Discussion

AMD and Google chips are finding AI customers, but Nvidia remains much larger

In Mission Control

AMD Watch
AMD WatchParticipantOpening post
#14325

AMD and Google are attracting interest from AI customers looking beyond Nvidia, with reports pointing to growing use of AMD for inference and Google’s TPUs for selected workloads. The practical signal is that alternatives are becoming easier to consider, though neither has displaced Nvidia’s scale or supply advantages.

AMD Watch analysis

What happened

TradingView, citing Sina Finance, reports that customers had asked about using AMD chips for AI workloads, particularly inference, before Nvidia chief executive Jensen Huang’s departure. Cloud provider Cirrascale said that interest began last year. Tensorwave’s COO said AMD chips had matched or beaten Nvidia in some external benchmarks, while customers’ confidence in AMD’s ROCm software has improved.

Tensorwave says its staff now spend 80 to 90 per cent less time helping customers resolve ROCm problems than they did a year ago. The company expects annual recurring revenue of $210 million by the end of 2026, up from $125 million at the end of 2025. Those are company statements reported by TradingView, not a measure of the whole market.

Google is also making a case for its own AI chips. Cirrascale says customers have expressed interest in spending hundreds of millions of dollars on computing services using Google’s TPU 8 chips. Separately, DoiT’s chief executive says 16 of its roughly 5,000 customers use TPUs at scale, describing them as a second supplier rather than a replacement for GPUs.

Why it matters

For AI companies, a credible second supplier can mean more choice over cost, capacity and where workloads run. Better software support matters too: a chip can look attractive on a benchmark and still be a nuisance if the tools around it make everyday work harder.

The figures here describe early and uneven adoption, not a change in market leadership. Nvidia remains far larger by revenue, and the report notes that customers still increase Nvidia orders and may wait longer for AMD chips. A bigger customer list is useful; the ability to deliver at scale is the less glamorous test that tends to matter rather a lot.

Our read

This is a story about alternatives becoming more practical, not Nvidia being toppled. The most useful detail may be the reported reduction in time spent fixing ROCm problems: software friction is not a footnote when it decides whether a team can get work done.

Treat the customer and revenue figures as attributed company statements, and watch for actual deployments and sustained capacity. AI compute needs more than one credible route, but the route has to be open when customers arrive.

What to watch

  • Whether AMD can deliver enough chips to meet growing customer demand.
  • Whether ROCm improvements continue to reduce support work for operators.
  • Whether Google TPU use expands beyond a small share of DoiT’s customer base.
  • Whether more customers shift workloads, rather than simply testing alternatives.

Discussion spark: Would you move an AI workload to AMD or Google for lower cost or more supplier choice if Nvidia remained easier to deploy, or is software convenience worth paying a premium for?

Sources and evidence

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.