Almost every business plan in this industry right now rests on one assumption: the model is the valuable thing.
That’s why capability gets announced like a moat. It’s why access to the best model gets sold as the product. And it’s why a lot of firms — law firms included — are quietly building their operations around one vendor’s model, as if which one they picked will still matter in five years.
I think that’s wrong. And I’ll put the prediction plainly enough that you can check it against reality later, because that’s the only way a prediction is worth anything.
The prediction: within a few years, the economic value of a language model as such will approach zero. Models will be commodities — interchangeable, abundant, nearly free at the point of use. The value won’t disappear. It’ll move down, to whoever owns compute. And under compute, to whoever owns power.
The open models are about three months behind
Start with how far behind the open-weight models actually are. It’s a lot less than people assume.
Epoch AI tracks this continuously with a composite capability index. Their current number: the best open-weight models trail the closed state of the art by roughly three months — about seven index points — with a 90% confidence interval running from about one month to about five. That’s the smallest gap since they started comparing, and it’s been shrinking.
Their caveats cut both ways, so you get both. Open models may hill-climb public benchmarks harder than closed ones and then underperform on private evaluations, which would mean the real gap is wider than measured. On the other hand, several recent open models don’t have enough evaluation coverage to be scored at all, which would mean the measured gap is too wide.
Take the whole range and you land in the same place. A lead measured in months isn’t a moat. It’s a release schedule. Nobody has ever built lasting pricing power on a head start the other guy closes before the customer’s annual contract renews.
The moat leaks every time somebody uses it
This one’s structural, and I don’t think people outside the field appreciate it enough.
You can transfer a lot of a frontier model’s capability into a smaller model by training the small one on the big one’s outputs. So the leading product can’t be sold without being exposed, and it can’t be exposed without teaching. Every API call is a lesson. The asset is a secret you have to keep publishing to earn anything off it.
I can’t think of another technology moat shaped like this. A drug company has patents. A chip fab has a fab. A frontier lab has a lead that wears away through the very act of selling it, and it wears away fastest when the most people are using it — which is to say, exactly when it’s most valuable.
Nothing on top of it is sticky
Ask what happens when you swap models in a system you’ve built.
For most well-built applications, you change a line of configuration and re-run your evaluations. No data migration, no retraining staff, no rewrite. It’s text in and text out, and that’s the same for every vendor.
That’s the definition of an interchangeable good. Interchangeable goods with falling marginal production costs end up in exactly one place, and the economics don’t care how impressive the factory is. When the switching cost is one line, commoditization is what winning looks like.
Good enough shows up before best
This one’s specific to our profession, and it’s the one I can speak to because I’ve done it.
Most legal work doesn’t need frontier capability. Classifying inbound mail, extracting dates from a scanned order, drafting a first-pass status letter, summarizing a deposition, tagging documents by issue — models that are already free, already runnable on a machine in your office, and already several generations old handle all of it competently. I’ve written about what running models on your own hardware actually costs, and even a year ago the answer was that a big chunk of real practice work never touches the frontier.
The frontier matters for a narrow band: hard reasoning, novel argument, pulling together a long and difficult record. That band is valuable. It’s also a small share of total tokens. When the commodity tier serves most of the demand, the premium tier isn’t a market. It’s a specialty.
Follow the money to the substation
If the model isn’t the asset, what is?
Look for what you can’t fork. You can copy weights in an afternoon. You can’t copy a gigawatt.
Hyperscaler capital spending is projected to top $600 billion in 2026, roughly a 36% jump over the prior year. The number isn’t the interesting part. What the money’s stuck on is. As of mid-2025, more than three dozen announced projects representing something like $162 billion in investment were blocked or seriously delayed. Not by chips. By power: interconnection queues, transmission, substations, permitting, generation, and local opposition. Industry analysts talk about needing tens of gigawatts of new capacity over two to three years. The American energy sector hasn’t built at that scale in decades.
That’s what a real constraint looks like. Chips are tight, but you can buy them with enough lead time. Electricity in a specific place, through specific wires, on a specific date, isn’t something a competitor can route around with better software or more money. It’s land, capital, regulatory process, and time — the four things that have always produced lasting economic rents, and the four things a clever model can’t stand in for.
So here’s the short version: compute is where the value lands, and power is why it stays there. Owning compute is a claim on something physically scarce, with a construction timeline measured in years and a permitting timeline longer than that. Owning a model is a claim on a three-month head start.
We’ve seen this movie. Operating systems were going to be the commanding height of the industry until an open one got good enough, and the value moved to silicon and then to the data centers. The railroads made their owners rich, but not because trains were hard to build. It was because right-of-way was impossible to duplicate. Value moves to the layer that’s scarce, capital-hungry, and can’t be forked, and it does it reliably enough to plan around.
How I’d know I blew this one
A prediction without failure conditions is just a mood. Here’s what would sink this one, roughly in order of how likely I think each is.
Compute stops being scarce. This is the likeliest way I’m wrong, and it goes straight at the heart of the argument. If the buildout overshoots demand — and giant capital programs overshoot all the time — then compute commoditizes too. Capacity gluts, prices collapse, and the value moves again, most likely to whoever owns distribution and customer demand. Watch for utilization falling while new capacity comes online. The 2001 fiber-optic buildout is the cautionary tale: the people who laid the glass mostly went bankrupt, and the value went to whoever bought it at ten cents on the dollar.
A breakthrough that won’t distill. If a capability shows up that genuinely can’t be transferred into a smaller model by training on outputs — something you need the whole apparatus to reproduce — then the frontier lead stops wearing away and the moat gets real. The leaky-moat argument is the one holding up the most weight, and it’s an empirical claim about a technique. It isn’t a law of nature.
Regulators restrict open weights. A serious legal regime around releasing frontier weights would re-monopolize the model layer overnight, by decree instead of by capability. That’s a policy fight, not an engineering one, and it’s live in several jurisdictions.
Inference-time compute becomes the whole game. If capability increasingly comes from spending more compute at the moment of use instead of from the weights, then “the model” isn’t really the thing anymore — the spending level is. If anything, that helps the compute owner: model quality would literally be purchased compute. But it would kill my side claim that open weights close the gap, because a free model on a small inference budget won’t match a free model on an enormous one.
Industry data turns into a moat. If, in particular fields, proprietary data and verification produce advantages no general model can match, then the value pools in that field instead of in the substrate. I think that’s partly true and mostly overstated, and law is a good test: our underlying corpus is largely public, which isn’t true of every industry.
If four years from now open-weight models are more than a year behind, or compute is cheap and sitting idle, I was wrong, and whatever I got wrong is probably on this list.
What I’d do with your firm
This part is for lawyers, not investors. And the funny thing is, the advice holds up whether I’m right or not.
Never couple your systems to one vendor’s model. Treat the model as a swappable part behind an interface you control. If switching would cost you a rewrite, you’ve accidentally bought a dependency on the one layer most likely to commoditize. Everything I’ve built for this firm can change models with a configuration edit. That’s not sophistication. It’s what you get by default if you build in the right order.
The durable assets are yours, not the vendor’s. Your data model, your prompts, your document templates, your retrieval corpus, and above all your evaluation set survive every model generation. They’re what makes a new model useful to you the day it ships, and they’re what actually piles up over the years. The model is a part. The description of your practice is the machine.
Don’t build a plan that needs inference to be expensive, or free. Cost per token has fallen relentlessly and I expect it to keep falling. If your economics only work because AI is costly, you’re betting against the trend. If they only work because it’s free, you’re betting the buildout finishes on schedule, and the interconnection queues suggest it won’t.
Assume local gets better every year. For confidentiality, the strongest position is client data that never leaves the building. Price is the least of it. A good-enough model you run yourself settles the confidentiality question before it ever comes up.
And then there’s the part that ought to be obvious to us and somehow isn’t.
Every industry has a layer you can’t fork. In AI it’s a substation. In ours, it isn’t the drafting, it isn’t the research, and more and more it isn’t the analysis. It’s the license, the judgment about consequences, the relationship with a person who’s scared, and the fact that a human being answers for the outcome. That’s our substation. It’s exactly where the value in this profession is going to pile up as everything above it gets cheap.
Which is the optimistic way to read all this. If intelligence becomes a commodity — abundant, nearly free, metered like electricity — then the cost of handling a legal problem competently falls toward the cost of the judgment and the accountability, and away from the cost of the hours. That’s the only way I know of that could close the gap between the number of people with legal problems and the number who can afford to do anything about them.
Whether the profession passes that savings along or pockets it as margin isn’t a prediction. It’s a choice, made one firm at a time. It’s also why everything in this section is free.
Sources
- Epoch AI, Open-weight models lag state-of-the-art by around 3 months on average — ~3-month lag, ~7 ECI points; 90% CI 1.1–5.3 months / 0–14 points; includes the authors’ caveats in both directions
- Epoch AI, Open models lag state-of-the-art closed models by 4 months — earlier measurement in the same series
- Data Center Knowledge, AI-First Hyperscalers: 2026’s Sprint Meets the Power Bottleneck — 2026 capex projected above $600B (+36% YoY); 36+ projects representing ~$162B blocked or delayed as of June 2025; analyst commentary from Dell’Oro and GlobalData on gigawatt-scale requirements
Speculative commentary on technology and the business of law. Not legal advice, not ethics advice, and emphatically not investment advice. The predictions above are the author’s opinion, offered with failure conditions so they can be tested rather than relied upon. No client information appears in this article.