AI’s Real Bottleneck: Infrastructure, Not Intelligence

The rapid deployment of AI in high-frequency trading isn’t just accelerating decisions--it’s quietly reshaping the entire structure of financial markets, talent pipelines, and competitive moats. Hudson River Trading’s experience reveals that the real bottleneck isn't compute power or chips, it’s access to power, space, and trust in a world where infrastructure scarcity is now the decisive edge. This isn’t about who has the best models--it’s about who can run them at scale, who can secure capacity before it’s gone, and who treats AI not as a tool but as a new operating system for finance. For investors, technologists, and policymakers, this signals a shift: the future of markets will be determined less by ideas and more by physical logistics and contractual leverage. The advantage goes to those who see AI not as software, but as a real-world coordination problem with geopolitical and economic consequences.

The Hidden Cost of Fast Solutions: Why Compute Access Is Now the Ultimate Moat

When Iain Dunning says Hudson River Trading isn’t constrained by chips but by powering the chips, he’s exposing a seismic shift: the competitive frontier in AI-driven trading has moved from algorithmic brilliance to real estate, energy procurement, and counterparty credibility. The bottleneck isn’t technical--it’s logistical. You can have all the GPUs in the world, but if you can’t plug them in, cool them, and convince a data center you’ll pay your bills for the next five years, you’re out of the game.

This changes everything. Most firms still think in terms of model performance or data advantage. But HRT’s reality is different. They’re in constant dialogue with hyperscalers, neo-clouds, and data center operators--not just negotiating compute, but committing to long-term contracts, sometimes paying half upfront. Why? Because capacity is so scarce that leases open up and vanish within a day. You don’t get to shop around. You take what you can get. “You’ve got a megawatt there? I’ll take it.” That’s not strategy. That’s survival.

And here’s the kicker: the real scarcity isn’t even the hardware. It’s trust. Data centers don’t just want customers--they want reliable ones. They’re asking about bond ratings. They’re worried about credit risk. They’re hedging against the possibility that a firm takes all their power and then goes bust, leaving them with a long-term vacancy. So HRT isn’t just proving its trading edge--they’re proving their financial stability. becoming a creditworthy tenant in a world where compute is now leased like commercial real estate.

"We've had everything from people being like oh you've issued bonds what's the rating on those to not wanting us to sell too much of one site because if we take all their power rights and then go bust they might have a long lead time with a tenant and fill that."

-- Iain Dunning

This creates a feedback loop. The firms that can secure capacity are the ones already trusted. Those trusted firms grow bigger. Their scale reinforces their creditworthiness. Which lets them secure more capacity. The moat isn’t built with code--it’s built with contracts. And it’s widening fast.

Where Immediate Pain Creates Lasting Moats: The Unpopular Bet on In-House Chips

While most firms wait for NVIDIA to deliver the next Blackwell or Rubin GPU, HRT--and others in their peer group--are quietly building hardware teams. Not to compete with NVIDIA on training--but to control inference. Why? Because inference is a “strictly simpler technological problem.” And in a world where every millisecond counts, owning the stack means owning the latency.

This is a delayed payoff. It’s not glamorous. It doesn’t make headlines. But it’s where the real advantage hides. Because if you’re relying on off-the-shelf inference solutions--you’re at the mercy of supply chains, pricing shifts, and architectural compromises. If you design your own chip? You optimize for your specific workload. Your specific data. Your specific trading rhythm.

And yes, Jensen Huang sleeps never. NVIDIA is acquiring Grok. Broadcom partnerships are the new status symbol. But Dunning’s point is subtle: the inference space is smaller. The design space is constrained. That makes it doable. Not for everyone. But for firms that are compute-hungry, capital-rich--and already operating at the edge of physical limits? It’s not just possible. It’s necessary.

This is the 18-month payoff nobody wants to wait for. Building a chip team isn’t a sprint. It’s a years-long investment with no guaranteed return. But for HRT, it’s not about selling chips. It’s about optionality. It’s about not being held hostage by a single vendor when the entire business runs on speed.

The 18-Month Payoff Nobody Wants to Wait For: Reimagining Talent in the Age of AI

One of the most revealing shifts HRT is making isn’t technical--it’s cultural. They’re rethinking what talent looks like. Because if AI can now implement ideas--then do you really need the engineer who can code the GPU kernel? Or do you need the theorist who can dream up the idea in the first place?

"If you can't implement your ideas how does that happen exactly well now claude presumably implement the ideas so trying to embrace that maybe we do accept more theorists more dreamers people who can come up with ideas trusting that the implementation work can be done by ai."

-- Iain Dunning

This flips traditional hiring on its head. For decades, firms prized the “shape rotator”--the engineer who could optimize memory access patterns or squeeze out microsecond gains. Now, HRT is betting on the “word cell”--the person who can describe what they want clearly. Because that skill--precise, unambiguous communication--has become the bottleneck. And it’s not evenly distributed.

They’re moving toward an “open book” interview philosophy. Let candidates use AI. Because that’s how they’ll work on the job. Pretending otherwise is theater. The advantage goes to firms that stop testing for outdated skills and start hiring for AI-amplified imagination.

But here’s the catch: this only works if you have the infrastructure to run those ideas at scale. You can hire all the dreamers you want--but if you can’t get the compute to test their visions? They’re just philosophers. HRT’s talent shift only makes sense because they’ve already solved the harder problem: access.

How the System Routes Around Your Solution: The Myth of Infinite AI Progress

There’s a delirium in the air. Dunning admits it. “Feel this every day--worry it’s some sort of AI-induced delirium.” Because the empirical measures are exponential. Model releases are coming faster. Capabilities are leaping. But here’s what the fever misses: the system fights back.

HRT isn’t the only one deploying AI. So are all their peers. And when everyone gets better at predicting markets--what happens? Margins compress. The edge erodes. The “highlander” moment everyone fears--where only one firm survives--might not come from innovation. It might come from exhaustion. From the point where the cost of compute exceeds the alpha it generates.

And yet. The system adapts. Firms don’t stop. They go deeper. They build chips. They secure power. They redefine talent. The feedback loop isn’t just technological--it’s economic, logistical, human.

But even then--control slips. Dunning admits: they use “magical models” they don’t fully understand. They see emergent clusters--meme stocks and crypto stocks grouped together--despite no fundamental link. They can’t explain it. But they trust it--because it works. And they sleep at night not because they understand the model--but because they’ve wrapped it in automated risk checks. That’s the real story: not AI replacing humans--but AI changing the nature of control.

Key Action Items

• Secure compute access now, not later -- Commit to long-term data center contracts even if capacity exceeds current needs. The ability to deploy at scale is becoming the primary competitive lever. Over the next 6--12 months, optionality in power and space will be worth more than model improvements.

• Invest in hardware teams for inference optimization -- While training remains dominated by NVIDIA and Google, inference is a tractable problem for elite trading firms. This pays off in 12--18 months, but only if started now.

• Shift hiring toward idea generation, not implementation -- Prioritize candidates who can articulate novel concepts clearly. Assume AI will handle coding. Begin piloting “open book” interviews with AI tools allowed within the next quarter.

• Treat creditworthiness as a strategic asset -- Build relationships with data center operators as if you were a real estate tenant. Issue bonds, secure ratings, and demonstrate financial stability. This creates optionality in a supply-constrained market.

• Design risk layers around “magic” models -- Don’t wait to understand AI decisions before using them. Instead, build automated, auditable risk checks that allow deployment without full interpretability. This enables speed without catastrophic failure.

• Monitor counterparty risk in neo-cloud contracts -- As firms rely on specialized compute providers, assess their financial stability as rigorously as they assess their own. A provider outage could be more damaging than a model failure.

• Accept that delirium is a feature, not a bug -- The pace of change is destabilizing. Lean into it. Build systems that work despite uncertainty--not because of clarity. That’s where the edge lives.