🧠 Food for thought

Implications, context, and why it matters.

Meta plans $70B+ in AI capex with unclear payoff; Google earns from AI cloud

• Meta raised its 2025 capex outlook to $70-72 billion and signaled even higher 2026 spend for compute and infrastructure 1. Shares fell 12.3% after the plan, despite strong earnings 2.
• Alphabet (Google’s parent company) said Google Cloud revenue rose 34% year over year to $15.15 billion in Q3 2025, which is direct AI monetization 3. A Meta shift of billions to Google tensor processing units (TPUs) would lift Google’s infrastructure business and invite fresh questions about Meta’s returns.

PyTorch teams could need TPU migration tools as big clouds look beyond Nvidia

• PyTorch (an open-source deep learning framework)/XLA (Accelerated Linear Algebra, a compiler that targets specialized chips like TPUs) can run on XLA hardware with small code changes 4. Moving from CUDA (Nvidia’s GPU programming platform) to TPU still needs device reassignment and careful handling of lazy tensor execution (queuing operations for deferred, optimized execution).
• The PyTorch/XLA project is shifting to code-generated operations (kernels produced from specifications rather than handwritten code) 5. vLLM (an open-source high-throughput inference engine for large language models) now offers experimental TPU support 6. GPU shops can switch backends with little config change. Providers can fill integration gaps with tooling, services, and training.