Google vs Amazon custom AI chips race and Nvidia challenge

Google’s eighth‑generation TPU 8t delivers three times the performance of the prior Ironwood chip, while Amazon’s Trainium 2 claims 30% lower cost per unit of performance, each positioning their firms as the most credible in‑house alternatives to Nvidia’s H100 GPUs [2].

Google’s vertical silicon push

Sundar Pichai told investors that Google “owns frontier models and owns the silicon,” underscoring a strategy that ties custom TPU hardware directly to its AI services. The new TPU 8t, now powering Gemini, processes over 16 billion tokens per minute via API and is slated to ship most of its $462 bn backlog in 2027, according to the CFO [2]. This vertical integration fuels a 63% surge in Google Cloud revenue to $20 bn, with the order backlog nearly doubling to $462 bn, highlighting strong demand for Google‑built chips.

Amazon’s merchant‑chip model

Andy Jassy framed Amazon’s approach as a “picks and shovels” business: Trainium 2, the latest generation of its custom AI accelerator, is 30% cheaper per unit of performance than comparable GPUs, and Amazon reports a $20 bn annual revenue run rate for its chip division, growing at triple‑digit rates year over year [2]. AWS growth of 28%—its fastest in 15 quarters—has been bolstered by multi‑gigawatt training contracts with Anthropic and OpenAI, suggesting that Amazon’s hardware is gaining traction beyond internal use. Jassy also hinted that Amazon plans to sell complete training racks within the next few years, expanding Trainium from a cloud‑only tool to a broader merchant product [2].

Competitive context with Nvidia

Nvidia remains the dominant AI‑chip supplier, with $47.5 bn in data‑center sales in 2024, up from $15 bn the year before, and analysts expect the figure to more than double in fiscal 2025 [1]. However, both Google and Amazon are scaling in‑house silicon to reduce reliance on Nvidia’s costly H100 GPUs, which sell for up to $57 000 each on secondary markets [1]. While Google’s TPU offers a performance edge, Amazon’s Trainium emphasizes cost efficiency and a merchant‑oriented distribution model, creating two distinct competitive moats against Nvidia’s pricing power.

What to watch

• Google TPU shipments: Monitor the 2027 timeline for the bulk of the $462 bn TPU backlog to gauge cloud revenue impact.
• Amazon Trainium 3 supply: Jassy described the next generation as “nearly fully subscribed”; its rollout will test Amazon’s ability to meet external demand.
• Nvidia H100 supply constraints: Continued shortages could accelerate adoption of Google and Amazon silicon, affecting Nvidia’s market share.

The race between Google’s vertically integrated TPU platform and Amazon’s merchant‑focused Trainium underscores a broader shift: large cloud providers are building custom silicon not just to cut costs but to lock in AI workloads, a trend that could reshape Nvidia’s dominance if supply constraints persist.