Wall Street analysis compared Google's TPU roadmap against Amazon's Trainium/Graviton family, finding both lag NVIDIA on established deployments but compete on cost-per-inference.
Google and Amazon both reported Q1 FY2026 results on April 29, 2026, each devoting significant portions of their earnings calls to custom silicon strategies. With NVIDIA still setting the industry pace, Google's TPU roadmap and Amazon's Trainium franchise represent the two most credible in-house alternatives in the cloud market.
**TPU Powers Gemini. Trainium Powers Everyone Else**
Google's silicon story is fundamentally vertical. Sundar Pichai told investors that "we own frontier models and own the silicon," and the business results bore this out. Cloud revenue grew 63% to $20 billion, with backlog nearly doubling to $462 billion. The new 8th-generation TPU 8t delivers 3x the processing power of Ironwood, and Gemini now handles more than 16 billion tokens per minute via direct API. Pichai acknowledged that "we are compute constrained in the near term."
Amazon is taking a different approach, positioning itself as a merchant vendor. Andy Jassy reported that the chips business topped a $20 billion annual revenue run rate, growing triple digits year over year, with total Trainium commitments exceeding $225 billion. AWS itself grew 28%—its fastest pace in 15 quarters—boosted by multi-gigawatt training deals from Anthropic and OpenAI. Trainium2 costs 30% less per unit of performance than comparable GPUs, directly pressuring NVIDIA's pricing power.
The strategic difference is stark. Google treats TPU as an internal competitive weapon that occasionally ships to select frontier AI labs and capital markets customers. Amazon treats Trainium as a core merchant product, with Jassy hinting that "we are going to sell racks in the next couple of years." One approach compounds margin. The other compounds distribution.
The critical question for Google is whether it can convert that $462 billion backlog into shipped TPU capacity, particularly given the CFO's statement that the majority of TPU hardware revenue lands in 2027. For Amazon, the pinch point is Trainium3 supply, which Jassy called "nearly fully subscribed." If memory costs continue to surge, gross margins on custom silicon could compress before economies of scale materialize.
On current valuation metrics, Google presents the cleaner near-term setup. TPU v7 Ironwood and Google's software ecosystem run frontier models at a tangible total-cost-of-ownership advantage, and a forward P/E around 25 is reasonable for 22% revenue growth. Amazon's forward P/E near 29 is harder to justify today, but Trainium's momentum and Amazon's distribution capabilities should establish it as the dominant non-NVIDIA hardware standard in the cloud infrastructure era ahead. Google has built the visible moat today. Amazon is laying concrete for tomorrow's.