Opinion: Opinion | The Great AI Rift: Why Meta, Palantir Are Fighting Against OpenAI, Anthropic, Google

Silicon Valley's sharpest quarrel this year is about who gets to hold the weights, the billions of trained numbers that make a model work. One camp rents intelligence through cloud APIs and keeps its frontier models sealed: OpenAI, Anthropic, and Google. The other wants weights published, so anyone can download a model, run it on their own servers, and change it at will.

The line hardened on July 24, when 25 companies, among them Nvidia, Microsoft, Meta, IBM, and Palantir, signed an open letter titled "Open Weights and American AI Leadership", urging Washington not to restrict open-weight models. Google, OpenAI, and Anthropic did not sign. Sam Altman later posted that he wanted America to win in both open and proprietary AI. The hedge was telling: OpenAI released open-weight models in 2025, alarmed by China's progress, but keeps its best systems behind an API.

Advertisement - Scroll to continue

Jensen Huang's case blends principle with arithmetic. Every downloaded model still needs a GPU to run, and most GPUs are Nvidia's. When Nvidia released the Nemotron 3.5 Lightning model that runs on a single GPU in August this year, he said that "Free AI should be great for chips. In March, Nvidia had assembled the Nemotron Coalition, eight labs including Mistral and Perplexity, to build frontier-grade open models. At CES in January, Huang claimed that 80% of startups build on open models.

Mark Zuckerberg's path has been more erratic. Meta built developer goodwill on Llama, then debuted Muse Spark in April as a closed flagship. On August 10, it swung back, releasing the 30-billion-parameter open-weight Muse Glimmer and promising Spark's weights too. His accompanying 6,500-word essay argued that concentrating AI in a few companies is itself a danger, and cast OpenAI and Anthropic as the foils.

Advertisement - Scroll to continue

Neither man is a pure idealist. Nvidia sells the hardware open models run on. Meta earns from advertising, not from selling model access, so giving weights away costs it little and undercuts rivals who do charge.

The closed labs' case is not frivolous. Altman put the core problem plainly in 2025: once weights are out, they cannot be pulled back. Neither can the safety guardrails a downloader strips away. A closed model can be patched, monitored, and switched off; an open one, once copied, cannot.

Geopolitics sharpens the worry. Washington fears that Chinese labs distil expensive American closed models into cheap open ones. Chinese open models from DeepSeek, Alibaba, and Moonshot have grown fastest in adoption, and Moonshot's Kimi K3 helped push Huang into the policy fight this summer. Notably, the White House's new framework for pre-release review of advanced models reportedly covers only the most powerful closed systems.

There is the commerce angle, too. Closed labs sell access to their models; publishing weights would give away the product. Their frontier models still lead on the hardest reasoning and agentic tasks, and that lead is what customers pay for.

The schism is reshaping how AI gets built. The capability gap that once settled the debate has narrowed sharply: by one widely cited estimate, the best open models now trail the closed frontier by about four months, and cost roughly a third as much per unit of capability. Yet enterprise use of open models has reportedly fallen, as large companies pay a premium for managed services, support, and the very best performance.

The result is a two-tier economy. Startups and researchers build on open weights because they can inspect, fine-tune, and run them cheaply; open models featured in 44% of single-model-family research papers this year, according to one August study. Large enterprises increasingly do both, routing routine, high-volume work to cheap open models and sending only the hardest tasks to a closed API.

That hybrid pattern is the quiet winner. It also means developers' loyalties are fluid, which is why both camps are fighting so hard to set the defaults now.
For the global community of users, the stakes go well beyond Silicon Valley's balance sheets. Open weights lower the price of entry. A hospital, a small business, or a government outside America can run a capable model on its own hardware, keep sensitive data at home, and adapt the model to local languages and laws. Countries without frontier labs gain a way to build rather than merely rent.

Openness also spreads scrutiny. Independent researchers can probe open models for bias and flaws in ways closed systems do not allow. When Hugging Face was hacked this year, it reportedly had to defend itself with a Chinese open model because closed models' guardrails blocked the security work.

The costs are real, too. The same freedom lets bad actors remove safeguards, and no one can recall a harmful model once it spreads. Users also face a geopolitical choice: if American policy restricts open weights, much of the world may default to Chinese models, along with whatever values and censorship they carry.

The open-versus-closed fight will not end in a knockout. Closed labs will keep the frontier for as long as they can afford to; open models will keep closing the gap a few months behind. The real question is who sets the rules in between: whether governments treat weights as dangerous exports or as shared infrastructure. Huang and Zuckerberg are betting on openness partly for profit, and the closed labs are guarding their models partly for the same reason. For the world's users, the healthiest outcome is a market where both survive, and no single company or country holds all the keys.

(Subimal Bhattacharjee is a policy adviser on digital tech issues and the author of 'The Digital Decades: Thirty years of the Internet in India')