The most valuable argument in artificial intelligence this month was made not about a model but about a price. Writing in the Washington Post, former venture capitalist Bill Gurley looked at the wave of powerful AI models being given away for free — the newest released in China days earlier — and reached a conclusion that cuts against most of the lobbying now reaching Whitehall and Washington. Open-weight AI is not a security threat to be licensed and contained. It is competition doing exactly what competition is supposed to do: driving price toward the cost of producing one more copy, which for software is almost nothing. For any UK organisation buying AI, and for any policymaker deciding how to spend the sovereignty budget, that reframing changes the maths.
The real story is a margin, not a menace
Anthropic and OpenAI are reportedly preparing to go public at valuations near a trillion dollars, and those valuations rest on a promise: extraordinary, durable profit margins on frontier models. Gurley’s point is that the fastest-growing challenge to those margins is not a rival lab with deeper pockets. It is free. A model you can download for nothing sets a competitive floor that no premium API can ignore for long.
This is not idealism or a foreign gambit. Open-sourcing is a business strategy with a three-decade record of building large, profitable companies, and the economics behind it are older still. Adam Smith and David Ricardo observed that in a competitive market, price is driven toward the cost of the next unit. For steel or wheat that floor stays high. For software, once the first copy exists, the millionth costs essentially nothing — so a price falling toward zero is not a market failing but a market working precisely as described. The genuinely odd thing, Gurley notes, is not that open models are free. It is that proprietary software held enormous margins for forty years behind an artificial scarcity of copyright and lock-in. Open source simply removes the fence.
Strategic Reality: The trillion-dollar valuations and the free models are the same story told from two ends. A market advertising the richest margins in software history is, by that very advertisement, summoning the largest crowd of competitors in software history.
| The economics at a glance | What it means for buyers |
|---|---|
| Marginal cost of a software copy ≈ £0 | Sustained premium pricing needs a fence, not just a better model |
| 30-year open-source track record | Red Hat sold to IBM for $34bn; MongoDB, Databricks, HashiCorp all built on downloadable code |
| ”Your margin is my opportunity” (Bezos) | The higher the advertised return, the larger the competitive response it calls forth |
| Increasing returns / network effects | Left alone, software markets tip to one or two firms — the risk open models offset |
What is really happening beneath the “China threat” framing
The proximate trigger for the security-threat lobbying is that the most capable open models increasingly come out of China — not only from state giants but from independently financed labs. Gurley cites DeepSeek, valued above $50bn in its first outside round; Moonshot, which has raised $2bn; and Zhipu, publicly traded in Hong Kong at a market capitalisation north of $100bn despite giving its models away. Whatever else one thinks of these labs, “openness has no business model” is not a claim their balance sheets support. At the 2026 World Artificial Intelligence Conference in Shanghai, President Xi Jinping urged countries to “encourage open source, openness, collaboration and sharing” — doubling down rather than retreating.
But the open frontier is no longer only a Chinese story. Reflection, founded by Google DeepMind veterans, has raised billions on an explicit promise to build an open American frontier. Thinking Machines, the lab founded by OpenAI’s former chief technology officer, released a capable open model of its own. Even Nvidia has joined in, releasing its Nemotron models freely on the bet that a cheap, abundant model layer sells more of the chips underneath. That last motive matters for the UK: it shows the open wave is driven by hard commercial self-interest across the stack, not charity — which is exactly why it will not simply stop.
Critical Context: The “wall off the foreign models” argument quietly assumes open weights are a Chinese tactic. They are not. They are a structural feature of software economics that American chipmakers, start-ups and enterprises are pushing just as hard.
Why nearly everyone in the AI economy wants an open foundation
Gurley’s most useful move for a UK audience is to count who benefits. Ownership of a model — the ability to download and run it behind your own walls — throws off advantage after advantage. It prevents lock-in, freeing you from any single vendor’s pricing, roadmap or survival. It keeps your data and methods on your own infrastructure rather than feeding someone else’s model. It lets you fine-tune and extend without asking permission. And, counterintuitively, it is often more secure: a model a million adversarial eyes can inspect gets stress-tested and hardened in the open, the same reason the most trusted code on earth is open.
Add up who that serves. Start-ups want open models to build without a gatekeeper. Enterprises want them so their intellectual property is not absorbed into a competitor’s system. Researchers, universities and security teams want them to see inside what they study and defend. Chipmakers want them to sell more silicon. Cloud providers without a model of their own want them. Foreign governments and militaries want them, to run systems no other nation can switch off. Nearly the entire field is on one side; a handful of very large incumbents whose fortunes depend on keeping the layer closed are on the other.
That map lands directly on Britain. In Britain in the 5 percent we argued that hosting American compute is not the same as controlling it, and in Trump’s Anthropic ban exposes Britain’s AI dependence we traced how quickly a closed dependency becomes a strategic vulnerability when the vendor’s home government can revoke access. Open weights are the most direct answer to both. A model no other nation can switch off is sovereignty you can actually hold, at a fraction of the cost of building a frontier lab from scratch.
SME Advantage: A downloadable model turns “which giant do we bet the business on?” into “which model do we run and tune ourselves?” For a mid-sized UK firm, that is the difference between renting capability and owning it.
The stake for UK buyers and the sovereignty budget
The uncomfortable question Gurley puts to the incumbents is the one UK policymakers should borrow. The large labs describe themselves as growing faster than any company in history, with the richest margins in software and models accelerating away from all rivals. Grant every word of it, he says, then ask: if all of that is true, why do you need the government’s help to restrict your competitors? A firm genuinely running away with the future does not need Washington — or Westminster — to hobble the free alternative.
| Stakeholder | What open-weight competition changes |
|---|---|
| UK enterprises | Downloadable models cap API pricing and keep sensitive data in-house; lock-in becomes a choice, not a default |
| UK SMEs | Ownership without a frontier-lab budget; tuning a competent open model beats renting a premium one for many tasks |
| Sovereign-AI policy | A far cheaper route to strategic autonomy than domestic frontier compute alone — the model no one can revoke |
| AISI and security teams | Open weights are inspectable; the safety case is evaluation and hardening in the open, not prohibition |
| Incumbent vendors | The free floor disciplines margins the IPO story depends on — hence the lobbying |
There is a genuine catch, and Gurley is honest about it. Software markets do not reward the many; they tip toward the few. Economist W. Brian Arthur called it the “tendency for that which is ahead to get further ahead,” and AI may intensify that loop like nothing before it, as a lab learning from hundreds of millions of users compounds a lead rivals cannot catch. Microsoft, Google and Meta came to own the internet’s core layers exactly this way. The one force that keeps the many in the game is a freely available model no one can corner or shut down. That is why the policy choice is not cosmetic.
What UK organisations should actually do
For buyers, the practical implication is to stop treating “open versus closed” as an ideological question and start treating it as a portfolio decision. The evidence in our own coverage — The AI feud your vendor strategy sits on — is that concentration on a single closed vendor is now a board-level risk, not a procurement convenience.
- Audit your lock-in. Map which workloads sit on a single closed API and what switching would cost in re-engineering. That number is your exposure.
- Pilot an open-weight model on a real workload. Not a benchmark — a task you actually run. The question is whether owning and tuning a competent open model beats renting a premium one for that job.
- Keep sensitive data on models you control. Where the data is the moat, an inspectable model behind your own walls is a security posture, not a compromise.
- Read the security-threat lobbying as competitive positioning. When a vendor argues its free rivals should be restricted, weigh that against its own claims of runaway superiority. Both cannot be fully true.
Take Action: Run one open-weight pilot this quarter on a workload where lock-in or data residency already worries you. The point is not to abandon closed models — it is to price your alternative, so your next vendor negotiation is conducted from leverage rather than dependence.
For policymakers, the reframing is sharper still. Antitrust law from the Sherman Act onward exists to prevent the very outcome that increasing returns produce — a market owned by one or two firms. Gurley’s warning is that the current lobbying inverts that purpose, asking government to ban the one force keeping AI markets from tipping to monopoly. The UK is already wrestling with where evaluation ends and prohibition begins; our analysis of what the UK already knows about AI self-regulation and of AISI’s evaluation framework both point the same way: the credible safety response to open weights is rigorous, transparent evaluation — the work AISI is built for — not a licensing regime that quietly hands the market to two American incumbents.
Four challenges the optimistic case understates
The competition-is-good framing is largely right, but it is not free of friction, and a UK organisation acting on it should plan for four things the op-ed glides past.
-
Open does not mean effortless. Running and tuning a model yourself trades vendor cost for engineering cost and operational risk. The saving is real but it is not automatic, and thin AI teams can spend the margin they saved on integration. Mitigation: cost the total, including the people, before declaring the open route cheaper.
-
“Free” models still carry provenance and licence risk. Downloadable weights come with terms, training-data uncertainty and, for some Chinese-origin models, legitimate procurement and due-diligence questions no clever economics dissolves. Mitigation: treat model provenance and licence terms with the same rigour as any supply-chain component.
-
Security-through-openness is a tendency, not a guarantee. “A million eyes” hardens code over time; it does not make every fresh open model safe on release day, and open weights can also be fine-tuned toward misuse. Mitigation: pair open adoption with the evaluation discipline AISI models, not blind faith in the crowd.
-
The market may tip before the remedy takes hold. Arthur’s increasing-returns warning is the real one. If closed incumbents compound their lead fast enough, the free floor may arrive too late to matter. Mitigation: the window for keeping the open option viable is now, which is exactly why the policy fight is urgent rather than academic.
The strategic takeaway
Gurley’s essay is best read as a correction to a category error. Britain has spent a year debating sovereign compute, vendor dependence and how to regulate frontier labs, often as if the only path to strategic autonomy ran through building or hosting the biggest closed models. Open-weight competition offers a cheaper, more resilient route hiding in plain sight — and the lobbying to restrict it should be read for what it is: incumbents asking government to remove the one discipline the market has produced for free.
Three things to hold onto. First, treat open versus closed as a leverage decision, not a belief: the value of the open alternative is that it prices your dependence on the closed one. Second, the security case against open weights is answered by evaluation, not prohibition — the work the UK’s own institutions already do best. Third, the clock matters, because increasing returns reward whoever gets ahead, and the free floor only helps if it is still legal and viable by the time it is needed.
- Identify one workload where lock-in or data residency is already a concern
- Pilot a competent open-weight model against your current closed provider on that workload
- Cost the total, including engineering and provenance due diligence, not just the licence
- Fold the result into your next vendor negotiation as priced leverage
The invisible hand, Gurley writes, is answering Arthur’s warning with Smith’s insight — competition disciplining even the mightiest firm. For UK organisations, the practical version is simpler. The free model is not the threat. It is the leverage you have not yet used.
Source: Bill Gurley, “Open-model AI is good competition for Anthropic and OpenAI”, The Washington Post, 20 July 2026.
Analysis by Resultsense — making sense of AI in the UK. For strategic guidance on AI vendor strategy and open-weight adoption, get in touch.