The instinct when a frontier AI model gets cheaper is to assume value is being destroyed. It isn’t. Margin in a supply chain behaves less like a fixed prize and more like a fluid: squeeze it out of one layer and it doesn’t evaporate, it moves to whichever layer is least willing to give it up. The most interesting bull case in AI infrastructure right now rests entirely on that single observation — and it explains a strategic posture from Nvidia that otherwise looks like charity.
Two Reservoirs: Value and Volume
The AI stack currently holds its money in two separate places. The volume pool is enormous and cheap — the vast majority of tokens processed today already come from inexpensive models, many of them open-source, running everything from autocomplete to bulk classification. The value pool is smaller and expensive — the premium tokens from the smartest frontier models, sold at inference margins north of 90%, doing the high-stakes work people will pay almost anything for.
The critical fact is that these two pools barely overlap. Most tokens are cheap; most economic value accrues to the models that are not. Volume lives in one reservoir, profit in the other. Nearly every disagreement about the future of AI economics is, underneath, a disagreement about whether those two reservoirs stay separate.
What Happens if the Reservoirs Merge
Suppose market share starts shifting away from the fat-margin frontier models toward cheaper ones — open or closed, it doesn’t matter which. Follow the money in sequence.
Intelligence per dollar rises for the customer, because they are getting comparable capability for far less. Better return on AI spend pulls in more usage — the Jevons bet, that cheaper intelligence gets consumed more than proportionally rather than less. And the margin that used to sit trapped as frontier-lab profit does not disappear. Part of it returns to customers as improved ROI. The rest flows down the stack, because every one of those newly affordable tokens still has to be computed on someone’s silicon. Each token is worth less; there are vastly more of them; and the profit that was concentrated at the model layer spreads across the layers beneath it.
Compressed to a sentence: lower margin percentage at the model layer means more margin dollars at the infrastructure layer, all else equal. That is the whole thesis. It is not a demand story — demand grows in almost every scenario. It is a story about where the demand’s profit ends up settling.
Why This Is Nvidia’s Actual Game
Nvidia’s public enthusiasm for open-source models reads like ecosystem goodwill. It is closer to margin engineering. Nvidia wins in direct proportion to total tokens processed, and open source maximizes that number by putting model-building in many hands and driving inference everywhere. But there is a second, quieter motive worth separating out.
The historic fear for a compute supplier is monopsony — not many buyers, but one dominant buyer, or a small cartel of them, with the leverage to squeeze pricing, design custom silicon, and route around the supplier entirely. A handful of frontier labs consolidating all compute demand is exactly that risk. Open source diffuses model-building across a crowd, and a crowd of buyers has no leverage. The plausible read today is that Nvidia is less worried about buyer concentration than it used to be — demand is broad and the buyers are many — which means the dominant reason to champion open source is no longer defensive at all. It is the upside: commoditize the layer above you, and the margin drains toward the layer you occupy.
The Compressive Force Is Indifference, Not Competition
Here is the part most analyses miss. What actually collapses frontier margins is not a better competitor. It is a competitor who does not need the model layer to be profitable in the first place.
A pure-play frontier lab has to defend its inference margin; it is the business. But a vertically integrated player that monetizes somewhere else is under no such constraint. Meta captures its value through advertising and can treat a state-of-the-art open model as a loss leader that commoditizes a rival’s crown jewel. The xAI orbit captures value through its broader platform and can price just as aggressively for the same reason. Neither is trying to win the model layer in margin terms — both are, in effect, willing to burn it down. That structural indifference is the compressive force, and it has never been better funded. Challenger models are now closing the capability gap on genuinely useful tasks at a fraction of the incumbents’ cost, which makes ranking them a notch below the frontier already conservative.
A frontier lab can out-engineer a competitor. It cannot out-engineer a competitor who is indifferent to whether the layer makes any money at all.
Who Wins Each Layer in That World
If the reservoirs merge, the winning trait at each layer changes.
- Infrastructure: the prize goes to the lowest cost per token. Once tokens commoditize, compute cost leadership captures the volume — and volume is where the redistributed margin now lives.
- Model layer: the winner is no longer the most intelligent model but the most token-efficient one — the most capability extracted per unit of compute. Raw intelligence stops being the moat; efficiency becomes it.
The Honest Part: It Isn’t Happening Yet
None of this describes the present. Cheap and open tokens already dominate volume, but the smartest, priciest models still capture the majority of the economic value. The reservoirs remain separate. The thesis is a forward bet that the value pool drains into the volume pool — and that bet leans on two assumptions doing quiet, load-bearing work.
The first is elasticity. If cheaper intelligence does not drive more-than-proportional usage, then cheaper tokens simply mean a smaller total pie, and there is less margin to redistribute anywhere — the infrastructure layer included. The second is that “all else equal” clause hiding at the infra layer. Redistribution only rewards the infrastructure incumbent if infrastructure itself does not commoditize in parallel — and custom accelerators, rival GPUs, and inference-specific silicon are all direct assaults on precisely the cost-per-token prize the thesis hands to the infra winner. “More margin dollars at the infrastructure layer” is really “more margin dollars for whoever wins infrastructure cost leadership.” That is a fight, not a birthright.
Strip it down and the bull case is a single claim: margin in the AI stack is conserved rather than destroyed, and it settles at whichever layer resists commoditization longest. The bet is that the model layer cracks first — and that Nvidia is standing at the drain.
Leave a Reply