No longer solely betting on NVIDIA, Meta spends billions of dollars to lease Google TPU

Meta, while promising to purchase millions of NVIDIA GPUs, is also renting Google TPUs for billions of dollars, entering a new phase of "diversifying suppliers" for AI computing power.

On February 26, The Information reported, citing sources involved in the negotiations, that Meta has reached an agreement with Google to rent Google AI chips TPU to develop new AI models over the next few years, with the deal valued at "billions of dollars." Meta is also discussing the possibility of purchasing TPUs for its data centers as early as next year, but the progress of the negotiations remains unclear.

Rare Moves on the Training Side: Meta is Not Just Looking for Alternatives in "Inference"

Notably, the news indicates that Meta plans to use TPUs for AI training. This is more sensitive to the market: most opportunities to challenge NVIDIA are typically considered to be in the inference stage, rather than in training clusters that require more stringent interconnect scale and hardware-software ecosystems. Therefore, the original consensus in the market was that training was an area where NVIDIA GPUs excel, and TPUs were merely alternatives for inference.

The report also mentioned that Meta announced a large deal with another NVIDIA competitor, AMD, this week, but sources indicated that Meta primarily plans to use AMD chips for running existing models (inference), rather than training new models. Meta is also continuing to develop its own inference chips to reduce costs and further diversify risks.

"It's Not That We Don't Use NVIDIA, But We Can't Rely Solely on NVIDIA"

Shortly before the disclosure of the Meta-TPU deal, NVIDIA announced a new collaboration with Meta: Meta stated that it would purchase millions of GPUs for its data centers in the coming years.

Putting these two pieces of news together points to the same conclusion—Meta still cannot do without NVIDIA's training ecosystem, but is shifting more training and inference loads to "second choices" to reduce the uncertainty of being "choked by a single supplier."

One of the backgrounds driving Meta's shift is the slow progress of its self-developed AI training chips; another practical factor is that last year, some clients, including OpenAI and Meta, encountered " technical failures and hardware complexity" issues when deploying NVIDIA's latest Blackwell chips on a large scale.

Google's Calculation: Making TPU a "Multi-Billion Dollar Revenue" External Business

Insiders say that Google is intensifying efforts to compete directly with NVIDIA in the AI training chip market, with TPU sales expected to bring " additional billions of dollars in revenue" to Google.

Someone within Google Cloud suggested that if they could " supercharge the TPU business," they might capture a share equivalent to about 10% of NVIDIA's annual revenue; according to reports, NVIDIA's annual revenue over the past 12 months was approximately $200 billion Google's approach to externalizing TPU is becoming more "financialized." In addition to reaching a deal with Meta, Google has also signed an agreement with a large unnamed investment institution to fund a joint venture project that will lease TPUs to other clients; Google is also in discussions with other private equity firms for more similar joint ventures. According to insiders, Google has at least signed a term sheet with one large investment institution.

At the same time, Google's corporate development team is discussing financing through a "special purpose vehicle (SPV)" to purchase TPUs for external leasing, with TPUs potentially being used as collateral for debt. Reports liken this to the "creative financing" structure that xAI and venture capital firm Valor have established around NVIDIA GPUs.

The Biggest Variable: TPU Supply, TSMC Capacity, and the Balance of "In-House/Outsourced"

The ramp-up of TPUs does not solely depend on demand. Google needs to balance multiple objectives: on one hand, it is challenging NVIDIA at the chip level, while on the other hand, Google Cloud is a major customer of NVIDIA GPUs—most AI developers still prefer the GPU ecosystem, and Google Cloud "cannot afford not to provide" NVIDIA servers, or it will affect its cloud competitiveness.

The supply side is also tight. Google's own Gemini model team also requires TPUs; at the same time, both TPUs and NVIDIA GPUs are produced by TSMC, meaning that both are "competing for the same type of capacity" within TSMC's factories. This determines whether Google can quickly replicate Meta-style orders with more large clients.

Risk Warning and Disclaimer

The market carries risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investing based on this is at one's own risk