Apple's Ternus calls the memory shortage 'absolutely a crisis'
Two days after Anthropic put a number on the AI trade, Apple's new chief executive put a number-free warning on the supply chain underneath it — and pointedly refused to rule out the option Washington least wants him to take.
John Ternus says Apple will not take Chinese memory chips off the table, calling the global shortage "absolutely a crisis" in an interview with TIME published Wednesday. Ternus, who replaced Tim Cook as chief executive on September 1, was responding to a bipartisan group of senators who wrote to Apple after the Financial Times reported in July that the company had begun testing memory chips from a Chinese supplier. "The constraints in memory are absolutely a crisis," Ternus tells TIME. "So our fundamental belief is that all options should be on the table." That is a notable piece of pushback from a chief executive two months into the job, from a company that has historically absorbed component cost increases rather than say much about them publicly — and it lands the same week the memory shortage reshaped phone pricing and memory makers reported sold-out capacity we covered the DRAM squeeze in September. Ternus also went out of his way to keep Cook, now executive chairman, in the frame: "He has great relationships and loves Apple. So wherever he can help, he is happy to help."
The subtext matters more than the quote. Apple's problem is not that memory is expensive — it is that the AI buildout has pulled DRAM and NAND capacity to hyperscalers at prices Apple's hardware margins cannot easily absorb, and the alternative suppliers that could add supply are the ones Congress is threatening Apple over. "All options should be on the table" is not a sourcing decision. It is Apple telling the Senate that its choice may be Chinese memory or delayed products, and that it has not agreed to the former being off-limits.
Huawei disclosed the engineering numbers behind its first logic-folded chip, and the headline figure is a 30-billion-parameter model running on the device in 2.5GB of memory. At a two-day technical briefing in Shanghai, Huawei semiconductor executives said the Kirin 9050 Pro uses more than 50 million hybrid bonding connections at a 1.5-micron pitch — needed because the chip's top metal layer has a 720nm pitch, so the bonding must stay under 2 microns to keep the ratio workable. Logic folding, which stacks circuit layers and joins them with short vertical connections, lifts CPU clock speed 12.7 percent, GPU performance 40 percent and NPU throughput 140 percent, with Huawei claiming energy-efficiency gains of 41 to 66 percent. Those performance figures are Huawei's own; the only independently verified element so far is the stacking geometry itself, confirmed by a teardown. We covered the first logic-folded part shipping this morning — Huawei ships its first logic-folded chip — and the briefing adds the detail that was missing.
The number that should interest developers is the memory figure. Huawei says a quantisation scheme combined with staged loading of model experts cuts memory use by 75 percent, letting a 30B mixture-of-experts model run locally on 2.5GB. If that holds up in third-party testing, it is a direct answer to the objection that killed on-device models this year: memory is expensive and phones are not. Huawei also said the next-generation "Kirin 2027" has been received and that all its test metrics met expectations, which is a roadmap claim no one outside the company can check yet.
Mueon, a Portland startup building modular compute blocks it calls Cubelets, raised $65 million in a Series A led by Maverick Silicon, with Intel Capital participating. The company said the round brings total funding past $80 million after a $15.5 million seed in 2025, and that the money moves Cubelets from platform development toward system validation and production readiness. Mueon's pitch is that compute, memory, power delivery and thermal management should be stacked into three-dimensional blocks rather than bolted into a standard 19-inch rack, and it is targeting up to 10x gains in power and cost efficiency in the same footprint, scaling toward megawatt-class density.
Treat the 10x with more suspicion than the funding round. Mueon has published no benchmark, no measured throughput, and no shipped product — the claim has been repeated across two funding announcements without data behind it, and even the phrasing drifts between rounds. Intel Capital's participation is the more interesting signal: the same investor led the seed, which means Intel has now written two cheques into the bet that the rack, not the chip, is the next bottleneck.
What to watch: whether Apple says anything binding about Chinese memory suppliers once the Senate letter gets a formal response, and whether anyone outside Huawei runs the 30B-on-2.5GB claim on a retail device.
If the memory crunch forces a choice between buying from a Chinese supplier and delaying products, which should Apple pick — and should Washington get a veto? Tell us in the comments.
