The model layer is commoditizing: sell the frontier, trade the token tax

I spent the last few letters arguing that the honest signal in the whole AI trade is the price of compute, and that the token-maxxing era, when spending on compute was worn as a badge of seriousness, was ending. The spot price of an Nvidia H100 roughly doubled to around USD 3 an hour in the spring, and I said to watch that line more closely than any earnings call. I still would.

This piece is the next question down. If compute is the cost, the model is the product built on top of it. And the product is getting cheap. The frontier itself keeps moving; what is collapsing is the distance between the frontier and free. That distance used to be measured in years. It is now measured in months. When the gap between the best model money can buy and the best model anyone can download for nothing closes that fast, the economics of the companies selling the closed model change in a way the private valuations have not yet admitted.

My conclusion sits in the title. For a general investor the frontier labs are the part of this story to be most careful about owning at today’s prices. The durable value is one layer down, in the compute and power that every model still has to rent. And the single most important thing to price over the next two years is political: who gets to write the first serious tax on AI usage, and whether that author is the industry itself.

Parity is arriving in months

What “frontier” and “open-weight” actually mean. A frontier model is simply the most capable AI system available at a given moment, the one the leading labs, OpenAI and Anthropic in the United States, charge the most to use. An open-weight model is one whose finished “brain”, the trained numbers that make it work, has been published for anyone to download and run on their own machines. Nobody has to pay the original lab to use it. For three years the frontier labs held a comfortable lead: the best open model you could get for free was roughly a year behind the best model you could buy.

That lead is compressing. The latest open-weight releases out of China, with Moonshot’s Kimi K3 the one drawing the most attention this summer, are landing within striking distance of the paid frontier within months of each new closed release, and at a fraction of the running cost. The exact benchmark gap deserves independent verification, but the direction is not in doubt, and it rhymes with a number OpenAI’s own CFO has put on the record: the cost of running a given level of intelligence fell about 97% in two years, from GPT-4 to GPT-4o. When the cost of yesterday’s frontier falls by that much, yesterday’s frontier becomes today’s giveaway.

OpenAI’s CFO framed the survival strategy plainly: the durable edge is “context and memory, not the model”. Read that as a confession. When the person building the most valuable AI company in the world tells you the model is no longer the moat, the model is no longer the moat. The intelligence is becoming a utility. The value migrates to whoever owns the scarce inputs underneath it and the customer relationship on top of it.

The dark-token problem

Half your bill is a scratchpad you never see. AI systems are billed by the token, roughly a word-piece, counting everything they read and everything they write. The newest models are “reasoning” models: before they answer, they generate a long private train of thought, working the problem out step by step. You never see most of that scratchpad. You pay for all of it. On the current generation of reasoning models a large share of the tokens on your invoice are these hidden “dark tokens”, and for the heaviest reasoning modes that share is reported to run past half. As a working number, more than one token in two that you are billed for is invisible thinking.

Here is why that matters for the closed labs. As long as the frontier was clearly ahead, customers accepted paying for the scratchpad because the answers were better than anything else available. Put a good-enough open-weight model beside it, one you run yourself at cost, and the calculus flips. The math is already sobering, and I laid it out last month: an agent left running continuously against a frontier API can cost around USD 300 a day, roughly USD 100’000 a year, for perhaps a fifth of a job. Ask a chief financial officer to keep paying that when a downloadable model does 90% of the work for the price of the electricity, and the pricing power of the closed API starts to leak. Microsoft was reported this spring to be throttling Claude usage, and Uber said its annual AI budget was gone by April. Demand for intelligence is fine. What those episodes show is the bill for it landing on a real income statement.

What this does to closed-lab economics

A commoditizing product, financed like a scarce one. The frontier labs are being valued, in private rounds and in the pre-IPO chatter, as if their lead were permanent. A pre-IPO multiple is just the price private investors pay for a slice before a company lists, and right now those prices assume the model stays special. Anthropic is reported to be near its first profitable quarter and a possible October listing. OpenAI is projecting a loss around USD 14 billion in 2026. The capital keeping this going is enormous and circular: Anthropic’s reported USD 45 billion, three-year compute arrangement; Amazon putting up to USD 25 billion into Anthropic; Nvidia committing up to USD 100 billion into OpenAI, which spends much of it back on chips. Money runs from chipmaker to lab to cloud and back again.

While liquidity expands, that loop reads as a virtuous cycle. The day it stops, it reads as vendor financing. And the product at the centre of the loop is the one thing in the chain that is losing its scarcity fastest.

I hold to the business-quality lens here, the one Bill Ackman would recognise: own defensible, cash-generative franchises, and refuse to pay a dream price for a story. A frontier lab today is a company with negative free cash flow, a product whose price floor is being set by a free Chinese download, and a balance sheet braided into its own suppliers. That is the profile an investor should be slowest to buy at a private-market premium, however good the technology is, and the technology is genuinely extraordinary. Anthropic’s own institute writes that its model “has gone from super helpful to superhuman in under a year”. Both things are true at once: the capability is real, and the business selling it is commoditizing. An investor is paid for the second fact, not the first.

Where the durable value sits

Own the constraint underneath, and let the abstraction on top go cheap. This is the same conclusion I reached from the compute side, arriving now from the model side, which is usually a sign it is right. When the model becomes a utility, the money accrues to whoever supplies the scarce inputs the utility runs on: the chips, the firms that make the machines that make the chips possible, the optical and networking layer, and above all power generation and the grid. One gigawatt of compute translates to roughly USD 10 billion of annual revenue for whoever supplies it, on OpenAI’s own math. That number does not care which model wins. It gets paid whether the frontier is a paid API from San Francisco or a free download from Beijing.

The model layer commoditizes. The layer beneath it, physical, permitted, slow to build, does not. For a general investor that is the cleaner place to hold AI exposure: the toll on the road rather than the car of the month. The exit signal is unchanged from what I have written before. Watch for a genuine slowdown in hyperscaler capex intent: delayed builds, paused orders, a pivot from building capacity to defending margins. Reports that a meaningful share of planned US data-centre builds are being delayed or cancelled are the first data point that could turn into that signal. It has not turned yet.

The liquidity that is holding it up

None of this breaks while money is still expanding. I want to be honest about what keeps the circular financing standing. US broad money is growing about 5.6% year on year, an expansion, and the Fed’s balance sheet is still near USD 6.75 trillion. Policy is mildly restrictive, real short rates are only about +0.2%, and the yield curve has turned positive again. In Michael Howell’s framework, refinancing and collateral, not the level of rates, drive global liquidity, and that tide is coming in, not going out. Rising liquidity is what lets a chipmaker fund a lab that funds a cloud that buys the chips, and lets the market call the loop a virtuous cycle.

So the commoditization I am describing is a structural fact working underneath a supportive liquidity backdrop. The frontier-lab economics erode regardless. Whether that erosion shows up as a gentle repricing or a sharp one depends on liquidity. A stall in money growth, or a stall in hyperscaler capex, is what converts a slow squeeze into a fast one. I hold that timing as a risk to watch, not a forecast.

The token tax

The next moat will be written in law. Here is the part I think markets are not pricing at all, and the reason I am writing this now rather than in six months. When a technology’s private economics collapse, the firms losing their moat rarely accept it quietly. They look for a new one. The most reliable place to find a moat that code cannot dissolve is regulation.

In April I walked through a striking academic paper, “The AI Layoff Trap” by Falk and Tsoukalas, whose conclusion was that only a Pigouvian tax on automation could fully correct the damage that firms racing to replace workers impose on everyone else. A Pigouvian tax simply sets the tax equal to the harm an activity does to people outside the transaction, the way we tax pollution so the factory feels the cost of the muck it dumps in the river. I found the argument serious then and I still do. What I did not spell out is who benefits if that idea becomes policy.

A tax on AI usage, a levy per token, a safety-licensing regime, a mandatory audit before deployment, would be sold to the public as protecting workers and containing a dangerous technology. It might even do some of that. But watch the second-order effect. A fixed cost of compliance, licensing, audits, a per-token levy, lands hardest on the competitor whose product is otherwise free. A free open-weight model with a USD 0 price and a mandatory USD 10 million compliance bill is no longer free. The frontier labs, whose pricing power is being dissolved by exactly those free models, have every incentive to help write that rule. This is regulatory capture: the firms a rule is meant to constrain end up drafting it so it protects them. A policy sold as a brake on AI becomes the moat the technology just took away.

That fight is a distinct, priceable event, and I want to treat it as its own node rather than as background noise. It behaves like a policy switch that ripples across many assets in a way you can position for in advance, the way the market learned to trade the Iran shock through terms of trade rather than as a vague risk. A serious token tax would widen the gap between the infrastructure layer, which gets paid either way, and the application layer, which pays the levy. It would be quietly bullish for the incumbent labs it is nominally aimed at, and bearish for the long tail of small AI-native companies built on cheap open models. And it would arrive dressed as consumer protection, which is what will make it hard to see coming.

I hold the direction of this with high conviction and the timing with low. The regulatory-capture incentive is structural and I am confident it is there. When the first serious proposal lands, and whether it clears any legislature, I cannot call.

What I am watching next

The tell will be authorship. When the first serious framework to tax or license AI usage appears, read who wrote it before you read what it says. A proposal drafted against the frontier labs, capping their prices or forcing their weights open, points one way. A proposal the frontier labs quietly helped draft, full of compliance floors that a free download cannot clear, points the exact opposite way, and it will be wearing the language of safety while it does. Same headline, opposite trade.

That authorship question is the whole thing. Kimi K3 told us the model is becoming a commodity. The next chapter is whether the industry lets the commodity stay cheap, or whether it goes to a legislature and buys back the moat that the code gave away. I will be reading the fine print, not the press release.

Markets research and commentary for a general readership, not personalised investment advice. Several figures are reported ranges or working numbers, flagged as such in the text, and should be verified independently before being relied upon.