$GOOGL $AMZN Battle for the Bench: Google TPU vs. AWS Trainium.
Google TPUs and AWS Trainium/Inferentia are best understood as vertically integrated AI infrastructure platforms rather than standalone accelerator products. Google is further along technically and institutionally: TPUs have been deployed in Google data centers since 2015, have been shaped around Google’s own model, compiler, networking, and data-center architecture, and now underpin Gemini, Search, Photos, Maps, and Google Cloud AI services. AWS is later in accelerator maturity but has a stronger cloud-operating abstraction layer through EC2, Nitro, Bedrock, and Neuron, which allows AWS to hide more of the silicon transition from customers and monetize capacity quickly once deployed. Google’s advantage is system-level performance and workload co-design; AWS’s advantage is commercial packaging, price-per-token positioning, and the ability to use Bedrock APIs and EC2 instance types to make custom silicon feel less like a developer migration.
The central investment conclusion is that Google has the higher-quality custom AI silicon stack, while AWS has the more explicit near-term monetization wedge via Anthropic and OpenAI commitments. Google’s TPU roadmap is more advanced in scale-up networking and system co-design, with TPU v7 Ironwood at 9,216 chips per pod, TPU 8t scaling to 9,600 chips per pod, and Virgo connecting 134,000 TPU 8t chips across data centers at 47 Pbps. AWS Trainium3 is narrower as an accelerator architecture but commercially powerful: it is positioned as a 3nm, 2.52 PFLOPS FP8 chip with 144 GB HBM3e, 4.9 TB/s memory bandwidth, 4.4x performance, 3.9x memory bandwidth, and 4x energy efficiency versus Trainium2.
The biggest portfolio-management implication is that the highest-confidence public equity beneficiary remains Broadcom, not Alphabet or Amazon directly, because the TPU externalization cycle, Meta MTIA, OpenAI ASICs, and AI networking create a more directly monetizable revenue line. JPMorgan estimates Broadcom’s FY27 AI backlog at >$150B, including ~6.5M Google TPU units in Broadcom’s FY27 order book, and expects Broadcom AI revenue near $60B in FY27, up ~3x Y/Y. However, Broadcom’s upside is paired with execution risk in advanced packaging, potential hyperscaler insourcing, and dependence on TSMC/CoWoS/HBM. MediaTek has materially higher convexity through Google TPU v8/v9 exposure, but its outcomes are more binary, with credible sell-side estimates ranging from roughly $5B-$18B in 2027 TPU-related revenue and a much wider 2028 dispersion depending on whether Humufish reaches volume production.
