Anthropic bought the dev tools startup that OpenAI, Google, and Cloudflare were already paying for. Read that sentence again. The company selling you the model now owns the workbench you use to build with everyone else’s model too. That is not vertical integration in the gentle sense. It is a frontier lab quietly making rent on its rivals’ developers while keeping a clear line of sight on what they are shipping next.
That deal is the cleanest single illustration of what this whole cycle is doing, which is concentrating. Capital is choosing silicon over nearly everything else it used to fund, and the choice is showing up in four separate places at once: in an acquisition, in trade data, in capital expenditure guidance, and in the order book of the public markets. Each gets reported in its own column inches. Together they are one trade recorded in four different ledgers, and the shape of that trade is the argument here.
What Anthropic’s Dev Tools Acquisition Buys
Start with the deal itself, because the logic is unusually explicit. Anthropic acquired the dev tools startup used by OpenAI, Google, and Cloudflare, which means the customer list came with the company. Those customers are competitors. They now pay a competitor for the tooling they use to build products that compete with that competitor’s models, which is an unusual position to find yourself in and an awkward one to unwind quickly.
If the model layer is effectively captured by a handful of parties, the application layer is the only place left where somebody outside that group can plausibly build something new. Owning the workbench at the application layer, meaning the development environment, the orchestration tools, and the evaluation harnesses, means owning visibility into where new entrants are pointing their effort. That is not a vanity acquisition, and it is not primarily a revenue acquisition either. It is a watchtower with a turnstile, charging admission to your competitors while showing you what they are building well before they ship it.
The $340 Billion Compute Import Number
The second ledger is physical, and it is the one least likely to appear in a technology column. Joseph Politano flagged that US imports of large computers surged amid the AI boom, running at roughly a $340 billion annual pace. Three years ago the number was a fifth of what it is today. That is physical machines crossing physical borders, paid for in physical dollars, cleared through customs, and installed in rooms that need power and cooling whether or not anything useful is running on them.
This is the detail that should reframe how the boom gets discussed. The AI story is told as a software story because software is where the demos live, and software has the margin structure that makes a narrative exciting. The trade data says it is a logistics story: containers, customs forms, and server rooms with the lights left on through the holidays. Software scales without incremental cost, which is the entire reason the industry has been valued the way it has for two decades. What is being built now does not have that property. It has a bill of materials, a lead time, and a depreciation schedule, and those three things behave very differently in a downturn than a codebase does.
Set the import number against the acquisition and the tension is informative. One is a capital-light move to occupy a strategic position. The other is the heaviest capital commitment the sector has ever made. Both are being financed out of the same conviction, and only one of them can be reversed cheaply if the conviction turns out to be early.
Cerebras and the Crowded IPO Window
The third ledger is the public market’s, and it rations more brutally than the private one. Cerebras pricing a blockbuster offering lifted the hype around SpaceX, OpenAI, and Anthropic while crowding out smaller players, and underwriters at those smaller AI shops are quietly redrawing their calendars because of it. There is a finite pool of generalist patience for any one sector in a given year. When a name with that kind of adjacency takes the oxygen, the rest of the room runs out of air, regardless of the quality of the businesses waiting in line.
That is the same concentration dynamic the acquisition displays, expressed through a different mechanism. In one case a lab absorbs the tooling layer. In the other, the market absorbs the entire year’s appetite into a handful of names. Neither requires anyone to act badly. Both produce the same outcome, which is a smaller number of surviving independent positions at the end of the cycle than at the start.
The capital expenditure side completes the picture. One of the largest platforms raised its 2026 spending guidance by ten billion dollars, to roughly $145 billion, and its finance chief admitted in plain English that the company keeps underestimating how much compute it needs. That admission is worth more than the number attached to it, because it is a statement about the shape of the forecast rather than its level. A company that keeps revising in the same direction is not making errors. It is discovering that the thing it is buying has no natural stopping point it can see from here.
The Toy Store Problem
Robert Cialdini tells a story about swearing off the toy store after Christmas and then finding himself back inside it in January, buying his son another expensive thing he had explicitly resolved not to buy. The mechanism, as he eventually worked out, was a subtle commitment trap the store had laid the month before. Read the admission about persistently underestimating compute alongside that story and the analogy does real work. The first build-out commits you. The second exists to defend the first. By the third you are explaining to shareholders that you simply must keep going, and the remarkable part is that you mean it, because by then it is true.
What this pattern leaves behind is the part worth watching. In any other decade, this much capital concentrating in one direction would have produced a wave of new entrants underneath it, because the leftovers of a boom historically become the seed capital of the next one. This cycle reads differently. Starting a competitive frontier effort now requires roughly the resources of a small national economy, and the tooling layer that a scrappy entrant would have used to move fast is being bought by the incumbents. The barrier to entry got tall enough to become a moat for the people who already own the moat.
None of this is alarming if you have made peace with concentration. Industries consolidate; that is a structural fact rather than a moral failure, and pretending otherwise usually produces worse analysis than accepting it. What is worth noticing is how fast this one is doing it, and how many apparently separate stories are actually one story wearing different clothes. The imported compute, the tooling acquisition, the crowded offering window, the guidance revision that keeps going the same direction: same trade, four ledgers.
The bill is being paid right now, in shipping containers, in allocations that never get made, and in dev tool invoices that quietly route to a new owner. The interesting part is not who is signing the checks. It is that the decision about who comes along has already been made, and the rest of the market is finding out the same way everyone else is, by reading the news. Intelligence at this scale does not have customers. It has a guest list.
Leave a Reply