Huang says Nvidia will sell twice as many chips next year
Two answers to the same question landed a day apart: how much compute is actually coming. One is a shipment forecast from the company selling most of it, the other is a hardware announcement from the one building it without American parts.
Jensen Huang says Nvidia will sell twice as many chips next year as it does this year, and he used the occasion to argue AI should be regulated as a product rather than a technology. Speaking to reporters before a summit with King Charles III in Scotland — attended by representatives from Google DeepMind, OpenAI and Anthropic — Huang tied the forecast to broad demand: "you can see that in almost every single country that we're in, people want to invest in AI," he said. It follows company guidance for 70% revenue growth in the fiscal year ending January 2028, or about $673 billion. Nvidia closed Thursday at $219.34 a share, up 2.54%, for a market capitalization around $5.29 trillion.
The footnote matters as much as the number. Nvidia has never disclosed total unit shipments, and "chips" here spans far more than data center GPUs — CPUs, switch silicon, optical networking parts, laptop and Jetson chips, and the processor inside Nintendo's Switch 2. Huang's only previous unit figure was 6 million Blackwell GPUs over four quarters, given last October. A doubling in units is not a doubling in revenue, and this was a remark to press, not updated guidance. The demand signal it corroborates is already visible in the secondary market, where Nvidia's B200 now resells for 58% above list — and lenders noticed.
His regulatory line is the argument the labs have been rehearsing for weeks. Huang told the summit that AI "is a technology, not a product," that the products are what should be regulated, and that when a product isn't safe "we should hold it back and keep engineering it." He has said the industry needs no new laws — that safety is an engineering problem, not a legal one. Demis Hassabis, the DeepMind co-founder in the room, offered the counterweight: "the probability of us making a mistake that leads to something catastrophic is definitely not zero," while saying he remains optimistic the problems are solvable. That split — engineering problem versus nonzero catastrophe risk — is now the industry's dividing line, and it maps almost exactly onto who is selling capacity and who is asking to be audited.
Huawei used the same week to show what its answer to export controls looks like in hardware: an Ascend 960 supernode it calls the first in the industry to use near-package optics, wiring 4,096 chips together with 5,500 of its Hi-ONE optical engines instead of conventional pluggable modules. Rotating chairman Wang Tao announced it at Huawei Connect in Shanghai, claiming 8 EFLOPS of FP8 compute (16 EFLOPS at FP4), a 7.2T-capacity optical engine with a built-in light source, unified memory addressing across the 4,096 chips, and system availability of 99.8%. Huawei says the 5,500 engines replace what would otherwise be more than 40,000 800G pluggable modules, cutting power draw by over 550 kilowatts and doubling time between failures. Supernodes link into clusters up to 512,000 cards, and with multi-track topology, up to a million.
The reported power and module counts come from Huawei's own briefing and are not independently verified. The strategic read is unchanged from yesterday's roadmap briefing — Huawei names its 2027 Nvidia challengers: the 960DT and Ascend 960PR — with one addition: the interconnect is where the engineering budget is going, because when you cannot buy the best chip, the only lever left is how tightly you can wire the ones you have. Pluggable optics are the part of a data center that fails and burns power, so moving them inside the package is the correct target. Whether it holds at a million cards is a claim, not a measurement.
What to watch: whether Nvidia puts a unit figure in its next earnings call, and whether any Ascend 960 supernode shows up in a customer's own power or availability numbers.
Does a chip-sales forecast from the seller settle the demand question, or do you want the buyers' numbers? Tell us in the comments.
