I can't forecast in a year or even six months what is going to be the top model. So as we build our businesses, one of the most important things to do is to preserve optionality.
— Andrew Ng
Optionality, Not Loyalty
Andrew Ng opens the open-versus-closed argument from an odd place: he is rooting for both sides. He wants OpenAI and Anthropic to have spectacular public offerings, and he says he uses their tools. Then he gives the advice he repeats most often to business leaders: “In fact, one piece of advice I often give to business leaders is I can't forecast in a year or even six months what is going to be the top model. So as we build our businesses, one of the most important things to do is to preserve optionality.”
Optionality is the word doing the real work. It sounds like a risk manager's caution, but it quietly assumes the thing the angle wants settled: no closed lab can be trusted to stay on top, so the rational buyer refuses to be locked in.
Ng makes the pressure explicit when he defines the alternative: “Open models are models that people have published on the internet, free for anyone to use. And those are extremely performant, very intelligent, and often much cheaper.” If the free thing is performant and often much cheaper, then the closed seller's edge is a forecast, not a moat. And Ng's forecast is deliberately no forecast.
He does not say closed labs die. What he says next is about the system, not the seller: “So, to sustain competitive advantage in America, one of the most important things we have to do is support and sustain open models.” The survival question has already moved. The buyer's hedge is to keep every model available, and the seller's margin is not part of that calculation.
Ng's loyalty is to optionality, not to any lab.
The Only Unproven Business Model Is Open
The angle asks whether closed AI must change to survive. Ng's answer is quieter than the lobbyists he criticizes. Pressed on open-model capital expenditure, he reaches for Red Hat, then stalls: “Red Hat built a great business by open sourcing the Linux operating system and many businesses have been built with open source. The details of how to do this with open models, I think, are still being worked out.”
Then the full admission: “So it's true that the capex of training open models is higher than the capex of writing traditional software, but I feel like their business models have yet to be worked out, and to the extent that open models give you a fundamental cost advantage, that's something to pay attention to.”
Read that against the angle. The only business model Ng calls unproven is open. He spends no time proving closed labs have a durable moat; he simply assumes both can coexist and wants choice preserved.
The closed seller's advantage, in his telling, is a running expectation plus a price for access, not a durable mechanism.
So Ng moves the valuable layer elsewhere. His internet analogy is that the applications built on top of the infrastructure became more valuable than the infrastructure itself.
His business advice is that buying a ChatGPT or Microsoft Copilot license never creates competitive advantage; use cases and people change management do.
That is not a closed-weights business model. It is an adoption story wearing a margin question.
The open model is the one without a business model.
The Shovel Seller's Reassurance
Jensen Huang takes the angle's question directly. Asked whether OpenAI and Anthropic are in trouble, he does not reach for a moat. He reaches for a total: “No, not even a little bit. Uh both companies are thriving. They're going to continue to grow. they're going to go public. These two companies will be the most uh successful IPOs in in human history.”
Then he supplies the mechanism, and it is not built on closed weights: “Whenever there's more use, you'll have to sell a lot more Nvidia computers.”
That is the tell. Huang's confidence is real, but it is the confidence of the company selling shovels to both sides, not the confidence of a lab economist.
In his model, open models are not erosion; they are demand generation. Cheap or free intelligence pulls people into AI, and every additional user eventually needs more inference, more data centers, more chips.
Whether a given company upgrades to a closed service or runs an open weight, Nvidia collects.
His coexistence passage points the same way. He says he uses closed services because they are convenient: “I see a future where the world uses tons of closed models. And I encourage everybody, including my company, to use OpenAI and Claude and Cursor and Cognition and Perplexity. Use everything that you can because it's out of the cloud, because it's just easier, and you build only what you must.” That is the buyer's convenience again, not a proof that the closed labs command durable margins. It predicts plenty of customers, not the margin those customers carry.
The shovel seller's boom is not the lab's moat.
The Harness, Not the Weights
The angle smuggles in a conclusion with three words: neo clouds are hosting it. If the data centers are hosting the closed model, the reasoning goes, the closed lab has a distribution path and therefore a future.
But neither source says that. Ng's version of hosting runs the other direction: “when you take an open model whether it's released from a Chinese lab or American lab and you run them on American infrastructure, you know it basically becomes — really American infrastructure can control that model.” That's about who controls the model, not who collects the margin. Hosting an open weight on American infrastructure puts the host in control; it does not hand the host the lab's revenue.
Jensen Huang makes the build-versus-buy case in the adjacent interview. He says closed models are cheaper because you don't have to build, train, fine-tune, maintain, guardrail, evaluate, and host them yourself.
That is a TCO argument, not a closed-weights argument. Renting an open model from a third-party cloud gives you the same build-versus-buy relief.
The hosting layer does not decide whether the closed lab survives; it decides which operations vendor gets paid.
The actual economic site has already moved. Jensen says: “Claude Code, for example, is a harness around um the Opus uh AI models. And so the large language model is the brain. The harness if you will turns it into an agent and into a thinking working um agent that can help you do things.” That is the product. Not raw tokens, not weights, but the integrated harness that turns a model into a working agent.
Ng is building the open rival harness—OpenWorker—which produces finished work rather than chat.
So the closed labs have already changed their business from selling access to selling agentic tools. The question is whether their harness can keep commanding a premium, not whether their weights are hosted.
Hosting can be a business, but it is a business built on switching, not on any single model's secret. The margin now splits among the application, the harness, and the multi-model host — and the closed weights themselves are the least protected layer.
The Closed Lab Has Already Changed
Ng never describes the closed labs as selling weights. His phrase is smaller and more truthful: “OpenAI, Anthropic, Google have done a great job training closed proprietary models that they sell access to for, you know, decently high price.”
Access, not a parameter file. The actual products he names as valuable are tools: “I'm actually excited about tools like Claude Code and ChatGPT Code and Gemini's agentic tools, which are all good tools.” Those are the things customers pay for—the harness, the agent, the finished work—not the tensor file.
This is where the outside numbers stop arguing with him.
By my own reading of reporting outside this conversation, Anthropic's annualized revenue in 2026 has passed $30 billion, with a large and growing set of customers paying more than $1 million a year.
That is not evidence of a closed-weights moat. It is evidence that the closed lab has already mutated out of the token business and into the application-and-service business Ng says will capture the value.
The survival question the angle asked—must closed AI change to survive?—has already been answered by the closed labs themselves: they have changed.
They no longer look like a vendor of proprietary weights. They look like an application company with a frontier model inside it.
The moat is the finished work, not the model.
Closed Was a Gap, Not a Business Model
The angle has been treating closed AI as an economic category with a future to defend. Ng's report on China collapses that category.
He says Chinese open models have rapidly accelerated and are approaching, maybe not quite at, parity with America's leading models. Not ahead, not equal—approaching.
But that trend is the point. If the open model is nearly at the frontier and far cheaper, then what the closed lab sold was never the weights. It sold the distance between its frontier model and the free one. Distance is a schedule, not a moat.
This changes the answer. Closed AI does not need to change to survive; it has already changed into an application company, because the old thing it sold—exclusive access to better intelligence—is being squeezed from underneath it.
Jensen's optimism did not run on a widening intelligence premium; it ran on every new user needing more compute. That optimism is real, but it is a claim about the size of the pie, not about the margin on a closed weight.
Ng's closing line points to the unit he actually cares about: “I want our great American businesses to win, but one of the best ways for America broadly, not just the AI sector, to win is preserving open models so that all American companies can do well.”
The subject is American companies, plural, not closed AI.
The bright future belongs to a business that resembles OpenAI and Anthropic the way a restaurant resembles a farm.
The closed weights may still be inside, but the sustainable business is the finished work, the harness, the service.
Closed was a gap, not a business model.
If the value is in the harness, the use case, and the finished job, then the companies that survive are not the ones that kept their weights secret but the ones that made themselves indispensable after the model ran—and that race is still open.
The Grounding
This article is built on four primary interviews, quoted verbatim:
- Andrew Ng — “China, Open Source & AI Competitiveness” — The Washington Post, Building America series (July 2026). Full interview
- Jensen Huang — “AI Doomers Have It Wrong” — Axios, Behind the Curtain (July 2026). Full interview
- Jensen Huang — Bloomberg Interview: Korea AI Golden Age — Bloomberg Television (July 2026). Full interview
- Alex Atallah, OpenRouter CEO — 20VC interview (August 2026) — consulted as counter-material for the multi-model host argument. Full interview
Top comments (0)