Margin Points - Arnold Engel<br>July 28, 2026 · [Essay 134]<br>Distilling frontier AI models<br>AI labs are discovering that preventing frontier models from being distilled is as important as the R&D that goes into the models.
Made possible by<br>Spacebar Studios<br>The #1 Newsletter Agency For B2B Companies. Work with us →
Missing a line item<br>Anthropic’s latest model has been ‘distilled,’ and now there are open-weight models out there that perform almost as well as Anthropic’s latest model at a fraction of the cost.
This is causing all sorts of consternation and political engagement. From a profit-and-loss perspective, we’ve all now seen that there is a major line item for frontier labs that we should have been accounting for but haven’t been.
Previously, we could break down OpenAI’s or Anthropic’s business by thinking about the inputs of oodles of R&D combined with gigawatts of compute to make state-of-the-art frontier models. After the models are made, they need to be served, and there’s more compute that goes with commercialization efforts. There is a cycle of training new models and supporting legacy ones that keeps going as long as a competitor keeps trying to surpass the state of the art.
What was missing in that simple understanding was countermeasures. The Louvre needs to pay for security, Coca-Cola needs steps to protect its recipe, and Disney needs to take steps to protect its intellectual property. What is different for the frontier AI labs is that their products are giving away tiny bits of the secret sauce with every output and response that’s provided to users.
For a lab to capture value from a state-of-the-art model, its model needs to be markedly better than alternatives. The difference over the competition comes just as much from protecting against distillation (or theft) as it does from novel research breakthroughs to make the model in the first place.
Anthropic, and likely OpenAI, have no choice but to develop ways to stop distillation from happening to future models. The current valuations for both companies assume that there is significant value capture from their frontier models. Frontier labs might still have months between release and when the model is distilled, but it’s naive to assume distillation won’t accelerate as distillers invest heavily to improve. Lengthening the time from releasing a model until it’s distilled will take considerable effort from the labs.1
How much will it cost to develop and deploy effective countermeasures? We don’t know yet. Disney spends much less to protect its IP than it did to develop and promote it. Coca-Cola might spend more to protect the recipe than it did to develop it, and the same would go for Chick-fil-A. The frontier AI labs could end up anywhere on the spectrum from incidental to uneconomical.
Obfuscation and deception might be a big part of how countermeasures are implemented. Cloak and dagger. Countermeasures are harder to do than political lobbying. We know this because we see Anthropic turning to lobbying.
We can infer that the costs of countermeasures aren’t incidental—at least until there’s a major breakthrough in approach. Fortunately, we’ve seen real breakthroughs in AI on the regular.
The objection that “it can’t be done—it’s impossible to prevent bad actors from distilling” is not a reason for inaction. The budgets might need to be bigger, and the measures more drastic.
The AI labs developing frontier models might not want to recognize that there’s a missing cost line that they haven’t accounted for. Time to get real.
Caught my eye:
Hasbro’s ‘Aging Up’ Strategy Continues With a Reattempt at Adult Play-Doh // Why Chili’s Isn’t Going ‘All In’ on AI // Meet the Companies Shelling Out for Top AI Models // Visa to Cut 7% of Workforce as CEO Seeks to Revamp Payments Firm // PayPal CEO Says Would Consider Opportunities for Greater Value // Granola launches an Apple Watch app | TechCrunch // Cursor makes its biggest India push yet ahead of SpaceX acquisition with localized pricing | TechCrunch // An Underground Nuclear Reactor Is Coming to This Kansas Town—and Dividing Locals // Apple Briefly Hits $5 Trillion Market Cap // JPMorgan’s Ohio State Deal Shows Evolution of College Sports
Or potentially a third party developing a solution could sell it to the labs ↩
Made possible by
Font sizeDark modeReading mode
© 2026 Arnold EngelRSS ReaderPrivacy Policy