Anthropic CEO Dario Amodei says the years-long fight between "open" and "closed" AI is centered on the wrong axis. According to The Decoder, Amodei argues that artificial intelligence is a centralizing technology by nature — and releasing model weights publicly doesn't undo that, it just moves the point of control from labs like Anthropic and OpenAI to whoever controls the underlying chips and data centers.

That's a pointed argument to make from inside a company that has built its entire business on keeping Claude's weights closed. Open-weight advocates have spent years framing model releases from Meta, Mistral, and DeepSeek as a democratizing counterweight to a handful of well-funded labs. Amodei's framing flips that: even a fully open model still needs somewhere to run, and that somewhere is an increasingly narrow list of GPU fleets, cloud providers, and chip fabs.

The claim lands at a moment when governments and enterprises are actively choosing between open and closed stacks for reasons that have nothing to do with idealism — cost, data residency, and vendor lock-in.

Why "open weights" doesn't mean "open power"

The mechanics of Amodei's argument are straightforward once you follow the supply chain. A model's weights are freely downloadable, but weights alone don't produce intelligence — they need inference compute, which today runs overwhelmingly on Nvidia GPUs, largely fabricated by TSMC, and largely deployed inside a handful of hyperscale data center operators (AWS, Microsoft Azure, Google Cloud, plus a growing set of neoclouds). Downloading Llama or DeepSeek's weights doesn't give an organization independence from that stack — it just changes which line item on the cloud bill grows.

• Training frontier-scale models still requires capital and chip access that only a handful of companies and states can secure
• Running open models at meaningful scale still depends on the same GPU supply that closed-model labs compete for
• Whoever controls energy and data center capacity increasingly sets the real ceiling on who can build with AI at all

Under this reading, "open" and "closed" aren't opposite ends of a power spectrum — they're two different licensing choices sitting on top of the same concentrated hardware layer.

What this means for teams choosing a model stack

For engineering teams evaluating open-weight versus API-based models, the practical takeaway isn't about ideology — it's about where the actual dependency sits.

• Self-hosting open weights doesn't remove your chip dependency — it just moves it up the stack, since you still need GPU capacity from a cloud provider or your own hardware budget
• Data sovereignty and cost control are the real, defensible reasons to run open models, not an assumption that self-hosting decentralizes AI power in some broader sense
• Chip and export-control policy is now a more direct constraint on your AI roadmap than which lab's license you choose — sanctions, export restrictions, or a shortage at your cloud provider can bottleneck an open-model deployment just as easily as a closed one

For startups and enterprises building products on top of any model, this argues for treating compute availability — not model licensing terms — as the primary risk to plan around when choosing infrastructure partners.

AiiN's takeaway

Amodei's argument is self-serving in an obvious way: it recasts Anthropic's closed-weight strategy as no worse, on the power-concentration question, than releasing weights openly. But the underlying observation about the hardware layer is hard to dispute — inference and training both funnel through the same narrow set of chip and data center suppliers regardless of a model's license.

In our estimation, the more useful framing for builders isn't open-versus-closed at all — it's which layer of the stack you actually want to depend on. A team chasing real independence from any single AI lab should be asking about multi-cloud GPU access and long-term chip supply commitments, not just which weights it can download.