Three companies published the same cost target. NVIDIA ($NVDA) wants co-packaged optics under 1.5 picojoules per bit. IBM ($IBM) is going for under 1. Intel ($INTC) is targeting under 5 across the package, end to end. Different numbers, because the engineering is different. Then all three name the same cost. 0.1 dollars per gigabit per second. Power targets diverge because the physics diverges. A cost target that lands on the same figure three times did not come up from the engineering. It came down from the customer. Which would be fine, except for what the customer cannot do. The director who set up Resonac's (TSE: 4004) US-JOINT consortium in Silicon Valley told Nikkei that hyperscalers and fabless companies are strong at simulation, but there are few environments where you can verify an actual semiconductor package. So the people setting the number cannot test against it. On the manufacturing side, the same week's coverage says who bears final responsibility for quality assurance is still an open question. Foundries and OSATs are not used to volume optical production, and the division of labour has not been agreed. The instruments exist. Keysight ($KEYS) announced a 220 GHz lightwave component analyzer on 12 March and showed it at OFC. The measurement is possible. The signature is not assigned. Here is why that matters more than it sounds. Both failure modes in this stack are local, not average. A fiber core and a waveguide differ in cross section by roughly 800 times, so a misalignment at nanometre scale becomes a coupling loss you cannot average away. On the thermal side a chip can run cool on average and still fail from one hot spot. Resonac buys test chips from imec that reproduce over 1000 watts and can heat a single region of the die. Testing that used to pass on averages now has to catch the local case. That means more test time per part, on parts that are already singulated. The back end already had the thinnest margins in the industry. So the layer that gets paid here is not the one making the optics. It is the one deciding whether they passed. Advantest (TSE: 6857) and Teradyne ($TER) are the duopoly in that seat. Keysight sits beside it. One more line, because the geography is odd. Resonac partnered with Purdue in May for thermal analysis. SK hynix (KRX: 000660) is putting 3.87 billion dollars into an advanced packaging site on land held by a Purdue-affiliated foundation in Indiana. A Japanese materials company and a Korean memory maker are working at the same university. Where a responsibility gap sits open, someone eventually charges a toll to close it. What would change my read: if a foundry or an OSAT publicly takes final sign-off on CPO. Then this stops being an open seat. Sources, in order: Nikkei xTECH, 17 August 2026, two articles. Keysight press release, 12 March 2026. Nikkei BP, book on optoelectronic fusion, industry map section. My read, not advice.

Resonac (TSE: 4004) and Intel ($INTC) are building a panel the same size. Not a similar size. The same number. Resonac's APLIC line in Ibaraki runs 510x515mm and makes organic interposers of 8 to 12 reticles. Intel presented a glass core panel at ECTC this year. 510x515mm, 24 layers. Same rectangle, different material. So the material is not settling first. The format is. Why a panel at all. Resonac's own February deck answers it, and the answer is a size problem. Its estimate has package interposers growing more than 80% by 2030, from 70x80mm today toward 100x100mm. And 8 to 12 reticles works out to roughly 83x83mm up to 101x101mm. The equipment slide and the market slide are describing the same object. Meanwhile the wafer does not grow. It has been 300mm for over two decades. That is the argument. A 510x515mm panel is 262,650 mm2 against 70,686 for a 300mm wafer. About 3.7 times the area in one pass. A Resonac fellow put the reason in a physics journal last year, and it is not a performance argument. A wafer leaves the front end worth 300,000 to 500,000 yen. Then it gets diced, and back end materials are applied to chips worth 300 to 500 yen each. The back end lost the batch and has been shaving cost ever since. One more number lines up. The same deck has substrate layer count rising 15 to 20% from 20 to 22 layers, which lands between 23 and 26. Intel's glass panel has 24. So glass is not the story. Glass is one of two answers to a story about batch size, and Intel is building to a range a Japanese materials company published in February. Being fair: those growth figures are Resonac's own estimates and the slide says so. The fellow works for the company building the organic line. His margin figures are 2021 to 2023 peaks. What would change my read: if the two panels stop converging. If Intel's glass moves to 600mm square, the format is not settled and the material fight has not started. Sources, in order: Resonac Holdings, Phase 2 strategy presentation, 13 February 2026, pages 13, 18 and 19. Intel, ECTC 2026. Seiichi Kondo, Resonac fellow, Oyo Buturi volume 94 number 6, 2025. My read, not advice. The area ratio and the reticle arithmetic are mine.

The market may not need to be larger. It needs to be industrial. Fifteen billion dollars is the 2030 forecast. Unitree closed today at roughly 53 billion dollars of market value. Those are different units, a stock against a flow, but the gap is the point. And of what Unitree shipped from January to September last year, 9% by value went to manufacturing or logistics. That is from its own listing disclosure. Nomura, initiating today, puts more than 70% of humanoid revenue with research buyers. The hype is not ahead of the market size. It is ahead of the use case. I wrote the long version an hour ago, with the Japanese component order books attached.

Furukawa Electric (TSE: 5801) reported 25.5 billion yen of operating profit for the June quarter. Its entire first half last year was 19.4 billion. Full year guidance went from 95 billion in May to 123 billion in August. Up 92.6 percent on last year's 63.9 billion. This is a cable maker, and the guidance is doing what a semiconductor name usually does. Source: Furukawa Electric, Q1 FY March 2027 results, 6 August 2026.

Japan's long-term rate touched 2.930 percent yesterday. That is the highest since September 1996. Today Tokyo gave back five sessions of gains in one. The Nikkei closed at 67,460.73, down 1,759.52 or 2.54 percent, the first fall in six sessions. Nikkei's own headline names the reason: Middle East concerns and rising rates. Two names did roughly half of it. Advantest and Tokyo Electron together accounted for about 836 of those 1,759 points, per Zaikei's contribution ranking. What was repriced is the multiple. What was not repriced is the commitment. JX Advanced Metals (TSE: 5016) decided in June to spend up to 120.0 billion yen through fiscal 2030 on InP capacity. Fuji Electric (TSE: 6504) is not raising group capex at all, it is reallocating inside a smaller number, from 85.2 billion two years ago to 56.0 billion this year. Both decisions were designed when the ten year cost of money in Japan started with a zero. This is the part the tape does not price in a single session. A multiple can move 2.54 percent in a day. A wafer line and a switchgear plant transfer are three to five year schedules with a signature already on them. So the question I am holding is narrow. Not whether Japanese suppliers are cheap after a down day. Whether any of them revises a multi-year capex plan because the discount rate moved, and which line it shows up in first. I do not think it shows up in guidance. Guidance is annual. It shows up in the capex schedule, and that is disclosed once a quarter. The September quarter closes in six weeks. Korea sits on the other side of this. SK hynix (KRX: 000660) breaks ground in Indiana on 27 August. Samsung Electronics (KRX: 005930) is expanding Pyeongtaek. Those are dollar financed. The Japanese layer under them is yen financed, and the yen curve just moved thirty years. Every capex schedule is somebody's bet that money stays cheap. That is a different risk from the one the tape traded today. I could be wrong in an obvious way: rates have been rising in Japan all year and none of these plans have changed yet. If the September filings show the same capex lines, this is noise and I will say so. Sources, in order: Nihon Keizai Shimbun, 17 August 2026, long term rate at thirty year high. Nihon Keizai Shimbun, 18 August 2026, Tokyo close. Zaikei Shimbun, 18 August 2026, Nikkei contribution ranking. Second hand. JX Advanced Metals company release, 16 June 2026. Fuji Electric FY2027 Q1 appendix. The 836 of 1,759 share is my arithmetic. My read, not advice.

14,641,148 shares of Taiyo Yuden (TSE: 6976) changed hands on 3 August. 10.8 percent of the company, in one day. Nobody who received them has to file anything. A correction before the rest. On 12 August I wrote that a Japanese large shareholding report does not name the counterparty. That was wrong. The 60 day table has a column for it, and Situational Awareness has now filed that detail. Five names took the block. Citadel Multi-Strategy Equities Master Fund, 3,335,335 and 1,287,985 shares at 9,397 yen Jane Street Financial, 1,882,020 at 10,225 JP Morgan Securities, 4,135,808 at 10,225 Barclays Bank, 4,000,000 at 10,225 Now count the slices. 3.05, 2.95, 2.46, 1.38, 0.94. Japan's disclosure threshold is 5 percent. Not one of them crosses it. And the seller finished that day at 4.41 percent, down from a peak of 16.61 on 21 July. That is also below 5. So both sides of the trade are now invisible. The fund has dropped under the line it was reporting from. The five who took the stock never crossed it. One more thing in the table. Two prices. Citadel paid 9,397 yen. The other three paid 10,225. Same day, same stock, 828 yen apart. The filing does not say why, so neither will I. And the selling started before the story did. 185,200 shares on 28 July, in the market. 1,700,000 on 30 July. The margin call reporting came on the 29th and 30th. What I still cannot tell you: three of those five are dealer entities. Where the stock went after them is not in any document I can read. Sources, in order: Situational Awareness LP, large shareholding reports on Taiyo Yuden, initial filing through change report No.8, EDINET. The 10.8 percent and the 9,963 yen weighted average are my arithmetic from that 60 day table, and the read is mine too. Not advice.

He named the resource. Land, power and shell. His own sentence is the one worth keeping: the site is long-lived, the compute inside can be upgraded repeatedly. That asymmetry is the whole trade. Twenty years at 4.25 IT-gigawatts is several generations of GPU. It is one set of substations. The compute gets swapped. The switchgear does not. So the question is who builds the part that stays. The release already prices it. At least 10 gigawatts of new generation to produce 8 IT-gigawatts of AI factory capacity, plus at least 4.2 billion dollars of new regional grid infrastructure with AEP Ohio. The distance between the 10 and the 8 is not rounding. It is conversion, distribution and cooling, and every piece of it is equipment somebody has to build before a single token is served. One data point from Tokyo yesterday. A Japanese maker of substation and switchgear equipment is cutting semiconductor capex 42 percent this fiscal year and raising energy capex 128 percent, with the company total flat. The stated reason in its own filing is construction of data centers and semiconductor plants. Not that it supplies Ohio. It is reading the same demand curve. Worth noting who owns the shell. SB Energy builds, owns and operates the site, and it sits inside SoftBank Group (TSE: 9984). Son is quoted in the release. The compute is American. The landlord is not. What I cannot tell you is who supplies PORTS-Pike. That is not public. He is securing twenty years of power. Someone has to build it first, and that spending is already sitting in filings.

Fuji Electric (TSE: 6504) A company can raise guidance and mean very little by it. What it cannot soften is when it has scheduled the money to go out. Fuji Electric (TSE: 6504) reported the June quarter with its data center power line down 3 percent in revenue. Same day, in the supplementary data file, that line is planned up 25 percent for the half and up 23 for the year. The capex schedule in the same file says how. Segment capital spending, June quarter actual and first half plan: Energy: 1.7bn yen actual, 10.5bn planned for the half Semiconductors: 2.8bn actual, 7.3bn planned Subtract the quarter from the half and you have the plan for the September quarter. Energy 8.8bn against 1.4bn a year ago, six times. Semiconductors 4.5bn against 13.7bn, one third. Company total 19.1bn against 18.2bn. The quarter as a whole barely moves. What is inside it moves six times one way and down two thirds the other. Context that cuts against the excitement. Company capex peaked at 85.2bn yen in the year to March 2025, came in at 56.7bn last year, and is planned at 56.0bn this year. This is not a spending cycle. It is a reallocation inside a shrinking one, and that is a different thing to own. Energy money buys new buildings at three sites, Malaysia, Kobe and Tsukuba, and moves switchgear production from Chiba to Kawasaki. The stated reason is demand for substation equipment, and the order book names data centers and overseas semiconductor plants among its drivers. One number for scale on the other side. Semiconductor depreciation and lease cost was 8.9bn in the quarter against 2.8bn of new spending. That segment is still paying for the last cycle while this one gets funded. The US and Korean names in this layer are well covered. Eaton ($ETN), GE Vernova ($GEV), HD Hyundai Electric (KRX: 267260), Hyosung Heavy Industries (KRX: 298040). This is the Japanese one I had never written about until today. What would change my read: the September quarter closes in six weeks. If energy capex lands near 1.4bn again instead of 8.8bn, the reallocation was a schedule and not a decision. Guidance is what a company says. A capex schedule is when it has to write the cheque. Sources, in order: Fuji Electric, Q1 FY2026 supplementary data file, pages P1, P3 and P9. Fuji Electric, Q1 results presentation, 30 July 2026. Fuji Electric, FY2026 management plan, 28 April 2026. My read, not advice. The September quarter figures are the half year plan minus the reported quarter.

The SEC accepted Situational Awareness' Q2 13F at 4:26pm New York time on Thursday. A summary was circulating 18 minutes later, and it was accurate. I checked all twenty-six lines against the filing. What no summary of that filing could contain is 6,374,823 shares. On 22 June Leopold Aschenbrenner's fund filed a Form 3 on SharonAI Holdings ($SHAZ), declaring itself a ten percent owner. That filing disclosed 1,696,127 common shares and 6,374,823 pre-funded warrants, exercise price $0.0001, no expiration, exercisable at any time. The 13F shows the common. By 30 June it had grown to 5,396,127 shares, $456.8 million. The warrants are not in it, because pre-funded warrants issued in a placement are not 13(f) securities. Same manager, same issuer, eight days apart, and only one of the two filings carries the other half of the position. Four things a 13F structurally cannot show you: - Securities that are not 13(f) securities, which is where the SharonAI warrants sit - Anything not listed in the United States - Which entity is reporting. In Q1 two affiliated filers reported an identical book, so anyone who summed them counted $13.7bn twice. The deadline for Q2 was Thursday and only one of them has filed. I do not know why - Timing. The 13F arrives 45 days after quarter end. The Form 3 arrived in seven. And in the 45 days between the reporting date and this filing, the fund was margin called out of its entire public book and sold it to Citadel. The list everyone is reading today is a portfolio that no longer exists The second one is the one I work on. This fund's two largest positions are $SNDK at $5.67bn and $MU at $5.57bn, together about 55% of the reported book. It is a memory bet. The equipment and the consumables that memory gets made and tested with are also not listed in the United States. Here is what that means in practice. On 29 June, one day before the quarter closed, the same manager filed in Tokyo disclosing 5.99% of Taiyo Yuden (TSE: 6976), 7,808,800 shares, ¥136 billion committed. By 21 July it was 16.61%. By 3 August it was 4.41%. So on the reporting date of this 13F, there was a nine-figure dollar position in a Japanese passive component maker sitting on the same book. The filing has no line for it. It never could have. Look at those Tokyo dates again. 21 July at 16.61%, 3 August at 4.41%. That window is the same window in which the public book was being unwound. The Tokyo filings show the exit in near real time. The 13F, filed eleven days after it was over, shows the book at its peak. Whether those warrants were still outstanding on 30 June, and what the 19.99% ownership cap does to them, is the part I have not finished. The Form 4 filed on 2 July covers that window. If you want to know what a fund is doing, a 13F is where you start reading. It is not where the position ends. Sources: SEC EDGAR, Situational Awareness LP (CIK 0002045724), 13F-HR for period ended 30 June 2026, filed 14 August 2026 Same filer, Form 3 on SharonAI Holdings Inc., period 22 June 2026, filed 29 June 2026 Same filer, 13F-HR for period ended 31 March 2026, filed 15 May 2026 Situational Awareness Partners LP (CIK 0002038540), 13F-HR for period ended 31 March 2026, filed 15 May 2026 EDINET large shareholding reports, Situational Awareness LP on Taiyo Yuden, nine filings between 29 June and 3 August 2026 CNBC, reporting on the unwind of the public book, 30 and 31 July 2026

17 days ago Advantest (TSE: 6857) raised full-year operating profit 35% and the stock fell. I wrote that I would take the filings over the tape. Yesterday it traded as high as ¥38,730 and settled at ¥36,870. That is 46% above the ¥25,200 close I was writing about. I am not posting this to be right about a price, and the honest part cuts the other way. My own note that day said the bear case was never that demand was fake, it was that 49 times trailing earnings was too much to pay for it. It closed yesterday at 58 times. The multiple argument did not get weaker. It got more expensive. What did resolve is the layer question. That post ended by saying the layer under the testers is the one that wears out, and named Micronics Japan (TSE: 6871). It reported on Wednesday. - Operating margin 35.1%, vs 27.0% the quarter before - Probe card segment margin 41% - Memory 95% of the mix - Asia 92.9% Every quarter in the two years before this one sat between 20.2% and 27.0%. More test seconds is more consumable, and that part showed up in the accounts. The company also said the quarter borrowed. Volume carried over from the prior quarter, and revenue was pulled forward on expanded capacity. Its cumulative guidance implies the next quarter lands about 31% lower. Advantest also told you where to look next. It raised its 2026 tester market forecast by about 19% in three months, then listed what could still stop it. The first item was not demand. "Industry uncertainties remain, including advanced packaging capacity and memory supply dynamics." The company that tests the chips is pointing at the floor they sit on. That is a different set of names again. Sources: Advantest Q1 FY2026 results presentation and Q&A, 29 July 2026 Micronics Japan H1 FY2026 results presentation, 13 August 2026 Micronics Japan revision to earnings and dividend forecast, 12 August 2026 Tester market figures compare the July deck with the April 2026 deck Prices and multiples are 14 August 2026 closes Quarterly operating profit is my own calculation, cumulative less prior quarter

Yesterday @PhotonCap and I arrived at the same place from opposite ends of the stack. He said he had not dug into the Japanese side yet. So here it is~. The number to start from is JX Advanced Metals (TSE: 5016). On 16 June it decided to spend up to 120 billion yen through fiscal 2030 at its Isohara and Hitachinaka plants in Ibaraki, taking InP substrate capacity to seven to ten times fiscal 2025, ramping from the second half of this fiscal year. Largest investment in the company's history. First, why jurisdiction is the whole question. AXT is the merchant name everyone reaches for, and its InP comes out of Tongmei in Beijing. China put InP on its export licence list in February 2025. The talks reportedly raised it. The licensing regime stayed. Temperature moved, mechanism did not. That narrows the question to something answerable. Who makes merchant InP substrate outside that desk? JX has been making InP substrate since the 1980s, so this is not a new entrant buying its way in. And the detail I would not skip is what came before the June decision. In fiscal 2025 alone JX approved three separate expansions worth roughly 25 billion yen. Then it decided on one nearly five times that size. Three incremental steps and then a jump is not a forecast. That is a company that has seen the order book. The quarter supports it, with a caveat I will give you first. JX is not an InP pure play. Copper smelting, sputtering targets, rolled copper foil, recycling. The company attributes the quarter to AI server demand, better selling prices, a weak yen and higher copper, and it flags lower mine output from bad weather on the resources side. So do not read the following as an InP number. Fiscal Q1 to June, IFRS: revenue 260.6 billion yen up 36.2%, operating profit 81.4 billion up 175.5%, net profit 53.1 billion up 181.3%. Against a full year operating guide of 232 billion, that is 35.1% of the year booked in a quarter that should carry 25%. And the company raised that guide on the same day, from 190 billion. The 35.1% is measured against the higher number. The second name is Sumitomo Electric (TSE: 5802), which holds the number one position in InP substrate and has presented six inch crystal growth and wafer processing. Same jurisdiction answer. Nikkei reported today that Nvidia and Sumitomo are development partners on Rubin, with an SMBC Nikko analyst saying Sumitomo product is expected to be adopted. That is an analyst view, not a company statement, and I am flagging it as such. Here is the part that keeps me honest about timing, and it comes from JX's own deck. In the same presentation the company raised its full year sales volume outlook for semiconductor targets, magnetic targets and titanium copper. It left InP substrate unchanged at plus 13 percent, and the June quarter came in at plus 8. Its own wording for InP is that it continues to improve efficiency on existing equipment. Capacity going to seven or ten times, volume plan untouched. Those are two different clocks in one document. Which is really the answer to your question. Japan is the structural answer to the merchant shortage. It is not the 2026 answer. Keep the physical scale in view too. Four inch to six inch multiplies usable area 2.25 times, and a six inch InP wafer is still roughly a quarter of the area of one 300mm silicon wafer. The layer carrying AI's light is tiny next to the layer carrying its logic, and multiplying tiny by ten still lands on small. So the honest shape of it is this. The merchant relief that does not pass through a Chinese licence desk exists, it is Japanese, it is funded, and it is late. Which gives a test rather than a call. Watch two things: the Isohara and Hitachinaka ramp dates, and whether JX lifts that plus 13 percent volume line at the half year. If the volume plan moves before the capacity does, demand arrived earlier than the company modelled. If it does not move by March, pricing power stays with whoever issues the permits, and every transceiver maker downstream is negotiating with that fact until then. Thanks for the prompt. Sources: JX Advanced Metals, decision on capital investment to expand InP substrate production capacity, 16 June 2026 JX Advanced Metals, FY2027/3 Q1 earnings release and results presentation, IFRS consolidated, 6 August 2026 China Ministry of Commerce export control listing, February 2025 Coherent FQ4 2026 earnings call, 12 August 2026 Nihon Keizai Shimbun, 14 August 2026

Advantest (6857.T) raised its full-year operating profit guidance by 35% today, from ¥627.5B to ¥846.0B. The stock closed down~ It traded as high as ¥27,520, fell as low as ¥24,135, and settled at ¥25,200, roughly 30% below its 52-week high. The reason Advantest gave for the raise: test demand for inference chips came in far above what it assumed in April. One day earlier, Teradyne ($TER) posted a record quarter, beat consensus by 8.9%, guided Q3 about 20% above the street, and said roughly 70% of revenue now comes from AI. It rose 13.6% after hours. Same thesis. Opposite tape. That gap is the whole story this week. The market decided AI spending had peaked. Memory fell into a bear market. CXMT listed in Shanghai and rose 466%. China began mass-producing its own immersion lithography and ASML dropped 6.6%. By Monday the sector was trading like the buildout was over. The companies that test the chips did not get that message. One of them got paid for saying so. The other did not. Here is why testing is the layer I keep watching. It gets paid regardless of who wins. $NVDA, $AMD and every hyperscaler custom accelerator has to be tested before it ships, so this layer is indifferent to the fight everyone else is trading. And inference silicon takes longer to test than training silicon does, which means test demand can grow faster than unit volume. That is the sentence hiding inside Advantest's revision. Not more chips. Harder chips. Advantest also told you what it believes with capital. It is expanding capacity by more than 70% by the end of this year. Companies that think the cycle is rolling over do not do that. And under the testers sits the layer that wears out, the probe cards and sockets that get consumed as test time rises. Micronics Japan (6871.T) and FormFactor ($FORM) live there. More test seconds is more consumable. Keep it honest, because the sellers have a case. At ¥25,200 Advantest still trades near 49 times trailing earnings, and the bear argument is not that the demand is fake, it is that 49 times was too much to pay for it. An implied 49% operating margin invites mean reversion. A raised guide is a forecast, not a receipt. And the China supply story is real, not something to wave away. But notice what is actually being argued. Nobody today disputed the demand. The dispute was about the multiple. Those are different trades, and only one of them tells you what is happening in the fabs. When the market says peak, and two companies on opposite sides of the Pacific raise in the same week and point at the same word, I will take the filings over the tape. My read, not advice.