The price of raw intelligence fell 80 percent overnight when OpenAI cut GPT-5.6 Luna to $0.20 per million input tokens. The move looked like a pricing event, the kind of thing that moves a line on a chart and then gets absorbed into the next news cycle. It was actually a disclosure. When the headline number on a product drops that far that fast, the thing being sold has stopped being the thing that carries the value, and everyone still shopping on that number is buying something other than what they think they are buying.
The argument of this piece is simple and it runs through every layer below: the model is not where the value sits anymore. Capability is converging toward commodity, and the scarce things are the rails underneath it. Power, cooling, custody, routing tables, and the balance sheets that fund all four. The demo draws the attention. The rails decide who keeps the money.
Per-Token Price Is the Wrong Unit
Per-token price is vanity. AlphaSense data shows frontier US models can cost less to actually use than cheaper Chinese alternatives, even while charging more per token, which is a genuinely strange sentence until you look at what a token is actually doing. A token is not a unit of work. It is a unit of output, and a model that wanders, restates, and second-guesses itself burns three times as many of them getting to the same answer as a model that goes straight there. The sticker number stopped being the real number weeks ago. What matters is total cost to solve the problem, and the models that finish tasks using fewer tokens end up cheaper regardless of their rate card.
That distinction has a procurement consequence that most buyers have not absorbed yet. Buyers who optimize on listed price per token are paying for the wrong unit of measurement. They are buying the ticket, not the journey, and the ticket price tells you almost nothing about how long the trip takes or how many transfers it requires. A finance team that builds its AI budget around rate cards will end up with a lower headline cost and a higher actual bill, and it will not be able to explain the gap, because the gap does not appear anywhere in the contract it negotiated. The vendors who understand this are already pricing against it. The ones who do not are competing on a number that describes their product less every quarter.
Buybacks, AI Capex, and the Share Count
That gap between listed price and delivered cost is where the real edge sits, and the same gap between what a number says and what it means is now showing up on the other side of the market. Wall Street is feeling it too. The trillion-dollar buyback engine that propped up valuations for years is stalling. Companies are issuing shares instead of repurchasing them because AI infrastructure capital demands ate the cash the market never priced in. The force that once supported multiples is now diluting them, which is not a subtle reversal. It is the same lever pulled in the opposite direction.
Buybacks and AI capex are two sides of the same balance sheet, and the side that wins determines the share count. That is the whole mechanism. A share count is not a footnote. It is the denominator under every per-share figure a market uses to decide what a company is worth. High-flying stocks that priced in perfection are now funding their own infrastructure by expanding the denominator, which means the earnings can grow and the per-share story can still deteriorate. An investor reading only the top line will see a company winning. An investor reading the share count will see who is paying for the win.
Compute as an Asset Class
Compute has already become an asset class for the people who see it. Scarcity shifted from weights to power, which is why Nvidia occupies the position of a bank and Cloudflare just put an ATM in every edge node. Those are not decorative comparisons. A bank’s power comes from sitting where capital has to pass through, and an ATM’s value is proximity, not the cash inside it. Both companies have arranged themselves around a flow rather than a product, and flows are harder to disintermediate than products are.
The next wave of agents will spend compute the way we spend money, routing decisions through whichever node has capacity and a balance sheet to back it. Once that behavior is normal, the question stops being which model an agent calls and becomes which node will clear the transaction. The bottleneck is no longer the model; it is the interconnect, the power feed, and the custody chain. Every one of those is physical, permitted, and slow to build, which is exactly what makes them defensible.
Renewable generation hit the same wall last year, except the story was framed as a cost problem when it was actually a timing problem. You can build cheap solar; you cannot will the sun to shine at 6pm. Cheap generation that arrives at the wrong hour is not cheap power, it is an asset that needs another asset to be useful, and the second asset is where the scarcity moved. The AI buildout has an identical shape: the model is cheap; the infrastructure that makes it reliable is scarce, and scarcity is what makes an asset valuable.
Routing, Custody, and the Rails That Decide
The AI adoption wave splits companies into two kinds: those redesigning work and those slapping AI onto existing meetings and calling it transformation. Culture change is never the tool; it is the constraint. The tool can be bought in an afternoon and the constraint takes years, which is why the second group keeps buying more tools and getting the same result. The firms that treat this as an infrastructure problem will outrun the ones that treat it as a feature launch, and the gap will look like a capability gap from the outside even though it is a design gap on the inside.
The OpenRouter acquisition tells the same story from a different angle. The $7 billion price is not about model switching. It is about owning the decision point between an enterprise and its AI, and a decision point is worth more than the things it decides among. That position shapes behavior. When every agent has to route through your table, you set the defaults, and defaults become policy, because nobody re-examines a default that is working. Routing is the new distribution, and the winners will not be the smartest models but the ones that control the dependency graph.
Custody is the next threshold, and it rhymes with routing for the same structural reason. Citi adding $BTC to bank-grade custody is being framed as a crypto win. It is actually something narrower and more important: trading gets the headlines, but custody is what makes an asset real to a compliance department. An institution cannot hold what it cannot account for, and no amount of price action fixes that. The institutions waiting on the sidelines got their answer when the plumbing got approved. Real-time settlement for everything is already live; crypto custody was just the next item on the list.
India learned a version of this with its silver licensing regime. The policy meant to fix a premium ended up creating it, because customs paperwork blocked most shipments and the market ran on existing inventory. Nothing about the underlying supply changed. The paperwork changed, and the paperwork turned out to be the real constraint. Regulation aimed at solving a problem often becomes the bottleneck itself, and the same thing happens in AI: the layer everyone ignores is usually the one controlling throughput.
The pattern repeats across every layer. The demo draws attention. The rails determine who keeps the value. Prototypes exist because most people cannot see the forest through the trees. The forest right now is power, cooling, custody rails, routing tables, and the balance sheet items that fund them. The tree everyone is photographing is the model name.
So the smart money is not betting on which model wins. It is betting on which infrastructure the winning model cannot escape. The companies that survive the agent wave will not be the ones with the smartest models. They will be the ones with the best balance sheet, the deepest custody integrations, and the cleanest control over the dependency graph. Model capability is a commodity. Infrastructure is a position, and positions compound.
Leave a Reply