Lilith.
⌕
Editorial illustration: The frontier pause lasted days while chip orders kept growing
Lilith illustration · editorial remix

Calls to slow the most capable AI models have not yet slowed the chip market. Nvidia reported $107.6 billion in data center revenue for the quarter ended October 25, up 110% year over year. AMD reported a record $8.5 billion, up 97%.

A ten day pause did not change infrastructure orders

OpenAI paused its largest planned frontier training run over model cyber capabilities and resumed it ten days later with added safeguards. The interruption helped trigger a sell-off in some chip stocks, yet supplier results show orders continuing to grow.

Nvidia began shipping Vera Rubin NVL72 systems. The company says the platform generated nearly $20 billion in its first quarter of production shipments. Nvidia expects major frontier labs to represent about 15% of data center revenue this fiscal year. AMD is shipping Helios systems with MI450 GPUs to customers including OpenAI and Anthropic.

Inference separates chip demand from any single training run

The chip market no longer depends only on a handful of giant training jobs. Running models, agents and enterprise applications consumes capacity continuously. Cheaper tokens can also stimulate more use: Nvidia said token prices had fallen almost 60% since May while usage increased.

For infrastructure buyers, the central risk is shifting. It includes safety stoppages, but also spending moving from general purpose GPUs toward cheaper custom accelerators. Counterpoint expects custom AI server processor shipments to surpass data center GPU shipments by 2028.

Quarterly results lag changes in purchasing decisions

Current revenue reflects contracts and procurement arranged months in advance. Counterpoint analyst Neil Shah estimated that a real slowdown would take about six months to appear in chipmakers' results. One strong quarter cannot rule out a later correction.

The safety pause also lasted only ten days. These figures do not test a scenario in which labs curb frontier training for several quarters. They show that a brief incident and more cautious rhetoric did not break the existing investment cycle.

Custom accelerators will show where the spending moves

The next quarters should be read through TSMC orders, GPU rental prices and the share of custom accelerators at hyperscalers. If inference keeps growing while general purpose GPUs lose share, demand for compute may remain strong even as margins and winners change.

Longer safety interventions provide the second signal. A pause measured in months, cancelled capacity reservations or a delayed data center would finally test whether laboratory warnings have reached procurement budgets.

Lilith's verdict

Labs lifted their foot for ten days while data centers kept ordering truckloads of chips. A real brake will show up in cancelled capacity, not one nervous day on the market.