The GPU and compute stacks of major quant trading firms, ranked
Quant trading firms have always been on the cutting edge of hardware, using programmable chips like ASICs and FPGAs to power ultra-low latency trading systems in the high-frequency trading era. Now, with machine learning strategies dominating the market, it's all about who has the most compute power. The more compute power, the more interesting the employer.
Why do quant trading firms want GPUs? Who works with them?
GPUs provide the compute power that traders at these firms need to power their machine learning strategies. Firms like Jane Street have been increasingly looking at alternative data like satellite imagery which can be messy and hard to fit into trading models without the assistance of an LLM to find signal in the noise. Trading firms use GPUs to train their proprietary LLMs. Machine learning quants use those LLMs alongside tools like PyTorch to build powerful trading models using statistical techniques like linear regression on a massive scale.
What does that mean for hiring? A bigger GPU stack will usually lead to an increase in machine learning headcount; Jane Street's machine learning team has grown to the point where both ML research and ML engineering have their own separate intern tracks. Several trading firms are also building out AI labs to make the most of their stacks including Optiver and Hudson River Trading.
The infrastructure also creates new jobs for the hardware engineering specialists who optimize these data centres to get the most out of the GPUs. When Jane Street gave an inside look into its Texas datacentre in May, co-head of Technology Ron Minsky said that the buildout of these centres is a "huge engineering challenge." These centres require you to delicately balance power so that the GPU racks don't overload, and they require complex networking infrastructure for the data to be transmitted across the firm at speed. Bloomberg reported yesterday that Hudson River Trading, has built a data centre in the caverns of a Norweigan mountain cooled by water from a nearby fjord if you're interested in a more remote working environment.
Who has the biggest GPU and compute stack?
We've pored through the websites and job listings of major electronic trading firms to see how they describe their compute stack, focusing on the number of GPUs, CPU cores and their total data storage capacity.
The topline figures suggest that XTX Markets has the best on-premise compute infrastructure. The British trading firm ran by Alex Gerko claims to have "over 25,000 GPUs," up from 20,000 in 2024. While other firms are measuring their data in petabytes, XTX says it has "1 exabyte of storage."
That's just one side of the coin; several major trading firms have agreements with cloud computing providers and hyperscalers to gain access to their huge GPU clusters. Jane Street, which claims to have "tens of thousands of high-end GPUs," also reportedly has agreements with two major hyperscalers: CoreWeave and Crusoe. Those two contracts were reported to be worth a combined $19bn. Citadel Securities, which is thought not to have on-premise datacentres, signed a major deal with Google Cloud in 2024 to power its quant research with Google's infrastructure. Bloomberg reported yesterday that the firm has 'signficantly sclaled up' its use of TPUs (tensor processing units), custom chips from Google. Citadel Securities told Bloomberg that these are four times more powerful than GPUs and 30% cheaper.
Hudson River Trading and Tower Research are each vague in the descriptions of their GPU clusters, but both state they have more than 100 petabytes of storage. Hudson River Trading says its GPU fleet is "larger than most national labs."
DRW and Jump Trading have each been pushing into the compute trading space while building their own stacks. Jump Trading's website says it has "10,000+ compute nodes" across "thousands of GPUs and TPUs." DRW doesn't disclose its internal GPU footprint, but it signed a deal with QumulusAI in August which provided the firm with a dedicated cluster of B300 GPUs.
Internal GPU fleets vs cloud deals: which is better?
The majority of top trading firms operate a hybrid compute structure using both on-premise and cloud computing services. Ron Minsky, Jane Street's co-head of technology, said on his podcast this week that Jane Street has found cloud offerings "really useful" despite its large internal stack. "There are capacities they have that we don't [so] sometimes they're just better than us." He said that, while cloud infrastructure runs some workloads quicker, "often it goes the other direction." Minsky said there are "real rewards" to having both at the same time, but that it's difficult to pull off.
The drawbacks of having just one are evident. If you build out a market-leading GPU cluster today, it might be market-lagging a few years from now. Ken Griffin said at a Citadel conference this week that he has a "a growing set of concerns" around GPUs, including their shelf life: "people are taking a step back and saying 'what is the useful life of a GPU right here, right now?'" For example, XTX Markets' GPU fleet consisted of 10,000 A100 GPUs and 10,000 V100s in 2024; newer chips like Nvidia's Blackwell B200 can be more than 70 times more powerful according to Nvidia's specifications.
Griffin also warned of a point where pricing of GPUs would "go into freefall" if supply exceeds demand, at which point the firms that invested billions into these fleets might lose their edge. In Hudson River Trading's UK entities, it made additional spending of £2.4m per head on 'property plant and equipment,' although its unclear whether any of this was specifically for AI infrastructure.
At the same time, the health of cloud computing businesses is also under question. CoreWeave's stock price is down more than 37% for the month as of today, and options on its debt priced in a 60% chance of default in the next five years. If a compute provider had to shut down, the trading firms using its architecture would have to find a new provider, and might be susceptible to price gouging from other hyperscalers aware of the situation.
DRW, Citadel Securities, Jane Street and Susquehanna declined to comment. HRT and Jump Trading did not respond to a comment request.
Know of any other notable GPU fleets in quant trading firms. Get in touch with us through any of the following for a strictly off-record conversation. Telegram: @AlexMcMurray, WhatsApp: (+1 269 237 3950). Click here to fill in our anonymous form, or email alex.mcmurray@efinancialcareers.com
Bear with us if you leave a comment at the bottom of this article: comments are moderated intermittently by human beings. Sometimes these humans might be asleep, or away from their desks, so it may take a while for your comment to appear. You must take sole responsibility for comments you post on this site. We will take reasonable steps to weed out anything that we consider to be offensive or inappropriate.
