Google said to collaborate with Meta to erode Nvidia's software

Google is developing a project called TorchTPU to make its Tensor Processing Units (TPUs) more compatible with PyTorch, the widely used AI software framework, according to people familiar with the matter.

The effort aims to help AI developers use Google’s chips without needing to rewrite code for Nvidia’s CUDA ecosystem, which dominates the market.

Google, which has relied on its own Jax framework internally, is seeking to attract more external customers to its TPUs by addressing compatibility issues with PyTorch.

Sources said Google may open-source some of the software to drive adoption.

Meta, which supports PyTorch, is working with Google on the project and has discussed gaining greater access to TPUs.

Google began selling TPUs directly to customer data centers in 2025 and named Amin Vahdat as head of AI infrastructure this month.

A Google Cloud spokesperson said the company is focused on giving developers flexibility across different hardware options.

Meta declined to comment.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

PyTorch-on-TPU maturity will determine if Google can dent Nvidia’s AI compute monopoly

TorchTPU aims to run PyTorch on Tensor Processing Units (TPUs) with less effort, since PyTorch Accelerated Linear Algebra (XLA) remains harder than Nvidia’s CUDA 1.
They bring 4x better inference cost than H100s 2 and 50-70% lower training costs 3. That could unlock share in the $255 billion inference market growing 19.2% annually 2.
Analysts expect a rental phase using hundreds of TPU pods (racks of networked TPUs) from 2026, then on-premises from 2027 4. CUDA still looms, so teams must revalidate kernels and retrain 4. The next 2 to 3 years will decide outcomes 5 1.

System integrators can capture TPU deployment revenue as Google opens on-premises sales

Direct TPU sales to customer data centers start in 2025. VAST Data, a data storage platform vendor, offers TPU integration with hybrid GPU clusters through Google Cloud Marketplace, plus reference architectures and joint validation 6 7.
System integrators (IT services firms that design and deploy infrastructure) should focus on the 2 to 6 month migration 2. Services include PyTorch code porting with the XLA toolchain 1, custom cooling and networking for on-premises TPU pods 4, and hybrid orchestration linking TPU clusters with existing GPUs 7.
Partners can cut risk with proofs of concept and cost models that compare $0.54 to $4.20 per chip hour 8 against H100 options. Turnkey packages position them for the buildout 4 as buyers diversify beyond Nvidia’s 80% share 2.