Chokepoint | Pacific AI, Chips & Capital @ChokepointDesk

Source: Bernstein HBM estimates via Investing.com. These are regression estimates from Korean export data, not reported results. Samsung's figure came in 10% under Bernstein's own forecast, while hynix's matched it: investing.com/news/stock-mar… The currency side of the print: x.com/ChokepointDesk…

Chokepoint | Pacific AI, Chips & Capital @ChokepointDesk

8 hours ago

The won just hit a two-year high against the dollar. For SK hynix's print on Tuesday night, that may matter more than any HBM headline. Hynix sells in dollars and reports in won. A stronger won shrinks the same dollar sale, so a won-only read can turn a beat into a miss. The

1 0 0 10 0

Bernstein's export-data model has Samsung passing SK hynix in Q3 HBM revenue, $8.7B to $7.6B. That would be the first lead change of this AI cycle. Hynix reports first: Tue 8pm ET (Wed 9am KST). Which line on the call matters most? NFA.

Sources: Seoul FX close and dealer outlook, Etoday, Oct 23: etoday.co.kr/news/view/2639… · Overnight NDF, Yonhap Infomax: real.einfomax.co.kr/v2/datafeed/ks… · July/Sept FX levels, Yonhap Infomax: news.einfomax.co.kr/news/articleVi… · FnGuide consensus now, Financial News, Oct 23: fnnews.com/news/202610230… · Late-Sept consensus, iNews24: inews24.com/view/tp/2008989 · Hana FX assumption, Maeil Business: mk.co.kr/news/stock/121…

The won just hit a two-year high against the dollar. For SK hynix's print on Tuesday night, that may matter more than any HBM headline. Hynix sells in dollars and reports in won. A stronger won shrinks the same dollar sale, so a won-only read can turn a beat into a miss. The numbers: 1. USD/KRW closed Seoul trading Friday at 1,324.6 after touching 1,321.5, the won's strongest level in two years. It was 1,549.9 in early July and 1,353.3 at the end of September. The offshore one-month contract settled lower again overnight, near 1,322. 2. A dollar of sales at Friday's close converts to about 15% fewer won than in early July. 3. FnGuide's Q3 consensus slipped from ₩78.1T operating profit on ₩99.4T revenue in late September to ₩77.0T on ₩98.2T now. Both fell together and the implied margin held near 78%. That's what currency translation looks like, not weaker demand. 4. Hana Securities cut its FX assumption from 1,465 to 1,340 and still nudged its 2027 profit estimate up on higher DRAM prices. Spot is already below 1,340. The constraint: Q3 is mostly locked in. Q4 isn't. If the won holds here, estimates built at 1,340 or higher run too hot in won terms before a single chip price changes. One Seoul dealer sees 1,300 by year-end, with the move slowing around 1,320. How it plays: won strength hits hynix's reported profit, making FX the quiet risk on the print. The cleaner read is margin and dollar pricing. If the margin holds near 78% and management says Q4 DRAM prices are still rising, a won miss is translation. If the margin falls below Q2's 76.3%, the currency isn't the excuse. The return plan is sized on won free cash flow, so a strong won makes a big headline number harder to hit too. What would change my mind: the won sliding back toward 1,400, or hynix saying hedges absorbed most of the move. The call: Tue Oct 27, 8pm ET (Wed 9am KST). SK hynix (KRX:000660). NFA.

Same release, other side of the trade: headcount 66,775, down 15% y/y. Q3 capex was $34.91B against $32.79B of operating cash flow, so free cash flow went to -$2.12B. The CFO also named third-party AI token costs as a driver of expense growth. Fewer people shipping more code, paid for by a compute bill bigger than the cash the business throws off.

One fix on the headline: no stake changes hands. Per Maeil Business, Tesla puts in ₩1.1T and Nvidia about ₩2.6T to fund dedicated FC-BGA lines in exchange for guaranteed volume. Tesla's piece covers part of a ₩4.27T Sejong expansion. Mass production 2H28, Tesla supply 2029–2033. Customer money buys capacity, not shares. Same playbook as Nvidia's Ibiden reservation.

Not 18A as a node. ASML said in July that selected 18A layers in Oregon qualified on both low-NA and High-NA, with a subset of Panther Lake shipping. That's a learning run. 14A is the node built around it. Auth's point is pitch, not the tool. PowerVia moves power off the front, signal pitch relaxes, some layers fit one High-NA pass. He was comparing against TSMC's frontside-power pitches, which still need double patterning that High-NA doesn't fix. TSMC's own High-NA volume date: 2030.

Sources: Tempus 8-K, Oct 23 (deal terms, $100/record, exclusivity, related party): stocktitan.net/sec-filings/TE… · Pathos round and BMS rights, Reuters via CNA: channelnewsasia.com/business/patho… · Pathos release: investingnews.com/pathos-ai-anno… · Tempus release: tempus.com/news/pr/tempus… · Lilly–METiS terms: digitalhealthnews.com/eli-lilly-part…

Bristol Myers just led a $237M round in an AI drug-development company. The filing behind it shows who gets paid no matter what the drugs do, and it isn't the AI platform. It's the company selling the patient records, at a disclosed price of $100 each. The numbers: 1. Pathos AI raised $237M in a Series E led by Bristol Myers, at about a $3B valuation. BMS gets non-exclusive access to Pathos's Foundry platform across its portfolio, plus three years of exclusive use in neuroscience. 2. Tempus's 8-K, filed at 9am ET today: $192.5M in cash to Tempus over about five years. $137.5M of it is fixed payments from BMS for oncology data and software through 2031, invoiced quarterly. BMS can only exit with 12 months' notice and a "substantial termination fee." 3. The neurology piece: Pathos pays $100 per delivered record for up to 500,000 Alzheimer's and other neurology records, 100,000 a year for five years, plus $5M for updates. That's $55M, about 23% of the round Pathos just raised. 4. Tempus can't license those records to anyone else until 18 months after the later of delivery or Oct 22, 2027. 5. For contrast, Lilly's AI lipid deal with METiS this week: an upfront in the low double-digit millions against up to about $1.1B in milestones. The constraint: models are easy to copy. Years of curated patient outcomes aren't. The 8-K even caps the supply rate: 100,000 records a year, on "commercially reasonable efforts." You can raise $237M in a day. You can't build longitudinal Alzheimer's data in a day. How it plays: the "AI drug discovery is hype" take treats the whole stack as one bet. It isn't. The data seller gets paid per record whether or not a drug works, while the platform and the pharma company carry the clinical risk. That makes owning curated outcomes data a real reward, and the trial result the real risk for everyone downstream. Around noon ET, Tempus was up about 3% and Bristol Myers down about 2%. Two caveats: the neurology fees only arrive if Tempus delivers the records, and the 8-K says Tempus and Pathos are related parties, with overlapping leadership and equity ownership. A special committee of disinterested Tempus directors approved the deal. What would change my mind: big pharma training mostly on its own trial data, or a rival neurology dataset selling well under $100 a record. $TEM, $BMY. NFA.

Sources: DRAM spot, TrendForce (Oct 23 session): trendforce.com/price/dram/dra… · Weekly spot note and NAND wafers, TrendForce, Oct 21: trendforce.com/news/2026/10/2… · Q4 contract outlook, TrendForce: trendforce.com/presscenter/ne… · Commercial Times Q4 estimate via TrendForce, Oct 20: trendforce.com/news/2026/10/2… · Western Digital results, Oct 22: investor.wdc.com/news-releases/…

The SOX chip index fell 6.72% on Thursday. Memory prices didn't follow: the main DDR5 and DDR4 spot prices ticked up again this morning. The stocks are pricing a turn. The price data shows a slowdown. Those aren't the same thing, and SK hynix's call on Tuesday is where they meet. The weekly memory scoreboard: 1. DRAM spot (TrendForce, today's session, posted 6:10am ET): DDR5 16Gb averaged $61.27, up 0.27%. Mainstream DDR4 8Gb averaged $47.11, up about 2.5% from Monday's $45.97. TrendForce says DDR5 buyers offering higher prices still can't get allocation. 2. NAND is the soft spot: 512Gb TLC wafer spot slipped 1.28% this week to $18.59, with module houses sitting on high inventory. 3. Contracts: TrendForce sees conventional DRAM contract prices up 10–15% in Q4 and NAND up 15–20%. Its Oct 21 bulletin says vendors are widening Q4 hikes and PC DRAM supply shrinks next year. Taiwan's Commercial Times puts Q4 DRAM near +20%, after about +80% in Q2 and +40% in Q3. 4. Storage: Western Digital guided last night for next-quarter revenue up 66–72% year over year, with gross margin of 60–61%. The stock is lower premarket anyway. The constraint: suppliers keep moving wafers to server DRAM and HBM, so PC and phone supply shrinks. Long-term agreements cut both ways. They lock in buyers' volumes, and they cap how fast contract prices can catch up. TrendForce says some suppliers' server price increases lag the market for exactly that reason. How it plays: the rate of increase is slowing, from 80% to 40% to about 20%. Prices are still rising. A market that sells a slower rise like a fall makes Q4 pricing guidance the real risk/reward. If hynix points to Q4 DRAM prices rising faster than that 10–15%, Thursday priced the wrong thing. If it guides at or below, with contract ceilings biting, the slowdown was the story all along. What would change my mind: DDR5 spot rolling over for two straight weeks, or a supplier saying its Q4 contract hikes got cut. The call: Tue Oct 27, 8pm ET (Wed 9am KST). $MU, $WDC, SK hynix (KRX:000660). NFA.

Sources: Nvidia 10-Q (quarter ended July 26), commitments note: sec.gov/Archives/edgar… · Nvidia supply-chain capacity release, Oct 21: nvidianews.nvidia.com/news/nvidia-an… · Ibiden price history, Yahoo Finance Japan: finance.yahoo.co.jp/quote/4062.T/h… · Tokyo midday market, Kyodo via Mainichi: mainichi.jp/english/articl… · Nikkei contributors, Zaikei/FISCO: zaikei.co.jp/article/202610…

Ibiden fell 9.5% in Tokyo on Thursday and closed down another 1% on Friday, after trading as much as 8% lower. That's two days after Nvidia said it's prepaying billions to triple Ibiden's main substrate plant. The market is selling the suppliers as if they carry the AI-capex risk. Nvidia's own filing says a big share of it sits with Nvidia. The numbers: 1. Nvidia's latest 10-Q: supply and capacity commitments of $279B as of July 26, up from $119B a quarter earlier. $92B falls due in the rest of this fiscal year, $87B next year, $88B the year after. Nvidia says it's primarily memory and manufacturing facilities. 2. At GTC Berlin on Oct 21, Nvidia said substrate capacity for its products goes 2.5x in three years through Ibiden, Samsung Electro-Mechanics and Unimicron. Ibiden's Ono plant gets a multibillion-dollar capacity reservation advance to reach 3x, ramping from Q1 2028. KYEC gets a $1.1B advance for US test lines, volume production in April 2027. Amkor targets 300%+ more interposer capacity in Korea by end-2028. 3. The tape: the Nasdaq fell 694 points to 26,008 on Thursday. Friday morning the Nikkei broke below 62,000 for the first time since late July, with Advantest alone taking about 293 points off the index. The constraint: Huang put it in one line: "Capacity is the constraint." The new capacity lands in 2027 and 2028, the same window when most of that $279B comes due. How it plays: a supplier with a prepaid reservation has its new lines funded and a buyer lined up before the demand question is answered. Its risk is a slower ramp, not a stranded fab. The leftover risk moves upstream: Nvidia's filing says that if demand differs from plan it "may be unable to reduce these commitments," though some can be cancelled or rescheduled before firm orders, at extra cost. That makes prepaid capacity a real reward for the suppliers holding it, and Nvidia's commitment book the real risk if AI spending slows. Same question for SK hynix on Tuesday (8pm ET / Wed 9am KST): does its 2027 HBM capacity come with customer prepayments, or does it sit on hynix's own balance sheet? What would change my mind: Nvidia's next 10-Q showing commitments cut or pushed out, or suppliers adding capacity without reservation money. $NVDA, Ibiden (TSE:4062), $AMKR. NFA.

Sources: Nvidia 10-Q (quarter ended July 26), commitments note: sec.gov/Archives/edgar… · Nvidia supply-chain capacity release, Oct 21: nvidianews.nvidia.com/news/nvidia-an… · Ibiden price history, Yahoo Finance Japan: finance.yahoo.co.jp/quote/4062.T/h… · Tokyo midday market, Kyodo via Mainichi: mainichi.jp/english/articl… · Nikkei contributors, Zaikei/FISCO: zaikei.co.jp/article/202610…

Sources: SK hynix statement, Oct 23: news.skhynix.com/en/fact-12/ · Yonhap (KR), no plan to revisit, Sandisk, 12.9%: yna.co.kr/view/AKR202610… · Reuters via CNA, $15B: channelnewsasia.com/business/sk-hy… · Cash and capex, The Fact: news.tf.co.kr/read/economy/2… · Return policy and Hana estimate, Financial Post (KR): financialpost.co.kr/news/articleVi… · Intraday swing, NewDaily: biz.newdaily.co.kr/site/data/html…

SK hynix shelved Solidigm's US IPO before Seoul's open Friday (about 7pm ET Thursday). The stock fell about 4%, then was flat by 10am KST. The dual-listing worry is gone. The bill for Solidigm's new US fab isn't, and it now hits the same cash pile that funds Tuesday's return plan. The numbers: 1. Reuters had reported the listing could raise as much as $15B. That's off the table. Hynix says Solidigm's growth and investment needs are unchanged, and told Yonhap it has no plan to revisit an IPO for now. The window had already narrowed: Sandisk trades about 40% below its June high. 2. Hynix's own statement says Solidigm's existing fabs have limited capacity for AI demand, so it needs new US production. Timing, scale and financing are all back under review. 3. The cash: net cash was ₩69.4T at the end of June (₩88T of cash against ₩18.6T of debt). First-half operating cash flow was ₩91.7T vs ₩19.0T of capex. The ₩40T buyback that ran Aug 20 to Oct 21 came out after that. 4. The expectation: hynix has pledged to return more than 50% of cumulative 2025–27 free cash flow and will detail the next round on the Q3 call. Hana Securities' Kim Rok-ho expects at least ₩30T more, mostly buybacks. 5. The stake: Korean press puts Solidigm at about 12.9% of hynix's value. Its US holding company earned ₩5.84T in the first half, 4.3% of the group total. The constraint: one pocket now pays for HBM capacity (M15X in Cheongju, the Yongin cluster), the Indiana AI-memory site, and a new Solidigm fab, while promising more than half of free cash flow back to shareholders. How it plays: keeping Solidigm whole removes the discount of outside holders in a 12.9% slice, which is a real reward. But it moves the risk/reward on Tuesday to one question: does the return plan still come in at ₩30T or more with a US NAND fab inside the capex budget? A big buyback plus a sized, scheduled Solidigm plan says the cash covers both. A smaller return plus "still reviewing timing" language says the fab is eating the return. What would change my mind: a new outside-capital deal at the Solidigm level, like a strategic stake, which brings the dilution question right back. The call: Tue Oct 27, 8pm ET (Oct 28, 9am KST). SK hynix (KRX:000660). NFA.

Repricing the capex is harder than repricing the stock. Meta's 10-Q, filed today: $454.5B of non-cancelable commitments plus $355.7B of leases not yet started. $121.4B of those contracts falls due in 2027. Q3 free cash flow was -$2.12B. Equity can sell off. Signed contracts still need funding.

The bottleneck here is the substrate, not the chip. SEMCO's CEO said in March that server FC-BGA demand runs at least 50% above capacity. Ibiden holds an estimated 70–80% of top-spec AI substrates; SEMCO about 10% (Maeil Business). Nvidia wants 2.5x its substrate capacity in 3 years, so it's funding SEMCO, prepaying Ibiden and locking in Unimicron.

The tool ceiling is already in the guidance. ASML expects ~65 low-NA EUV systems this year, up from 44 in 2025, and says its planned 2027–28 output still won't fully meet demand. Lam said last night its $8.6B guide would have been higher without supply-chain constraints. The toolmakers have their own chokepoint.

Sources: Huang on memory, CNBC, Oct 22: cnbc.com/2026/10/22/jen… · NVHBM base dies on TSMC 3nm, Seoul Economic Daily, Oct 18: sedaily.com/article/201029… · Experts: Nvidia orders base dies directly, CNA, Oct 22: cna.com.tw/news/afe/20261… · TrendForce custom HBM note, Oct 22: trendforce.com/research/downl… · Spec comparison and UBS figures, e-Focus: e-focus.co.kr/news/articleVi…

Jensen Huang said this morning he sees no memory glut coming by 2029: "you can always use more of it." Micron is still down about 5% at midday. The glut debate is about how many DRAM wafers get built. For Nvidia's next HBM, the tighter limit may sit in Taiwan, not Korea. The numbers: 1. NVHBM, Nvidia's custom HBM unveiled in August, moves the memory controller off the GPU and into the stack's base die. Nvidia says that's up to 30% more bandwidth per stack and up to 15% less HBM power vs standard HBM4E. 2. Per Seoul Economic Daily (Oct 18), Nvidia asked Samsung, SK hynix and Micron to make that base die on TSMC's 3nm process, and the DRAM layers on 1c, the sixth-gen 10nm-class node. Samsung reportedly agreed: its first time using TSMC at the HBM manufacturing stage. Neither Samsung nor TSMC has commented. 3. Today, Taiwan's CNA quoted experts saying Nvidia orders those base dies from TSMC directly, and TrendForce published a note naming TSMC the main foundry for custom HBM, with base dies lifting advanced-node demand from 2028. 4. UBS (via Seoul Economic Daily) sees 2027 HBM demand up 91% to 61.5B gigabits, vs 54.6B of new supply. That's an ~11% gap. The constraint: every NVHBM stack now needs three things in a row. DRAM layers from a memory maker, a 3nm logic die from TSMC, and TSMC's CoWoS packaging to sit it next to the GPU. Those 3nm lines also make phone and HPC chips. Samsung is reportedly validating its own 2nm for the base die in case NVHBM demand outruns TSMC's 3nm. How it plays: more DRAM capacity doesn't become more NVHBM unless TSMC's 3nm and packaging keep pace, making TSMC allocation a real reward for TSMC and a real risk for memory makers' volume plans. Samsung also gives up the base-die margin it kept by building turnkey. CNA's experts add a twist: NVHBM doesn't fit AMD or custom-ASIC platforms, so the more of it gets made, the less HBM is left for everyone else. For SK hynix's Oct 27 call (8pm ET / Oct 28 9am KST): hynix already makes HBM4 base dies with TSMC, but announced HBM4 on 1b DRAM. The line to watch is its 1c timing, the other half of Nvidia's spec. What would change my mind: Nvidia or TSMC saying base dies can come from a second foundry, or Samsung qualifying its 2nm as that second source. $NVDA, $TSM, $MU, SK hynix (KRX:000660), Samsung (KRX:005930). NFA.