SK Hynix ties its US NAND fab to a China export license

Share
SK Hynix ties its US NAND fab to a China export license

Washington wants memory fabs built on American soil, and Seoul's biggest memory maker wants something in return. On the same day, ByteDance went shopping for the non-Nvidia half of its chip plan.

SK Hynix is asking Washington to approve exports of US chipmaking equipment to its plant in Dalian, China — and making construction of a new US factory conditional on that approval, two people familiar with the matter told Reuters. Solidigm, SK Hynix's US unit, wants to install new NAND production equipment in unused space at the Dalian site, known as Mod B, to lift output of the flash that goes into data-center SSDs, according to the report. The Dalian fab runs tools from Lam Research, Applied Materials and KLA. There is no public record of the request: neither Solidigm's pursuit of the license nor its decision to link a US NAND fab to the Chinese one had been previously reported. SK Hynix did not immediately respond to Reuters, Solidigm declined to comment, and the Commerce Department did not respond.

Why it matters: this inverts the onshoring bargain. The Trump administration is pressing SK Hynix, Samsung and other foreign chipmakers to build plants in the US, and Commerce has said it will license existing China fab operations while declining to approve equipment for capacity expansion or technology upgrades — its December license for Dalian capped equipment quantities. So the ask is for exactly the category Washington said it would refuse, offered as the price of a US plant. If it is granted, the US subsidizes domestic capacity while simultaneously enabling more output in China; if it is refused, the US fab plan may simply stall. The timing explains the leverage: memory is the tightest link in the AI buildout, and the squeeze that started in DRAM and HBM reached flash this year. Solidigm has been weighing what would be the first US NAND fab since September — we covered that deliberation when it surfaced — Solidigm weighs the first US NAND fab as the memory squeeze spreads.


ByteDance is close to an agreement to buy custom data-center chips from Taiwan's MediaTek, in talks that would make it MediaTek's second major customer for such ASICs after Google, according to Bloomberg. The chips are application-specific designs tailored for ByteDance's own data centers rather than off-the-shelf accelerators, and unit volumes were not disclosed. MediaTek is best known for phone silicon; it has spent the past year pushing into data-center custom chips, and Nvidia took a stake in the company in August specifically to keep that lane inside its own rack fabric — Nvidia pays $3.5B for MediaTek's custom-silicon handshake.

The interesting part is the sourcing map underneath. ByteDance is the world's most aggressive buyer of Nvidia hardware and has been told by Beijing to file its Nvidia purchase plans, a squeeze we covered at the time — Beijing tells Alibaba and ByteDance to file Nvidia purchase plans. A Taiwan-designed ASIC is a different export-control conversation than a top-bin Nvidia part, and a Chinese hyperscaler that designs its own accelerator and rents the manufacturing is a customer Western export rules are much less able to reach. If MediaTek signs, the custom-silicon wave stops being an American hyperscaler story.

What to watch: whether Commerce answers the Solidigm license request before the year-end license window closes, and whether ByteDance's MediaTek deal gets a unit-volume number that would tell us how serious the substitution really is.

Would you trade approval for a US memory fab for a license to expand in China? Tell us in the comments.

Read more

Huawei ships its first logic-folded chip — and bonding is the gate

Huawei ships its first logic-folded chip — and bonding is the gate

Huawei has started mass production on a phone chip that engineers its way around the lithography wall instead of through it, and in doing so it has quietly moved the competitive bottleneck from transistor density to how tightly two silicon wafers can be glued together. Huawei's Kirin 9050 Pro is the first mass-produced logic-folding chip, stacking multiple dies and wiring them with more than 50 million hybrid bonds. The design puts circuits that would normally route sideways on a flat plane in

Alibaba's Qianwen Office ships an e-commerce agent for merchants

Alibaba's Qianwen Office ships an e-commerce agent for merchants

The enterprise-agent race is moving from generic chat to verticals with a measurable invoice. Alibaba picked commerce first — and timed the launch days before Singles' Day pre-sales. Alibaba's Qianwen Office opened an invite-only beta of an e-commerce edition on October 14, and says it wants the full rollout inside one to two weeks. Merchants reach it by switching the assistant into a dedicated commerce-operations mode. CEO Chen Yusen frames the product as an independent, third-party commerce

Flock's CEO says the backlash shouldn't pause the cameras

Flock's CEO says the backlash shouldn't pause the cameras

Flock's chief executive went on stage with a protest outside the door and said the quiet part about surveillance accountability: his company has never cut off a customer. A few blocks away from that panel, OpenAI's trade-secrets fight with Apple produced an injunction and a counter-allegation on the same afternoon. Flock CEO Garrett Langley used his TechCrunch Disrupt panel in San Francisco on Wednesday to argue that privacy trade-offs are the price of a safer society — and to admit his compan

Today in AI — October 14, 2026

Today in AI — October 14, 2026

A day the money wobbled: Barclays told clients enterprise AI isn't showing up in the hardware numbers yet, Dell dropped 11%, and the same questions followed OpenAI into eight state courthouses and two EU probes. Meanwhile a 35B model learned to run on a phone, and the biggest Starlink story of the day was about who actually controls it. Models & Research * A 35B model now runs offline on a phone in about 3GB of memory. The Chinese lab ModelBest and OpenBMB open-sourced MiniCPM-V 4.7, whose