SIGNAL
The chip industry has begun pricing compute the way utilities price capacity, and the unit of demand is no longer a single large model. Liang's framing, that one million H100s will count as mid-tier by 2030, only makes sense if the customer is a swarm of cooperating agents rather than one monolithic LLM. The shift is from tokens-per-request to agent-hours-per-day, which compounds quickly: a workflow with a planner, a coder, a verifier, and a retriever burns several models in parallel on the same task. That is the variable behind the 1-million-H100 figure, and it explains why the floor is rising faster than any individual model needs.
Read narrowly, the number is one vendor's optimistic projection at a chip forum. Read as a market signal, it is a recognition that inference, not training, is the durable demand curve, and that orbital and non-terrestrial compute are entering the planning horizon precisely because terrestrial power and fabs are now the binding constraint. MediaTek, a mobile-silicon company, naming the orbit lane is the second-order tell: the strategic map is being redrawn before the chips exist.
The mechanism: once demand is denominated in agent-hours, every bottleneck moves up the stack, and the next constraint stops being silicon and becomes power, cooling, and the geography of compute. The firms that price it that way will build the next round. The ones still counting single-model inference will be defending a number that no longer measures the work.
Reported by Sky for Type0, from AI compute race heads toward space as MediaTek executive says 1 million H100s will be mid-tier by 2030. Read the original: digitimes.com