Jane Street Runs 4,000 Nvidia GPUs as Traders Race for AI Compute

By Sophie Lindqvist

One executive estimates that compute spending on Wall Street's fastest trading desks is now ten to twenty times higher than it was before ChatGPT launched. That figure, reported by Bloomberg, captures a structural shift in high-frequency trading: firms that once competed on microseconds are now competing on computing power, and they're paying for it the way big tech companies do.

The biggest quantitative trading firms — Jane Street, Hudson River Trading, XTX Markets — are spending billions on cloud contracts, chip startups, and their own data centers, Bloomberg reports. The reason is simple. These firms increasingly rely on machine learning to predict market moves, and machine learning runs on hardware that is expensive, scarce, and quick to age.

Why are traders suddenly building data centers?

High-frequency trading has always been an arms race. First it was about fiber routes and microwave towers between exchanges. Now the edge comes from models, and models need GPUs — lots of them.

Jane Street runs a data center in Texas equipped with more than 4,000 Nvidia GPUs. XTX Markets, the London-based firm, is building its own data center in Finland, a location chosen for the kind of cheap cooling and power such facilities demand. These are no longer side projects. They're core infrastructure for firms whose trading strategies depend on machine learning systems that must be trained and continuously retrained.

The scale of the spend tells the story:

• Hudson River Trading spends $1 billion a year on AI, according to the firm's AI chief, Iain Dunning.
• Jane Street invested $1 billion directly in CoreWeave, the GPU cloud provider, and separately committed $6 billion to its computing services.
• Hudson River signed a multiyear deal with CoreWeave for its own compute needs.
• One executive puts the industry's compute spending at 10 to 20 times pre-ChatGPT levels.

That's big-tech money moving into a sector that, until recently, measured its infrastructure edge in colocated servers next to exchange matching engines. The takeaway: for elite trading firms, compute has become as strategic a purchase as the trading algorithms themselves.

What does the CoreWeave money actually buy?

The two firms took noticeably different paths to the same problem. Hudson River Trading rents. Its multiyear CoreWeave deal buys access to GPU capacity without the firm having to own and operate the machines, and Dunning's $1 billion annual AI budget signals how central that capacity is to the firm's research and trading.

Jane Street did both. The firm invested $1 billion directly in CoreWeave — an equity stake in the GPU cloud company — and committed $6 billion to CoreWeave's computing services. That structure gives Jane Street two things at once: a financial interest in the compute supplier whose fortunes track AI demand, and a guaranteed claim on the computing capacity its models need.

It's a hedge, but also an admission. If compute is the bottleneck for machine-learning-driven trading, then owning a piece of the bottleneck is one way to make sure you're never last in line. In plain terms: Jane Street paid to skip the queue.

Why did Jane Street back a chip startup?

The chip angle is where the story gets more forward-looking. Jane Street also led a funding round for Etched, a chip startup, and became the first customer for its AI chip. The chip is designed to deliver more performance by running at a lower voltage — a technical choice that trades one kind of engineering difficulty for efficiency gains, which matters enormously when your power bill scales with your GPU count.

Becoming a chip startup's first customer is a bet that specialized silicon can outperform the general-purpose GPUs the rest of the market fights over. It's also an early-mover position: if Etched's approach works, Jane Street holds the reference-design advantage that comes from shaping the product's development around real trading workloads.

What are the risks?

According to Bloomberg, the industry itself sees two real risks in this spending spiral. The first is rising cost — the 10-to-20-fold increase in compute budgets is not a one-time jump but a new baseline that keeps climbing. The second is fast-aging chips: hardware that counts as cutting-edge today depreciates quickly as Nvidia and its rivals ship new generations, meaning today's multibillion-dollar data center is tomorrow's stranded asset.

Those risks are real enough that they show up in how executives frame the decision. Jan Boomaars, CEO of Amsterdam-based high-frequency trader Optiver, has made his position clear: doing nothing is more dangerous. The cost of falling behind on machine learning, in that view, exceeds the cost of overpaying for hardware.

That logic is worth pausing on. It's the same reasoning that drove big tech's AI capex boom: in a winner-take-most technology race, the downside of underinvesting looks larger than the downside of overinvesting. Trading firms, whose entire business model is pricing risk precisely, have concluded that the riskiest position on AI compute is having none.

Does this change the market?

There are broader stakes here beyond the trading desks. Quant firms joining the queue for GPUs, cloud contracts, and chip startup equity adds another deep-pocketed bidder to an already tight market for AI computing — one where hyperscalers and AI labs have been absorbing most of the supply. Every billion that Jane Street commits to CoreWeave is capacity the rest of the market doesn't get.

It also marks a maturity point for machine learning in finance. A decade ago, machine-learning-driven trading was a research curiosity at most firms. Today, the industry's cost structure is being rebuilt around it, from Texas data centers to Finnish construction sites to equity stakes in chipmakers. The firms that once optimized network latency now optimize the full AI stack.

The competitive logic is unforgiving, and it's the quiet conclusion of Bloomberg's reporting: if machine learning is now the edge in trading, then compute spending isn't discretionary anymore — it's the price of staying in the game, and firms like Optiver's Boomaars have decided the only unacceptable move is standing still.