Nvidia shares fall as Meta eyes Google’s AI chips

Nvidia’s shares fell after reports that Meta was in talks to use Google’s AI chips, called tensor processing units (TPUs), in data centers from 2027.

Meta, one of the world’s largest spenders on data centers and AI, may also rent Google chips through its cloud services in 2026, according to the Information.

Alphabet, Google’s parent company, saw its stock rise as much as 2.7% in late trading, while Nvidia dropped 2.7% at one point.

An agreement could help establish TPUs as an alternative to Nvidia’s chips, which currently dominate the AI hardware market.

Google previously agreed to supply up to 1 million TPUs to Anthropic, an AI startup.

Analysts say Meta’s potential use of TPUs signals large companies may look beyond Nvidia for AI accelerators.

Nvidia remains the leading provider of chips for AI development.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Google is pitching tensor processing units (TPUs) for on premises use, a shift from cloud-only access, and Meta is discussing 2027 integration; AI platform teams must judge if gains justify adopting Google’s proprietary stack 1.
TPU workloads depend on the Accelerated Linear Algebra (XLA) compiler to turn framework code into machine code 2. Dynamic tensor shapes cause trouble because XLA targets fixed shapes, so changes can trigger recompiles or halt execution, and many models need engineering to use fixed shapes 2.
Meta’s interest signals buyers are exploring alternatives to Nvidia, while Google Cloud execs say TPU uptake could capture up to 10% of Nvidia’s revenue; that hinges on easing PyTorch/XLA migration hurdles that limit some PyTorch-based production workloads 1 2.
PyTorch/XLA, the bridge that links PyTorch code to Google’s XLA compiler, often forces engineers to hand-tune code 3. Teams swap device-specific APIs, trim logging on TPU tensors (the multi-dimensional arrays used in deep learning), then adapt data loaders. That drives demand for automated migration tools.
Independent software vendors and consulting firms can offer services with XProf, a TPU performance profiler 2 4. These services spot code that triggers recompilation on TPUs, ship optimization packages that reshape tensor dimensions to multiples of 128.
Meta and other hyperscalers (large cloud plus internet companies that run massive data centers) are evaluating TPU adoption. Machine learning operations (MLOps) platforms can ship benchmarking suites that compare PyTorch performance on Nvidia GPUs versus TPUs to guide IT procurement before market consolidates.