Hardware for AI: Google's entry will increase competition in market

Meta is reportedly considering deploying Google’s custom tensor processing units (TPUs) in its artificial intelligence (AI) processing data centres in 2027, while renting TPU capacity in Google Cloud centres from 2026. Supplying chips to Meta (and other customers) marks a major shift in strategy for Google, which has hitherto used its TPUs only in its own data centres, while renting out time to companies. This positions Google as a rival to Nvidia, which currently has 80 per cent market share when it comes to the supply of AI-specific graphic-processing units (GPU).

 

However, Google’s recent launch of Gemini3 has served as a good advertisement for the capability of its TPUs, which were used to develop the new model. Initial reports say Gemini3 is better in some respects than rivals like ChatGPT, from OpenAI, and Claude, from Anthropic. Gemini has around 650 million monthly users, while the market leader, ChatGPT, has around 800 million but Gemini3 is rapidly catching up in terms of popularity. Meta is currently Nvidia’s biggest customer with a capex budget of $72 billion on AI infrastructure. If Google picks up a large chunk of that capex, it would be money that flows directly out of Nvidia’s pocket and the deal would be a validation for Google’s hardware ambitions.

 

The stock market reacted sharply to the announcement of the Meta-Google deal. The Nvidia stock was sold down by around 2 per cent while the Alphabet (Google’s parent) stock was bid up by roughly 8 per cent. The stakes are enormous. Some analysts reckon TPUs could quickly capture up to 10 per cent of Nvidia’s annual revenues, which amounted to over $165 billion in the financial year ended June. Nevertheless, Nvidia would remain the market leader. The other major player in the AI- GPU market is AMD. However, as demand for AI explodes, Google is not the only company looking to develop in-house hardware. There’s a long waiting-list for high-end GPUs, and Amazon and Microsoft, among others, are developing their own chips in the hope of faster deployment at lower costs. Despite being rivals, in a scenario that’s typical of big tech’s “frenemy” model, Google and Nvidia also cooperate on many projects and Google is a major customer for Nvidia’s GPUs.  Anthropic, which runs Claude, a rival to Gemini, is also exploring the possibility of deploying Google TPUs. Apple also trains its Apple Intelligence models on TPUs.

 

Google’s chips are very different from Nvidia’s. They are not like-for-like. Nvidia is a pure-play hardware company. The GPUs are designed for large number crunching with a software platform thrown in to help users to tailor codes efficiently. GPUs split tasks into many pieces and run the calculations side-by-side. They are more flexible than TPUs, which only do matrix mathematics for deep learning. But deep learning is a key research area and foundational to large-language models and at this specific task, TPUs may be more cost-effective. Google’s reputation and AI-branding come from its development of AI software and algorithms in services like “search”, “map”, and “translate” for many years and it started designing its TPUs in-house (in a collaboration with Broadcom) a decade ago. Its entry should provide some healthy competition and if other giants also get into the act, we could find a sea change in the AI ecosystem.