Almost every business plan being written in this industry right now rests on one assumption: that the model is the valuable thing.
That assumption is why capability is announced like a moat, why access to the best model is sold as the product, and why a great many firms — including law firms — are quietly organizing their operations around a particular vendor’s particular model, as though which one they picked would still matter in five years.
I think it is wrong, and I want to state the prediction plainly enough that it can be checked against reality later, which is the only way predictions are 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 will not disappear; it will migrate downward, to whoever owns compute. And beneath compute, to whoever owns power.
Here is the mechanism, and then the specific things that would prove me wrong.
Mechanism one: the gap is measured in months
The most direct evidence is simply how far behind open-weight models actually are, and it is much less than the discourse assumes.
Epoch AI tracks this continuously using a composite capability index, and their current measurement is that 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 is the smallest gap in the history of their comparison, and the trend has been toward compression.
Their caveats deserve to be repeated rather than buried, because they cut both ways. Open models may hill-climb public benchmarks more aggressively than closed ones and underperform on private evaluations, which would mean the real gap is wider than measured. On the other side, several recent open models lack the evaluation coverage to be scored at all, which would mean the measured gap is too wide.
Take the whole range seriously and the conclusion is still the same. A lead measured in months is not a moat. It is a release schedule. No durable pricing power has ever been built on a head start that the follower closes before the customer’s annual contract renews.
Mechanism two: the moat leaks by being used
The second mechanism is structural and, I think, underappreciated outside the field.
A frontier model’s capability can be substantially transferred into a smaller model by training the smaller one on the larger one’s outputs. Which means the leading product cannot be sold without being exposed, and cannot be exposed without teaching. Every API call is a lesson. The asset is a secret that must be published continuously in order to earn anything.
There is no prior technology moat with quite this shape. A pharmaceutical company has patents, and a chip fab has a fab. Here the frontier lab has a lead that decays through the very act of monetizing it, at a rate set by how many people are using it — which is to say, fastest exactly when it is most valuable.
Mechanism three: nothing above it is sticky
Ask what happens if you change models in a system you have built.
For most well-built applications, the answer is that you change a line of configuration and re-run your evaluations. There is no data migration, no retraining of staff, no reimplementation. The interface is text in and text out, and it is the same interface for every vendor.
That is the definition of an interchangeable good. Interchangeable goods with falling marginal production cost have exactly one destination, and the economics do not care how impressive the manufacturing process is. Commoditization is not a failure of the model layer. It is what success looks like at a layer where the switching cost is one line.
Mechanism four: good enough arrives before best
This one is specific to our profession and it is the one I can speak to from having done it.
Most legal work does not 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 — these are handled competently by models that are already free, already runnable on a machine in your office, and already several generations old. I have written about what running models on your own hardware actually costs, and the honest answer even a year ago was that a large fraction of real practice work does not touch the frontier.
The frontier matters for a narrow band: hard reasoning, novel argument, long-context synthesis across a difficult record. That band is genuinely valuable and it is also genuinely small as a fraction of total tokens. When the bulk of demand is served by the commodity tier, the premium tier is not a market. It is a specialty.
Where the value actually goes
If the model is not the asset, what is?
Look at what cannot be forked. You can copy weights in an afternoon. You cannot copy a gigawatt.
Hyperscaler capital expenditure is projected to exceed $600 billion in 2026, roughly a 36% increase over the prior year, and the striking thing is not the number — it is what the money is now blocked on. As of mid-2025, more than three dozen announced projects representing something on the order of $162 billion in investment were blocked or significantly delayed. Not by chip supply. By power: interconnection queues, transmission, substations, permitting, generation, and local opposition. Industry analysts describe requirements measured in tens of gigawatts of aggregate new capacity over two to three years, which is a scale of buildout the American energy sector has not demonstrated in decades.
That is what a real constraint looks like. Chips are constrained but purchasable with enough lead time. Electricity in a specific place, delivered through specific wires, on a specific date, is not something a competitor can route around with better software or more funding. It is land, capital, regulatory process, and time — the four things that have always produced durable economic rents, and the four things a clever model cannot substitute for.
So the sharpened version of the thesis is this: compute is where the value lands, and power is why it stays there. Owning compute is a claim on a physically scarce resource with a construction timeline measured in years and a permitting timeline measured in longer than that. Owning a model is a claim on a three-month head start.
None of this is a new pattern. Operating systems were going to be the industry’s commanding height until an open one became good enough, and the value moved to silicon and then to the data centers. The railroads made their owners rich not because trains were hard to build but because right-of-way was impossible to duplicate. Value migrates to the layer that is scarce, capital-intensive, and non-forkable, and it does this reliably enough that you can plan around it.
What would prove me wrong
A prediction without failure conditions is just a mood. Here are the specific things that would falsify this one, in the order I think they are likely.
Compute stops being scarce. This is the most plausible refutation and it cuts directly at the core of the thesis. If the current buildout overshoots demand — and enormous capital expenditure programs routinely do — then compute commoditizes too, capacity gluts, prices collapse, and the value migrates again, most likely to whoever owns distribution and customer demand. Watch for utilization rates falling while capacity comes online. The 2001 fiber-optic buildout is the cautionary case: the people who laid the glass mostly went bankrupt, and the value went to whoever bought it at ten cents on the dollar.
An architectural discontinuity that does not distill. If a capability emerges that genuinely cannot be transferred into a smaller model by training on outputs — something that requires the full apparatus to reproduce — then the frontier lead stops decaying and the moat becomes real. Mechanism two is the load-bearing one here, and it is an empirical claim about a technique, not a law of nature.
Regulation restricting open weights. A serious legal regime around releasing frontier weights would re-monopolize the model layer overnight, by fiat rather than by capability. This is a policy question rather than a technical one and it is genuinely live in several jurisdictions.
Inference-time compute becomes the whole game. If capability increasingly comes from spending more compute at the moment of use rather than from the weights themselves, then “the model” is not really the artifact anymore — the spending level is. Note that this does not refute the thesis so much as intensify it: it would mean model quality is literally purchased compute, which puts the compute owner in a stronger position, not a weaker one. But it would falsify the sub-claim that open weights close the gap, since a free model with a small inference budget would not match a free model with an enormous one.
Verticalized data moats. If, in specific domains, proprietary data and verification produce advantages that no general model can match, then value pools in the vertical rather than in the substrate. I think this is partly true and mostly overstated, and law is a good test case: the underlying corpus is largely public, which is not true of every industry.
If four years from now open-weight models are more than a year behind, or if compute is cheap and idle, I was wrong and the mechanism I got wrong is probably one of the above.
What a firm should do about it
This section exists for lawyers rather than investors, so here is the operational translation, which is the part that actually matters and is oddly insensitive to whether the prediction is right.
Never couple your systems to one vendor’s model. Treat the model as a swappable component behind an interface you control. If switching would cost you a rewrite, you have accidentally bought a dependency on the one layer most likely to commoditize. Everything I have built for this firm can change models with a configuration edit, and that is not sophistication — it is the 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 are what makes a new model useful to you on the day it ships, and they are the actual accumulation. The model is a component; the description of your practice is the machine.
Do not build a plan that depends on inference being expensive, or on it being free. Cost per token has fallen relentlessly and I expect it to keep falling. If your economics only work because AI is costly, you are betting against the trend. If they only work because it is free, you are betting on the buildout finishing on schedule, which the interconnection queues suggest it will not.
Assume local gets better every year. The strongest confidentiality posture is data that never leaves the building, and the reason to care about the commodity tier is not price. It is that a good-enough model you run yourself resolves the confidentiality analysis before it starts.
And then the part that ought to be obvious to us and somehow isn’t.
Every industry has a layer that cannot be forked. In this one it is a substation. In ours, it is not the drafting, and it is not the research, and it is increasingly not the analysis. It is the license, the judgment about consequence, the relationship with a person who is frightened, and the fact that a human being is accountable for the outcome. Those are our non-duplicable layer, and they are exactly where the value in this profession is going to concentrate as everything above them gets cheap.
Which is, in the end, the optimistic reading. If intelligence becomes a commodity — abundant, nearly free, metered like electricity — then the cost of competently handling a legal problem falls toward the cost of the judgment and the accountability, and away from the cost of the hours. That is the only mechanism 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 saving through or keeps it as margin is not a prediction. It is a choice, made one firm at a time, and it is the reason everything in this section is published 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.