Google's TPU push threatens Nvidia's 86% AI chip market share

Google's custom Tensor Processing Units could cut AI compute costs by 30%, threatening Nvidia's pricing power in a market where it controls 86% of data center chip revenue.

Google's in-house Tensor Processing Units can handle AI workloads at up to 30% lower cost than rival processors, threatening Nvidia's dominance in a market where it commands an 86% share of data center revenue, according to company disclosures and analyst estimates.

"Custom processors designed for specific AI models deliver meaningful cost advantages at scale," Sundar Pichai, chief executive officer of Alphabet, said on the company's most recent earnings call.

Alphabet is spending as much as $190 billion in capital expenditures this year, with management expecting a "significant increase" in 2027. The company announced a joint venture with Blackstone to deploy 500 megawatts of TPU capacity by 2027, with plans to scale further. Google also plans to rent some of that capacity to other tech companies through a neocloud model — a market that could capture 20% of total AI cloud spending by 2030, according to industry estimates.

Nvidia's gross profit margin of about 74% reflects pricing power built on years of limited competition in AI accelerators. If Google's TPU rental business captures meaningful share, it could compress those margins and reshape the competitive landscape for the $50 billion-plus AI chip market.

The Custom Chip Wave Is Building

Google is not alone in building proprietary AI silicon. OpenAI unveiled Jalapeño, its first custom inference chip built with Broadcom, on June 24. The chip moved from design to tape-out in nine months and targets deployment by late 2026. Amazon is exploring direct sales of its Trainium processors to third-party data centers, while Microsoft's Maia 200 accelerator already powers GPT-5.2 models on Azure.

The structural shift matters more than any single product. When every major AI customer designs its own hardware, Nvidia's negotiating position weakens. The company's CUDA software ecosystem and high-speed interconnects remain competitive advantages in training workloads, but inference — the stage where trained models generate responses for users — is where custom chips can deliver the fastest cost savings.

DeepSeek, the Chinese AI startup valued at more than $50 billion after its June funding round, is also developing its own inference chip, according to Reuters. The project targets inference rather than training, a decision that reflects both the cost sensitivity of large-scale AI deployment and the constraints of U.S. export controls on advanced manufacturing tools.

What It Means for Investors

Nvidia shares, trading at roughly 35 times forward earnings, fell 1.6% in premarket trading on July 7 after Reuters reported DeepSeek's chip plans. The broader selloff in memory stocks — Micron Technology dropped 4.7% and Western Digital fell 6.3% — suggests investors are beginning to price in a more competitive chip landscape.

For Alphabet, the payoff is twofold: lower internal AI compute costs that improve margins on its Gemini model and cloud business, plus a new revenue stream from renting TPU capacity to third parties. The company's $190 billion capital expenditure plan implies confidence that custom silicon will deliver returns that general-purpose GPUs cannot match.

The risk for Nvidia is not that any single custom chip outperforms its best GPU on every metric. It is that the cumulative effect of dozens of customers building their own alternatives erodes the pricing leverage that has made Nvidia the most valuable semiconductor company in history.

This article is for informational purposes only and does not constitute investment advice.