PaLM and the 540B Frontier

By Satwik · February 20, 2026

PaLM, Google's Pathways Language Model, was a 540B-parameter dense transformer trained across thousands of accelerator chips using the Pathways system. It was one of the largest densely activated models publicly described at the time and served as a demonstration of both engineering scale and capability breadth.

What made PaLM notable was not size alone but performance discontinuities. On a range of reasoning and language tasks it showed sharp gains, and it became a common reference point in discussions of so-called emergent behavior - tasks where performance stayed near chance until a scale threshold, then jumped. PaLM was also the vehicle for early chain-of-thought results, where prompting the model to reason step by step unlocked multi-step arithmetic and commonsense performance.

The security angle sits in two places. First, capability opacity: a model whose behavior changes abruptly with scale is hard to red-team ahead of deployment, because the dangerous capability may not be visible in a smaller proxy. Second, infrastructure. Training at this scale concentrates enormous value into a single artifact and a single pipeline, making model weights and training infrastructure high-value targets and raising the stakes of supply-chain and insider risk.

PaLM stayed largely internal, accessed through controlled research channels rather than open release. That gating was itself a governance choice. For us the lesson is that frontier capability and frontier access control arrived together, and the debate over who can run a 540B model foreshadowed the release-policy arguments that dominated the rest of the year.