Last month in this section I made a prediction: the economic value of a language model as such would head toward zero, and the durable value would move down the stack to compute and, under that, power. Back then the gap between the best open-weight models and the closed state of the art measured roughly three months and was shrinking.
Now I’m taking that prediction to the place people still think is safe, and then following it into the policy fight it explains.
Claim one: even the frontier has no durable value. Not the trailing tier. The leading edge itself. Being three months ahead is worth real money for exactly three months, and then it’s worth nothing, and the vendor has to buy the next three months with capital spending that grows every cycle. That’s not an asset. That’s a treadmill with a subscription attached.
Claim two: the push to restrict open-weight models is protectionism. When every technical moat melts on contact with the market, the only moat left is one the government builds for you. The safety framing on these proposals deserves to be taken seriously once, checked against the evidence — which the government’s own study already did — and then whatever’s left over, the part the evidence doesn’t support, should be called what it is.
I’m a paying customer of a frontier lab. This firm’s entire operation runs on a closed frontier model, by choice, today. That’s the point. I pay for the frontier because right now it’s the best tool, and I keep the frontier vendor’s pricing in check — every month, at renewal — because the open alternative exists. Take away the alternative and you haven’t made me safer. You’ve made me captive.
The frontier is a release schedule
Start with the measurement, which hasn’t changed direction. Epoch AI’s running comparison puts the best open-weight models about three months behind the closed state of the art on their composite capability index — the narrowest gap in the history of the comparison, with their caveats cutting both ways. Wherever the true number sits in that band, it’s counted in months.
So what’s a three-month lead worth?
For a handful of tasks at the true edge — the hardest research problems, the most complicated agentic work — it’s worth a premium, briefly. But most economically valuable work, and nearly all legal work, sits comfortably below the edge. Drafting, summarizing, extraction, first-pass review, correspondence: those topped out a while ago. A model three months behind the frontier does them just as well, and you can’t tell the difference. So the frontier premium is shrinking onto a thinner and thinner slice of tasks, while the cost of staying at the frontier — the training runs, the data centers, the power contracts — keeps compounding.
And the lead leaks every time somebody uses it. A frontier model’s capability transfers into smaller models trained on its outputs; every API call is a lesson. The industry spent early 2025 arguing over whether a Chinese lab’s near-frontier open release had been distilled from American closed models. Whoever’s right about that, the leading product can’t make money without being exposed, and it can’t be exposed without teaching. The frontier lab’s crown jewel loses value fastest exactly when it’s most popular.
Follow the money all the way down. The frontier tier is a release schedule — a promise to keep being first, bought with escalating capex, throwing off margins that burn down on a three-month fuse. The durable assets in this industry are the ones I named last month: compute, power, and distribution. A model, frontier or not, is inventory. Inventory with a sell-by date.
If you run a frontier lab, you know this better than anyone. Which gets us to the policy fight.
What you do when every moat melts
A business whose technical lead goes stale in months has four options. It can out-execute forever, which is exhausting and leaves no margin once everybody catches up. It can buy its way down into compute and power, which the biggest players are plainly doing. It can build distribution and switching costs, which is what every “platform” and “ecosystem” announcement is. Or it can reach for the one moat that doesn’t melt: law.
A regulatory moat is the best moat there is. It doesn’t depreciate. Competitors can’t distill it. And it always shows up dressed in the language of protecting the public, because nobody in history has lobbied for a barrier to entry by calling it a barrier to entry.
So look at what’s actually on the table in 2026, and ask the same thing of every piece: does this scale with risk, or does it scale with being a competitor?
The evidence baseline got set two years ago, and I’ll keep insisting on it, because it came out of exactly the process skeptics should want. The federal government’s own study — NTIA’s July 2024 report on dual-use foundation models with widely available weights, commissioned under the 2023 executive order — went looking for the case that open weights were dangerous enough to restrict and reported that the evidence wasn’t there. Monitor for marginal risks, it said, but don’t restrict the availability of open model weights. That’s what a real safety process looks like. It asks for evidence, and when the evidence comes up short, it doesn’t ban anything.
Since then, the models got better, the evidence of catastrophe kept not showing up on schedule, and the restriction proposals got bigger. California’s SB 1047 — vetoed in September 2024 — would have imposed liability and shutdown obligations that open developers couldn’t meet, since by definition you can’t recall a model once it’s released. The veto message itself said the bill regulated by developer size instead of demonstrated risk. This year’s sprawling House framework requires “large frontier developers,” defined by revenue, to hire licensed independent verification organizations to audit their risk frameworks. A two-hundred-person lab can staff that. A graduate student releasing weights can’t. In other words, it’s a fixed cost that works as a market-share floor for whoever’s already big. And Washington spent the summer openly weighing bans on foreign open-weight models — restrictions its own analysts admit can’t be enforced against files that have already been downloaded a hundred million times — while the enforcement machinery quietly grows into a tool for restricting any open model.
See the pattern. Thresholds keyed to revenue and compute, not to demonstrated capability for harm. Compliance regimes priced in millions per year. Liability for downstream uses that only a closed, API-gated model could even in theory control — which makes “closed” the only legally safe way to build, which makes the incumbents’ design the one the law requires. None of it lines up with the actual risk models safety researchers publish. All of it lines up with the competitive interests of companies whose product goes stale in ninety days.
And look at who wouldn’t play along. In July, twenty-five companies — Nvidia, Microsoft, Meta, IBM, Mozilla, the Linux Foundation, Hugging Face among them — published an open letter against “premature restrictions” on open models that would “stifle competition or drive innovation overseas,” two days after nearly two hundred startups told the White House the same thing. Days later, the frontier lab with the strongest safety reputation in the industry — the one whose model this firm runs on — published its own position. It said it has never advocated a ban on open-weight models, that open models without dangerous capabilities are “a public good,” and that the real levers are chip export controls, action against industrial-scale distillation, and capability testing applied to open and closed models alike. When even the most safety-minded frontier lab won’t back the ban, the ban has no safety constituency left. Subtract the measures the safety community actually supports from the restriction agenda, and what’s left is the protectionism.
That subtraction is the test, and here’s how it works: take any proposed restriction, remove every provision that applies equally to closed models, and look at what is left. Capability testing for everything powerful enough? Hits both. That’s safety policy, and I’m for it. Export controls on chips? Doesn’t care whether a model is open or closed. That’s national-security policy, and you can argue it on the merits. Liability only a gated API can satisfy, verification regimes priced for incumbents, thresholds drawn around revenue? Those land on openness as such — and openness as such isn’t a risk category. It’s a distribution model. Regulating a distribution model is how you protect distributors.
Lawyers have seen this costume before
Our profession ought to spot this move faster than anybody, because we invented a version of it.
Every unauthorized-practice-of-law rule gets defended as consumer protection, and some of it really is. But for a century the profession has also worn that costume to keep out competitors whose real offense was being cheaper: form preparers, publishers of self-help materials, software. The test there is the same subtraction. Take the UPL enforcement action, strip out the part that addresses demonstrated consumer harm, and look at what’s left. A lot of the time, what’s left is a guild protecting its rates from people who were serving clients the guild had priced out entirely.
I practice in Minnesota, where thousands of people walk into court every year with no lawyer at all. They didn’t choose that. The market the profession built just can’t serve them at a price they can pay. The technology in this essay is the first thing I’ve seen in my career that could really change that math — and it changes it fastest at the bottom, through the cheap and open tier: the models a legal-aid office can run on its own hardware, the assistants a solo can hook up to free public law without a subscription. That’s the tier the restriction agenda would freeze. Not the frontier. The incumbents will always be able to comply with their own moat. What gets frozen is the tier that was about to make competent help cheap.
So when the profession gets asked to weigh in on AI policy — and it will — carry the question into the room that we should have been asking about our own rules all along: does this restriction scale with demonstrated harm, or does it scale with being a threat to somebody’s margins?
What would prove me wrong
A prediction without a way to lose is worthless, so here’s how I’d know I blew each of these.
The frontier’s value: if, three years from now, frontier labs are holding high margins on model access itself — not on compute, not on enterprise integration and distribution, not on regulatory position, but on the model as a model — while open alternatives stay months behind, then the depreciation argument failed and I’ll say so. Watch pricing power at renewal time. Does the premium for “best” hold once “second-best, self-hosted, free” is good enough for ninety-five percent of tasks?
Protectionism: if the people pushing restrictions settle on measures that pass the subtraction test — capability-based, evidence-gated, the same for open and closed — then the safety framing was sincere and I’ll retire the costume line. The NTIA evidence standard and the targeted-measures position show what that looks like. If the proposals keep drifting toward revenue thresholds, incumbent-priced compliance, and liability only closed designs can carry, the costume is the policy.
I have a stake in this, and I’ve said so: my firm’s costs fall as the open tier rises, and my clients’ access to justice gets better with it. The incumbents have the opposite stake, and you should read their filings the way you’d read any interested party’s. Lawyers do that with every witness who profits from being believed.
The model will be free. The frontier will be a release schedule. The only question Congress actually controls is whether the free tier gets to everybody — the legal-aid office and the solo and the client who could never afford either of us — or gets held at the border while the people who couldn’t afford the closed tier wait for permission to be helped. Every article in this section is free for the same reason this one is. The thing standing between people and justice was never the technology. It was the price. Don’t let anybody put the price back and call it safety.
Sources
- Epoch AI, Open-weight models lag state-of-the-art by around 3 months on average — the gap measurement and its two-directional caveats, as previously cited in this section
- NTIA, Fact Sheet: NTIA AI Report Calls for Monitoring, But Not Mandating Restrictions of Open AI Models (July 30, 2024) and the underlying Report on Dual-Use Foundation Models with Widely Available Model Weights
- Governor of California, SB 1047 veto message (Sept. 2024) — regulation keyed to developer scale rather than demonstrated risk
- CNBC, Nvidia, Microsoft, Meta warn against ‘premature restrictions’ of open-weight models (July 24, 2026) — the 25-signatory “Open Weights and American AI Leadership” letter; see also Tom’s Hardware’s coverage noting the concurrent debate over banning Chinese open models
- Anthropic, Our position on open-weights models (July 27, 2026) — “never advocated for a ban”; open models without dangerous capabilities as “a public good”; chips, distillation, and testing as the proposed levers
- Cybersecurity Dive, Sprawling new House AI bill includes frontier model oversight, open-source security grants — the revenue-defined “large frontier developer” category and licensed independent verification organizations
- Companion essays in this section: The Model Will Be Free. The Electricity Won’t. and The Direction of the Integration
Speculative commentary on technology policy and the business of law — the opinions, predictions, and stated falsifiers are the author’s. Not legal advice, not investment advice, and not a comment on any pending legislation’s constitutionality. The author’s economic interest is disclosed in the text: this firm benefits from a competitive, partly open model market. No client information appears in this article. Questions about anything here: Send us a message or 612-470-6529.