Report
Pierre FerraguÊ

The cluster economics of Nvidia GPU vs. Google TPU: How different? Implications for market structure?

Google is deploying billions of dollars of TPU, followed by Amazon deploying Trainium and Inferentia, and the chip development initiatives of Meta and Microsoft are widely reported as well. The phenomenon goes beyond hyperscalers, with Tesla bringing online the in-house “Dojo” supercomputer, and start-ups like Cerebras. Everybody wants to develop an alternative to Nvidia for datacenter A.I.
Leveraging recent disclosures by Google, we analyzed the TCO (total cost of ownership) of a TPU cluster vs. an Nvidia cluster. We find that the former can indeed be 4-6x better than the previous generation of Nvidia chips, but probably in the ballpark of what can be achieved with the H100, the latest generation being rolled out today.
This means being successful with an alternative in-house architecture requires being able to keep up with the pace of innovation of Nvidia, in addition to being able to manage the reduced flexibility and more limited ecosystem support implied by the use of a more specialized architecture.
Our conclusion: Alternatives to Nvidia GPUs will continue to develop, but will remain a minority of the market. Recent newsflow and current growth dynamics seem to support our view. This doesn’t mean next-gen A.I. chips won’t do well… we see them as a $5-10bn market in 2026, up from $1.3bn in 2021!
Underlying
NVIDIA Corporation

NVIDIA engages in graphics processing unit (GPU)-based visual computing and accelerated computing platforms. The company has two segments, GPU and Tegra Processor, which are based on a single underlying architecture. The company's GPU product brands are aimed at markets including GeForce for gamers; Quadro for designers; Tesla and DGX for artificial intelligence data scientists and big data researchers; and GRID for cloud-based visual computing users. The company's Tegra brand integrates an entire computer onto a single chip, and incorporates GPUs and multi-core CPUs to drive supercomputing for autonomous robots, drones, and cars, as well as for game consoles and mobile gaming and entertainment devices.

Provider
New Street Research
New Street Research

Provided by our team of experienced analysts, our work is idea driven, based on independence of thought, sector expertise, and firmly focussed on fundamentals and valuation.

New Street Research is an independent, partner-owned, research firm specialising in equity and debt research. Our equity research embraces the following sectors:

  • Pan European Telecom Services and Cable
  • US Telecoms, Cable, Satellite and Towers
  • Global Emerging Market Telecoms
  • Japan Telecoms
  • Asian Internet
  • Pan European Healthcare
  • Global healthcare thematic research

We provide debt research on:

  • Pan European Telecom Services and Cable


Analysts
Pierre FerraguÊ

Other Reports on these Companies
Other Reports from New Street Research

ResearchPool Subscriptions

Get the most out of your insights

Get in touch