Nvidia is king in AI chips, but Google and Amazon want to catch up by making their own

Nvidia reported blowout profits as demand for its GPUs powering AI surged—shipping about 6 million Blackwell GPUs over the past year and selling server racks with 72 GPUs for roughly $3 million each while shipping ~1,000 racks weekly—but the AI hardware landscape is rapidly diversifying. Major cloud providers and Big Tech are investing heavily in custom ASICs (Google TPUs, AWS Trainium, Microsoft’s Maia, and OpenAI’s planned Broadcom-assisted chips) and in-device NPUs and FPGAs that offer better price-performance for inference and edge use cases (Anthropic plans to train on up to 1 million TPUs; AWS says Trainium delivers 30–40% improved price performance), albeit with high upfront design costs that favor hyperscalers. Broadcom, Marvell and TSMC are central to this shift, and while bespoke silicon could reduce hyperscalers’ long-term reliance on Nvidia, Nvidia’s scale, capacity and entrenched CUDA software ecosystem make it difficult to displace in the near term.

Analysis

Nvidia reported blowout results driven by AI demand, shipping roughly 6 million current‑generation Blackwell GPUs over the past year and selling server racks with 72 GPUs for about $3 million apiece while shipping roughly 1,000 racks per week, underscoring near‑term capacity strength and the stickiness of its CUDA software ecosystem. GPUs remain the primary workhorse for training large models because of their parallel compute architecture, and Nvidia's scale makes near‑term displacement difficult despite high per‑unit prices (up to ~$40,000).

Hyperscalers and Big Tech are aggressively deploying custom ASICs and alternative accelerators: Google TPUs (up to 7th gen), AWS Trainium (30–40% claimed price‑performance edge and a Trainium3 cadence expected by December), Microsoft’s Maia 100, and OpenAI’s planned Broadcom‑assisted ASICs (commercial rollout noted from 2026). ASICs and NPUs lower per‑unit inference cost and on‑device chips (Qualcomm, Apple) shift some workloads away from data centers, though ASIC development carries high upfront costs (tens of millions) favoring large cloud players.

Broadcom, Marvell and TSMC are strategic infrastructure beneficiaries—TSMC’s Arizona capacity and Broadcom’s design partnerships (including with OpenAI) were highlighted—while FPGAs and edge NPUs add diversification. The key investor axis is timing: Nvidia’s entrenched developer ecosystem and manufacturing ramp give it near‑term dominance, but accelerating ASIC and edge adoption represent a credible medium‑term erosion risk to GPU share and pricing.

Key Decisions for Investors

• Maintain positive near‑term exposure to NVDA to capture ongoing demand and revenue strength but position size for medium‑term substitution risk from ASICs and edge chips
• Add selective exposure to Broadcom (AVGO), Marvell (MRVL) and TSM (TSM) as beneficiaries of hyperscaler ASIC programs and fabrication capacity, keeping allocations tactical
• Monitor leading indicators—TPU/Trainium deployment rates, Trainium3 and OpenAI/Broadcom timelines (noted 2026), GPU rack pricing and weekly shipment cadence—and reduce NVDA exposure if ASIC adoption accelerates materially
• Consider modest exposure to edge AI plays (QCOM, AAPL) to capture growing on‑device AI demand while recognizing most dollars remain concentrated in data‑center infrastructure