"Patronage-style financing" with a valuation of 120 billion yuan
A low-key piece of financing news has sent a huge shockwave across the entire AI industry.
Recently, foreign media outlets broke the news that AI infrastructure company Fluidstack has completed a $1.5 billion financing round, with a post-money valuation of $18 billion (about 120 billion RMB). The lead investor is Jane Street, the quantitative trading giant that has frequently hit media headlines this year.
It is worth noting that Fluidstack just closed a $750 million financing round two months ago, with a post-money valuation of $7.5 billion. That means the company's valuation has multiplied by 2.4 times within 60 days.
What is even more incredible is that this company barely has any of the popular Silicon Valley elements: it has no large models, does not produce GPUs, has no To C products, and has no legendary founder narrative.
If you have to find one relevant point, it is that it holds an auditable, multi-year huge contract endorsed by Anthropic. To put it more vividly, just as the market teases that Coreweave is "taken under the wing" by NVIDIA, Fluidstack is essentially a company "taken under the wing" by Anthropic.
From reselling idle graphics cards to a valuation of 120 billion
Back in 2017, two Fluidstack founders, Gary WuF (Gary Wu, British-Chinese) and Jamie Cox met when they were studying at the University of Oxford.
Gary studied economics and management and taught himself programming, while Jamie majored in classical literature. The two observed a very intuitive contradiction: AI researchers pursuing doctorates and small laboratories can hardly get sufficient GPUs quickly from AWS, Microsoft, and Google Cloud, while the cloud resources of large manufacturers have long queues and are extremely expensive. On the other side, a large number of game players, Internet cafes and small laboratories only use their high-priced graphics cards for two or three days a week, and most of the time the devices are left idle.
So the two came up with an idea: to build an "Airbnb for GPUs", aggregate the idle graphics cards of individuals and small institutions, and lease them in packages to researchers who need computing power, so that players can earn pocket money with their idle GPUs. In 2017, Fluidstack was officially established in Oxford, with its headquarters located in London.
Their early business was not limited to GPU leasing, they also aggregated idle bandwidth of global users to provide CDN content distribution services, with clients including the established video website Dailymotion. In 2018, Jamie dropped out of Oxford and received the Thiel Fellowship, and the two began to devote themselves full-time to entrepreneurship.
However, due to the inherent defects in the business model of the P2P idle GPU market, many scientific research users abandoned the platform after trial, so the company got off to a bad start, with slow growth and meager revenue.
The turning point came in 2020, when the third co-founder César Maklary joined. He holds a master's degree in aeronautical engineering from Imperial College London, and previously worked as an aerodynamic engineer in the Scuderia AlphaTauri F1 team, specializing in large complex systems, heat dissipation, power and hardware engineering.
Soon the COVID-19 pandemic arrived. During this period, reselling scattered graphics cards could not generate high profits, so the three shifted their business to lease GPU pools in batches from enterprises and university laboratories, and then sublease them to AI companies. At the same time, the team developed the self-developed bare-metal scheduling system Atlas OS for unified management and operation. No one could have predicted that this capability would become the company's greatest moat.
After the launch of ChatGPT, orders flooded in like a tide. In 2022, Fluidstack's revenue was 1.8 million US dollars; in 2023, it reached 30 million US dollars; in 2024, the audited report of the UK-registered company showed a revenue of 66.2 million US dollars.
The core selling point of the company is that Fluidstack can deploy and launch a cluster of thousands of GPUs within 48 hours, while other players need several months to deliver such products. The most well-known case is that it once deployed 2500 GPUs for Poolside at one time. Its customer list is also expanding rapidly, and cutting-edge AI startups such as Mistral AI, Character.ai, Poolside and Black Forest Labs have signed contracts one after another.
Even so, it was still a "small but beautiful" existence in the European AI circle. In February 2025, French President Emmanuel Macron announced at the AI Action Summit that Fluidstack would participate in the construction of a 10 billion euro, 1 GW AI data center in northern France. Overnight, this little-known company became the hope of European AI sovereignty.
Soon, a dramatic scene took place. In November 2025, Anthropic announced a $50 billion US computing power investment plan. Fluidstack was selected as the constructor and operator of the custom data center, and also deeply participated in Anthropic's plan to deploy up to 1 million Google TPUs, making it the first publicly known TPU data center operator outside Google.
A few months later, with the US contract secured, Fluidstack withdrew from the French project, and its global headquarters was moved from London to New York. Capital votes with its feet, and the geographic location follows the contract. Since then, with the reputation of Anthropic, Fluidstack has become a frequent guest in news reports, but its founding team remains low-key and rarely appears in public.
The "zero-chip" model sought after by capital
Although Fluidstack was established as early as 2017, before its valuation took off, its financing track record was so poor that it did not look like a star company at all.
The first real financing in its history took place in 2019. In this year, Fluidstack completed a $3 million seed round of financing, with investors including Seedcamp, Mercuri and 7 Global Capital. At that time, the company was still engaged in P2P idle GPU and CDN business, with extremely low popularity.
The second financing did not take place until five years later. In 2024, during the critical period when the business shifted to enterprise GPU clusters, Fluidstack obtained SAFE (Simple Agreement for Future Equity) and $37.5 million in debt financing, and the investors of this round were not announced.
UK registration files show that by the end of 2024, Fluidstack had completed a total of only $30 million in equity financing, plus $38 million in debt. In February 2025, it secured a $200 million Series A financing led by private equity firm Cacti, after which Fluidstack's financing frequency accelerated.
Money began to chase the company this year. At the beginning of the year, it completed a $450 million Series A+ financing led by the Situational Awareness fund under former OpenAI researcher Leopold Aschenbrenner. In July, the financing amount of this round increased to $830 million, the lead investor and valuation remained unchanged, while new investors including BlackRock, Google and Jane Street were added.
Then comes the latest $1.5 billion new financing led by the quantitative trading giant Jane Street. The company's valuation has soared to $180 billion, with a total equity financing of about $26 billion, and hundreds of billions of dollars in supporting debt financing for data center construction. This financing also made the three founders successfully rank among the "youngest self-made billionaires in Europe", but unfortunately they have moved to New York.
At this moment, in the eyes of capital, Fluidstack is not a cloud vendor, nor a customer agent of NVIDIA, but an asset-light platform for "computing power matching and delivery":
Most of the shells, electricity and chips of the data centers are leased. It locks in nearly 1 GW of power in the next ten years from transformed Bitcoin miners such as TeraWulf, Cipher Mining and Hut 8, signs long-term agreements with real estate and capital parties, and then packages them into long-term computing power leases to Anthropic, Meta, Mistral, Poolside and Black Forest Labs.
On the revenue side, according to an investment memorandum of an institution, Fluidstack will operate 1.3 GW of computing power at more than 10 sites in 2026, with a revenue of $660 million. This figure is more than twice the expected value of $200 million in 2025, and 10 times the actual value in 2024.
The key point is that it does not make money from the appreciation of hardware, but the service fee of "connecting 1 GW of electricity within the correct time frame". Its Atlas OS platform can automatically launch bare-metal GPU clusters within a few days, and the modularly built computer rooms are powered on in about 6 months, while the construction period of traditional data centers is 18 to 36 months.
Another path for Jane Street
If calculated according to the traditional price-to-sales ratio, the 180 billion valuation corresponds to the estimated 660 million revenue in 2026, which is 27 times. For a business with construction risks, power approval risks and high single-customer concentration, this figure is aggressive in any normal market.
But from another pricing framework, this is not pricing for revenue, but pricing for contracts. Once the multi-year contract at the level of 50 billion US dollars is locked, the remaining problem is engineering execution, which is exactly what institutions like Jane Street are best at evaluating with probabilistic language.
Therefore, to understand Fluidstack's valuation, we must first understand the person who wrote the check. We have elaborated on the legendary story of this company in the previous article "With an average annual salary of 18.25 million and no CEO in the company", so we will not go into details here.
It is the world's largest ETF market maker, and the Financial Times called it "an extremely profitable anarchist commune". In 2025, Jane Street's net trading revenue reached 39.6 billion US dollars, and its adjusted EBITDA was 31.2 billion US dollars, setting a historical record on Wall Street.
The performance in the first quarter of 2026 is even more impressive, with a single-quarter trading revenue of 16.1 billion US dollars and a profit of 10.3 billion US dollars. Its own capital (member's equity) is about 45 billion US dollars, which has increased by nearly 20 times compared with 2016. In Chinese public opinion, this company has got a conventional title "the US version of HFM Quant".
However, different from Liang Wenfeng's choice, Jane Street did not choose to train large models, but used capital to bet on every layer of the AI industry chain. In 2026, the company for the first time transferred a large amount of money from its profit pool from trading to investment.
Let's take a look at what happened to it this year:
On April 15, CoreWeave announced that Jane Street signed an AI cloud platform contract of about 6 billion US dollars, and at the same time bought 1 billion US dollars of CoreWeave stock at a price of 109 US dollars per share, locking in the priority right to use NVIDIA Vera Rubin systems.
On August 18, inference chip company Etched announced a 700 million US dollar financing with a valuation of 21 billion US dollars, led by Jane Street. It is also the first real paying customer of Etched. It had invested more than 100 million US dollars in Etched before July, accompanying the valuation from 5 billion US dollars all the way to 21 billion US dollars.
In early September, Jane Street was exposed to have signed a five-year cloud contract of about 13 billion US dollars with Crusoe. On the same day, Jane Street also led a 1.5 billion US dollar financing for Fluidstack.
Looking through Jane Street's investment territory, we can see a clear main line: it holds full-stack assets covering the model layer, computing power layer and infrastructure layer, and the targets it invests in also appreciate rapidly. Specifically, in the model layer it has Anthropic and Thinking Machines, in the chip layer it has Etched and MatX, in the cloud layer it has CoreWeave, Nscale and Crusoe, in the infrastructure layer it has Fluidstack, and in the power layer it has X-energy.
Therefore, the industry believes that this is not the logic of a venture capital portfolio at all, but a position sheet for cross-asset market making.
As one of the largest buyers in the AI computing power market, Jane Street has a clearer view of the real supply and demand of each layer than any VC. By leading the investment in Fluidstack, it essentially converts its profits into futures of computing power.
Jane Street has also encountered pitfalls. In February, the liquidation administrator of the TerraUSD crash sued it for using insider information obtained from a former intern to sell off in advance. In July, Reuters reported that Situational Awareness, the AI hedge fund it held heavily, lost about 15 billion US dollars in a single month due to high-leverage bets on the collapse of AI hardware stocks, and its investment portfolio was sold to Citadel at a large discount.
This is a certainty: genius is always accompanied by controversy.
Finally, back to this financing itself, some market views believe that this is a typical "under-the-wing" financing. The more academic name is anchored contract financing or circular financing, that is, the party that provides money is also the customer, guarantor, or even the upstream supplier of the party that receives the money.
This is not Fluidstack's invention, not even an invention of this era.
What Jane Street bought is not Fluidstack's execution capability, nor even the future of AI. What it bought is a cash flow certificate guaranteed by Anthropic's balance sheet. The valuation that has multiplied by 2.4 times in two months does not correspond to a company, but a contract.
Of course, the cost is also obvious: money circulates in the system and turns into revenue, then who will pay for the final real demand of the system? The ultimate buyers of computing power are highly concentrated in a limited number of players. In case the IPO process or revenue fails to meet expectations, the pyramid will collapse directly from the bottom.
This article is from the WeChat official account "ChinaVenture", author: Zhang Xue, published by 36Kr with authorization.