Annapurna Labs and the Hidden Israeli Silicon Inside AWS
Amazon Web Services runs the world's largest cloud platform on silicon designed in Tel Aviv. Two senior departures in the past nine months suggest the moat is being tested.
Amazon Web Services runs the world's largest cloud platform on silicon designed in Tel Aviv. That fact has, until recently, been one of the cleanest competitive moats in enterprise computing. Two senior departures from Annapurna Labs in the past nine months suggest the moat is being tested.
Annapurna Labs — acquired by Amazon in January 2015 for approximately $350 million — designs the four chip families that define AWS infrastructure: Graviton, the general-purpose ARM-based server processor; Trainium, the custom accelerator for AI model training; Inferentia, the inference counterpart; and Nitro, the security and virtualization controller that sits between every customer workload and the physical hardware. The largest Annapurna engineering site in the world is in Tel Aviv. A second Israeli site, in Yokneam, complements it. Amazon also operates Annapurna labs in Austin, Texas, and Toronto, but the design center of gravity has remained in Israel.
Amazon does not disclose chip revenue. Industry estimates, widely reported during 2026, place the chips business at an annual revenue run rate of roughly $20 billion — calculated as internal sales from Annapurna to AWS — with quarter-over-quarter growth near 40 percent in the most recent reporting period. If the chips business were a standalone public company, analysts have estimated its standalone run rate would approach $50 billion. The acquisition is, by Amazon's standards, the most successful small purchase in the company's history.
The talent question
The current question is whether Amazon can keep the people who built it.
Rami Sinno, a senior leader inside Annapurna's chip team, left for Arm Holdings in August 2025. Gadi Hutt, the unit's director of product and customer engineering and one of the longer-tenured Annapurna executives, left Amazon in March 2026. The departures are unusual for an organization that has historically retained its senior people through Amazon's compensation cycles. Israeli press reports have linked the moves to broader competition from NVIDIA, Arm, and well-funded AI silicon startups — many of which are themselves hiring out of Tel Aviv at compensation levels Amazon's pay structure was not built to match.
The Anthropic commitment
The departures arrive as Annapurna's commercial position has never been stronger.
In April 2026, Anthropic and Amazon announced a $100 billion, ten-year commitment by Anthropic to use AWS silicon — Trainium and Graviton — across multiple chip generations, securing up to five gigawatts of capacity to train and deploy Claude. The deal is one of the largest single-customer commitments to non-NVIDIA AI silicon in the industry. Trainium3 began shipping in early 2026 and, according to AWS, is largely sold out across its initial customer base. Trainium4 is in development, with first parts expected in 2027.
The Israeli stake
For Israel, Annapurna remains the cleanest example of how Israeli engineering capital can drive a category-defining outcome at a U.S. parent company without ever becoming a household name. Most consumers will never hear of Trainium. Most enterprises will never see a Graviton chip directly — they will rent EC2 instances powered by one. Yet the architectural choices being made in Tel Aviv now determine, in part, how the world's largest cloud platform structures its cost basis against NVIDIA, how Anthropic's Claude is trained at scale, and how AWS competes against Google's TPU strategy and Microsoft's emerging Maia silicon.
The structural position is enviable and exposed in equal measure. Annapurna's deepest competitive moat is the tight integration between its silicon and AWS's hyperscale software stack — a vertical play that NVIDIA, for all its market power, cannot replicate. The exposure is talent. Annapurna's senior engineers are now among the most poached people in the global semiconductor industry, and Amazon's compensation structure was not designed to compete with NVIDIA stock at current valuations.
The next twelve months — Trainium3 ramp, Trainium4 design, the Anthropic capacity build — will be executed largely from Tel Aviv. The hidden Israeli chip business inside AWS is, at this point, no longer particularly hidden. Whether it can hold the team that built it is the question that defines the next chapter.
