Frontier AI as HFT: Compute Is the Edge

Frontier AI is usually framed as a research race. The better analogy is HFT: compute, infra, data, capital, and execution. The moat is not the idea; it is the machine that makes intelligence cheaper.

The business model of frontier AI labs is often described as a research race. That framing is too narrow. Research matters, but it is unlikely to be the durable moat by itself. Researchers move. Papers diffuse. Architectures get copied. New techniques are quickly absorbed by competitors. A better analogy is high-frequency trading: the visible strategy may not be impossible to understand, but the durable advantage comes from the full machine around it — data, infrastructure, latency, capital, simulation, execution, and relentless optimization. Frontier AI increasingly looks similar. The moat is not a single algorithmic insight. It is the vertically integrated system that discovers small edges, validates them faster, serves them cheaper, and monetizes them at massive scale.

In HFT, the spread compresses over time. Firms compete to quote tighter, react faster, manage adverse selection better, and capture smaller and smaller units of edge. What begins as an algorithmic opportunity eventually becomes an infrastructure race. The winners are not simply the firms with the cleverest trading idea. They are the firms with the best data, the best simulation, the best execution stack, the lowest latency, the lowest cost, and the deepest capital base. The edge moves from “the strategy” to the machine that repeatedly finds and exploits marginal advantages.

Frontier AI has the same shape, but with one important difference: the “spread” is not simply compressing. Intelligence-per-dollar is rising. The frontier keeps moving outward, and each fixed level of capability becomes cheaper to serve over time. So the competitive question is not merely “who has the smartest model?” It is: who can deliver a given level of intelligence at the lowest cost, and who can compound that advantage fastest?

That is why compute is the edge. This point is still underrated by the market. Compute is often treated as raw capex, but at frontier scale it is much more than that. Compute determines how many experiments a lab can run, how quickly it can run them, how large a model it can train, how efficiently it can serve inference, how quickly it can collect feedback, and how aggressively it can reinvest usage back into the next training loop. Compute is not just an input to the business. It is the substrate of iteration speed.

The latency race in frontier AI therefore has two meanings. The first is research latency: the speed from hypothesis to training run to eval to deployment to user feedback to the next model. The second is economic latency: the speed at which a lab can reduce the cost of serving the same level of intelligence. If two labs have comparable model quality, the one with better utilization, better kernels, better networking, better routing, better distillation, better memory management, and better inference scheduling can quote the tighter “AI spread.” In that world, infra is not backend plumbing. Infra is product strategy.

This also means research itself is less defensible than the industrialized research loop. A new architecture, post-training method, or RL recipe can diffuse. But the full loop is much harder to copy: large-scale clusters, distributed training reliability, data pipelines, eval harnesses, deployment telemetry, inference fleets, enterprise feedback, product surfaces, and capital access. Just as an HFT trader can leave one firm but cannot easily take the entire execution machine with them, an AI researcher can leave a lab but cannot carry away the full compute-data-deployment engine.

The likely result is consolidation, but not necessarily monopoly. HFT is structurally closer to zero-sum: firms fight over a finite pool of market microstructure edge. Frontier AI is still expanding-sum. New demand is being created across coding, office work, customer support, education, research, agents, media generation, and enterprise automation. Fast market growth can support multiple winners for longer, even when their cost structures differ.

There are also natural forces against a single-player endpoint. Enterprises do not want one intelligence supplier. Governments do not want one corporate or national choke point. Hyperscalers want strategic independence. Open-weight models create pricing pressure from below. Products are differentiated by trust, UX, ecosystem, privacy, latency, compliance, and workflow fit. So the most likely endpoint is not monopoly, but oligopoly: a few frontier-scale labs at the top, open models and specialized vertical players underneath, and infrastructure suppliers capturing a large share of the economics.

This brings the hardware layer into focus. If compute is the edge, then the strongest position is not merely access to capacity. It is control over the full-stack compute system: accelerators, rack architecture, networking, compiler, kernels, training framework, inference scheduler, memory hierarchy, and utilization software. That makes NVIDIA and AMD strategically important. Their GPUs and rack-scale systems are not just commodity hardware. They are the arms dealers for the frontier AI race. For an independent frontier lab, NVIDIA or AMD infrastructure is strategically cleaner because these suppliers do not directly own the competing model layer in the same way hyperscalers do.

This is what makes Anthropic’s infrastructure position such an interesting thought experiment. Anthropic has relied heavily on AWS, including Trainium, and has also used Google TPU capacity. That may be rational in the near term because the immediate constraint is simply access to enormous compute. But in an HFT-like framework, it is strategically counterintuitive. If compute is the edge, then relying deeply on AWS Trainium or Google TPUs means part of your edge is mediated by partners who also want to capture the AI platform economics themselves.

Google’s decision to share TPU capacity is equally interesting. If TPU is a genuine infrastructure edge, why share it? The answer is probably that Google is optimizing for more than Gemini alone. By selling TPU access, Google Cloud monetizes its infrastructure, amortizes investment, pulls strategic customers into its ecosystem, and prevents important labs from becoming exclusively dependent on AWS, Microsoft, or NVIDIA. But the tradeoff is real: Google converts some of its internal edge into platform revenue, while also leaking part of that edge to a model competitor.

So the closing question is this: in frontier AI, who ultimately captures the economics — the model labs, the cloud platforms, or the rack-scale hardware suppliers? If the industry is truly HFT-like, the deepest profits may accrue not simply to whoever has the best model at a point in time, but to whoever controls the machine that makes intelligence cheaper, faster, and more scalable. That machine may sit inside a frontier lab. It may sit inside a hyperscaler. Or it may increasingly be defined by NVIDIA and AMD at the rack level.