GPT-6 Astra draws AGI claims and squeezes the memory chip trade

Nvidia chief executive Jensen Huang declared that AGI has arrived following OpenAI’s release of GPT-6 Astra^1,3,4. Huang wrote on X that the model was trained on around 100,000 Nvidia Grace Blackwell NVLink72 server racks, with plans already in place to bring roughly four times that amount of hardware online^2,9.

OpenAI’s own executives went partway with him. President Greg Brockman characterised the model as potentially an early sign of artificial general intelligence⁷ and described its release as a significant threshold⁸, saying the model can navigate the web and write⁶. Chief executive Sam Altman called the rollout messy as enterprise users waited for access⁵.

The trade that actually moved

The market reaction was in memory rather than in models. SK Hynix shares jumped more than 7% in Seoul trading on Monday after the debut, tightening memory supply⁶, with a separate report putting the move at 8.3% as inventories fell below ten days⁹. Morningstar described the launch as having reignited the memory chip trade outright⁸.

The supply picture behind the move is tight. Samsung and SK Hynix have dropped below a ten-day supply as HBM4 devours capacity⁷. FinanceFeeds noted hyperscaler capital expenditure plans holding at $1.3 trillion through 2027⁹, which is the demand side of the same equation.

Why it matters to investors

The AGI label is contested and unfalsifiable in the near term; the hardware numbers are not. A training run on roughly 100,000 accelerator racks with a plan to quadruple that² is a capital commitment that flows directly into semiconductor order books, whatever the model turns out to be capable of.

The memory squeeze is the cleaner signal. Inventories below ten days at the two dominant suppliers^7,9 leave no buffer against a demand step-change, which is why a single product launch produced a single-day equity move of that size⁶. It also means pricing power sits with the memory makers rather than their customers for as long as the condition persists. The counterweight is execution: Altman’s own description of the rollout as messy⁵ is a reminder that capability announcements and deployed capacity are different things.

What to watch

Watch whether the larger training cluster Huang flagged is actually built², whether memory inventories recover above ten days⁷, and whether the hyperscaler capital expenditure path holds at the level reported⁹. On the product side, the test is whether enterprise access catches up with the consumer rollout Altman apologised for⁵.