SEATTLE — The $25 billion partnership between Amazon and Anthropic has a dirty little secret: Amazon is using Anthropic's own models to build competitors.
According to a report from The Information, Amazon engineers are actively distilling Anthropic's Claude models — feeding Claude's outputs into smaller, cheaper models to replicate its capabilities at a fraction of the cost. The practice is reportedly widespread enough that Amazon has explored alternatives to Anthropic, including OpenAI and its own Nova models.
This is not corporate espionage. This is not a terms-of-service violation. This is, apparently, exactly what the contract allows.
According to a person familiar with the matter, Amazon has "certain rights to use Anthropic's models" for distillation — similar to Apple's arrangement with Google Gemini. Amazon even offers a distillation service on its Bedrock cloud platform, though Claude models aren't available there — only Amazon's Nova and Meta's Llama models are supported.
The irony is thick enough to cut with a knife.
The Distillation Arms Race
Distillation is the AI industry's worst-kept secret. It's the process of training a smaller, cheaper model to mimic a larger, more expensive one. You feed the big model thousands of prompts, capture its outputs, and use that data to train the small model. The result: 80% of the capability at 20% of the cost.
Everyone does it. Apple distills Gemini. xAI admitted to distilling OpenAI. Alibaba allegedly conducted the "largest known distillation attack" according to Anthropic's own accusations. The only rule is: don't get caught doing it without permission.
Amazon has permission. It paid $25 billion for it.
The question is whether Anthropic understood what it was selling. When Amazon negotiated its investment, did Anthropic's lawyers foresee that Amazon would use Claude's outputs to build Claude-killers? Or did they assume the $25 billion was a straightforward cloud computing deal — Anthropic uses AWS, Amazon gets equity, everyone wins?
The Information report suggests the relationship is already strained. Starting next year, Amazon will pay Anthropic based on tokens processed rather than compute hours — a change that could dramatically increase Anthropic's revenue, but also gives Amazon a powerful incentive to reduce token usage. Amazon's spokesperson claims the changes won't raise costs. Anthropic points to lower prices relative to performance.
Both statements are technically true and strategically meaningless. The real story is that Amazon is preparing for a future where it doesn't need Anthropic at all.
Why This Matters
Amazon's distillation effort is not just about saving money. It's about control.
Anthropic's Claude models are among the most capable in the industry, but they are also among the most expensive. For a company like Amazon — which processes billions of tokens across Alexa, AWS, internal tools, and customer-facing AI features — the cost of relying on third-party models is unsustainable. At scale, even a 10% reduction in per-token cost translates to hundreds of millions in savings.
But the strategic risk is bigger than the financial risk. Every company building on top of OpenAI, Anthropic, or Google is building on someone else's foundation. If the model provider changes pricing, restricts access, or goes bankrupt (see: Stability AI), the downstream company is exposed.
Amazon learned this lesson the hard way with its dependence on Intel chips for AWS. It spent years and billions building its own Graviton processors to reduce that dependency. Now it's doing the same thing with AI models.
The distillation playbook is identical: use the incumbent's product to train your replacement, then switch when your version is good enough. It's how AMD challenged Intel. It's how ARM challenged x86. It's how Amazon will challenge Anthropic.
The Meta Parallel
Amazon is not alone in its paranoia. The same Information report reveals that Meta has restricted its engineers from using Claude Code and OpenAI's Codex — not because the tools are bad, but because Meta is worried about its own outputs leaking into its training data.
Meta's concern is the mirror image of Amazon's strategy. Amazon wants to distill Anthropic's models. Meta wants to prevent Anthropic from distilling Meta's models. Both companies understand the same thing: in the AI era, your outputs are your intellectual property, and anyone with access can clone you.
Meta is reportedly building its own coding assistant, MetaCode, to replace Claude Code and Codex. The company is on track to spend billions on internal AI use this year alone. The policy is clear: no AI outputs for test tasks, no AI outputs for code analysis, human review required for everything.
This is the new normal. The companies that build AI models are terrified of being copied. The companies that use AI models are desperate to copy them. The result is an industry-wide distillation arms race where every partnership is secretly a Trojan horse.
🔥 Hot Takes
1. Anthropic's $25 billion Amazon deal is the most expensive self-destruct button in tech history. Amazon didn't invest in Anthropic to support a competitor. It invested to buy time — time to distill Claude, replicate its capabilities, and replace it with a cheaper in-house alternative. The $25 billion is not a partnership. It's an R&D subsidy with a sunset clause. Anthropic got cash and cloud credits. Amazon got the blueprint for Claude's replacement. In 3 years, Anthropic will be a footnote in Amazon's AI strategy, and Dario Amodei will be giving TED talks about "what we learned from the Amazon experience."
2. Distillation is the Napster moment for AI — and the record labels are still pretending it's not happening. The music industry spent a decade fighting file-sharing before accepting that the business model had changed. AI model providers are making the same mistake. Distillation is not a bug. It's the feature that makes AI economically viable at scale. Every company with enough compute and enough data will distill the best models. The only question is whether the original creators get paid for it. Amazon's contract apparently says yes. Most companies don't have that contract. The result will be a wave of lawsuits, settlements, and eventually, a new licensing model that looks a lot like music royalties. But by then, the distillation cat will be so far out of the bag that the original models will be commodities.
3. The real winner here is not Amazon or Anthropic — it's the open-source community. Every distillation effort, every leaked model, every reverse-engineered capability eventually finds its way into open-source projects. Llama, DeepSeek, Qwen — all built on the shoulders of proprietary giants, either through distillation or parallel research. Amazon's distilled Claude will not stay internal forever. It will leak, or be replicated, or inspire an open-source equivalent that is 90% as good and 100% free. The $25 billion Amazon-Anthropic deal is a temporary alliance in a war that open source is already winning. In 5 years, the best models will be free, and the companies that charged for them will be wondering where it all went wrong.
The Bottom Line
Amazon's distillation of Anthropic models is not a scandal. It's a strategy. And it's a strategy that every major tech company is pursuing, either openly or in secret.
The AI model business is facing the same commoditization pressure that hit every software market before it. Operating systems became free (Linux). Databases became free (MySQL, PostgreSQL). Cloud infrastructure became a race to the bottom (AWS, Azure, GCP). AI models will follow the same path — the only question is how fast.
Anthropic's bet is that its models will stay ahead of the distillers through continuous innovation. Amazon's bet is that it can distill fast enough to catch up. The open-source community's bet is that neither company can outrun a million developers working for free.
History suggests the open-source community wins in the long run. But in the short run, Amazon's $25 billion distillation project is a masterclass in corporate strategy: pay for the best, copy it cheaply, and leave the original holding an empty bag.
The fence, in this case, is the contract. And Amazon just built a very expensive ladder.