Why retail traders staring at TradingView indicators are competing with a firm that runs hundreds of thousands of GPUs, FPGAs wired directly to exchange fiber, and signs $6 billion compute contracts like it’s ordering office chairs.
If you still think Jane Street is just another Wall Street market maker with smart traders and fast machines, you are roughly a decade behind the actual story. The firm that booked $20.5 billion in net trading revenue in 2024 and posted a single quarter of $10.1 billion in Q2 2025 is no longer competing with banks. It is competing, infrastructurally and architecturally, with frontier AI labs like OpenAI and Anthropic. The only difference is that its model outputs are not chatbot replies. They are bids, offers, and hedges across 200+ venues in 45 countries.
This post breaks down how Jane Street is actually built underneath the trading desk, drawing from a recent walkthrough with Ron Minsky (co-head of technology) and Dan Pontorvo (head of physical engineering) at one of their Texas data centers. If you build trading infrastructure for a living, or you trade against firms like this in NSE F&O, the architecture matters more than the P&L headlines.
What is the actual time horizon Jane Street trades on?
Jane Street operates across multiple time horizons simultaneously, from sub-100 nanosecond FPGA decisions wired into exchange fiber to multi-hour model inference runs, with completely different hardware stacks for each regime.
There is no single “Jane Street trading system.” There is a portfolio of trading systems calibrated to entirely different physics. At the fastest end, Minsky describes a regime where if you put an oscilloscope on the inbound and outbound network wires, you would see the packet starting to leave before the inbound packet finished arriving. At that speed, the choice of programming language is irrelevant. CPUs are too slow. The decision logic lives on an FPGA hardwired into the network card, and the “model” is whatever simple computation can be expressed in gates.
At the other end, there are models that have an hour or a full day to think, and inference can happen on a GPU cluster in a data center hundreds of kilometers away from the exchange. In between sit microsecond and millisecond decisions that run on CPUs or specialised silicon.
The tradeoff is brutally simple. Smarter decision equals slower turnaround. The trick is building an ensemble that picks the right altitude for the right kind of question.
What do Jane Street’s models actually predict?
The simplest and most important target is fair value, what an instrument is currently worth, which then composes into many different downstream trading processes across asset classes.
This is worth pausing on, because it is genuinely different from how most retail traders and even many quant shops frame the problem. Most retail strategies predict direction or breakout. Jane Street’s primary prediction target is a number, the fair value, and the trading logic around that prediction is a separate composable layer. The same fair-value signal can drive a market-making quote, an ETF arbitrage, or a hedge sizing decision. The model does not need to know what the trading strategy is. It just needs to be right about what the thing is worth.
Minsky has been doing variations of this since 2000 with linear regression. The math is older than most algo traders. What changed is the size and depth of the models that now sit behind that same prediction target.
Why does Jane Street need its own data centers when colo space exists?
Because at sub-microsecond regimes you measure the physical length of the fiber spool, and at large model regimes you cannot fit enough power, cooling, and GPU density inside a third-party colocation cage next to the exchange.
This is the part most fintech bloggers miss. Yes, Jane Street has colocation racks next to NSE, NASDAQ, CME, and every other major venue. For the FPGA layer, that proximity is non-negotiable, and they will literally measure the length of fiber runs to shave nanoseconds.
But the moment you want to run a real machine learning model with serious parameter counts, colo space is the wrong tool. Third-party colos limit power per rack, dictate cooling configurations, and force you to spread one GPU across multiple racks because each draws too much wattage. You cannot run liquid-cooled NVL72 systems in someone else’s facility on someone else’s rules.
So Jane Street builds. Its own facilities. Its own power, cooling, generator strategy, networking, and storage tiers. The Texas data center exists because the firm needs to control every variable that affects training throughput.
How big is Jane Street’s compute footprint?
The firm currently runs in the range of tens of thousands of GPUs and is on track to scale to hundreds of thousands, backed by a roughly $6 billion CoreWeave commitment plus a $1 billion equity investment in CoreWeave at $109 per share.
To put that in perspective:
The April 2026 CoreWeave deal gives Jane Street access to NVIDIA Vera Rubin generation compute across multiple facilities, with dedicated connectivity, custom storage configurations, and integrated software services. CoreWeave’s own SVP of Revenue, Max Hjelm, described Jane Street as operating “like a frontier lab.” That is not a marketing line. Jane Street’s 2024 net income of $13 billion is comparable to what it costs to train several frontier language models from scratch.
This is a market maker with the balance sheet to buy compute at the same scale as OpenAI and Anthropic, and the discipline to actually use it.
Is Jane Street training one giant model like a foundation lab, or many smaller ones?
Jane Street trains many specialised models with diverse architectures, optimised for different data sources, data rates, latency budgets, and the unique byte-to-flop ratio of financial data, which is far noisier and far higher volume than text.
This is a critical architectural decision. The frontier AI labs converged on the idea of one general model that does everything. Jane Street did not. Their data pipeline forces specialisation.
Why? Because financial data is, byte for byte, less informative than language data. A token of English carries more semantic signal than a tick of NSE order book data. To compensate, Jane Street ingests vastly more data, but the data is noisier and the optimal model size is often smaller. The byte-to-flop ratio is fundamentally different. That changes everything downstream: how you store data, how you batch it, how you shape the model, how you design the inference path.
The result is a research culture where teams are constantly trying very different architectures in parallel. Faster iteration time for researchers turns out to be one of the highest-leverage investments the firm can make.
What is the actual inference workload like compared to an LLM company?
Latency matters more, batching dynamics are different because models are often disaggregated per symbol, and the sequential data rate from any single feed like the NASDAQ direct feed is dramatically higher than what any single LLM user generates.
Think about it this way. An LLM provider like Anthropic or OpenAI has massive aggregate data rates from millions of users, but each individual user types slowly. A single conversation is not a high-bandwidth firehose.
Jane Street’s inference path looks the opposite. A single market data feed, the raw bytes coming out of NASDAQ or NSE, is causally sequential and extremely high bandwidth. The model has to consume it in order, in real time, without falling behind. That changes the entire engineering optimisation surface. The same fundamental questions about throughput and latency apply, but as Minsky put it, “all the constants are twiddled to different places,” which leads to different architectural choices.
Why did Jane Street give up its “one CPU architecture” simplification?
Because power constraints forced compute to be distributed across multiple data centers, and NVIDIA’s roadmap now requires ARM support, so the old “x86-64 everywhere” shortcut collapsed under scaling pressure.
For years, Jane Street took advantage of an aggressive simplification: pretend the world only has one CPU architecture. Pretend you only have one research data center and one storage cluster. Both assumptions made the engineering surface dramatically smaller.
Both assumptions are now dead.
You cannot wire enough power into a single data center to run modern AI training. You have to build across geographies. Once compute is geographically distributed, storage scheduling and compute scheduling become intertwined, and moving training data between sites becomes a non-trivial engineering problem in its own right. Jane Street has been building its own large-scale object store internally to handle this, because vendor products do not quite fit the research scale and multi-site requirements.
And NVIDIA’s Grace-based ARM products are now too compelling to ignore, which kills the x86-only assumption.
Is human judgement still in the loop, or is it all automated?
Humans are always watching, and they matter most during phase transitions, the volatile days when weird things happen and models trained on previous regimes start to misbehave, which is often when market makers earn the most.
This is one of the more counterintuitive parts of the interview. Jane Street’s most profitable days are often the chaotic ones, when the world goes sideways, liquidity dries up, and providing markets becomes risky. Those days pay well precisely because they are hard.
And on those days, humans matter more, not less. Minsky’s framing is that models work well within a regime but humans work better through phase transitions. The meta-judgement of “today is different from the last 500 days, and here is how” is still something humans do better than machines. So even the largely automated systems have humans watching, ready to override.
This has a quiet implication for the AGI debate. Even at a firm with this much compute, this much modelling sophistication, and this many PhDs, the human element has not been engineered out. If anything, the senior engineers there say humans are more valuable than ever, because everything they touch now leverages more capital and more compute.
Where are the actual hardware bottlenecks in 2026?
Generators, transformers, and liquid cooling components are the binding constraints right now, with lead times that change every two weeks, which is why Jane Street stockpiles fungible long-lead items in warehouses and is moving toward modular pre-built infrastructure.
Pontorvo’s part of the conversation was the most underrated. The headline AI bottleneck story is always about GPUs and HBM memory. The real bottleneck inside a hyperscale build right now is the boring industrial equipment. Generators have multi-month lead times. Transformers too. Liquid cooling gear is in heavy demand because rack densities keep climbing.
Jane Street’s response has been pragmatic: warehouse fungible long-lead items so they can be deployed across any of their facilities. For non-fungible items like industrial-scale generators, work the markets earlier and stage them carefully. And increasingly, push toward modular pre-built infrastructure that can be shipped to site as close to plug-and-play as possible.
One quietly interesting decision they have made: stripping back generator backup so that only the most resiliency-critical parts of the data center are generator-backed. The tradeoff is that GPUs land six months sooner. As Pontorvo put it, that may not be the best engineering decision, but it is the best business decision, and at this scale that distinction matters.
What does this mean for Indian quants and retail algo traders?
You are not competing with traders. You are competing with a vertically integrated stack that includes custom silicon, FPGA networks, proprietary data lakes, multi-site GPU training clusters, and a research culture that out-iterates almost every other firm in finance.
Let us be honest about what this means for someone running an OpenAlgo setup, a Pine Script strategy on TradingView, or even a serious mid-sized prop desk in Mumbai.
When you take liquidity in NSE options, you are quite often interacting with a fair-value surface that has been continuously updated by a model trained on more data than most quant firms have ever seen, running on hardware that costs more than the GDP of a small district. Your edge does not come from outpacing that infrastructure. It comes from operating in pockets where this kind of infrastructure has no edge: smaller instruments, slower time horizons, structural inefficiencies, or behavioural patterns that big market makers do not bother to model because the strategy does not scale.
The firms quietly winning in Indian markets, the ones building thoughtful systematic strategies in F&O, equities, and commodities, are the ones who understand which fights to pick. You do not beat Jane Street at fair-value prediction in liquid Nifty options. You build something that exploits structure they cannot or will not capture. Volatility regime detection. Cross-asset correlation breaks. Event-driven setups. Microstructure quirks specific to Indian markets like the expiry-day dynamics or the cash-futures basis behaviour around dividends.
The era of “I built a moving average crossover, where is my edge” is over. The era of “I understand the market structure better than a generic global model” is still wide open.
The takeaway
Jane Street is the clearest example in finance of what happens when a firm fully internalises the AI infrastructure stack instead of treating compute as a vendor problem. They own the FPGAs at the exchange edge. They own the data centers. They own the storage layer. They build the models. They train them across distributed compute. They run their own inference. And they have $6 billion of forward compute commitments lined up to do more of it.
For Indian fintech builders, the lesson is not “try to be Jane Street.” That is a dead-end aspiration. The lesson is that the gap between a properly built systematic trading firm and a casual algo trader is no longer a 10x gap. It is a 1000x gap, and it is widening every quarter. The only sane response is to build at a layer where that infrastructure does not give your competitor a decisive edge, and to be ruthlessly honest about where it does.
If you are running OpenAlgo, AmiBroker, or a custom Python stack, your edge has to come from market understanding, asset selection, and disciplined risk management. The technology is necessary but no longer sufficient. The frontier of trading technology now lives at a scale that most firms in the world cannot afford to chase, and a small handful are simply running away with it.
Jane Street is one of those firms. Plan accordingly.
