Open vs Closed AI: Why Nvidia, Meta and China Are Picking Sides
Jensen Huang's pro-open-source letter split the AI industry. Here's why Nvidia backs open models, Anthropic didn't sign, and who actually profits.
What is the open source vs closed source AI debate about?
The debate centers on whether the most powerful AI models should be published openly for anyone to inspect, modify, and run, or kept locked inside a handful of companies that sell access to them. It came to a head when Nvidia CEO Jensen Huang published a letter backing open-source AI, co-signed by a wave of major tech leaders. Anthropic did not sign. The split isn’t just philosophical. It reflects different bets on where the money in AI actually gets made.
TL;DR
- Jensen Huang’s letter pushed a pro-open-source position and picked up co-signatures from most major tech CEOs, with Anthropic notably staying silent and continuing to warn publicly about the risks of open AI.
- Open source and closed source AI are built the same way technically. The only real difference is the business model: one gives the trained model away, the other sells access to it.
- Closed source models currently lead on quality, largely because OpenAI and Anthropic got there first, scaled fastest, and reinvested massive revenue into better training runs.
- Open source AI has closed the gap fast, with Moonshot AI’s Kimi K3 landing near the top of independent leaderboards alongside Claude and ChatGPT.
- Nvidia benefits either way, but especially when open source thrives, because more competing model builders means more GPU demand across the entire stack, which is why Nvidia committed $20 billion toward open ecosystem efforts.
- China leads in open-weight model releases partly out of necessity: it lacks the top-tier chip access the US has, so publishing open models is a way to build influence and adoption without owning the compute advantage.
- Jevons paradox suggests cheaper AI won’t shrink the market, it should expand it, since falling costs tend to drive so much more usage that total spending rises anyway.
Why did Anthropic not sign Nvidia’s open-source letter?
Anthropic has built its public identity around AI safety, and the company has spoken openly about the risks of releasing powerful models without control over how they’re used. Unlike OpenAI, Meta, and other major labs that added their names to Jensen Huang’s letter, Anthropic stayed out of it entirely, and around the same period pushed messaging in the opposite direction: that open models can be dangerous in the wrong hands.
This isn’t just a safety stance, it’s also a business stance. Anthropic sells Claude as a closed product. Its revenue model depends on being the exclusive gatekeeper to its own intelligence. A pro-open-source position would cut against the company’s own commercial structure, which makes its absence from the letter unsurprising even setting aside genuine safety concerns.
Are open source and closed source AI models built differently?
No. There’s no technical difference in how an open or closed model is trained, fine-tuned, or served. Both rely on the same kinds of data, compute, and architectures. The difference shows up only after training: a closed model stays behind an API that a company charges for, while an open model gets published so anyone can download it, inspect it, modify it, and run it on their own infrastructure.
That distinction matters because it means “open source AI” isn’t a separate technology track chasing catch-up. It’s the same technology, released under a different set of rules.
Why has closed source AI stayed ahead so far?
Two compounding advantages explain the current gap. First, the foundational breakthroughs, and the initial scaling of those breakthroughs, came from closed source companies. ChatGPT arrived first and was closed from day one. Anthropic followed shortly after with the same model. That head start let them define the pace of the field.
Second, being first let them make money first, and that revenue got reinvested into bigger data centers, larger research teams, and more compute for the next generation of models. Each new closed model built on the gains of the last, and revenue kept the flywheel spinning.
Open source AI, by contrast, arrived almost by accident. Meta’s LLaMA model leaked before Meta had fully committed to an open strategy, and the company leaned into open releases only after the fact. There was no proven open-source AI business model to justify equivalent investment, so the two closed labs, OpenAI and Anthropic, ended up absorbing most of the industry’s revenue, compute, and top researchers.
Why is it hard to make money on open source AI?
Training a frontier model requires enormous upfront investment. In most software industries (Android, Apache, the early open web), a company can release something for free and still build a sustainable business around services layered on top. AI adds a second cost on top of that: even after spending huge sums to train a model, a company still needs either heavy infrastructure to serve it efficiently at scale, or a strong application layer built on top of it, to turn that model into revenue.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
That means an open-source AI lab effectively needs to fund two expensive efforts (the model itself, and the business built around it) while a competitor can skip the first cost entirely. Anyone can take a published open model and build only the data center layer, or only the application layer, without ever paying to train the underlying intelligence. That asymmetry is a big part of why open-source AI labs have struggled to out-earn closed ones so far.
How good is open source AI right now?
Closer than most people assume. On independent evaluations like the Artificial Analysis leaderboard, closed models from OpenAI and Anthropic still occupy the top spots, but Moonshot AI’s Kimi K3, a Chinese open-weight model, ranks right behind them, ahead of many other closed offerings. There’s no law of physics keeping open models behind. The gap today reflects funding and timing, not some inherent ceiling on open-source quality.
Why does China favor open source AI?
China currently produces some of the strongest open-weight models available, and a major reason is compute scarcity. Chinese firms don’t yet have access to the same tier of chips that US companies get from Nvidia, AMD, Cerebras, and Groq. What China does have is a large base of AI researchers, significant electricity generation capacity, and strong engineering talent. Releasing open models is a way to compete for global adoption and influence even without matching the US on raw chip supply, turning a constraint into a strategy.
Why does Nvidia benefit from open source AI winning?
Nvidia sits underneath the entire AI stack, selling the chips that both open and closed model providers train and run on. More competitors in the model layer, open or closed, means more total demand for GPUs. But open source specifically multiplies that effect: it invites far more companies into the game (startups, researchers, regional players) all of whom still need chips.
Nvidia backed this logic with real capital, committing $20 billion toward building out open-source AI efforts. When more people build on open models, efficiency improvements spread faster because more engineers are studying and optimizing the same publicly available code. Lower costs then tend to drive more usage rather than less, a pattern known as Jevons paradox: as a resource gets cheaper and more efficient, total consumption of it tends to rise, not fall. Applied to AI, cheaper intelligence doesn’t shrink the market, it expands who uses it and how often, which ultimately means more tokens processed and more chips sold.
Frequently Asked Questions
What’s the actual difference between open source and closed source AI?
Open source AI publishes the trained model publicly so anyone can download, inspect, modify, or run it. Closed source AI keeps the model behind a paid API or product. The underlying technology and training process are essentially the same.
Is open source AI as good as closed source AI?
It’s close. Closed models from OpenAI and Anthropic currently lead independent benchmarks, but open-weight models like Moonshot AI’s Kimi K3 rank just behind them, ahead of many other closed alternatives.
Why did Meta end up building open source AI models?
Meta’s LLaMA model leaked publicly before the company had settled on an open strategy. Once people started downloading and using it, Meta leaned into releasing models openly going forward.
Does Nvidia make more money from open or closed source AI?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Nvidia profits from both, since it sells the chips underneath the entire stack. But open source tends to expand the number of companies training and running models, which increases total chip demand, one reason Nvidia has publicly backed open-source efforts with major investment.
Why does China release so many open-weight AI models?
Chinese AI labs currently lack access to the top-tier chips available to US companies, so open releases let them compete for global influence and adoption through research talent and engineering strength rather than raw compute advantage.
