The Post That Made Retraining Look Cheap

  • AI
  • November 16, 2026
  • 0 Comments

Jacob Coxon spent his working hours on pretraining, the stage where a model is fed enough text to learn what language is. When he left Anthropic, he did not go quietly onto a competitor’s payroll. He went public with a claim about arithmetic: algorithmic efficiency is improving on a super-exponential curve, and the compute required to retrain a model of the same standard is falling sharply.

Anthropic has a name for the phenomenon, Closing the Loop, and the name is the interesting part. It describes a research program in which better models help design better training runs, which produce better models, and each improvement pays for the next one.

For anyone who sells compute by the hour, that is a threat wearing the costume of a research note. A lab that can reach the same capability with a smaller run needs fewer chips to hold its position, and a lab that holds its position more cheaply has less reason to keep expanding the cluster.

What the Curve Does at the Margin

Super-exponential is a strong phrase, and it should be read as a claim about compounding rather than a forecast. Efficiency gains in training do not arrive as a single discount. They arrive as a series of tricks, each of which makes the next one cheaper to find: better data mixtures, better schedules, better ways of deciding where to spend a step.

The compounding matters because the industry has priced its future on the opposite assumption. Capital spending plans, power contracts, the multi-year leases behind new data halls, all of it rests on the idea that the next model costs more than the last one and needs a bigger building to train. A curve that bends the other way does not end demand. It changes who can afford to stay in the race.

There is a second-order effect that cuts against that conclusion. Cheaper training has historically expanded the number of people who train, and a technology that gets cheaper tends to get used more rather than less. The labs themselves are the proof: the same organizations announcing efficiency gains are also the ones signing the largest compute commitments in the industry.

The Other Balance Sheet

Coxon’s post landed in the middle of a separate argument about Nvidia, one that has less to do with algorithms than with trust. The company has spent years acting as something close to a central bank for the AI economy, selling chips and quietly helping finance the customers who buy them. That role works while everyone believes the lender is solvent, patient and permanent.

Permanent is the part under question. With Jensen Huang gone from the chief executive’s chair, the exposure his tenure created has become visible: a $99 billion equity stake in the company, $36 billion in cloud agreements, and guarantees behind both. If that exposure contracts, the pressure travels into the balance sheets of the customers who leaned on it, and the hole it leaves is a liquidity squeeze of more than $50 billion.

The politics around the succession have not helped. The Financial Times reported that Huang may return to the job, while Washington has been pushing the board toward an outside hire rather than a restoration. A company that is simultaneously a supplier, a financier and a symbol is not an easy thing to hand to a stranger.

Two Curves, One Question

Put the two stories on the same page and the problem sharpens. One curve says the cost of a given capability is falling. The other says the commitments made on the assumption of rising cost are enormous, long-dated and partly unhedged.

If the first curve is real and the second set of commitments is fixed, the gap has to close somewhere. It can close through demand: more companies training more models, absorbing the capacity at lower prices. It can close through consolidation, with the smallest labs renting rather than building. Or it can close through a repricing, which is the version nobody in the industry wants to discuss on the record.

What makes the argument hard to settle is that both claims can be true at once. Efficiency can improve on an aggressive curve while the absolute amount of compute consumed keeps climbing, because the frontier keeps moving and someone always pays to be at the front of it. The disagreement is not about whether the trick works. It is about which number grows faster.

The Man Who Left the Room

Coxon cannot test his own claim from where he now sits. The evidence that would settle it lives inside the training runs of the labs that employ people like him, and those numbers appear in public only after they have been packaged for investors.

So the post functions as a marker. It says that the cost of a capability point is falling faster than the people selling the hardware would like to admit, and it says so from inside one of the companies that benefits most from the old arithmetic. The next earnings season will be the first place where the two curves are forced to sit next to each other, and the executives presenting those decks will be answering a question Coxon raised in a farewell note: whether the industry is buying intelligence, or only the machinery for making it cheaper.

Related Posts

  • November 20, 2026
  • 16 views
Sandisk Completes Eight-for-One Split as NAND Outlook Darkens

Eight for one. Sandisk’s stock split took effect on Nov. 19, and the shares began trading on the adjusted basis on Nov. 20, a change that divides the price by…

  • November 20, 2026
  • 14 views
A Lawmaker’s Privacy Suit Puts Model Design on Trial

For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed…