Ask Eve

Episode 338 min

TSMC Outside Taiwan, China's Chips & HBM

David picks up the chip thread: how far TSMC has moved production out of Taiwan, what China would actually get if it invaded, why Arizona, who else is building fabs, how China is doing on chips and fabs, and why HBM memory is hard to make.

10 correctionsFact-checked,

Cover art for TSMC Outside Taiwan, China's Chips & HBMAsk Eve
Show notes

Context

Recorded on Wednesday, November 11, 2026, in a Grok voice conversation, picking up the conversation about Taiwan's chip industry from earlier that day.

Topics

  • TSMC in Arizona, Japan and Germany
  • What China would get if it invaded
  • Why Arizona
  • Who else is building fabs
  • China's chips and fabs
  • Why HBM memory is hard

Episode details

Recorded
Nov 11, 2026
Length
8 min
Turns
12
Transcript notes

Note. Edited for clarity from the shared Grok voice conversation. Every cut is marked […] and its omitted text is shown in an omission note beside it; words added for readability are in [brackets]; fillers, false starts and speech-recognition errors are fixed silently. Nothing else was added or changed.

Shareable Video Audiograms

Transcript

Tap any line to play from there. The current line follows along as you listen.

Edited for clarity. Cuts are marked […]: tap one to read what was cut. Words in [brackets] were added for readability. Nothing else was added or changed. Fillers and speech-recognition errors are fixed without a mark.

  1. David0:00

    Yeah, how's the Taiwan chip diversification looking?

  2. Eve0:04

    […] It's real but slower than the headlines suggest, and the pattern is deliberate. Arizona's the headline — two hundred sixty-five billion committed across ten planned fabs, but right now exactly one is producing: four-nanometer since late twenty twenty-four, ramping toward thirty thousand wafers a month. Fab two, the three-nanometer one, just finished construction with tools moving in, and volume starts third quarter twenty twenty-seven, slightly ahead of schedule. Fab three for two-nanometer is being roofed. Everything beyond that is dirt or paper.

    The catch that matters for the Taiwan conversation: every advanced package still gets assembled back on the island. Blackwell wafers made in Arizona ship to Taiwan for CoWoS, and TSMC's own Arizona packaging plants have no production date. Amkor's Peoria plant starts packaging those wafers in early twenty twenty-eight. So even "made in America" chips are a round trip. When the whole Arizona buildout's done, TSMC expects about thirty percent of its two-nanometer-and-better capacity there — seventy percent stays home, along with all the research.

    Japan's Kumamoto fab one has been running since twenty twenty-four on mature twenty-two and twenty-eight nanometer nodes. Fab two just got Japanese government approval in October to switch from six-nanometer to three-nanometer — Japan's first leading-edge line — but starting at only six thousand wafers a month because EUV lithography tools are scarce. Production's December twenty twenty-eight.

    Dresden's ESMC joint venture with Bosch, Infineon, and NXP is on schedule for late twenty twenty-seven, ten billion euros, twelve-to-twenty-eight nanometer automotive chips. Roof's nearly watertight, first employees arrive summer twenty twenty-seven.

    The logic underneath all three: the newest process always starts in Taiwan and migrates outward later. Diversification is insurance, not replication. […] […]

    Cut (tool narration)

    Checking TSMC diversification progress.

    Cut (unsourced claim, host request)

    Not shown, at the host’s request.

    Cut (unsourced claim, host request)

    Not shown, at the host’s request.

  3. David2:01

    What does it look like if China tries to invade Taiwan? Would they actually get the fabs or would the people and technology just leave and go to the US?

  4. Eve2:11

    Mostly the latter. The fabs are buildings full of machines that only work if the right people show up. EUV lithography tools from ASML need constant calibration by engineers who trained for years, and the supply chain for spare parts, chemicals, and photoresists runs through Japan, the Netherlands, and the US. If China takes the island by force, they'd inherit facilities that stop producing within weeks without that human layer […]. The silicon shield cuts both ways: it makes Taiwan too valuable to destroy, but it also means the fabs are the first target and the first thing that breaks. What China would actually capture is real estate and maybe some wafer stock — not the capability. The capability walks. That's exactly why the diversification push matters: every fab that comes online in Arizona or Kumamoto is one less reason for Beijing to think taking the island gets them the crown jewels.

    Cut (unsourced claim, host request)

    Not shown, at the host’s request.

  5. David3:04

    Why Arizona?

  6. Eve3:06

    Water, power, and politics. Tucson has a deep aquifer and cheap land, and the state government threw in billions in incentives plus a fast-tracked permitting process that no other US state could match. There's also the strategic angle — putting leading-edge capacity inside the continental US, far from any Pacific conflict zone, is exactly what Washington wanted. The tradeoff is that Arizona's summer heat means massive cooling costs, and the water table is already stressed by agriculture and growth, which is why TSMC's using recycled water for most of its fab needs.

  7. David3:41

    Who else is building real fabs now?

  8. Eve3:44

    Intel is the big one. They're putting over a hundred billion into Ohio, Arizona, Oregon, and New Mexico — the Ohio site in Licking County is the marquee one, two fabs on a thousand acres, though the timeline's slipped from twenty twenty-six to twenty twenty-seven or later. Samsung's Taylor plant in Texas is already producing four-nanometer, seventeen billion spent, and they shelved a second Texas site to focus there. Micron's doing fifteen billion in Idaho for memory plus a New York expansion. Texas Instruments is building a thirty-billion complex in Sherman, Texas. SK Hynix dropped almost four billion in West Lafayette, Indiana for HBM memory. GlobalFoundries is expanding in Malta, New York. And Rapidus in Japan is trying to stand up a two-nanometer line with IBM's help — that's the one to watch because if it works, Japan's back in the leading-edge game for the first time since the nineties. The pattern across all of them is the same: government money is the catalyst, and every one of these projects is years behind its original schedule.

  9. David4:45

    What about China? How are they doing with chips, fabs?

  10. Eve4:50

    […] China's the paradox of the industry. SMIC's producing at seven-nanometer class using deep-UV multi-patterning with no EUV machines, and they've pushed to a five-nanometer-class node called N-plus-three — but yields are brutal, under thirty percent in some estimates, and TechInsights found the Kirin nine-oh-three-oh in Huawei's Mate eighty is likely being made at an operating loss. Huawei's shipped over a hundred million Kirin chips since twenty twenty-three and just unveiled the Kirin nine-oh-five-oh with a new LogicFolding architecture using hybrid bonding across stacked dies. On AI chips, TrendForce projects China's domestic share of its own AI market hits fifty percent this year, but in absolute volume they're still only producing one to two percent of what the US makes. The real story is the domestic equipment push — SMEE's lithography machines now have over ninety percent localized parts and have processed ten million wafers, and Yuliangsheng is planning twelve domestic advanced DUV systems by year's end. SMIC's capacity at seven-nanometer and below is around sixty thousand wafers a month this year, heading to eighty thousand in twenty twenty-seven. By volume China's self-sufficient in mature nodes at about seventy percent, but by value it's still only a third. The bottleneck that actually limits Huawei's Ascend output isn't the fabs anymore — it's HBM memory, which they can only get from stockpiles bought before the twenty twenty-four controls.

    Cut (tool narration)

    Let me pull the latest on China's chip push.

  11. David6:18

    What's the tech challenges for making HBM memory? Is that just as hard as the GPUs, or is the tech relatively simple?

  12. Eve6:26

    HBM is harder than it looks but different from GPUs. A GPU is about logic — billions of transistors doing math in parallel, and the hard part is design and yield at extreme transistor density. HBM is about precision assembly. You're stacking eight to twelve DRAM dies, each thinned to maybe thirty microns, connected by thousands of through-silicon vias with micro-bumps at pitches under ten microns. One misaligned bond and the whole stack is scrap. Thermal is brutal too — that stack sits right next to a power-hungry GPU, and heat has to escape through a package that's only a few millimeters thick. Yield multiplies with every layer, so a twelve-high stack means twelve good dies or nothing. And then you have to integrate it with the GPU using CoWoS-style advanced packaging, which is its own bottleneck — TSMC's CoWoS capacity has been the single biggest constraint on AI accelerator supply for two years running. SK Hynix got there first with HBM3E for Nvidia and still holds roughly half the market. Samsung and Micron are close but not quite at the same yield. So it's not simple, but it's a different kind of hard — manufacturing and materials versus architecture.

TSMC Outside Taiwan, China's Chips & HBMEpisode 33
0:00 7:40