Microsoft and Google scramble for scarce AI memory as HBM crisis deepens

This piece crystallizes a structural reality: in 2025, the AI race is as much about memory as it is about model architecture. HBM and cutting‑edge DRAM are now strategic chokepoints controlled by a tiny oligopoly of suppliers. If Microsoft and Google are indeed sending senior executives to South Korea with a ‘secure capacity or lose your job’ mandate, it shows just how quickly commodity components have become board‑level geopolitical assets.

For AGI timelines, constrained memory changes the optimization problem. Labs can’t just throw more GPUs at the frontier; they must design models and training runs around a hard memory budget that won’t ease until at least 2027. That will push innovation in memory‑efficient architectures, sparse and mixture‑of‑experts designs, and algorithmic tricks that trade compute for bandwidth. It also raises the value of vertically integrated players—think NVIDIA or potentially in‑house chip efforts—that can co‑optimize silicon, packaging and model design.

The broader implication is that scaling laws are colliding with supply chains. Whoever locks in HBM for the next 2–3 years will have disproportionate ability to train frontier models and serve high‑context inference, while laggards will be forced into smaller, narrower systems.