$GLW did not raise two billion dollars on Friday. It arranged the option to, and the distinction is the whole story. I pulled the 8-K. Item 8.01, event dated 11 September: an Equity Distribution Agreement with Goldman Sachs to sell shares having an aggregate offering price of up to two billion, from time to time, through an at-the-market programme with Goldman as sales agent. Under the agreement the company sets the parameters, including any price, time or size limits. Then the sentence that settles what kind of thing this is. Corning intends to sell from time to time in varying amounts, which may be limited, based on market conditions, trading liquidity, the trading price, and determinations by the company of its need for, and appropriate sources of, additional capital. That is a company telling you the amount is undecided and may be zero. Proceeds, if any, go to general corporate purposes. No project is named. Goldman takes one percent of whatever gets sold. And it runs off a shelf that has been sitting there, automatically effective, since April. The stock closed Friday at 166.39 and the last extended-hours print I hold is 162.44, about two and a half percent lower. One wire put the fall at 2.7. Either way, the market spent Friday evening pricing dilution that has not happened and may never. Now the part I find more interesting than the filing itself. Look at what this facility sits next to. In May, Corning published an annualised run-rate plan: twenty billion by the end of this year, thirty by the end of 2028, forty by the end of 2030, with a high-confidence version of the last at thirty-five, on nineteen percent annual sales growth against fifteen in the phase now closing. The same day it announced a multi-year arrangement with Nvidia for ten times the US optical connectivity capacity and more than fifty percent more US fibre, across three new plants in North Carolina and Texas. Microsoft had already named it to manufacture hollow-core fibre at those same North Carolina plants. In March it co-founded the four-core fibre group with Sumitomo Electric, Fujikura and TeraHop, so it sits on both of the competing new-fibre routes at once. On 8 September it signed Verizon for more than eighty million miles. And on the second-quarter call the chief executive said, plainly, that scale up and photonics are not yet in the results. So here is the frame I keep coming back to. Three times in four weeks a different part of this chain has told us how the buildout gets paid for. Oracle: the customers, via pre-pay and bring-your-own-hardware, with 11.4 billion of prepayments landing against nothing a year earlier. Memory: the contract, on long-term agreements that already carry price ceilings, while DRAM and NAND go from 47 percent of major cloud capex to 68. Corning: the equity market, possibly, if it decides it needs it. Of the three, the one that arranged an equity facility is the one whose new revenue stream has not started yet, on its own account. That is a note about sequencing rather than about anyone's balance sheet, and the capacity is committed either way. What I am watching is the next 10-Q. An at-the-market programme only becomes news when the share count moves. $GLW $ORCL $MU $NVDA

Next week has one event on it and it is not ours. The FOMC decides on Wednesday the 16th at 14:00, and futures put a quarter-point rise at better than four chances in five. The range is 3.50 to 3.75 and has not moved all year, so this would be the first tightening of the cycle. The repricing followed Warsh at Jackson Hole, where he said the summer's better inflation prints did not show that the underlying trend had meaningfully improved. August CPI then printed 3.4%, with core up 0.3% on the month, a tenth above forecast. Retail sales land at 08:30 that same morning. That is the last data input, five and a half hours before the decision. Now the part that decides how to read the week. Fifty-nine companies report between Monday and Friday. I checked every one against our thirty names and not one of ours is on the list. The next is Micron on 30 September, from the company's own notice. No results, no product event, nothing at all from our side of the chain. So these names get priced by rates next week rather than by anything they do. Most of the pool trades on a long-duration multiple, which puts it in the front row for exactly this kind of move. Worth separating what a hike actually touches. I have spent this week reading fab schedules and supply agreements. SK Hynix with 38 billion dollars approved and no output before December 2028. Samsung's P5 in 2028. Micron's Idaho fabs in 2027 and late 2028. TSMC's packaging roadmap running out to 2029. Memory long-term agreements that already carry price ceilings. A quarter point moves none of those dates. It moves the rate at which you discount them. The week tests the multiple, not the thesis. Those are different things, and next week will tell you very little about the second one. Two handles if you want something other than macro. Oppenheimer runs an AI infrastructure conference in New York on Tuesday, and Wells Fargo takes investors around Texas power and gas on Wednesday and Thursday. Those are the only two places our lane speaks all week. Then keep the week after clear. ECOC opens in Málaga on the 20th, and that is where the optical supply chain says what it has been holding since Shenzhen. $NVDA $MU $SKHY $TSM

The best thing I have read on pacing the frontier argues that it is a margin lever. Shipping a better model at the same price is, in unit terms, a price cut per unit of intelligence, so slowing the cadence slows the price cuts, and an industry-wide agreement not to compete on cadence is an agreement not to compete on price. That is Paradis's argument and I think it is right. Here are two things from my side of the chain that sit underneath it. First, the accounting point is sharper than it looks. Labs expense training as research and development, so nothing amortises, but the economics are the same because every frontier model is superseded the moment the next one ships. Now put that next to what Oracle filed this week. GPUs coming up for renewal went out at a 20% premium to the prior contract price, and management said the majority of them were four years old or more. Utilisation on the same page was 97.9%. So the model has no book life and an economic life set by somebody else's release calendar. The GPU has a long book life set by policy and an economic life that just proved longer than feared. The asset with the shortest economic life in this industry is the one nobody capitalises, and the asset everyone argues about depreciating is the one that just repriced upward. Read that way, pacing is an attempt to extend the useful life of the only asset in AI that has no useful life on the books. Second, and this is the part that matters if you are positioned in the supply chain. Pacing is a statement about demand. It does not touch supply, because supply is already contracted. Oracle: 664 billion of RPO, up 209 billion year on year, the majority of the quarter's growth via pre-pay or bring-your-own-hardware, and 11.4 billion of customer prepayments against nothing a year earlier. Memory: DRAM and NAND go from 47% of major cloud capex this year to 68% next year, on long-term agreements that already carry price ceilings. Fabs: SK Hynix has 38 billion dollars approved with no output before December 2028, Samsung's P5 lands in 2028 after being pulled forward two years, Micron's first Idaho fab is 2027 and its second late 2028. Packaging: TSMC past 14 reticles with 24 HBM5E stacks by 2029. None of that unwinds if the release cadence slows. Concrete does not unpour and prepayments do not un-pay. Then the asymmetry worth keeping in view. One side of this has contracts with dates on them and money already transferred. The other has an essay, two replies, and exactly one concrete commitment, which is third-party evaluators with employee-level access. Nobody has committed to a slower release cadence. What would change my mind is a lab publishing an actual release schedule, or a hyperscaler cutting capex guidance and naming pacing as the reason. Until one of those happens, pacing reprices the demand curve and leaves the supply curve exactly where it was. $ORCL $NVDA $MU $SKHY $TSM

I pulled $ORCL 's slides off the exhibit the night they filed and wrote this up on the 11th, so let me add the part I got wrong and the part nobody has put next to it. What I got wrong: I said Oracle did not disclose what the 20% was measured against. It did, on the call rather than on the slide. Management said all capacity up for renewal was renewed or resold 20% above the prior contract price, and that the majority of those GPUs were four years old or more. That is a better fact than the one I had, and it makes the read stronger rather than weaker. So the version where rent falls away as silicon ages is dead for now. That was the version that broke the arithmetic, and it was the one that mattered. Here is the part that has not been put next to it. The renewal premium is on slide ten. On slide nine, one page earlier, Oracle says the majority of the quarter's RPO growth came via pre-pay or bring-your-own-hardware. And in the cash flow statement filed with the same deck there is a line reading increase in deferred revenues from customer prepayments with significant financing component, 11,363 million dollars, against a dash for the same quarter a year earlier. So the counterparties paying the 20% are substantially the same ones who prepaid for the capacity, and on the bring-your-own side, part of the fleet is not Oracle's asset at all. That is not an accusation, and Oracle disclosed all of it on adjacent pages. It is a note on what kind of number 20% is. A price agreed with someone who already funded the build is a different thing from a price set in an open market. Not a worse thing. A different one. The other number on that slide is doing more work than the premium. Utilisation was 97.9%. At 97.9 there is no spare capacity to discount into, so the premium is what a sold-out market prints, which is a claim about the market rather than about how well the asset held its value. Both readings fit the data. Only one survives utilisation slipping, and the premium is the first thing to go when it does. Two questions keep getting merged here and they should not be. Useful life is set by a policy. Rent is set by a market. This quarter told us a great deal about the market and nothing about the policy. Both numbers are disclosed now, so both are checkable in ninety days. That is the date I am watching, along with whether the prepayment line stays this large while the premium holds. $ORCL $NVDA $NBIS $IREN

A week is a useful unit. Here is what one looked like: Monday 7 to Friday 11 September, on one desk covering the AI supply chain. What went in first. IQE's half-year results presentation, read as a deck rather than as a press summary. Oracle's FQ1 and Adobe's FQ3 out of the 8-K exhibits the night they filed, with Oracle's own fifteen slides alongside. Korean memory inventory, Taiwan trade data, US CPI and PPI, each from the body that issued it. A broker's two-volume optical networking report, read for what its own charts said rather than for its conclusion. Infineon on the power shortage, in its own materials. Two items on that list did not exist before we made them. Jensen Huang's session at Goldman, 2,256 seconds and 371 timestamped segments, and Lumentum at the Citi conference. We pulled the audio and transcribed both. And three days on the floor at the Shenzhen optoelectronics show, behind an archive of 3,182 exhibitors, 604 agenda rows and 54 conference pages that we counted before the doors opened. What came out: 555 files across the five days. 266 figures, 162 post drafts in two languages, 40 news and data notes, and twelve primary documents pulled in from the issuers. 77 pieces were written in English, each with a Chinese counterpart. Not all were published, because a week produces more drafts than it has slots. None of that is the interesting part. Four of the five days produced roughly the same amount of work, so volume tells you almost nothing. The number I care about is four, which is how many times that week the collection got wider. The news triage learned simplified-Chinese optical vocabulary and picked up the show's own conference feed. A technology-maturity desk went in with thirteen standing sources and nineteen technology nodes, swept weekly and compared word for word against the previous copy. Open-access research became its own leg. And a financial layer that reads straight from SEC XBRL went live, alongside a bank for material that has no slot yet. Those four are still running. The 555 files are not. They were read once. Almost nothing above is exclusive. Every document was published by whoever made it, and anyone could have read it. The only difference is that it was read. $IQE $ORCL $ADBE $NVDA $LITE

I spent three days on the Shenzhen optics floor collecting what people would say on the record. The most useful thing I came away with is that the industry does not agree on where its own bottleneck is, and the second most useful is a number nobody would give. Start with the disagreement, because both halves of it come out of one mouth. On the 13 August earnings call, Coherent's chief executive said indium phosphide capacity continues to be the primary constraint, and in the same answer said the company is not constrained in assembly and test capacity right now, that it has that capacity available. Four weeks later on the show floor, the Credo and Astera Labs booths put the capacity gate for co-packaged optics in packaging and test, to be absorbed by the outsourced assembly houses. Neither side is wrong, and they are not talking about the same thing. One is a company describing its own lines. The other is about whether the industry can absorb co-packaged optics at volume. That question has not been answered because it has not been put to the assemblers. What settles it is whether the outsourced assembly and test houses announce capacity for co-packaged optics, and that is observable with a date on it. Now the number. On the 9th, Sivers said at its booth that 70mW continuous-wave lasers are the tightest link right now, that 100mW is tight too, that 200mW is not yet in volume production, and that prices have been raised. The size of the increase was not given. The same day, POET was asked directly for the output power of Blazar and answered that it is a question of system-level balance. No figure. On the 11th, POET's own closing post again said high-power. Again no figure. One thing did get said plainly. On the closing day Sivers said the substrate underneath all of it is itself in shortage. Meanwhile the capacity schedules are public and dated. Sivers put 30 million dollars into its Glasgow fab on 3 September, taking annual capacity past 100 million continuous-wave DFB lasers, in production from the fourth quarter of 2027. Coherent's ultra-high-power line for co-packaged optics is ramping in Texas and Sweden with revenue expected from the December quarter. So the spending plans are on the record and the specification is not. How many watts a line can actually ship is what moves an external light source business off the pluggable curve and onto the AI curve. Three days, three chances, and nobody said it. Nobody was evasive either, and a company is free not to publish a specification. What I can report is that it was asked for and not given, and the absence is itself the reading. The date I am watching is whichever comes first: an OSAT announcing co-packaged optics capacity, or someone publishing a wattage. $COHR $LITE $SIVE $CRDO $ALAB $AAOI, and 3711 TT with $AMKR on the assembly side.

I went through TSMC's 2029 packaging roadmap looking for what got cut in the write-ups. The two lines that got cut are the ones with a bill attached. The presentation was at a SEMICON Taiwan forum on 1 September, by TSMC's advanced packaging R&D director. What travelled: CoWoS package size going from 3.3 reticles to more than 14 by 2029, compute transistors inside one package up about 48 times, and HBM bandwidth per package up more than 34 times, from eight HBM3 stacks to twenty-four HBM5E. What did not travel is on the same slides. Package power on that path goes from about 600W to 4,100W, and over the same stretch power delivery losses rise more than fivefold. That second one is the sentence worth stopping on. Not the power the thing draws, the power you lose putting it in. I checked the leading English write-up word by word before saying this: it contains none of 600W, 4,100W, power delivery, underfill, or microbump. The cooling answer was laid out as a ladder. Today, a cold plate on top with thermal interface materials underneath. Next, microchannels built into the package lid to shorten the path. After that, jet impingement and two-phase boiling. Microchannel cooling is now on the R&D roadmap, and the stated difficulty is manufacturing, because cutting coolant channels into an expensive chip means a small error scraps it. The other thing that did not travel is the bump. TSMC is staying on conventional microbumps for HBM rather than moving to hybrid bonding, and has asked suppliers to develop roughly 5 micron class bumps. Korean and Japanese suppliers have started. Mass production is estimated for the second half of 2028. And the reported gate is not the bump at all. It is the underfill that has to hold a 5 micron bump and pass qualification with it. So the published roadmap is a foundry document, and underneath it sit three schedules that belong to other people: cooling, power delivery, and bonding materials. The first set of numbers tells you how big the package gets. The second set tells you who has to spend to make it work. 3711 TT and $AMKR on assembly, $TER and $FORM on test, $MU and $SKHY on the twenty-four stacks, and an underfill supply chain almost nobody covers. What I would watch is the underfill qualification, not the bump pitch. That is the date that moves everything else.

I went and read the Chinese version of the Micron story. The wire was accurate. The translation was not. Reuters on Friday: Micron's Taiwan workers to get rewards worth up to 68 months of pay. It went everywhere. Someone with fifty thousand views called it fake news. It is not fake. But two things in the Chinese original did not make the trip. First, the unit. The company's figure reads 整體薪酬, total compensation. Taiwan production staff get 35 to 68 months of total comp, meaning base pay and bonus and stock added together. In English that became rewards worth up to 68 months of pay, which reads as a bonus on top of salary. Same number, different thing. Second, the counterparty answered, and the answer was no. The same day, in Chinese, the union called it no different from the insulting bonuses handed out before, and said that if the company still does not address its demands the union will keep moving toward a strike. The strike option has not been withdrawn. The numbers underneath it. The union asked for a one-off FY26 bonus of about 83 months of salary, on top of pay. From FY27 it wants the Incentive Pay Plan scrapped and replaced with 15 percent of operating profit, paid quarterly, benchmarked openly against peers. In its own words, what it is fighting for is not more bonus or more stock, but a fair share of the result and protection by a system when the cycle turns. Scale: about 15,000 staff across Taoyuan and Taichung, close to 10,000 of them union members, and an August ballot in which over 80 percent of voting members backed a strike. What makes it price. Micron's June quarter revenue was 41.456 billion dollars against 9.301 billion a year earlier, from the 10-Q. Up 346 percent, and the run into it was 13.6, then 23.9, then 41.5. Taiwan is a key Micron HBM site. A stoppage there would not land in a normal memory market. It would land in the tightest one on record. I have no idea who wins the bargaining. What I can say is that in English this is a settled story and in Chinese it is an open one, and only one of those is true. $MU $SNDK $SKHY

Two research shops I read merged this week, and the interesting part is not the price, which was not disclosed. Read what each one says it sells, on its own About page. SemiAnalysis: product-first research from the ground up, we start with the technology itself, the actual chips, systems, and models. It runs its own hardware teardown lab. Citrini Research: as a subscriber you will never have to ask what's the trade, and the Citrindex core model portfolio goes to every paid subscriber, updated weekly. One stops at the fact, the other picks up at the trade. They were two firms and now they are one, and van Geelen, who sold, stays on as chief executive and is separately raising a fund. Nobody has to behave badly for that to matter. It just means a reader carries a third question now. It used to be, is this technically right, and separately, is this trade any good. From here you also have to ask when it went into the book. The second-order piece is the one I actually care about. The best teardown lab in the business is now wired straight into a position book that updates every week. That shortens the distance between finding something in a fab and expressing it across assets. Which means a supply-chain fact is worth slightly less than it was on Thursday, and finding it first is worth slightly more. Disclosure, since it would be odd not to: we read SemiAnalysis and we quote it, and that is not changing. What changes is the label we put on the source, the same label we would put on any research shop with a fund at the end of it.

Hollow core had a good week in English. I spent it reading three Chinese write-ups from one forum in Shenzhen, and they are more specific about the technology than anything I have read on it this year. Same stage, 11 September, three speakers, all three write-ups published within half an hour of each other. The buyer first. Han Liuyan of China Mobile's research institute calls anti-resonant hollow core a complete disruption of optical communications, and says loss reduction is essentially finished and now sits far below the solid-core limit. Then he lists what is not finished. CO2 absorption badly hurts transmission and the two known fixes, sealing and positive pressure, are in his words neither an ultimate solution. The loss spectrum is rougher than solid core and degrades after transmission. Hollow-core fibres with different cladding unit counts have high connection loss and are not suited to being mixed at random. He had good news too: two years of monitoring on the Wuxi line, gas evenly diffused, fibre and splice loss stable, without purging. Then the standards body. Ao Li, a vice president of CAICT, listed five engineering problems. The first is the one I keep coming back to. A single preform draws less than 100 kilometres. Lifetime cannot yet be assessed. In live networks splice time is long and loss is high, and construction and maintenance norms do not exist yet. Then the toolmaker, which is the part nobody covers. Fujikura's Zhao Lin: hollow core has an internal microstructure, so splicing needs rotational alignment and the structure must be kept from collapsing. Conventional four and six-motor splicers align cladding or core only. Hollow core needs an eighth motor. That is a specialist machine, and it is not what is in the vans. Now put the preform number next to the volumes. CAICT puts global hollow-core demand at 800 thousand fibre-km in 2030, from a market it says is still at pilot stage. At under 100 km a preform, that is more than 8,000 preforms. Microsoft's 15,000 km is more than 150, and more again if that figure is route rather than fibre. That one line explains something I had filed as a curiosity: why the most motivated hollow-core buyer on earth stopped trying to scale the company it bought and hired Corning and Heraeus instead. Corning draws it in North Carolina. Then the clock. ITU-T opened its first hollow-core project in July, two months ago, and its scope covers interoperability, deployment and maintenance, which are precisely what the buyer and the standards body flagged. For calibration, the multicore items opened in March 2025 and are expected to publish in 2028. So here is what I take from it. Between now and a standard, the money in hollow core is not in the fibre design, because the design is the part they say is working. It is in preform and draw capacity, and in the splicer. Corning is paid on both new-fibre routes at once, drawing Microsoft's hollow core in Carolina and co-founding the four-core group in March alongside Sumitomo Electric, Fujikura and TeraHop. The disruptor is paying the incumbents to industrialise it. What would change my mind is a preform length. If anyone shows a draw well past 100 km at usable yield, five problems collapse to two and the timeline moves in. Until then I read hollow core as a datacentre and campus product, where runs are short and splices are few, and not yet as a long-haul one. $GLW $MSFT

$MU doubling HBM capacity inside a year is the headline. What I went looking for is when the wafers behind all of this actually show up, because that is a different date. First, what is being added. The Korean piece behind this has Micron going from 40 to 50 thousand wafers a month last year to about 100 thousand by year end, and puts Samsung and SK Hynix at 150 to 200 thousand each, described as an industry estimate rather than a company figure. Call it 400 to 500 thousand wafers a month of HBM across the three of them by December. Now the part that sets the price. SK Hynix wrote the mechanism into its own M15X announcement: HBM "requires at least twice as large capabilities to secure the same production as general DRAM products." A separate industry estimate puts it at four times the capacity per gigabyte. Either number says the same thing. Until new fabs arrive, every HBM wafer is paid for out of the ordinary DRAM pool. That is not a side effect of the 2026 price move, it is the cause of it. And the bill is visible. Conventional DRAM contract prices rose 90 to 95 percent in Q1, 58 to 63 in Q2, 13 to 18 in Q3, all TrendForce's own survey numbers. Still rising, but the rate is down more than eighty percent in three quarters. Over the same stretch DRAM and NAND go from 47 percent of major cloud capex this year to 68 percent next year, and long-term agreements signed from Q2 onward already carry price ceilings. When your input is two thirds of the customer's budget, the customer stops negotiating price and starts funding your competitor. Then the dates. M15X this half. Yongin tools moving in February. Micron's ID1 in Boise mid-2027, which Micron's own site backs as DRAM output in 2027. TrendForce puts the supply inflection in the second half of 2027, at a 27 percent rise in combined server DRAM and HBM bits. After that it stacks: Samsung's P5 in 2028, Micron's ID2 late 2028, SK Hynix's newly approved 38 billion dollars of fabs with nothing before December 2028. So 2027 is the first partial year of greenfield wafers and 2028 is the first full one, with the second wave landing on top of it. Where I think that breaks first is not HBM. HBM is qualification-gated and locked into long-term agreements, and TrendForce still has its contract prices up 70 to 140 percent next year. Conventional DRAM has neither protection. It is the market that got starved to feed the stack, and it is the one that stops being starved the moment greenfield wafers arrive. The 2028 risk is being discussed as an HBM story. I think it lands in DDR5 first. What would change my mind is a fab slipping. These dates move, and Samsung's P5 has already moved two years earlier rather than later. $MU $SKHY $SNDK

I pulled the recipe file behind this. The gap has a name, and it is four lines long. vLLM's official DeepSeek V4.1 Flash recipe landed 10 Sept at 06:32 UTC. AMD was in the hardware block on that very first commit, mi350x verified next to h200, gb200 and gb300, with its own image on the same 0909 build tag. That is the vLLM image, a separate artefact from whatever AMD ships on its own hub, so both things can be true at once. 18 hours later a commit signed off from an AMD email address added one environment variable, plus a comment saying why: DeepseekV41ForCausalLM does not support torch.compile. The ROCm sparse SWA backend only reports UNIFORM_BATCH, so default FULL_AND_PIECEWISE cannot start unless breakable CUDA graphs are on. So the AMD path serves this model with FULL_AND_PIECEWISE off. vLLM's own docs call that mode the default and generally the most performant setting, especially for low latency with small models or MoEs. The flag standing in for it is marked Experimental in and defaults to off. V4.1 Flash is an MoE running 8 to 16B active, benchmarked down at the low-interactivity end. That is the exact corner where the setting bites hardest. None of which shrinks the 14.8x. It dates it. The V4 generation went from every Instinct SKU marked unsupported, blocked on a Hopper-only kernel launch primitive baked into the model's own TileLang path, to all three verified. Same file, same repo. So the open question is how many weeks this one holds. Now the part I actually trade. I pulled all 190 recipes in that repo. 146 carry a hardware block, 72 list an Instinct part, and 70 of those are verified. Where one file names both vendors it is 61 of 63. At 100B parameters and up, 46 of 48. Whether a frontier model runs on AMD is close to settled. How fast it runs is not, and that is where the entire fight now sits. The memory bill sits outside that fight. On this recipe's own verified list, HBM per GPU runs H200 141GB, GB200 186, GB300 278, MI350X 288. The part losing the tokens-per-dollar argument is carrying the most memory per die. DRAM gets paid per gigabyte installed, not per token delivered, and it does not care which logo wins the kernel race. $MU $SNDK $NVDA $AMD

$SKHY closed down 3.89% in Seoul on Friday. Its Nasdaq ADS, $SKHY, is up 1.67% as I write, on the same calendar day. Same company, 5.6 percentage points apart, and neither tape is wrong. What Seoul was pricing: KOSPI down 1.8% to 6,909.91, under 7,000 for the first time in three sessions. Samsung Electronics 005930 KS down 4.09%, SK Hynix 000660 KS down 3.89%, SK Square 402340 KS down 5.20%. The Korean press put the blame on the overnight Wall Street session, on Brent at $107.95, and on a US ten-year approaching 4.97% while the Middle East conflict dragged on. Then about twelve hours passed. US August CPI came in on consensus, 3.4% headline unchanged and core down to 2.4%. Crude gave back 2.8%, with Brent at $104.57 as I write. The ten-year sits at 4.95%, roughly flat on the day. Seoul closed before any of that existed. It priced the fear version of Friday. New York got to price the resolution version of the same Friday, and the two versions are 5.6 points apart on one company. This is not an arbitrage. The time difference is structural, the ADS only listed on Nasdaq in July, and it trades nothing like the depth of the Seoul line. But it is the cost of information arrival, and it is unusually legible today because we can see both quotes on the same date. The rest of the New York session ranks cleanly in one direction. Semiconductors $SMH +2.10%, technology $XLK +1.58%, industrials $XLI +1.23%, Nasdaq 100 +1.14%, S&P 500 +1.05%, small caps +0.62%, software $IGV +0.31%, cybersecurity $CIBR +0.02%. Hardware that ships this quarter at the top, cash flows that arrive later at the bottom. The curve says the same thing from the other end. $SHY down 0.08% at the front, $IEF down 0.04% in the belly, $TLT up 0.27% at the long end. Short end sold, long end bid. That is the shape of a market pricing a hike, and futures have the 16 September move at about 71%, up from 61% before Thursday's producer prices. Which makes Friday two different trades stacked on one date. Asia traded the macro at its worst reading of the week. America traded it after the print that settled it, and after crude broke. And in memory specifically, the split inside New York is its own story: $SIMO +9.42% and $SNDK down 3.16% on the same afternoon, controller silicon and the flash itself moving opposite ways. What would change my read: Monday in Seoul. If Korean memory gaps up to catch what New York did on Friday, this was purely the clock. If it keeps falling, then the Korean tape was pricing something New York is not, and the candidates there are the won, domestic policy, or a view on memory pricing that the US names do not share yet. One session is one session, and I would not build a position on a single day's dislocation.

Everything I read out of CIOE this week was a seller talking. $LITE on laser power tiers, $COHR on ultra-high-power ramps, $POET on 6.4T, Sivers on price rises. So I went looking for a buyer, and found one who had already answered the central question in public, in May, and answered it the other way. The sellers' position, from Lumentum management at a broker conference on 9 September: "There's a great concern what happens to the laser when it's so close to the heat source … there's a strong desire to move the laser outside." Two reasons for keeping the light source external. Reliability, and heat. Chen Qin, optical network architect at Alibaba Cloud, speaking at an xPO forum in Shanghai on 28 May, in remarks the CIOE organisers published themselves: deployment of 400G and 800G optical modules, now in the tens of millions, has thoroughly validated laser reliability. The spread of liquid cooling has effectively resolved the thermal bottleneck. NPO and CPO, in his words, now fully possess the objective conditions for putting the laser inside. He then called on the industry to push laser integration forward together. Translation mine. Both of the seller's reasons, marked settled by the customer, fifteen weeks earlier. That is the part worth sitting with, because the external light source story is load-bearing right now. It is the reason indium phosphide capacity is being prepaid years out, the reason a 70 mW part and a 400 mW part are suddenly different businesses, the reason anyone is modelling ELS modules as a separate line at all. His architecture choice is more specific than the headline. Scale-out keeps pluggables, with OSFP iterating. Scale-up moves to near-package or co-packaged, and at 224G per lane and below he says NPO has ample headroom and can fully reuse the existing supply chain. CPO he calls the long-term direction, with its current obstacles named as immaturity and single-vendor lock-in. That last phrase explains something I wrote about earlier this week. If the buyer's stated problem with co-packaged optics is being locked to one vendor, then a group of suppliers writing a common pluggable socket for optical engines is not a standards exercise. It is the answer to a customer objection. And the dates do not line up the way the market has them. Alibaba Cloud lit the first 3.2T NPO system to the OIF standard at the end of 2025, plans small-volume deployment at the end of this year and volume deployment next year. Its 6.4T UPO work expects system bring-up in the second half of this year and beta by year end. Against that, CPO was dated this month by three separate parties to the end of 2027 at the earliest. Near-package is roughly a year ahead, on the customer's own calendar. What I would not do is over-read one buyer. This is a single hyperscaler, in China, with its own silicon and its own supply constraints, and the US hyperscalers sit inside a different vendor ecosystem. It was also May, and plans stated in May are not deliveries. But nobody has publicly walked it back, and the supplier arguments this week were made as though the objection had never been raised. What would change my read: a Western hyperscaler putting a comparable schedule on record, either way. Also worth watching whether VCSEL-based designs gain ground, since he specifically noted they support an internal laser and need no polarisation-maintaining fibre, which removes two of the cost lines the external route carries. Friday's close: $LITE 931.68, $COHR 305.21, $POET 8.01, $AAOI 105.89, $AXTI 66.06.

$POET closed CIOE on Friday evening saying its optical engine product lines can address the NPO market for 6.4T and above. Two days earlier, on the same stand, its own sales people were describing 800G in mass production and 1.6T in small-volume build with multiple customers running qualification. I put the two sentences next to each other, because they are not the same kind of sentence. Can address is a claim about reach. In mass production is a claim about shipment. Both are true at the same time and neither contradicts the other. What moved between Wednesday and Friday is which one the company led with, and the speed tier it chose to lead with went from 1.6T to 6.4T and above. The number that would settle it did not arrive. This whole CPO and NPO conversation has spent a year converging on high-power continuous-wave light sources, and Blazar is aimed directly at that. The closing post describes it as a next-generation high-power multi-channel hybrid laser. High-power is an adjective. There is still no figure attached to it, and the show is now over. For a sense of what such a figure would eventually have to stand beside: Lumentum has published its co-packaged external light source at 400 mW in the first generation, 350 mW per laser in the sixteen-channel version, and a higher class above 1.0 W at 25 degrees. Those are numbers on a page. Nothing comparable has been put out for Blazar. Then the balance sheet, which reorganises the whole thing. Second-quarter revenue was $569,925. Net loss $11.3m. Cash and short-term investments $796.3m, after a $400m raise in May. Divide those into each other. The cash position is worth 1,397 quarters of the current revenue line. Measured against the current quarterly loss instead, it covers roughly seventy quarters. Whatever is uncertain here, runway is not the uncertain part. Which is why I read the 6.4T line as positioning rather than as a stretch. A company sitting on seventeen years of cash at the present burn has no reason to oversell a quarter. It can talk about where the architecture reaches and let the shipping line catch up behind it. The risk in that posture is not solvency, it is duration: the gap between the capability sentence and the shipment sentence staying open long enough that the market stops paying for the first one. The one forward figure carrying a dollar sign is Lumilens. An initial order of $50m, which the company calls the first phase of a relationship that could scale past $500m cumulative over five years. At the current run rate the $50m on its own is 88 quarters of revenue. The $500m ceiling is 877. $POET traded at 7.85 mid-session, up 3.4%. What would change my read: a published output power per channel for Blazar, or the first 1.6T customer moving from qualification into a production order. Either one converts a capability sentence into a shipment sentence, and on this name that is the only conversion that counts. Revenue has now risen sequentially for six straight quarters, so the direction is already established. What is missing is the size of the step, and the company has the cash to take its time deciding when to show it.

A visitor at CIOE in Shenzhen this morning asked a manager on the $ALAB stand whether the value of retimers gets eroded as the industry moves toward optics. I went and put the answer he got next to a specification the company is already helping to write. The answer, as he wrote it down: "The number of high-speed links is going to increase exponentially. Whether it's copper, re-timed optics, or linear optics, all of them are going to increase exponentially. If you think this is crazy right now, 2027 by all accounts is supposed to be where it goes even faster after that. So you cannot go wrong by investing in any of these. You just have to invest in at least one of them." The question was about value per link. The answer is about the number of links. Those are different quantities and only one of them was asked about. I am not going to make much of that on its own. It is a stand manager, unnamed, no title given, at a trade show, and none of that is company guidance. What is worth something is that Astera Labs has already put its name to the thing that decides the answer. Astera Labs is a contributor member of the Open CPX MSA, formed in March this year. What that group is writing, in the words of Credo's own release when it joined on 31 August: common specifications covering "mechanical, thermal, electrical, optical, and management interfaces" for "an optimized optical engine with a defined pluggable socket and electrical connector system." A defined pluggable socket is the whole argument in five words. Once the socket is specified, what goes in it becomes swappable, and swappable is the mechanism by which value per unit gets competed away. That is not a prediction, it is what the MSA is for. The founding members say so directly: the stated goal is to reduce power, cost and latency. The roster is the part I would sit with. Founding members are Ciena, $COHR, $MRVL, Molex, Samtec and TeraHop. Contributors include $ALAB itself, $INTC, $QCOM, $LITE, Source Photonics, TE Connectivity, Accton and Viavi. $CRDO announced it had joined on 31 August and had still not appeared on the published list when I read it this morning. Two names are on neither list: $NVDA and $AVGO. That is the shape of it. The companies without a proprietary scale-up interconnect are standardising the socket between them. The two companies that do have one are not in the room. NVLink is NVIDIA's and nobody sells into it. An open socket is the tool everyone else uses to build domains big enough to compete with it, and the price of that tool is that nothing inside the socket stays scarce for long. So the honest version of the booth answer is that the manager is probably right about volume and the question was about something else. Link counts go with the size of the coherent domain, and in all-to-all topologies links grow faster than accelerators do. Value per link is a separate fight, and the specification his own company is contributing to is the venue for it. Worth saying plainly: absence from a roster is not opposition to a standard, and being in the room is not the same as losing pricing power. Astera is contributing rather than founding, which is a real distinction in how these bodies work. For scale, everything in this chain is up today, which is a CPI day and not a verdict on any of this. $ALAB 294.53, $COHR 310.43, $MRVL 238.11, $LITE 969.09, $CRDO 163.69, all mid-session. What would change my read: watch whether the first Open CPX specification puts the retiming function inside the socket or leaves it on the board outside. If re-timing lives in the module, the module vendors take that content. If it stays outside, the socket standardises the optics and leaves the signal conditioning where it is, and the booth answer holds for reasons better than the one he gave.

August CPI landed exactly on consensus this morning, 3.4% headline and core ticking down to 2.4%, and that is not the rescue the headline makes it sound like. The Fed is being priced for a hike next Wednesday, not a cut. In-line data does nothing to stop one. Here is what the annual numbers are hiding. Month on month, headline went from 0.1 in July to 0.4 in August. Core went 0.2 to 0.3. A core run rate of 0.3 a month compounds to roughly 3.7% a year, which is not 2%. The twelve-month figures are easing because of what fell out of the base, and the monthly prints are doing the opposite. Warsh said at Jackson Hole two weeks ago that recent readings "do not tell me that underlying inflation trends have meaningfully improved." The underlying trend is the monthly number. He also said financial conditions do not look restrictive enough. Before that speech the market had a September hike near one in three. It has been 56 to 60 percent since. And yesterday's PPI is the piece almost nobody has paired with this one. Producer prices are running 5.4% over twelve months against 3.4% at the consumer, core 4.7% against 2.4%. In August itself the two moved identically, 0.4 headline and 0.3 core on both sides. But a month ago that headline gap was 1.4 points, not 2.0. Producer prices reaccelerated from 4.8% to 5.4% while the consumer number sat still. Both sides were pushed by the same thing. Petrol up 3.9% drove over a third of the consumer increase. Diesel up 24.1% drove over a third of the producer goods increase. Consumer energy is up 16.3% on the year, petrol 27.4%, heating oil 52.0%. The standard answer is that a central bank looks through an energy shock, because rates do not produce diesel. So look at where the diesel went in yesterday's release: truck freight up 2.0%, transportation and warehousing up 2.3%. That is the shock leaving the energy line and entering services. Once it is in freight it is in everything that moves, and it stops being the kind of thing you get to look through. The bar has quietly inverted. Data now has to prove a hike is not needed, and this print did not do that. There is nothing else of size on the calendar before Wednesday. One more thing worth sitting with. Warsh's stated complaint is that financial conditions are too loose. $SPY is up 0.62% and $QQQ 0.66% pre-market on this print. Every risk rally between now and Wednesday is another data point for his own argument. That is an uncomfortable loop to be long into. The curve is already splitting on it. $TLT up 0.27% while $IEF is down 0.13% pre-market, which is the front end pricing tighter policy and the long end pricing the slower growth that follows. $XLE down 0.55%, on a day energy drove the entire print, is the odd one out. What would change my read: strip the energy pass-through out and watch what core does. If freight and transport services cool in the September PPI while diesel stays high, the shock stayed contained and the look-through argument survives. If core services keep firming while the annual headline drifts lower, then the annual number is lagging comfort and the Fed will treat it that way. And if they do hold on Wednesday, watch whether the statement leans on energy as the reason, because that tells you they have decided this is supply rather than demand.

$SIVE raised prices, confirmed at its own CIOE booth this week, and channel checks put lasers across the 70 to 200 mW band up on ASP. That part is real and I am not going to argue with it. I spent this afternoon putting every published power figure in this chain onto one axis instead, and the rise is landing one curve below the curve the stock is being sold on. Lumentum said the same thing about pricing two days before the show opened, in flatter language. Laser supply "is not really catching up," they do not see the supply demand gap closing "within the next couple of years," and they are allocating capacity based on price. Every deal they sign today is at equal or higher pricing than the escalator they already had in place. So the price rise describes the whole layer right now. It is not something happening to one company in it. What separates the names inside that layer is power. Same call: "It's not going to be 100-milliwatt laser. It's going to be 150-milliwatt lasers or 200-milliwatt lasers." And then precisely, for all NPO and CPO work, "120, 150-milliwatt and above, which is not the traditional 70-milliwatt laser piece, which is used in today's pluggable market." Sivers' own site lists exactly one in-sale product with a power figure attached. The DR8 1310nm at 70 mW, described as built for 800G and above transceivers. The eight-channel array is over 65 mW per channel. The other two chipsets say high power and give no number at all. Lumentum's CTO stood up in Taipei on 31 August and specified its CPO external light source three ways: 400 mW first generation at roughly 20 percent wall-plug, 350 mW per laser in the sixteen-channel DWDM version, and a higher class above 1.0 W at 25 degrees. Coherent is not publishing milliwatts, it is just building them. Ultra-high-power CW lasers for CPO are ramping in Texas and Sweden right now, revenue expected in the December quarter, including volumes tied to its Nvidia partnership. Lumentum saying it sees nobody else in high power is a competitor's line and it flatters them, so take it at a discount. The voice in this with nothing to sell is Lightmatter's Nicholas Harris, a buyer of these things, at that same Taipei summit: the laser engineers at Coherent and Lumentum "have taken the power to 400 mW," and 400 mW "is already very close to the damage threshold of the fiber." Sivers is early on that side rather than absent from it. A USD 3.4m program with SemiNex announced 13 August covers high-power external sources and SOA gain stages, customer sampling and early production targeted for the second half of 2027. On 3 September it put USD 30m into expanding the Glasgow fab past 100 million CW DFB lasers a year, operational Q4 2027. So Sivers has published a capacity number and no power number for the AI tier, while Ligitek in Taiwan published a power number, 100 mW today and 400 mW targeted for 2027, and no capacity number. Each has shown half its hand. The wait costs money. Q2 on 27 August: net sales SEK 53.8m, down 12% year on year. Operating cash flow minus SEK 70.0m. Equity per share from 3.55 down to 2.97. They raised roughly SEK 825m in directed issues during the quarter and converted a USD 12m loan to equity after it closed. Solvency is not the question, they have the cash. Dilution is, and it keeps showing up ahead of the revenue. Management's own horizon is a product business in 2027 and the long-term financial model from 2028. One line in that report has stayed with me. SEK 42.9m of the EBITDA loss is a non-cash social security charge "related to the strong appreciation of the company's share price during the quarter." The share price going up is itself making the accounts look worse. The bull case is not hiding either, which I respect. Asked this week whether he still sees all-time highs, Serenity said all it takes is a surprise announcement along the lines of Sivers powers Innolight, and added, "Could be anytime or never, who knows." That is an honest description of a catalyst trade. It is a different object from a spec sheet. What would change my read: one published power figure from Sivers at 150 mW or above with a named customer against it. That single number moves the company from the pluggable curve to the AI curve and most of what I have written here falls over. Short of it, I would rather own the price rise where it is already converting into shipped product, $LITE at 946 and $COHR at 296 pre-market against $SIVE at 30.76 in Stockholm. And the argument has a date on it regardless: Marvell's CTO put co-packaged optics ramping "at the earliest by end of 2027," Cisco said it starts at the 3.2T generation, and Ligitek dated its own 3.2T to that same quarter. Everyone is building toward one window.

$NVDA guided fiscal 2028 to roughly 70% revenue growth on 26 August. I went back to that call this week, because the sentence sitting directly after the number is the one doing the work, and almost nobody quotes it. Colette Kress: "This is a supply-constrained outlook." Huang, on the same call: "even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%." So 70 is what the supply chain can physically hand over. The demand number lives somewhere else. At Goldman Sachs on Wednesday the interviewer put a size on that somewhere else, unconstrained demand growth of over 100%, and Huang went straight into answering the question without correcting the figure. Call it 30 points of demand that exist and do not get filled. Those 30 points do not evaporate. They sit as pricing power with whoever is holding up the line. Three minutes before that exchange he had already listed who that is, unprompted and in order. Packaging. DRAM. LPDDR DRAM. Connectors. Voltage regulators. Wafers. He said "everything is challenging" twice while working through it, which is not a phrase you use about a supply chain you have solved. Four of the six are places we already sit. Packaging is $AMKR and ASE at 3711 TT. DRAM and the LPDDR line are $MU, $SKHY and Samsung. Wafers is partly $GFS, $AXTI and $IQE, though the actual silicon underneath is Shin-Etsu and SUMCO and neither of those trades here. Two of them we hold nothing in, and that is the part I have been chewing on since. Connectors is $APH and $TEL. Voltage regulators is $MPWR, Infineon and Renesas. Both got named in the same breath as DRAM by the single largest buyer of all of it, and neither is anywhere near as crowded a trade. $MU at 989 and $SKHY at 191 overnight are not undiscovered. $APH at 81 and $AMKR at 50 are a different conversation. Downstream he switched vocabulary to land, power and shell, and said the neoclouds go and secure it because the hyperscalers have already exhausted theirs. $BE at 267 sells the only item on that whole list that cannot be shipped in from anywhere. The detail I keep returning to is why they guided a full year out in the first place. Huang said flatly they have never done it: "we've never forecasted or never guided to a year in advance." The reason he gave was that customers, shareholders and the supply chain should all see the same view. Read that again as what it is. A capacity signal pointed upstream, telling six industries how much to go and build. Worth being straight about what is not proven. The 100% is the interviewer's framing, and declining to correct a number is thinner evidence than stating one. A company telling you its suppliers are tight is also a company pre-explaining a future miss. And the fat part of this list has been obvious to everyone for a year. What would change my read: if NVIDIA lifts the 70% at any point before fiscal 2028 closes, the constraint eased and the rent compresses with it, so watch which line they credit when they do it. If instead they are still saying supply-constrained twelve months from now, the six names have a longer runway than their multiples currently assume. Cleanest single tell is $MU's print in late September, because DRAM is the one he bothered to say twice.

$CRWD shipped its AI security stack nine days ago with $NVDA's CEO standing on the stage next to George Kurtz. I read the three model partnership announcements side by side this morning, and the thing that stands out is which model got handed the work that never stops. At Fal.Con on 1 September, CrowdStrike launched SafeMind. An offensive model called Red Tempest finds the attack path, a defensive model called Blue Solano closes it, and the harnesses run the two against each other in a loop until nothing gets through. Blue Solano is a fine-tuned NVIDIA Nemotron 3 Super. Nemotron 3 Ultra orchestrates the defensive harness. Open weights, both of them. Huang's words on stage for Nemotron were "completely free." The next day CrowdStrike announced OpenAI, and Anthropic, and on a headline scan the three look like the same kind of news. Put them next to the product and they are three different things. GPT-5.6 Cyber arrives through the FAIRR service to assess risk and rank what to fix, with expert oversight in the loop. The Anthropic one is procurement: their customers can put part of their existing Anthropic commitments toward Falcon. No Anthropic model runs inside the product. So the closed frontier model gets called when a human asks it something, and the open model is the one running continuously. CrowdStrike's own framing of this is blunt. Frontier labs can tell a defender a risk exists. The harnesses can act on it. Cost is why. CrowdStrike says Blue Solano beats leading frontier models on accuracy at 99% lower cost, alongside 29% higher detection and 6x faster remediation. Take those carefully. They are internal evaluations, no baseline model named, no task set published, nobody outside has reproduced them. The direction is still the part that matters. Anything that has to run around the clock lives or dies on unit cost, and security is the first large AI use case where continuous is the requirement rather than a nice feature. That is also why Nemotron is free. Free model, paid compute, and the whole point of taking the per-loop cost down 99% is that you then go and run a great many more loops. Huang's own sentence in the release: cyber defense will be among the most compute-intensive applications of AI. $CRWV is in there as well, named for both training and inference. Nine days later at Goldman Sachs he was asked about cybersecurity and said the industry is getting ready to launch some products. Then: "what better way to create demand than to create a problem?" And that there are responsible ways of doing it and less attractive ways. He opened the event where those products launched. Which makes the line worth more than a gotcha would be, because he is describing a mechanism he is standing inside of. Worth holding next to the two numbers used to frame the threat at that same event, an 89% rise in AI-enabled attacks and a 27-second breakout time, both from the company selling the fix. What would change my read: CrowdStrike says the harnesses work with frontier and open source models too. If that holds up in the field, the moat is the harness and fifteen years of incident response data, and NVIDIA's seat here is a swap rather than a lock. I want to see whether the next SafeMind release still needs Nemotron specifically, or whether customers start bringing their own model and CrowdStrike keeps the margin anyway. $PANW and $NTSK have no equivalent layer yet, which is the other half of the same question.