Nvidia closed up 0.75% at 183.22, but the slow creep higher belies a sharper risk narrative: its biggest customers are building their own accelerators. With Google, Amazon, Microsoft and OpenAI doubling down on custom silicon—and Meta buying its way in—hyperscalers are turning from price-takers into platform competitors, a shift that threatens Nvidia’s pricing power even if unit demand stays hot.
The AI chip land grab remains Nvidia’s to lose, but the customer mix is changing in ways that matter for margins. Google’s long-running TPU program is moving beyond internal workloads and, per recent reports, into external sales. Amazon is scaling Trainium2 under Project Rainier, with Anthropic already a flagship user. Microsoft is pushing Maia despite a slower start, and OpenAI is designing its own parts with Broadcom. Meta’s move to acquire Rivos adds more in-house firepower. None of these companies needs to sell chips at scale to dent Nvidia’s leverage. They just need enough viable, cheaper alternatives to keep their own data center economics intact—and to keep Nvidia honest on price and delivery.
JPMorgan pegs custom chips at 45% of the AI accelerator market by 2028, up from 37% in 2024. That mix shift is critical. Hyperscalers say they make less renting Nvidia GPUs than they could on their own silicon because they control the full stack, from compiler to networking. Cheaper, workload-tuned chips are good enough for a rising share of training and inference, especially as software abstractions improve. That is why analysts talk about death by a thousand cuts. Even if Nvidia continues to grow units into a larger total market, a steadier drip of internal alternatives can clip the premium pricing that has defined this cycle. The short-term result is healthy revenue; the longer-term risk is normalized gross margin as hyperscalers wield credible substitutes.
Google’s reported move to physically sell its TPUs to an outside cloud provider is the cleanest signal that Nvidia’s most sophisticated customer is now a competitor. The TPU line has matured through multiple generations, with the newest Ironwood extending a decade of iteration that has narrowed the performance and efficiency gap. Analysts argue demand would exist among frontier labs and cost-sensitive enterprise AI, particularly where workflows are already optimized for Google’s software stack. If Google scales external TPU systems, it resets expectations on performance-per-dollar and power-per-rack—benchmarks that influence procurement far beyond a single vendor. It would also validate the business model for custom silicon as more than an internal cost hedge, inviting AWS and Azure to push their own hardware harder into third-party workloads.
Expect 2026 to be a busy year for non-Nvidia silicon. Seaport’s Jay Goldberg sees accelerated activity across the custom chip supply chain, and the roadmaps suggest why. AWS has Trainium2 in production deployments and is investing in capacity under Project Rainier. Microsoft, behind on first-gen Maia, is incentivized to catch up to protect Azure margins and developer lock-in. Meta is leaning on Rivos to bolster its AI accelerator plans after mixed results with earlier efforts. OpenAI’s partnership with Broadcom points to a dedicated path for its most compute-intensive workloads. These are not vanity projects. They are the logical outcome of hyperscalers refusing to be captive to one supplier in a capex cycle measured in the hundreds of billions.
Nvidia’s counter is simple and, for now, effective: it sells entire AI factories. Jensen Huang’s pitch is that Nvidia is more than a GPU vendor. It integrates Blackwell GPUs with in-house Grace CPUs and networking gear into turnkey racks, a full-stack offering that reduces deployment risk and time-to-value for customers racing to ship AI features. That matters because the bottlenecks in AI are as much about networking, software and orchestration as raw compute. By controlling the system and the CUDA-led software ecosystem, Nvidia defends its moat where rivals are thin. Hyperscalers can build a chip; building a coherent, scalable system with a vast developer base is harder. That is the edge investors are paying for.
Some on Wall Street argue the custom-chip surge will not derail Nvidia because the market is still expanding faster than any one rival can fill. Bank of America and others frame it as a share-of-a-growing-pie story: Nvidia keeps finding ways to expand demand, from training massive frontier models to the coming wave of inference and data processing. That can remain true while pricing power normalizes. The critical detail for valuation is not shipments but the sustainability of premium margins. If custom silicon captures 40 to 45% of accelerators by 2028, even without stealing Nvidia’s crown, it can anchor expectations for lower unit ASPs and more aggressive discounting on big cloud deals. Nvidia still wins on volume and platform, but the shape of the win changes.
Custom silicon does not just displace Nvidia. It redistributes profit pools. Broadcom (AVGO) and Marvell (MRVL) are baked into hyperscaler roadmaps as design and manufacturing partners. Memory makers like SK Hynix and Micron ride every configuration that craves high-bandwidth memory, regardless of the logo on the accelerator. Networking remains a choke point that benefits whoever solves bandwidth-per-watt and low-latency scaling at the rack level, where Nvidia is strong but where merchant silicon and in-house designs are rising. If Google’s TPU systems ship externally, expect a ripple across optical interconnects and power infrastructure as the industry optimizes around multiple reference architectures, not just Nvidia’s. Winning the AI data center is increasingly a question of supply chain control, not just chip leadership.
Near term, watch Google’s external TPU cadence and pricing, AWS’s Trainium2 availability and benchmark transparency, Microsoft’s Maia roadmap, and any tangible milestones from OpenAI’s Broadcom partnership. Those will show whether hyperscalers intend to push their chips beyond internal workloads into revenue-generating services for customers. For Nvidia, keep an eye on how Blackwell systems are bundled, whether software entitlements become a bigger lever, and how aggressively it prices multi-year capacity reservations with cloud giants. AMD (AMD) remains the most direct alternative in the GPU lane; if hyperscalers push their own silicon for mainstream inference, AMD’s upside likely leans toward enterprise and specialized training niches unless it lands repeat, large-scale wins. As of the latest close, Nvidia shares are holding up, but the tone in research and the slow drumbeat of customer announcements point to a tighter margin story ahead. The bull case endures on platform depth and execution. The bear case grows on mix and pricing.
Nvidia is not ceding the AI era. It is redefining the battlefield around systems, software and speed of iteration while the hyperscalers chip away at the parts of the stack where they can best control cost. That makes today’s modest stock gains plausible and the long-term debate sharper. If Big Tech’s custom silicon hits its stride by mid-2026, Nvidia’s unit growth can stay strong while its take-rate trends lower on the largest deals. For investors, the question is less whether Nvidia stays on top and more what premium multiple matches a world where its customers are also its most capable rivals.
