Hyperscalers Unleash AI Chip War Against Nvidia

The relationship between Nvidia and major hyperscalers Amazon, Google, and Microsoft has shifted from a friendly supplier-buyer dynamic to a high-stakes competition. Each player is trying to commoditize the layer the other serves, with Nvidia's Jensen Huang having a strong balance sheet and 75% gross margin business. The hyperscalers have come to the same conclusion: they don't want to pay Nvidia's margin forever.

With hundreds of billions and potentially trillions spent on AI chips annually, every layer of the AI stack wants to capture as much value as possible. Training is where Nvidia's moat is strongest, but inference is where most future volume is going. Inference workloads are known and stable, making custom ASICs a cost-effective option over general-purpose GPUs.

Google, Amazon, and Microsoft have all invested heavily in chip design and production. Google produces 3 million TPUs annually and has partnered with Broadcom for next-gen TPU capacity starting in 2027. Amazon sees its semiconductor design business as a key pillar of the company's future and has already achieved $25 billion in annualized revenue with Trainium. Microsoft is also developing its own chip, Maia, which aims to offer a 30-40% cost advantage over Nvidia chips.

The hyperscalers are framing their AI spending as an option rather than a must, committing to long-term CapEx but buying chips only months before data centers come online. This allows them to slow down or cancel chip purchases if demand changes, giving them more flexibility and independence in the market.