Inside FinRAG
4,032 Nvidia GPUs in one building, at a trading firm, not an AI lab
Jane Street has shown a Texas data center with 4,032 GPUs and committed about $6 billion to AI cloud. Here is what trading firms appear to want from all that compute, and who else is spending.
In this article
If you saw a photo of 4,032 Nvidia GPUs in 56 liquid-cooled racks, you might guess it belonged to an AI lab. This one belongs to the world of market making.
The site is a Jane Street data center in Texas, shown in a video the firm released this year. And it is only part of the picture. Jane Street's co-head of technology, Ron Minsky, has said the firm runs tens of thousands of GPUs and expects to reach hundreds of thousands.
So what does a trading firm want with all of that?
The numbers
- 4,032 GPUs, 56 racks, about 8,000 km of fiber at one retrofitted Texas site, with most cooling now liquid-based.
- About $6 billion committed to CoreWeave's cloud, plus $1 billion of CoreWeave stock at $109 a share, according to CoreWeave's April 15 announcement.
- 100 to 200 MW of possible new capacity. Bloomberg first reported early talks in June. Jane Street has not publicly confirmed a project, and no site has been named.
CoreWeave's Max Hjelm put it this way: "Jane Street operates like a frontier lab." He is a vendor describing a customer, so read it with that in mind. The hardware makes the comparison hard to dismiss.
How it can afford this
Reuters reported that Jane Street made a record $39.6 billion in net trading revenue in 2025, above JPMorgan's $35.8 billion, with about 3,500 employees. Reuters cited documents and people familiar with the matter. Jane Street declined to comment.
The firm is not only buying compute. It is also investing around it. Reuters later reported a record $16.1 billion first quarter, helped in part by gains on stakes in AI companies including Anthropic and CoreWeave. It also led a funding round for chip startup Etched and is the first customer for its chip.
What the chips are for
This is the part people get wrong. These GPUs are not, as far as public reporting shows, racing orders to an exchange.
Reported uses center on training language models and custom architectures built for trading problems and proprietary data. One analysis says ultra-low-latency systems still handle execution, while AI works on slower tasks such as research, pricing and pattern detection.
Compute is scarce enough inside the firm that it reportedly runs an internal currency to auction it. And Minsky has said experimentation there is limited by how much compute it has.
Our reading: the speed race now has a second front. Nanoseconds still decide who gets the trade. Compute may increasingly decide how good the models behind that trade are. Jane Street shares limited detail, so nobody outside can say how much of its profit comes from these models.
It is not just Jane Street
| Firm | Reported move |
|---|---|
| Hudson River Trading | About $1 billion a year on AI, per its AI chief, plus a multiyear CoreWeave deal |
| XTX Markets | Building a data center in Finland |
| Citadel Securities | Relying on hyperscalers and Google's TPU infrastructure |
| Optiver | CEO argues that doing nothing is the bigger danger |
One executive told Bloomberg, as reported by The Decoder, that compute spending is 10 to 20 times what it was before ChatGPT launched.
The risk
The same reporting flags rising costs and fast-aging chips. A GPU bought today depreciates quickly, and there is no public evidence yet that more compute reliably produces more trading profit. Jane Street's record year also came with heavy market volatility, which Reuters cites as a driver. A correlation between big revenue and big clusters does not prove one causes the other.
What this means for everyone else
Few firms can spend at this level, and most do not need to. But the questions apply at any size:
- Which of your decisions actually improve with more compute?
- Is your limit compute, or is it data and research time?
- Does your edge depend on models, on speed, or on neither?
What would you do with 4,032 GPUs?
References
- CoreWeave announcement via SEC filing, April 15, 2026
- Reuters: Jane Street's $40 billion trading haul, April 24, 2026
- Reuters: Jane Street's record first quarter, May 2026
- The Decoder, citing Bloomberg, October 26, 2026
- Disruption Banking: Jane Street's 200 MW plan, July 20, 2026
- Dealroom: Inside Jane Street's Texas data centre
- Dealroom: Inside Jane Street's GPU empire
- iTechGuides: why trading firms are buying AI compute, October 26, 2026
FinRAG.tech has no connection to these firms. Revenue figures are as reported by Reuters and were not confirmed by Jane Street. The interpretation is ours. This article is commentary, not investment advice.
Worth sharing?
Pass this article along to someone who would find it useful.
The work continues in the studio.
More build notes, walkthroughs, and perspectives from the FinRAG team.
Explore all stories
