

The scarce thing in AI right now is not GPUs, it is somewhere to plug them in. Grid interconnects, permitting, and construction stretch a typical hyperscale build past two years, which is an eternity when model roadmaps move in quarters. Fluidstack sells its way around that: it takes on power sourcing, site selection, and the whole physical buildout so a lab can commit compute today and have it energized in months.

This is not a cloud for a weekend project. Fluidstack sells to the handful of organizations training frontier models and to governments building sovereign compute, the customers who order thousands of GPUs at a time and care more about lead time than list price. Its own case study makes the profile clear: it had 2,560 GPUs live for the AI lab Poolside within 48 hours of signing.

Fluidstack splits into two offerings that share a supply chain. Managed GPU clusters ship with zero customer setup, built-in monitoring and remediation, and a fifteen-minute engineering response SLA, so a lab gets compute and not a pager. Custom data centers are the heavier lift: greenfield facilities Fluidstack sites, powers, and builds, the same machine now pointed at Anthropic's Texas and New York sites.

Fluidstack does not sell by the hour like a commodity GPU cloud. It signs large, multi-year contracts to build and operate dedicated capacity for a single anchor customer, then finances the GPUs against the contract. Revenue grew from $1.8 million in 2022 to about $66 million in 2024, and the model now runs on multi-billion-dollar deals like the $50 billion Anthropic partnership, with GPU-backed debt lines reportedly up to $10 billion funding the hardware.

Fluidstack competes with the neocloud pack (CoreWeave, Lambda, Crusoe, Nebius) and, at the edges, with the hyperscalers building their own AI capacity. Its edge is not a cheaper GPU-hour. It is being the dedicated builder for a marquee tenant, owning the physical buildout, and moving faster than anyone on power and construction.

The receipts are unusually concrete for a company this young. It stood up 2,560 GPUs for Poolside within 48 hours of signing, grew revenue from $1.8 million in 2022 to roughly $66 million in 2024, and landed a $50 billion buildout with Anthropic plus a reported multi-billion-dollar French government supercomputer project. SemiAnalysis tracks it among the neoclouds it rates in its ClusterMAX GPU cloud rating system.

The recurring compliment is velocity. Poolside called Fluidstack's lead times and support the fastest of any partner it worked with, and industry analysts frame it as proof you can build a massive infrastructure business as the dedicated builder for one anchor tenant. That same framing is the worry: a business leaning this hard on a single customer and on GPU-backed debt is a bet on both staying healthy.

Fluidstack's cap table reads like a who's who of the AI-compute thesis. It raised a $200 million Series A in February 2025 led by Cacti, then roughly $700 million at a $7.5 billion valuation in December 2025, and by April 2026 was reportedly in talks for $1 billion at an $18 billion valuation with Jane Street co-leading. Backers include Situational Awareness, the Collison brothers, Nat Friedman, and Daniel Gross, with Google reportedly weighing a $100 million stake. Separately, it has arranged up to $10 billion in GPU-backed debt through Macquarie to finance the hardware.

Fluidstack has the deals, the backers, and the debt lines to build at a scale that would have been unthinkable for it a year ago. The hard part is now physical: energizing gigawatts across Texas and New York on schedule, through permitting, supply chains, and construction, is a very different challenge than reselling idle GPUs. If it delivers, it becomes core AI infrastructure. If Anthropic's needs shift or the build slips, a business this concentrated and this loaded with debt has little margin for error.

Fluidstack was founded in 2017 out of Oxford and led by co-founder and CEO Gary Wu, with Cesar Maklary as president. It frames its mission in unusually big terms, arguing that whoever deploys frontier compute fastest gets to decide whether AI expands human freedom or shrinks it, and it hires people who buy that framing.

The day job is heavier than most AI startups. Teams span software and hardware, and the company is staffing up power engineers, construction managers, and data center operators alongside platform engineers as it builds physical sites across Austin, New York, San Francisco, Seattle, Buffalo, and beyond. Engineering roles expect fluency with modern AI tooling, from LLM APIs to Claude Code and Cursor, in daily work.

Fluidstack discloses base ranges on its US roles and pairs them with stock options; postings note a total comp package of salary plus equity, retirement, and health, dental, and vision, with PTO in line with local norms. Engineering bands are the deepest sample and run wide, roughly $90,000 to $350,000, reflecting everything from junior to senior infrastructure roles. Operations and G&A roles disclose comparable spreads, and pay is quoted in USD for US positions with a smaller number in GBP for London.

For a company selling to governments and frontier labs, security posture is part of the product. Its job requirements repeatedly reference SOC 2, ISO 27001, and FedRAMP, along with the physical-plant safety standards (NFPA 70E) that come with running data centers. The registered entity is Fluidstack Ltd., a UK company, though the operating headquarters is moving to New York.

Fluidstack spent years as an obscure GPU marketplace before its Anthropic deal made it front-page infrastructure news. The coverage below traces the arc: the September 2025 Forbes profile of a tiny UK neocloud breaking into the big leagues, the November 2025 Anthropic announcement, and the 2026 funding reports that followed.