FluidStack
Last reviewed
Jun 8, 2026
Sources
23 citations
Review status
Source-backed
Revision
v1 ยท 2,005 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
23 citations
Review status
Source-backed
Revision
v1 ยท 2,005 words
Add missing citations, update stale details, or suggest a clearer explanation.
FluidStack (stylized Fluidstack) is an artificial intelligence cloud company that provides on-demand access to large clusters of graphics processing units (GPUs) for AI training and inference, and that has expanded into developing and operating dedicated AI data centers for frontier AI labs. Founded in 2017 as a marketplace for idle GPU capacity, the company is one of the more prominent firms in the so-called neocloud sector, which supplies GPU compute specialized for AI rather than general-purpose hyperscale workloads. FluidStack rose to wider attention in November 2025, when Anthropic named it as the developer and operator partner for a reported buildout of up to 50 billion US dollars in American AI data centers, beginning with sites in Texas and New York.[1][2] In April 2026 the company was reported to be in talks to raise roughly 1 billion US dollars at an 18 billion US dollar valuation, only months after a reported 7.5 billion US dollar valuation.[3][4]
FluidStack operates an AI cloud platform that aggregates and runs GPU capacity, and increasingly builds and operates the underlying data centers itself. The company describes its central capability as delivering "gigawatts of compute in 6 months rather than the industry's 18 to 24 months," acquiring power, designing and constructing facilities, and operating them with combined hardware and software teams.[5] It markets clusters built on Nvidia GPUs including the H100, H200, B200, and GB200, connected over InfiniBand fabrics for large single-job training runs, and says it can stand up more than 1,000 GPUs in under 48 hours.[6] By late 2025 the company reported managing more than 100,000 GPUs.[7]
The firm positions itself as a specialized alternative to general-purpose clouds, building infrastructure tuned to the training and inference patterns of specific AI systems rather than offering a broad menu of services. This focus, together with its ability to secure power and deliver capacity quickly, underpinned its selection by Anthropic for one of the largest infrastructure programs in US history.[1][5]
FluidStack was founded in 2017 as a spinout from the University of Oxford. Public records and profiles consistently list Gary Wu, who serves as chief executive, and Cesar Maklary among its founders, alongside several others.[8][9] (The notion sometimes circulated that a "Csaba Kiss" co-founded the company is not supported by available sources and appears to be in error.) The original business was a marketplace that connected users needing cheap compute with idle GPU resources, an approach that earned it the label "Airbnb for GPUs." In its early years the company remained small, reporting roughly 1.8 million US dollars in revenue in 2022.[8]
The launch of ChatGPT in late 2022 and the resulting surge in demand for GPU compute reshaped the business. FluidStack pivoted from a marketplace toward building and aggregating dedicated infrastructure, contracting capacity from independent data centers and operating clusters for AI developers. Revenue rose sharply, reaching about 66.2 million US dollars in 2024.[8] Originally based in London and a notable name in the European AI scene, the company said in December 2025 that it would relocate its global headquarters to New York, a month after announcing its Anthropic partnership.[10][11]
In March 2026, FluidStack withdrew from a high-profile data center program in France valued at around 10 billion euros (roughly 11.5 billion US dollars), including a planned one-gigawatt facility at Bosquel in northern France and a second project near Paris tied to French AI firm Mistral and developer Eclairion. The French commitment had been a centerpiece of President Emmanuel Macron's national AI investment package; FluidStack's exit reflected its decision to concentrate on larger US opportunities.[12][13]
FluidStack's platform is built around software that automates the operation of large GPU fleets. Its custom operating environment, marketed as Atlas OS, lets administrators launch GPU clusters in a few clicks and automates maintenance tasks such as restarting servers, reducing the manual effort of running clusters at scale.[7] The company says its hardware is validated to deliver at least 95 percent of theoretical performance, and that its InfiniBand networking can interconnect more than 12,000 GPUs for a single training job.[6]
Over time the company has served a range of AI developers. Reported customers and former customers include Anthropic, Google, Meta, Mistral, Character.AI, Poolside, Black Forest Labs, and Midjourney, with the company's own materials also citing relationships with vendors such as Nvidia, Samsung, Dell, and Honeywell.[1][7][14] This customer base, spanning model developers and enterprises, reflects FluidStack's role as a supplier of frontier-scale compute capacity.
A defining feature of FluidStack's growth has been the financing structure behind its data center leases. Beginning in 2025, Google agreed to backstop a series of FluidStack lease obligations to data center and former-cryptomining operators, agreeing to cover payments to lenders if FluidStack failed to meet them and, in exchange, retaining an option on the underlying power and capacity. Reports described Google backstopping roughly 3.2 billion US dollars of obligations tied to a TeraWulf campus at Lake Mariner, New York, and taking an equity stake of about 14 percent in TeraWulf, along with backing for additional projects and a stake in Cipher Mining.[15][16] In February 2026 Google was reported to be exploring a direct investment of around 100 million US dollars in FluidStack itself.[3]
On November 12, 2025, Anthropic announced that FluidStack would develop and operate a new wave of custom-built American data centers, with the program reported at up to 50 billion US dollars and the first sites in Texas and New York.[1][2] Anthropic said the buildout would create roughly 800 permanent jobs and about 2,400 construction jobs, with facilities expected to come online throughout 2026.[1][17] Anthropic cited FluidStack's speed and its ability to bring large amounts of power online on short timelines as reasons for the selection. Chief executive Dario Amodei framed the sites as infrastructure to "build more capable AI systems" supporting scientific and other breakthroughs.[17] The deal marked a strategic shift for Anthropic, which had previously leaned on a multi-cloud, multi-chip approach using Google Cloud TPUs and Amazon's Trainium accelerators.[17]
The 50 billion US dollar figure refers to the broader program rather than a single signed contract, and the physical capacity is being assembled through FluidStack-arranged deals with several data center developers, many of them former cryptocurrency miners that have pivoted to AI hosting:
| Partner | Site / location | Initial capacity | Notes |
|---|---|---|---|
| Hut 8 | River Bend, Louisiana | 245 MW | 15-year, triple-net lease reported at about 7 billion US dollars, scaling to up to roughly 2.295 GW and as much as 17.7 billion US dollars if options are exercised[18][19] |
| TeraWulf | Lake Mariner, New York | ~360 MW | Long-term hosting; Google backstop of about 3.2 billion US dollars[15][20] |
| TeraWulf | Abernathy, Texas | 168 MW | HPC joint venture; Google backing reported around 1.3 billion US dollars[15] |
| Cipher Mining | (Texas pipeline) | ~244 MW | 10-year AI hosting agreements; Google-backed[16][20] |
Reuters and other outlets reported that the Hut 8 arrangement alone could deliver at least 245 MW and up to about 2.295 GW of AI data center infrastructure for Anthropic, with clusters operated by FluidStack.[18] These structures illustrate how FluidStack ties together power, real estate, debt financing, and a creditworthy backstop to deliver capacity at a scale far larger than its own balance sheet would otherwise support.
FluidStack's reported valuation climbed steeply alongside its data center commitments. According to Bloomberg, the company raised around 700 million US dollars at a 7.5 billion US dollar valuation in December 2025, shortly after the Anthropic announcement.[3][4] In April 2026, multiple outlets reported that FluidStack was in talks to raise roughly 1 billion US dollars at an 18 billion US dollar valuation, a figure that would more than double its worth in a matter of months. Jane Street and the Situational Awareness fund, associated with Leopold Aschenbrenner, were reported to be in talks to co-lead the round, with Morgan Stanley advising; earlier backers reportedly included Stripe co-founders Patrick and John Collison, former GitHub chief executive Nat Friedman, and investor Daniel Gross.[3][4] As of mid-2026 the 18 billion US dollar round was described as in talks rather than closed, and these figures should be treated as reported and subject to change.
Sources differ on FluidStack's cumulative funding, with some company profiles citing more than 1 billion US dollars raised across several rounds.[8] The rapid revaluation underscores how a single anchor contract with a well-capitalized AI lab can transform a neocloud's financing position.
FluidStack competes within the broader neocloud market against firms such as CoreWeave, Crusoe, Nebius, Lambda, Together AI, and Nscale. By revenue, the segment is led by CoreWeave, which surpassed 5 billion US dollars in annual revenue faster than any prior cloud platform, along with Crusoe, Lambda, and Nebius; FluidStack has historically sat among a long tail of smaller specialists by revenue, even as its contracted future capacity rivals far larger names.[21][22] The neocloud sector as a whole exceeded 25 billion US dollars in revenue in 2025 and is forecast by some analysts to grow toward several hundred billion dollars by the early 2030s.[23][22]
FluidStack's significance lies less in its current revenue than in what its rise reveals about the economics of frontier AI. The Anthropic buildout illustrates the extreme capital intensity of training and serving large models, the scramble for electrical power and data center real estate, and the novel financing arrangements, such as Google's lease backstops, that channel hyperscaler balance sheets into third-party capacity. In this sense FluidStack is a case study in how the AI data-center boom is reshaping infrastructure markets, with specialized operators acting as intermediaries between frontier labs that need gigawatts of compute and the developers and financiers able to build it.