June 18, 2026 — Deep inside Amazon’s cloud empire, a quiet mutiny against silicon orthodoxy is underway. The company’s third-generation Trainium AI chip is essentially sold out. Demand for next year’s fourth iteration is already surging. And CEO Andy Jassy has placed the tech world on notice: within two years, there is “a good chance” Amazon will sell full racks of its proprietary chips directly to third-party data centers.
The Deal That Isn't Signed Yet — And Why It Already Changes Everything
The conversations are already happening. Peter DeSantis, who orchestrates Amazon’s AI, custom silicon, and quantum computing efforts, has confirmed active discussions to sell Trainium accelerators externally. This marks a profound departure from Amazon’s traditional model of restricting access strictly through AWS.
Born from the 2015 acquisition of Annapurna Labs, Trainium already powers heavy-duty workloads for OpenAI, Anthropic, and Uber. Crucially, revenue commitments tied to the chip have eclipsed a staggering $225 billion. The third-generation silicon, shipping on advanced 3nm nodes since early 2026, is nearly fully subscribed.
This isn't merely a product expansion. It is a tectonic shift in strategic posture—a leap from renting out cloud compute to selling the bedrock hardware that makes it possible.
Ten Years of Infrastructure Theology, Now Coming Due
Wall Street is misreading the moment, treating this as a sudden pivot. In reality, Amazon has been executing this exact playbook for a decade: acquire Annapurna, build Nitro cards for virtualization, deploy Graviton CPUs to undercut Intel, and finally unleash Inferentia and Trainium for AI.
The Graviton precedent is the blueprint. Amazon leveraged its massive internal scale to iterate the chip, offered undeniable price-performance gains, and systematically made the AWS control plane matter more than the merchant silicon humming beneath it. Trainium elevates that strategy to the most expensive layer of the data center.
Amazon’s custom silicon unit already generates an internal run-rate exceeding $20 billion, growing nearly 40% sequentially in early 2026. If spun out, it could command $50 billion in addressable revenue. With Trainium claiming up to a 40% price-performance edge over comparable Nvidia GPUs, the endgame is clear. The goal isn't to become the next Nvidia. It is to turn Nvidia into the Intel-inside-the-cloud: a tolerated necessity, no longer allowed to dictate the terms.
The Valuation Disconnect Professionals Should Not Miss
For capital allocators, the mispricing is hiding in plain sight. Amazon trades at 29x forward earnings with a $2.63 trillion market cap. Nvidia commands a 31.7x multiple at $5.08 trillion. The market assigns Amazon almost zero structural premium for owning the custom chip, the cloud control plane, the customer billing relationship, and $225 billion in locked AI demand.
But the reflexive bull case—extrapolating a clean $50 billion standalone chip business—is analytically reckless. External hardware sales carry margin drag, warranty burdens, and channel complexities absent in pure cloud computing. Trainium should be valued as massive capex deflation and strategic leverage, not as a pristine semiconductor P&L.
The most dangerous misallocation on the board is AMD. Trading at an eye-watering 173x P/E, it is priced as the default Nvidia alternative. Yet the true existential threat to Nvidia isn't another GPU vendor. It is proprietary hyperscaler silicon like Trainium, Google’s TPU, and Microsoft’s Maia. AMD is the crowded second-source trade in a market where the marginal dollar is fleeing to ASICs it will never manufacture.
The Announcement That Will Matter More Than the Revenue
Within twenty-four months, Amazon will likely formally signal external Trainium rack availability. Initial volumes will be carefully rationed for strategic partners, not flooded into the open merchant market. AWS simply cannot feed Anthropic, OpenAI, internal Bedrock inference, and external buyers simultaneously without painful allocation tradeoffs.
When that announcement drops, it will be a narrative earthquake before it generates material hardware revenue. Sell-side analysts will inevitably model the external sales too aggressively. But the symbolic weight will trigger a multiple-compression event across the AI hardware complex.
Nvidia is not a clean short; its 70% gross margins and CUDA software moat are formidable. But its valuation is now hostage to customers who are violently motivated to break their dependence. Hyperscaler AI capex is projected to hit $770 billion in 2026—eating 100% of operating cash flow. At those extremes, custom silicon stops being a science project and becomes a financial imperative.
Amazon will modestly rerate. Nvidia will face a grueling margin test. And the real pain will concentrate in legacy hardware alternatives, as the market realizes the future of AI silicon is being forged inside the cloud, not outside it.
not investment advice
Sources: https://www.aboutamazon.com/news/aws/why-ai-startups-choose-amazon-trainium-chips