Cloud Giants Accelerate Custom Chip Development While Setting New Nvidia Procurement Records
Alina Collins
Amazon, Google, and Microsoft are all racing to build their own AI chips — yet all three are simultaneously buying more Nvidia GPUs than ever. Custom silicon and record procurement running in parallel is the core tension in Nvidia's investment case right now.
How far along is Amazon's custom chip effort?
Amazon's custom chip business — spanning Graviton CPUs, Trainium AI chips, and Nitro networking chips — hit an annualized revenue run rate above $20 billion in fiscal Q1 2026.
CEO Andy Jassy said that if the chip unit operated independently and sold externally, annualized revenue would reach $50 billion, placing it among the world's top three data-center chip companies.
Yet Amazon plans roughly $200 billion in 2026 capex, with a large share still flowing to Nvidia-GPU infrastructure. This means → the company furthest ahead in custom chips is also one of Nvidia's biggest buyers.
What does Google opening up TPUs mean?
Google has released its eighth-generation TPU and, for the first time, is shifting from internal-only to external access. Blackstone formed a joint venture with Google, committing an initial $5 billion to offer TPU capacity as a rentable cloud service, with 500 MW of compute planned for 2027.
Google also gave AI lab Anthropic access to up to one million TPUs and reportedly signed a lease deal with Meta. This means → TPUs are no longer just an internal tool — they now compete directly with Nvidia in the open market.
But this same week, Google signed a multi-year cloud deal with SpaceX involving roughly 110,000 Nvidia GPUs. In plain terms = Google is building its own chips and still buying Nvidia at scale.
Where does Microsoft stand?
Microsoft trails the other two. Its Maia accelerator — a custom chip designed for AI workloads — only recently went live in select data centers in its second-generation form (Maia 200), primarily serving Microsoft 365 Copilot and some OpenAI models.
The vast majority of AI workloads on Azure still run on Nvidia GPUs. Maia is more a tool to gradually claw back some external procurement spending than a full replacement.
Microsoft expects roughly $190 billion in 2026 capex. Azure revenue grew 40% year-over-year last quarter, yet capacity remains tight through the year. This reflects that even with custom silicon coming online, Nvidia remains Azure's compute backbone in the near term.
Can Nvidia's "parabolic" growth continue?
Nvidia posted 85% year-over-year revenue growth to $81.6 billion in fiscal Q1 2027. Data-center revenue rose 92%, with hyperscalers still accounting for roughly half.
CEO Jensen Huang called demand "parabolic" and pointed to AI startups, enterprise customers, and sovereign governments as a fast-growing new buyer pool — "these customers don't build or design their own chips."
In plain terms = old customers are building in-house, but new customers are flooding in — Nvidia's total pie is still growing.
What is the real test for the bull-bear debate?
The bear case: the four largest customers plan a combined $725 billion in 2026 capex, up roughly 77% year-over-year — but as their custom-chip share expands, Nvidia's slice of that spending pool will narrow, and these customers have strong incentives to reduce single-supplier dependence.
The bull case: the overall AI spending pool is expanding fast enough that Nvidia can still grow revenue even as its share edges down.
This means → the real test arrives when AI capex growth slows. Only then will we see whether custom-chip erosion can be offset by incremental demand from new buyer cohorts. As long as spending accelerates, the contradiction stays dormant; the moment growth decelerates, the share battle begins in earnest.
Content is for reference only, not financial advice.