Nvidia faces a new kind of pricing pressure as hyperscalers build more custom AI chips

Nvidia’s dominance in AI compute has helped it capture a surge in data center spending, but a recent market-focused commentary suggests the company’s harder problem may be upstream: customers increasingly have incentives to stop paying Nvidia’s prevailing prices for every layer of the AI stack.

The article points to Nvidia’s latest reported momentum. It cites Nvidia’s Q1 fiscal 2027 results, including revenue of about $81.6 billion, up roughly 85% year over year, and data center revenue of about $75.3 billion. It also cites a market positioning that includes networking revenue growing about 199% year over year to roughly $14.8 billion, indicating that demand is not limited to GPUs alone.

Still, the central claim is about bargaining dynamics. The commentary says roughly half of Nvidia’s data center revenue comes from hyperscalers, and that those same customers are actively funding alternative silicon. It links this strategy to a push for proprietary chips designed to control cost and improve negotiating leverage.

Amazon, Google, Microsoft, and Meta are specifically mentioned as building custom AI accelerators, including references in the commentary to Amazon Trainium, Google TPU, Microsoft Maia, and Meta MTIA. In that framing, Nvidia’s customers are spending billions, but attempting to shift the mix of what they buy from Nvidia toward what they develop in-house.

The piece also argues that Nvidia is trying to defend the value of its broader platform beyond the GPU. It highlights Nvidia’s interconnect approach, including NVLink Fusion, as an effort to keep networking and system integration sticky even when hyperscalers shift portions of compute to custom chips.

At the same time, the commentary implies Nvidia may still have to concede some battles over raw GPU demand as alternatives scale. It positions the situation as a pricing pressure problem rather than a demand collapse, suggesting customers want Nvidia’s performance while finding ways to pay less overall for the components they can internalize.

The market discussion adds that Nvidia is in a period of platform change, referencing product cycle items such as Blackwell Ultra ramping and the announcement of Rubin, and it also reiterates the sense that Nvidia’s “AI factory” buildout theme remains central for buyers and suppliers. But the argument keeps returning to one practical question: how much of the full AI training and inference workload customers will choose to source from Nvidia versus their own chips.

Nvidia did not disclose in the commentary any customer-by-customer pricing details, contract terms, or explicit guidance for how custom silicon might change gross margin in future quarters. It also does not quantify how much of the “customization” is expected to replace Nvidia GPUs versus wrap them with different system configurations. For now, the debate is framed as leverage and economics, not as a confirmed revenue loss.