FluidStack has closed a $1.5 billion private-equity round led by Jane Street Capital at an $18 billion valuation. The figure highlights an unusual feature of the business: the company selected by Anthropic as its anchor infrastructure partner does not own a single chip.

While CoreWeave, Nebius and almost every other major neocloud have built their valuations on Nvidia hardware ownership, FluidStack designs and operates data centres, develops the software that runs inside them and leaves the actual silicon to its customers. That model has now been valued at $18 billion.

The round takes FluidStack's total funding beyond $2.6 billion and makes it the second-largest AI deal of the past week, behind rival Crusoe's $3 billion Series F. Investor appetite for specialised AI computing infrastructure is accelerating into a new category: companies that own the power and platform, but not the chips, are commanding valuations comparable to those of publicly traded GPU cloud operators.

Gary Wu and James Cox founded FluidStack at Oxford University in 2017. Its original concept was modest: aggregate underused graphics-processing units at data centres and research laboratories worldwide and rent spare capacity to developers — an "Airbnb for GPUs". During the pandemic, the company expanded into renting entire GPU blocks from corporate and university laboratories and secured early customers including AI start-ups Poolside, Mistral and Character.AI.

From Oxford project to $18 billion neocloud

The arrival of ChatGPT in late 2022 transformed the business. Demand for large-scale AI computing quickly made a marketplace model insufficient: frontier AI laboratories needed dedicated, custom-built capacity rather than spare marketplace hardware. FluidStack pivoted to a full-stack neocloud, designing, building and operating dedicated GPU clusters for AI companies.

Revenue rose from $1.8 million in 2022 to $66.2 million in 2024, the last year covered by the company's publicly disclosed UK corporate filings. A $200 million Series A in February 2025 accelerated its expansion.

In December 2025, FluidStack moved its global headquarters from London to New York, placing its future focus on the US AI infrastructure market. By February 2025, it had become an unexpected centrepiece of French President Emmanuel Macron's AI Action Summit, with a signed memorandum of understanding for a €10 billion data-centre project in France.

That European chapter ended abruptly. After signing its landmark Anthropic agreement in November 2025, FluidStack withdrew completely from the French project in March 2026 and shifted its focus exclusively to the US market.

The Anthropic deal that changed everything

FluidStack's most significant turning point was its agreement with Anthropic, signed in November 2025, under which the AI company committed $50 billion to American AI infrastructure. FluidStack is building custom-designed data centres in Texas and New York for Anthropic's Claude models, with sites scheduled to come online throughout 2026.

The agreement is among the largest infrastructure contracts in technology history and the biggest independent infrastructure commitment Anthropic has made outside its relationships with Amazon Web Services and Google Cloud.

It also marked a vote of confidence in FluidStack's modular construction model. Traditional data centres can take 18 to 36 months and require hundreds of millions of dollars in upfront capital. FluidStack assembles facilities from prefabricated, factory-made modules that can be transported to a site and powered up in roughly six months. For AI laboratories whose computing requirements can double from year to year, that speed is central to the product.

The deal reflects the view that hyperscalers, despite their capabilities, are not designed for the specialised and continuous demands of frontier-scale model training. Shared cloud infrastructure built for general-purpose computing cannot be customised sufficiently for the workloads Anthropic requires.

An investment memorandum shared by one of FluidStack's venture-capital backers projects that revenue will reach $660 million in 2026, more than tripling from $200 million in 2025. The memorandum also forecasts that FluidStack will operate sites with more than 17 gigawatts of capacity by 2030 — greater than the combined current US data-centre capacity of Amazon and Microsoft, according to Jefferies estimates.

Atlas OS, Lighthouse and the case for not owning chips

FluidStack's technical distinction is not only what it builds but also the software it operates inside its facilities. Its in-house Atlas OS is a bare-metal platform that automates server imaging and provisions Kubernetes or Slurm environments in days rather than the weeks typically required for standard cloud onboarding. It enables cloud-like scaling without customers having to configure hardware manually.

Its Lighthouse monitoring platform provides real-time performance optimisation, automatically restarting failed training jobs and switching to replacement hardware to keep workloads running.

Both Atlas OS and Lighthouse are chip-agnostic. FluidStack's infrastructure supports Nvidia and AMD hardware, including H100, H200, B200 and GB200 GPUs, as well as AMD's MI300X and MI355X. All have been validated to deliver at least 95% of theoretical performance.

Clusters operate as single-tenant environments by default, isolated at the hardware, network and storage levels and connected through InfiniBand for high-performance communication between nodes. There are no shared clusters or competing workloads from neighbouring customers.

The strategic implication is significant. A platform that supports Nvidia and AMD equally, without a financial interest in either company's success, could become more valuable as competition among chipmakers intensifies. CoreWeave and Nebius remain structurally tied to Nvidia's continued dominance, while FluidStack avoids that exposure. As Amazon's Trainium chips, Google's TPUs and AMD's MI series become more credible for AI training, FluidStack's neutrality may become an advantage.

Google's role as financial backstop

Google has played a role in FluidStack's business not only as a customer but also as a financial backstop. Through agreements with former crypto miners TeraWulf and Cipher Mining, which are converting Bitcoin-mining sites into AI campuses, Google has backed billions of dollars in FluidStack's colocation lease obligations in exchange for equity stakes in the two companies. The structure was disclosed in TeraWulf's filings with the US Securities and Exchange Commission.

At TeraWulf's Lake Mariner campus in western New York, FluidStack has contracted approximately 360 megawatts of critical IT load under two 10-year hosting agreements. The contracts represent roughly $6.7 billion in revenue, with Google backing $3.2 billion of FluidStack's obligations.

At Cipher Mining's Barber Lake site in Colorado City, Texas, a separate 10-year agreement covers approximately 207 MW of critical IT load worth about $3.8 billion. Google is backing $1.73 billion of FluidStack's obligations.

FluidStack has also obtained up to $10 billion in GPU-backed debt-financing capacity from Macquarie Group. The senior debt facility is intended to finance GPUs for deployment at a renewable-energy-powered data centre in Iceland.

The arrangement gives FluidStack unusual financial firepower for a company that does not own the chips operating in its facilities. It also gives Google influence over where and how capacity is deployed, as well as a financial interest in FluidStack's ability to deliver.

Why Jane Street invested

Jane Street's role as lead investor is closely tied to its broader AI strategy. The quantitative trading firm generated $16.1 billion in trading revenue in the first quarter of 2026, largely through AI-powered strategies, and has been building one of the most aggressive AI infrastructure positions among private capital firms.

In April 2026, Jane Street signed a $6 billion AI cloud agreement with CoreWeave and invested $1 billion in the company's equity. It also holds a stake in Anthropic, whose valuation has risen above $800 billion.

The FluidStack investment completes a strategic circle. Jane Street is both a major buyer of computing capacity — paying CoreWeave and reportedly Crusoe for GPU access — and now an equity investor in the main infrastructure builder for Anthropic, an AI company it has backed through several funding rounds.

The investment is not simply a bet on continued demand for AI infrastructure. It is a bet that a company building dedicated, customised and chip-agnostic facilities for one of the world's most demanding AI laboratories occupies a particularly defensible position in the infrastructure stack.

Nvidia is notably absent from FluidStack's investor list. The chipmaker has provided billions of dollars to GPU cloud rivals including CoreWeave, Nebius and Nscale, but has not backed FluidStack. The difference appears partly deliberate: a platform built around chip neutrality does not benefit from Nvidia's strategic alignment in the same way as a neocloud that owns Nvidia hardware.

Jane Street's decision to lead the round without Nvidia's participation suggests that investors increasingly view chip neutrality as a feature rather than a gap.

Does FluidStack own enough to justify its valuation?

The neocloud sector has attracted a series of enormous financings in 2026. Crusoe raised $3 billion this week at a $30 billion valuation. Lambda raised more than $1.5 billion, Firmus raised $2 billion in August and TensorWave raised $350 million as it positioned itself as an AMD-based alternative.

Publicly traded CoreWeave, which is expected to approach 1.5 gigawatts of total power capacity by the end of the year, has a market capitalisation of approximately $47 billion. Nasdaq-listed Nebius has a market capitalisation of about $57.4 billion.

FluidStack differs from most competitors because it does not buy chips and rent them out. It builds and operates physical data-centre facilities and leases space to customers that provide or contract for their own chips. That model allows FluidStack to move faster by keeping capital equipment off its balance sheet. It also leaves revenue more exposed to customer demand than to the GPU scarcity that has boosted chip-owning neoclouds.

McKinsey's analysis of the neocloud sector found that bare-metal-as-a-service operators generate gross profit margins of 14% to 16% after labour, power and depreciation. FluidStack avoids some of that pressure because it does not carry GPU depreciation on its balance sheet. It therefore does not face the same margin compression when each new Nvidia generation reduces the value of older hardware.

The trade-off is customer concentration. Approximately $50 billion of FluidStack's long-term contracted revenue comes from a single relationship with Anthropic.

Is FluidStack's funding enough to secure its position?

Execution risk remains substantial and affects the industry as a whole. Data Center Watch reported that opponents blocked or delayed at least 75 projects worth roughly $130 billion in the first quarter of 2026 alone, the highest quarterly total since tracking began. The number of active opposition groups more than doubled, from 396 at the end of 2025 to 833 by March 2026.

FluidStack's planned 10-site US buildout, concentrated in Texas and New York, is taking place in this environment.

Patrick Moorhead, chief executive and chief analyst at Moor Insights and Strategy, summarised the challenge: "The neocloud model works brilliantly when demand is surging and capital is cheap. The real test comes when you have to deliver physical infrastructure at scale while managing supply chains, permitting, and community relations."

With a $50 billion anchor contract, $2.6 billion in total funding, an $18 billion valuation and customers including Anthropic, Meta, Google and Mistral, FluidStack has moved from relative obscurity as a European start-up to the centre of the US AI infrastructure race in less than two years.

Investors are betting that a company building custom-designed, chip-agnostic and modular data centres for demanding AI laboratories occupies a position that hyperscalers cannot quickly replicate and that chip-owning neoclouds cannot easily match.

Whether that bet succeeds will depend on execution. Designing a $50 billion infrastructure programme is not the same as delivering it. Community opposition, permitting, power procurement and construction management across 10 US sites are challenges that $1.5 billion in fresh capital cannot automatically resolve. The decisive test is expected in 2027 and 2028, when Anthropic's first facilities are due to become fully operational.

The model's main strength also creates a specific vulnerability. If Anthropic's demand for custom computing capacity slows, FluidStack has no chip inventory to redeploy to other customers. Jane Street is betting that Anthropic's demand will remain structural and long-term. Its $1.5 billion investment reflects confidence that it will.

Frequently Asked Questions

What does "chip-agnostic" mean, and why does it matter?

FluidStack builds and operates data-centre facilities, including the buildings, power systems, networking and cluster-management software, but does not purchase or own the GPUs inside them. Customers such as Anthropic bring or contract for their own chips. This distinguishes FluidStack from CoreWeave and Nebius, which acquire Nvidia GPUs and rent them out.

The model shields FluidStack's balance sheet from GPU depreciation as new generations are launched, but makes the company more dependent on continued customer demand. Atlas OS and Lighthouse support both Nvidia and AMD hardware, allowing FluidStack to benefit if competition among chipmakers intensifies.

Why did Nvidia not invest in FluidStack?

Nvidia has invested billions in GPU cloud companies including CoreWeave, Nebius, Nscale and Lambda to ensure that its chips have well-capitalised distribution channels. FluidStack's model does not fit that strategy as neatly: its platform supports AMD chips as readily as Nvidia chips and does not depend on owning Nvidia hardware. The absence of Nvidia backing therefore reflects FluidStack's deliberate neutrality as much as any rejection of the company.

How does Jane Street's investment fit its wider AI strategy?

Jane Street is both a major private-market AI investor and a major buyer of AI computing capacity. It signed a $6 billion CoreWeave cloud agreement in April 2026 and took a $1 billion equity stake in the company. It has also backed Anthropic in several funding rounds.

By leading FluidStack's $1.5 billion round, Jane Street has invested in the principal infrastructure builder for Anthropic. It now has exposure to the AI model company, the chip-cloud provider and the dedicated-facility builder, reflecting a view that AI infrastructure value is distributed across several layers of the technology stack.

What are the biggest risks facing FluidStack's US expansion?

Two risks stand out. The first is community opposition. Data Center Watch said opponents blocked or delayed at least 75 US data-centre projects worth approximately $130 billion in the first quarter of 2026, while the number of active opposition groups rose from 396 to 833 across 49 states between the end of 2025 and March 2026.

The second is customer concentration. About $50 billion of FluidStack's long-term contracted revenue comes from Anthropic. If demand from that customer slows, infrastructure optimised for its systems may be difficult to redeploy quickly. Both risks are manageable — Anthropic's demand shows no sign of slowing, and Google is backing key lease obligations — but they remain material.

Originally published on Tech Times