Meta Google TPUs instead of Nvidia: Explanation
The news that Meta is negotiating the use of Google's Tensor Processing Units (TPUs) has influenced the stock prices of Nvidia, Alphabet, and Meta. This indicates a shift in the AI accelerator market, as Google gains ground in the competition with Nvidia.
Source: Own representation
Meta & Google TPU Deal
Meta is negotiating a multi-year, billion-dollar deal with Google for Tensor Processing Units (TPUs), which could also run in Meta's own data centers starting in 2027. This is reported by Reuters . Until now, Google's TPUs were primarily available via Google Cloud as rented capacity. Now, it is being considered that Google will bring its AI chips as "Customer Premise Hardware" to third-party hyperscale data centers for the first time, as SwingTradeBot reports. It adds that Google is discussing the use of its tensor chips with Meta and other customers, which put pressure on Nvidia and AMD in after-hours trading. Investor’s Business Daily adds that Google is discussing the use of its tensor chips with Meta and other customers, which put pressure on Nvidia and AMD in after-hours trading.
If Meta, a major buyer of Nvidia GPUs for models like Llama, shifts part of its future capacity to Google TPUs, Nvidia loses some of its status in the high-end AI segment, according to datacentremagazine.com. In parallel, Meta has already concluded a six-year cloud contract worth over $10 billion with Google Cloud to procure additional AI infrastructure ( datacentremagazine.com). A few weeks later, Meta announced it would invest at least $600 billion in AI-optimized data centers, energy projects, and local programs by 2028 ( datacentremagazine.com). Meta does not want to base its massive AI investment program on a single chip supplier, and Google is seizing the opportunity to bring its TPUs to market.
TPUs vs. GPUs
Nvidia currently dominates the training of large language models with GPUs such as the A100 and H100 series ( Google Cloud Documentation). Google pursues a different approach with its Tensor Processing Units (TPUs). TPUs are specialized ASIC chips optimized exclusively for tensor computations in machine learning workloads ( Google Cloud Documentation). The official Google documentation describes TPUs as custom-built accelerators that are particularly well-suited for large matrix operations in models like Transformer networks (
A recent Google blog post summarizes the difference: CPUs are universal all-rounders, GPUs are massively parallel accelerators for graphics and AI, and TPUs are even more tailored for AI computations and run in Google data centers for services like Search, YouTube, and DeepMind models ( blog.google). Technically, benchmarks show that Google's TPU pods achieve high efficiency per watt and per dollar for certain training workloads ( YouTube). These metrics are crucial for players like Meta, which need to scale a fleet of AI systems for billions of users ( datacentremagazine.com).
Source: note.com
Google TPUs (Tensor Processing Units) and NVIDIA GPUs (Graphics Processing Units) are the leading architectures for AI workloads, each with specific strengths.
Google's current high-end generation for training, Cloud TPU v5p, is designed as a cluster of up to 8,960 chips per pod and offers significantly more high-bandwidth memory per chip than the previous generation (95 GB HBM vs. 32 GB for v4), as Google describes in its cloud documentation ( Google Cloud Documentation). A TechRadar-Artikel concludes that TPU v5p trains up to 2.8 times faster than TPU v4 and could be on par with or superior to Nvidia's H100 in rough comparisons. According to the company, Nvidia's H100 is up to four times faster than the A100 generation in training and offers up to 80 GB of HBM3 memory per GPU for AI workloads ( Google Cloud Documentation).
The question of which AI chip is better is context-dependent. For companies heavily invested in the Nvidia ecosystem, the H100 remains the obvious choice ( Google Cloud Documentation). For workloads tuned for TensorFlow and JAX running in Google Cloud, TPU v5p can offer advantages in throughput and cost per training epoch ( Wikipedia). Google optimizes entire data centers around TPU pods, including networking, storage, and scheduling, underlining the entire "AI Supercomputer" concept ( YouTube).
Source: YouTube
AI chip shortage
A report by Bain & Company warns that the AI boom could once again tip the balance in semiconductor supply chains. Describes how "runaway AI demand" combined with limited production capacity and new trade barriers is leading to persistent shortages of certain AI-related components and driving up prices. For data center operators, this means that capacity is not just a matter of budget, but also of lead times, tariffs, and geopolitical risks ( Sourceability describes how the "runaway AI demand" combined with limited production capacity and new trade barriers is leading to persistent shortages of certain AI-related components and driving up prices. For data center operators, this means that capacity is not just a matter of budget, but also of lead times, tariffs, and geopolitical risks ( sourceability.com).
Meta plans to invest at least $600 billion in AI data centers by 2028, but is also selling data center assets worth around $2 billion to utilize more flexible financing models for AI infrastructure ( datacentremagazine.com). The rationale: AI data center buildouts are so capital-intensive that even tech titans have to increasingly rely on leasing, co-location, and cloud partnerships ( datacentremagazine.com). A Meta deal for Google TPUs fits this picture. Instead of exclusively buying Nvidia GPUs, Meta can secure additional capacity via Google TPUs in the future, thus cushioning supply and price spikes ( Investors).
Impact on companies
For companies considering their own AI infrastructure, the Meta-Google move shifts parameters. A large corporation that has so far relied exclusively on Nvidia GPUs in a hyperscaler region can now plan a multi-sourcing model: training large models on TPUs in Google Cloud, inference on GPUs in another cloud or on-premises, combined with classic CPU compute for less time-critical workloads ( Google Cloud Documentation). On-prem TPUs will become an additional option, especially for highly regulated industries, if Google implements its plans ( Reuters).
Source: gigazine.net
OpenAI's decision to use Google's TPUs underscores the growing importance of alternatives to NVIDIA's GPUs and the diversification of AI hardware supply chains.
Start-ups are particularly affected by GPU shortages. Increased competition between Nvidia GPUs, Google TPUs, and other specialized chips like AWS Trainium or AMD Instinct could alleviate price pressure in the medium term ( Medium). For developer practice, the Meta-Google TPU move also means: More teams will be forced to keep their stack portable. Those who consistently build models and pipelines using frameworks like PyTorch/XLA, JAX, or well-abstracted serving layers can switch between GPU and TPU backends without having to rewrite everything ( Google Cloud Documentation).
Strategic dimension
The video "AI Hypercomputer with Cloud TPU v5p | Google Gemini" illustrates how Google is concentrating its own AI power on TPUs and is now gradually opening this platform to customers ( YouTube). In parallel, formats like "NVIDIA vs Big Tech : Who Wins The AI Chip War?" analyze the growing competition from hyperscalers who are developing their own chips, thereby aiming to take market share from Nvidia in the long term ( YouTube).
Source: user-added
Modern AI hardware like TPUs and GPUs requires sophisticated cooling systems to dissipate the enormous waste heat from the high-performance chips.
From the perspective of the entire AI ecosystem, the Meta-Google TPU deal is another step away from a world where one manufacturer alone dictates the rules for hardware, prices, and roadmaps. Google strengthens its role as a complete infrastructure provider, Meta increases its negotiating power and reduces dependencies, and other players like AWS, AMD, or specialized ASIC manufacturers will have to measure themselves against this new reality ( Medium).
The combination of Meta's multi-billion dollar investments in AI data centers and Google's strategic move to no longer use TPUs exclusively for its own needs marks a turning point in the AI chip race ( datacentremagazine.com). For Nvidia, this does not mean the end, but the end of its comfortable monopoly phase. For Google, it opens up the opportunity to establish its own AI hardware as a real alternative in the market. For Meta and other hyperscalers, it's a lever to better manage chip shortages, energy demand, and capital costs ( CIO).
For companies, startups, and developers, this means: The question of which AI chip is better will be answered less often with a name – and increasingly with a choice of architecture and strategy. Those who prepare for a hybrid, portable setup early on will have significantly more options in this multipolar AI hardware world than the current H100-dominated reality suggests.