From DowJo — where this research becomes practice. Train your judgment before you risk your money.
Course Research · Episode 5
The Boring Component That Became the Bottleneck
The least interesting part of the machine — how did it come to set the price of everything in the chain?
The evidence-led takeaway
A bottleneck is not a property of a component. It is a moment — the moment demand arrives faster than a physical thing can be built — and while it lasts, the price of everything in the chain is set at the narrow place.
Ask Jo about this researchJo reads this note before answering. Sign in to ask Jo with this evidence attached — you come straight back here.
In brief
What this research concludes
A bottleneck is not a property of a component. It is a moment — the moment demand arrives faster than a physical thing can be built — and while it lasts, the price of everything in the chain is set at the narrow place. The economics do not stay where the story is. They landed on the unglamorous memory in a police radio, not on the famous memory in an accelerator; they landed on the company that owns the constrained capacity, not on the company that performs the constrained step. And it always ends: every previous version of this ended, and the mechanisms that end it are already visible and dated.
The judgment skill this hands over
Find the constraint, then ask the two questions that decide whether it is investable: how long does it take to relieve, and what would relieve it that nobody is watching for. A thesis you cannot break is a thesis you do not understand.
What we investigated
The least interesting part of the machine — how did it come to set the price of everything in the chain?
This is episode 5 of Investing in the AI Era, a 14-episode course. The film tells the story; this note is the research behind it, kept inspectable.
The research question the team set itself: Which constraint should this episode put at its centre? The curriculum names HBM/memory as the case study, and the founder's direction was to research the 2026 landscape first and not force an old example if the evidence had changed.
Who and what this looks at
How we tested it
Researched to be refuted, not confirmed
Candidate claims went to an independent pass instructed to refute them against primary sources. Survivors became evidence; casualties became the refused list below.
The method, in the research team's own words
Twelve independent research lenses (HBM as the canonical constraint; the spillover into conventional memory; advanced packaging and foundry; power and grid equipment; networking and optics; the upstream tools and materials one level above the famous names; pricing power and margin migration; historical bottleneck cases and how each one ended; adversarial; lead-time physics; who-pays-whom; and a dated frontier window) run in parallel against live sources, each returning candidate claims with primary sourcing. Every one of the 310 candidate claims was then sent, in batches, to an independent fact-checker instructed to REFUTE it against the primary source itself rather than to confirm it, and to default to UNSUPPORTED when it could not personally reach a source that states it. 223 came back SUPPORTED, 85 CORRECTED and 2 UNSUPPORTED; where a claim was corrected, the corrected sentence is what the episode may narrate and the original is discarded. Numeric series that go on screen were compiled separately and directly from the SEC XBRL company-concept API and from company earnings releases, not quoted from any lens's prose. Every researcher was told explicitly that the model's own knowledge ends around May 2026 and is therefore not current, and that nothing may be stated from memory. A completeness critic run over the whole campaign then found the pass's largest defect — no share price, multiple or return anywhere in ~250 claims — and the six price and valuation facts added in response were each sent to two further independent checkers, which corrected two of them.
What was screened before the case was chosen · 8 candidates
Ordinary DRAM — the memory in a police radio, a cash machine, a school computer
SELECTED as the spine. It is the episode's title made literal: HBM is famous now, so it is not the boring component. The wafers moved to HBM, and the shortage landed on the memory next door. Nineteen companies with no AI business wrote 'memory costs' into an SEC filing this year, against none in each of the three years before, and three of them state the causal chain themselves.
High-bandwidth memory as the constraint
USED, but as the CAUSE rather than the subject. Sold out a year early, sixteen take-or-pay contracts, $18bn of cash paid before shipping, and NVIDIA committing $279bn 'primarily memory'. It is the strongest bottleneck evidence in the packet — and by mid-2026 research reports HBM profitability trailing ordinary DDR5. An episode built on HBM would have been right about the physics and late about the money.
Advanced packaging / CoWoS
USED as the forecast-humility beat, not as the subject. TSMC said in July 2023 that tightness would be 'somewhat released' by end-2024 and in July 2026 that packaging was still limiting customers' growth. No CoWoS capacity number has ever appeared in a TSMC filing, so there is no series to chart — and the packaging houses' own margins (16.8%, 27.3%) make the sharper point about where rent lands.
Power, turbines and grid equipment as the migrated constraint
DEMOTED to one beat, deliberately. The supply-side evidence is strong (GE Vernova signed 20 GW and shipped 3 in a quarter) but the campaign found NO operator saying power has delayed an AI deployment — Microsoft's own July 2026 call names supply and never names power. The episode may say that supply chain is sold out; it may not say power is the binding constraint on AI compute. Grid physics is Episode 6's subject.
Optical interconnect / indium phosphide lasers
USED as one line in the accumulation beat. Coherent's chief executive locates the constraint precisely — not the transceiver, the wafer that makes the light — and the margin pattern across Lumentum, Fabrinet and Arista is a clean illustration. Too narrow to carry an episode aimed at a general learner.
NAND flash
NOT the spine, and the critic pass was right that this was arguable. SanDisk was the best-performing name in the chain over the year, and it appears in the episode as exactly that — the top bar in S11B, the least glamorous component of all. Building the whole episode on it would have needed NAND fundamentals the campaign did not verify to filing grade.
The relay — one constraint handing off cleanly to the next
REFUTED BY THE EVIDENCE and rebuilt. This was the episode's original design. Packaging never let go, memory became acute, optics and turbines joined. The instrument was changed to accumulate rather than move, and the discarded design became the scene about why forecasts of relief keep being wrong.
An episode with no valuation input at all
REJECTED LATE, by a completeness critic over the finished research. Across roughly 250 refereed claims there was not one share price, multiple or return, which would have taught 'find the constraint' with no price attached — the error Episode 2 exists to teach against. S11B was written in response and its figures verified twice more.
What the evidence says
57 claims survived refutation
Each carries its source, the date the thing happened and the date it was said — two different facts — and its limits, stated by the research team rather than left for you to discover.
- FactFrontier · Sep 2026Primary sourceF01
Nineteen different companies filed thirty-eight annual, quarterly or current reports with the SEC containing the phrase 'memory costs' in the twelve months to 1 September 2026 — against zero such filings in each of the three preceding twelve-month periods.
Event 1 Sep 2026 · Published 1 Sep 2026
Basis: Exact-phrase query "memory costs" restricted to root forms 10-K, 10-Q and 8-K, run on 1 September 2026 across four consecutive twelve-month windows: 2022-09-01 to 2023-08-31 (0), 2023-09-01 to 2024-08-31 (0), 2024-09-01 to 2025-08-31 (0), 2025-09-01 to 2026-09-01 (38 filings, 19 distinct filers). The method was validated with the control phrase "supply chain disruptions", which returned 9,144 / 7,262 / 6,991 / 7,689 across the same four windows, confirming the date and form filters behave correctly. The nineteen filers: Akamai, Alpha & Omega Semiconductor, Calix, Cisco, Diebold Nixdorf, Flex, GoPro, HP, Illumina, Inseego, Motorola Solutions, NCR Atleos, NETGEAR, Pacific Biosciences, Roku, Sirius XM, Sonos, Xerox, Zebra Technologies.
Limits: This counts one exact phrase. Companies writing 'DRAM costs', 'component costs' or 'commodity costs' are missed — HP's own 10-Q says 'memory and storage costs' and is NOT in the count — so nineteen is a floor, not a census. EDGAR full-text search covers 2001 onward, so the zeroes are genuine, but absence of a phrase can be stylistic. Filing counts do not measure the size of the economic effect.
Research note
Falsifier: the same query in twelve months returning near zero would show the shock had passed out of company disclosure.
- FactFrontier · Sep 2026Primary sourceF02
SiriusXM reported that its equipment revenue fell twenty-two per cent year over year, primarily because of higher memory costs in its hardware modules.
Source: Sirius XM Holdings Form 8-K Exhibit 99.1, second-quarter 2026 earnings release (opens sec.gov)
Event 30 Jun 2026 · Published 30 Jul 2026
Basis: Verbatim in the Q2 2026 release for the three months ended 30 June 2026: equipment revenue 'declined 22% year-over-year, primarily due to higher memory costs associated with our hardware modules.' Total company revenue was $1.62 billion, roughly flat, with subscriber revenue of $1.51 billion.
Limits: Equipment revenue is a small line inside a $1.6 billion quarter, so a 22% decline is not material to SiriusXM overall. The company does not split the decline between volume, price and deliberate de-emphasis of low-margin hardware, and does not explain the mechanism by which higher costs reduce revenue rather than margin.
- FactFrontier · Sep 2026Primary sourceF03
Motorola Solutions told the SEC it is experiencing higher costs for memory in its products, and named the cause: substantial demand in the market driven by AI.
Source: Motorola Solutions Form 10-Q for the quarterly period ended July 4, 2026 (opens sec.gov)
Event 4 Jul 2026 · Published 5 Aug 2026
Basis: Verbatim MD&A: 'experiencing higher costs for memory in our products which is a result of substantial demand in the market driven by AI.' Motorola Solutions makes land mobile radios, body cameras and command-centre systems sold principally to police, fire and public-safety agencies.
Limits: The filing does not quantify the increase in dollars or basis points of margin, and bundles memory with tariffs and export regulation as sources of supply-chain volatility. No claim is made that any equipment delivery was delayed.
- FactFrontier · Sep 2026Primary sourceF04
Diebold Nixdorf reported that higher memory cost partly offset an improvement in the gross margin on the ATMs it sells.
Source: Diebold Nixdorf Form 10-Q for the quarterly period ended June 30, 2026 (opens sec.gov)
Event 30 Jun 2026 · Published 29 Jul 2026
Basis: MD&A: product gross margin rose 110 and 80 basis points in the three and six months ended 30 June 2026 'primarily due to favorable geographic mix of ATM machines sold in the period, tariff recoveries of $10.0 and improved pricing, partially offset by higher memory cost and stronger EPOS mix.' A separate note attributes part of a $74.9 million inventory increase to higher memory costs. A second ATM maker, NCR Atleos, reported absorbing net tariff and higher memory cost impacts of roughly $11 million.
Limits: Memory is one of several named offsets and is not separately quantified in the margin bridge. Both ATM makers still grew product gross margin, so the effect was absorbable.
- FactFrontier · Sep 2026Primary sourceF05
Raspberry Pi said the price of the memory it puts in its cheapest computers had risen sevenfold in a year, and in April 2026 raised board prices by up to a hundred dollars while introducing a three-gigabyte Raspberry Pi 4 at eighty-three dollars seventy-five.
Source: Raspberry Pi Ltd company announcements by chief executive Eben Upton (opens raspberrypi.com)
Event 1 Apr 2026 · Published 1 Apr 2026
Basis: The 1 April 2026 post cites 'a seven-fold increase over the last year in the price of the LPDDR4 DRAM' and raises prices by $25 (4GB), $50 (8GB) and $100 (16GB Pi 5), introducing a 3GB Raspberry Pi 4 at $83.75. The 2 February 2026 post cites an 'unprecedented rise in the cost of LPDDR4 memory, thanks to competition for memory fab capacity from the AI infrastructure roll-out'. On 1 December 2025 the company had introduced a 1GB Pi 5 at $45. Both posts state the company expects to reverse the increases when memory prices normalise.
Limits: The sevenfold figure is Raspberry Pi's own procurement experience for one specific legacy part (LPDDR4) at its own volumes, not a published market index. A small buyer of an older memory type is unusually exposed, and LPDDR4 supply was shrinking for technology-transition reasons as well as HBM reallocation.
Research note
Falsifier: Raspberry Pi reversing the increases, which it has publicly committed to doing once component costs stabilise, would date the end of this episode precisely.
- FactFrontier · Sep 2026Primary sourceF06
NETGEAR told the SEC that the memory cost challenge is expanding to other parts of the bill of materials, and that it is experiencing modest production delays.
Source: NETGEAR Form 10-Q for the quarterly period ended June 28, 2026 (opens sec.gov)
Event 28 Jun 2026 · Published 6 Aug 2026
Basis: Verbatim MD&A: 'The memory cost challenge is expanding to other parts of the Bill of Materials ("BOM"), and we are also experiencing modest production delays given the tightening environment.' Consumer net revenue fell $8.3 million, or 9.4%, partly on 'electronics pricing pressures from rising memory costs'.
Limits: 'Modest' is not quantified and no product is named as late. NETGEAR does not identify which other components are inflating, so the spread is directional rather than measured. Consumer weakness is also attributed to general softness in home networking retail.
- FactPrimary sourceF07
Micron's own filing says that because of its more complex manufacturing process, high-bandwidth memory needs more wafers and more cleanroom space to produce the same number of bits as ordinary memory on the same technology node.
Event 28 May 2026 · Published 25 Jun 2026
Basis: Verbatim risk-factor language: 'Due to the higher performance and more complex manufacturing process, HBM requires a higher number of wafers and more cleanroom space to produce the same number of bits as conventional DRAM in the same technology node. If demand for HBM weakens and suppliers shift capacity to conventional DRAM, this could result in a significant increase in conventional DRAM supply.' SK hynix's US listing prospectus makes the parallel statement that HBM requires greater wafer input.
Limits: Neither filing gives a number. This is the filing-grade qualitative version of the trade ratio; the quantitative version comes from a conference presentation and a research firm, not from an SEC document. The same paragraph warns the mechanism runs in reverse if HBM demand weakens.
- FactFrontier · Sep 2026JournalismF08
Micron told an engineering conference in August 2026 that high-bandwidth memory currently consumes roughly three times the wafer area of ordinary DDR5 memory for the same capacity, and that the gap is widening with every generation.
Event 23 Aug 2026 · Published 24 Aug 2026
Basis: Reported from the Hot Chips 2026 presentation: asked whether newer parts narrow the gap, Sreeramaneni said it is 'definitely not getting better', putting the overhead at roughly three times the wafer area of DDR5 for the same capacity. Supporting technical detail reported in the same talk: an HBM4 die carries 256 banks against 32 on a DDR5 die.
Limits: Trade-press reporting of a conference talk, not a filing or a peer-reviewed paper. The three-times figure is a spoken approximation whose comparison base is HBM3E against DDR5, and 'wafer area for the same capacity' is not identical to 'wafers consumed per bit shipped', which also depends on yield and stack height. It must be narrated and captioned as 'roughly three times, per Micron' and never as a precise ratio.
Show the remaining 49 pieces of evidence
- FactFrontier · Sep 2026ResearchF09
The three largest memory makers are expected to put about twenty-two per cent of their memory wafers into high-bandwidth memory by the end of 2026, and to get back about nine per cent of their memory bits.
Source: TrendForce, 'Tight DRAM Supply Gives Suppliers Greater Pricing Power in HBM' (opens trendforce.com)
Event 2 Jun 2026 · Published 2 Jun 2026
Basis: TrendForce estimates for HBM wafer input among the three largest suppliers as a share of total DRAM wafer input: approximately 18% at end-2025, 22% at end-2026, 30% at end-2027. HBM bit supply as a share of total DRAM bit supply: approximately 8% in 2025, 9% in 2026, 13% in 2027. The gap between the two shares is the industry-level expression of the trade ratio.
Limits: Research-firm estimates covering the top three suppliers only, excluding CXMT and others. Wafer input is measured at wafer start and therefore leads shipped bits by the cycle time, so comparing the two shares within one year is an approximation, not an accounting identity. The 2027 figures are forecasts and are not narrated.
Research note
Falsifier: TrendForce revising the series, or a step change in HBM die efficiency or stack height.
- FactFrontier · Sep 2026Primary sourceF10
SK hynix's own US listing prospectus states that the industry-wide reallocation of manufacturing capacity to high-bandwidth memory significantly constrained the supply of PC, mobile and consumer memory, producing structural undersupply in the traditional memory market from the third quarter of 2025.
Source: SK hynix Inc. SEC-filed US listing prospectus (opens sec.gov)
Event 30 Jun 2026 · Published 29 Jul 2026
Basis: Prospectus disclosure describing the causes of the traditional DRAM price recovery since the third quarter of 2025, listing structural undersupply arising from the reallocation of capacity to HBM among them. This is the supplier's own account of the mechanism, filed with the SEC under liability.
Limits: A prospectus is written to explain a company's own market to prospective investors and has an interest in describing demand as structural. It attributes the recovery to several factors together, not to capacity reallocation alone.
- FactFrontier · Sep 2026Primary sourceF11
SK hynix said in October 2025 that it had already secured full customer demand for its entire next year of memory production.
Source: SK hynix Newsroom, 'SK hynix Announces Q3 2025 Financial Results' (opens news.skhynix.com)
Event 30 Sep 2025 · Published 29 Oct 2025
Basis: Company earnings release for the quarter ended 30 September 2025 states that the company 'has already secured full customer demand for its entire DRAM and NAND production for next year' and that negotiations with key customers on 2026 HBM supply were completed.
Limits: 'Secured full customer demand' is a company characterisation, not a disclosed contract schedule: it does not reveal price, volume, cancellation terms, or how much is binding rather than indicative. It covers all DRAM and NAND, not HBM alone.
- FactFrontier · Sep 2026Primary sourceF12
Micron has signed sixteen take-or-pay agreements with customers, typically running five years, covering roughly a fifth of the memory it expects to make over that period.
Event 24 Jun 2026 · Published 24 Jun 2026
Basis: Prepared remarks: 'we have completed 16 SCAs with customers across the data center, consumer and auto market segments'; 'Typically, these agreements have a five-year term, from calendar 2026 through the end of calendar 2030'; 'The 16 signed agreements represent roughly 20% of our DRAM volume and a third of our NAND volume over this period.'
Limits: Volume percentages are management estimates over a five-year horizon and depend on Micron's own future supply, which is not fixed. Customers are not named and the take-or-pay terms are not published. The agreements cover DRAM broadly, not HBM specifically.
- FactFrontier · Sep 2026Primary sourceF13
Micron expects to receive twenty-two billion dollars of customer deposits and related commitments under those agreements, about eighteen billion of it in cash — money paid before anything ships.
Source: Micron Technology Form 10-Q for the quarterly period ended May 28, 2026 (opens sec.gov)
Event 28 May 2026 · Published 25 Jun 2026
Basis: Verbatim: 'In connection with these strategic customer agreements, we expect to receive cash deposits and related financial commitments of $22 billion for agreements concluded to date. Approximately $18 billion of these commitments will be in the form of cash deposits.' The agreements are described as take-or-pay with binding commitments for specific volumes.
Limits: This is an expectation of amounts to be received, including agreements signed after quarter end, not cash on the balance sheet at 28 May 2026. Micron states the deposits are unrestricted, are NOT prepayments, and will be returned to customers in the latter half of the agreement term — so this is customer-funded working capital, not revenue. It must never be conflated with the separately disclosed $5 billion of remaining performance obligations, which is a different measure on a different basis.
- FactFrontier · Sep 2026Primary sourceF14
NVIDIA disclosed that its supply and capacity commitments rose from a hundred and nineteen billion dollars to two hundred and seventy-nine billion in a single quarter, and stated that they are primarily for memory and manufacturing facilities.
Source: NVIDIA Corporation Form 10-Q for the quarter ended July 26, 2026 (opens sec.gov)
Event 26 Jul 2026 · Published 26 Aug 2026
Basis: Commitments note in the 10-Q filed 26 August 2026: supply and capacity commitments increasing 'from $119 billion last quarter to $279 billion as of July 26, 2026', described as commitments for data-centre infrastructure systems, 'primarily memory and manufacturing facilities'. The same series runs $16.1bn (January 2024), $30.8bn (January 2025), $95.2bn (January 2026), $119bn (April 2026), $279bn (July 2026).
Limits: The same disclosure states that in certain instances these agreements may be cancelable, rescheduled or adjustable before firm orders are placed — so this is a commitment ceiling, not a contracted floor, and the episode narrates that caveat in the same breath. 'Primarily memory and manufacturing facilities' is not a memory-only figure.
Research note
Falsifier: NVIDIA's next quarterly filing, expected late November 2026.
- FactFrontier · Sep 2026Primary sourceF15
In the quarter to May 2026 Micron's revenue rose from nine point three billion dollars to forty-one and a half billion — while the cost of making that memory rose from five point eight billion to six point four.
Event 28 May 2026 · Published 24 Jun 2026
Basis: GAAP consolidated results for the 13 weeks ended 28 May 2026 against the 13 weeks ended 29 May 2025. Revenue $9,301M to $41,456M (+346%); cost of goods sold $5,793M to $6,400M (+10.5%); gross margin 37.7% to 84.6%.
Limits: One company, not the industry, and it blends HBM with conventional DRAM and NAND. Cost of goods sold is not a pure unit-cost measure: it reflects mix, depreciation timing, and previously written-down inventory being sold at high prices, which flatters near-term margin. It proves the size of the rent, not who ultimately paid it.
Research note
Falsifier: a subsequent quarter in which cost of goods sold rises roughly in proportion to revenue, or gross margin falls back toward 35-40%.
- FactFrontier · Sep 2026Primary sourceF16
Micron's own filing says why: in that quarter its memory prices rose in the low two-hundred-and-sixty per cent range from a year earlier, while the number of bits it shipped rose in the low twenties.
Event 28 May 2026 · Published 25 Jun 2026
Basis: MD&A year-over-year DRAM comparison: revenue up 343%, driven by average selling prices up in the low-260% range against bit shipments up in the low-20% range. The sequential comparison against the February 2026 quarter is starker still: DRAM revenue +67% on ASPs up in the low-60% range and bit shipments up only in the low-single-digit percentage range.
Limits: Micron discloses ASP and bit changes as qualitative bands, not exact figures, so these cannot be reconstructed to a precise number and must be narrated as ranges. They reflect Micron's own shipments, not industry output.
- FactFrontier · Sep 2026Primary sourceF17
Micron's gross margin went from minus thirty-two point seven per cent in the quarter to March 2023 to plus eighty-four point six per cent in the quarter to May 2026.
Event 28 May 2026 · Published 25 Jun 2026
Basis: GAAP consolidated gross profit divided by GAAP consolidated revenue per fiscal quarter, from Micron's own XBRL. Quarter ended 2023-03-02: revenue $3,693M, gross profit −$1,206M = −32.7%. Quarter ended 2026-05-28: revenue $41,456M, gross profit $35,056M = 84.6%. Both from filed statements of operations, not from non-GAAP reconciliations.
Limits: A margin percentage does not by itself prove the cause of the swing — it is consistent with price increases, cost reduction, mix shift, or the absence of the inventory write-downs that depressed the 2023 figure. It says nothing about whether the level is sustainable. Micron's fiscal Q4 2026 contains 14 weeks rather than 13, so future period comparisons are not uniform.
- FactFrontier · Sep 2026Primary sourceF18
Micron's most profitable business in that quarter was not the one that sells high-bandwidth memory to AI data centres. It was phones and PCs, at an eighty-six per cent operating margin against seventy-eight for the cloud unit.
Event 28 May 2026 · Published 25 Jun 2026
Basis: Segment table, fiscal Q3 2026, operating income as a percentage of segment revenue: Cloud Memory $13,769M revenue / $10,793M operating income = 78%; Core Data Center $11,524M / $9,519M = 83%; Mobile and Client $11,521M / $9,873M = 86%; Automotive and Embedded $4,634M / $3,493M = 75%. In the year-ago quarter the same four ran 46%, 20%, 15% and 21%.
Limits: Micron allocates indirect operating costs to segments by cost-of-goods-sold or forecast wafer share and does not allocate stock compensation, so segment margins reflect an internal allocation convention rather than standalone economics. HBM's higher wafer intensity per bit means the cloud unit's lower margin may reflect capacity cost rather than weaker pricing.
- FactFrontier · Sep 2026Primary sourceF19
Micron shut down its Crucial consumer brand after twenty-nine years, saying it did so to improve supply and support for its larger, strategic customers.
Event 3 Dec 2025 · Published 3 Dec 2025
Basis: Company press release, verbatim: 'The AI-driven growth in the data center has led to a surge in demand for memory and storage. Micron has made the difficult decision to exit the Crucial consumer business in order to improve supply and support for our larger, strategic customers in faster-growing segments.' Crucial shipments continued through February 2026.
Limits: A brand exit is a corporate decision with several possible motives, including channel economics and warranty cost, not solely allocation. It does not quantify how many bits were redirected.
- FactFrontier · Sep 2026Primary sourceF20
Zebra Technologies told the SEC it fully mitigated increased memory costs through price realization.
Source: Zebra Technologies Form 10-Q for the quarterly period ended July 4, 2026 (opens sec.gov)
Event 4 Jul 2026 · Published 4 Aug 2026
Basis: MD&A: gross margin rose to 53.0% from 47.6% a year earlier, 'primarily due to the favorable impacts of IEEPA tariff recoveries and foreign currency. We also fully mitigated increased memory costs through price realization.'
Limits: 'Fully mitigated' is Zebra's own characterisation and is not separately quantified; the same quarter also benefited from tariff recoveries and currency. Enterprise buyers on service contracts are structurally easier to pass costs to than consumers, so this is not a pure test of pricing power.
- FactFrontier · Sep 2026Primary sourceF21
HP sold one point eight billion dollars more computers than a year earlier, and kept seven million dollars of it.
Source: HP Inc. Form 10-Q for the quarter ended July 31, 2026, segment information (opens sec.gov)
Event 31 Jul 2026 · Published 27 Aug 2026
Basis: Personal Systems net revenue $11,767M against $9,931M a year earlier, up $1,836M or 18.5%. Personal Systems cost of net revenue $10,250M against $8,421M. Implied segment gross profit $1,517M against $1,510M, up $7M; segment gross margin fell from 15.20% to 12.89%.
Limits: Segment gross profit is derived by subtracting the disclosed segment cost of net revenue from segment revenue — HP does not present a segment gross profit line, and the derivation is stated on screen. Commodity costs include storage and other components, not memory alone. Currency was a favourable tailwind of roughly 1.8 points on revenue.
- FactFrontier · Sep 2026Primary sourceF22
HP's average selling price for a PC rose forty point eight per cent, and it sold fifteen point eight per cent fewer of them.
Source: HP Inc. Form 10-Q for the quarter ended July 31, 2026 (opens sec.gov)
Event 31 Jul 2026 · Published 27 Aug 2026
Basis: Three months ended 31 July 2026 against 31 July 2025: Personal Systems ASP +40.8%, PC units −15.8%; Commercial ASP +41.1% with units −13.6%, Consumer ASP +37.1% with units −19.4%. HP states separately that it 'experienced higher inflationary pressure in memory and storage costs and supply constraints in our Personal Systems business, which we anticipate will continue.'
Limits: HP attributes ASP growth to pricing actions, currency and mix together, so the 40.8% is not purely memory pass-through, and the unit decline is partly a deliberate focus on higher-value units. HP writes 'commodity costs' in the margin bridge rather than naming memory at that specific line.
- FactFrontier · Sep 2026Primary sourceF23
In one company, in one quarter, on one income statement: Samsung's chip division earned eighty-nine point two trillion won while Samsung's phone division lost seven-tenths of a trillion — which Samsung attributed to rising component costs.
Source: Samsung Electronics second-quarter 2026 results announcement (opens news.samsung.com)
Event 30 Jun 2026 · Published 29 Jul 2026
Basis: Quarter ended 30 June 2026. Consolidated revenue KRW 171.5 trillion, operating profit KRW 89.5 trillion. Device Solutions division: revenue KRW 127.5 trillion, operating profit KRW 89.2 trillion (70.0% margin). MX and Networks: revenue KRW 33.2 trillion, operating LOSS KRW 0.7 trillion, which the company attributes to rising component costs alongside solid Galaxy S26 sales.
Limits: Samsung's divisional results are internal management segments inside one consolidated entity, and inter-segment transfer pricing conventions affect where profit lands. Samsung does not disclose the price at which its chip division sells memory to its phone division. Component costs include more than memory, and the phone business also faces competitive and currency pressure independent of memory.
- FactFrontier · Sep 2026Primary sourceF24
SK hynix reported a seventy-six per cent operating margin in the June 2026 quarter — against an operating loss of seven point seven trillion won for the whole of 2023.
Event 30 Jun 2026 · Published 29 Jul 2026
Basis: Consolidated K-IFRS. Q2 2026 revenue KRW 79,318,746 million; operating profit KRW 60,542,608 million, a 76.3% margin. Q2 2025 comparison: revenue KRW 22,231,952M, operating profit KRW 9,212,851M — revenue +256.8% and operating profit +557.2% year over year.
Limits: The 6-K states these are preliminary results that may differ from final results. K-IFRS is not US GAAP, so this margin is not strictly comparable to Micron's, and operating margin sits after R&D and SG&A while gross margin does not — the two must never share an axis. Won-denominated figures embed exchange-rate movement against the dollar in which memory is priced.
- FactPrimary sourceF25
Micron broke ground on its Boise fab in September 2022. Construction of the fab began in October 2023. It expects the first memory wafer out of it in the middle of 2027.
Source: Micron Technology Form 10-Q for the quarterly period ended May 28, 2026 (opens sec.gov)
Event 30 Sep 2022 · Published 25 Jun 2026
Basis: Direct filing text: 'in September 2022, we broke ground on a leading-edge memory manufacturing fab in Boise, Idaho. Construction of the fab began in October 2023, with first DRAM wafer output projected in mid-calendar 2027.' Approximately 57 months from groundbreaking to projected first wafer; 13 months between the groundbreaking and the start of construction.
Limits: 'First DRAM wafer output' is not saleable volume production or revenue — wafers must still be yielded, the product qualified with customers, and volume ramped. The mid-2027 date is a company projection, not an achieved milestone.
- FactFrontier · Sep 2026Primary sourceF26
The whole world received forty-eight extreme-ultraviolet lithography machines in 2025 — four more than the year before.
Source: ASML Holding N.V. 2025 Annual Report (US GAAP), audited (opens ourbrand.asml.com)
Event 31 Dec 2025 · Published 1 Feb 2026
Basis: Audited annual report: 'EUV systems recognized (in units): 48; 2024: 44'. The operating results section specifies four EXE (High-NA) and 44 NXE systems recognised in 2025, against two EXE and 42 NXE in 2024.
Limits: 'Recognised in sales' is a revenue-recognition count, not units physically shipped or installed in that year. EUV unit count is not a direct proxy for capacity: newer models have materially higher throughput, so 48 systems is not simply 9% more capability than 44.
- FactFrontier · Sep 2026ResearchF27
The industry association SEMI projects memory equipment spending to rise twenty-nine per cent to fifty-two billion dollars in 2026 — and memory wafer capacity to rise from four point one million wafers a month to four point two.
Source: SEMI, 300mm Fab Outlook, 2Q26 edition (opens prnewswire.com)
Event 29 Jun 2026 · Published 29 Jun 2026
Basis: SEMI 300mm Fab Outlook published 29 June 2026: worldwide 300mm memory fab equipment investment rising 29% to $52 billion in 2026 and a further 11% to $57 billion in 2027; worldwide 300mm memory capacity reaching 4.1 million wafers per month in 2026 and 4.2 million in 2027. SEMI explicitly notes that 'effective capacity growth remains moderated by technology migration and process complexity, including advanced-node DRAM, HBM, and higher-layer NAND transitions.'
Limits: Forecasts from an industry association's subscription report, not results. Installed wafer capacity is not effective bit output — SEMI itself flags that technology migration and HBM's wafer intensity absorb capacity. The 2027 figures are forecasts and the episode narrates only the 2026-to-2027 comparison SEMI itself publishes, labelled as a projection.
- Historical analogyFrontier · Sep 2026Primary sourceF28
In July 2023 the chief executive of TSMC said he expected the packaging shortage to be somewhat released, probably towards the end of the following year. In July 2026 he said packaging capacity was so tight it was limiting his customers' growth.
Event 16 Jul 2026 · Published 20 Jul 2023
Basis: On 20 July 2023 C.C. Wei said of CoWoS tightness: 'we expect these tightness somewhat be released in next year, probably towards the end of next year.' On 16 July 2026, roughly 19 to 21 months past that target, the same executive said packaging capacity was 'so tight that now it's limiting my customers' growth.'
Limits: Not perfectly like for like: 'somewhat' is a hedged word, and demand in 2024-2026 rose far beyond what anyone forecast in 2023, so the miss reflects a demand surprise as much as a supply shortfall. The 2026 quote comes from a commercial transcript rather than TSMC's own posted document, and two independently produced transcripts render the sentence identically. A forecast being wrong is not evidence of bad faith.
Research note
This is the episode's forecast-humility beat: the reason to distrust today's confident predictions of when memory relief arrives.
- FactFrontier · Sep 2026JournalismF29
The chief executive of Coherent said on an August 2026 earnings call that indium phosphide capacity continues to be the company's primary constraint, and that it is not constrained in assembly and test capacity.
Source: Coherent Corp. fiscal Q4 2026 earnings call transcript (The Motley Fool) (opens fool.com)
Event 12 Aug 2026 · Published 19 Aug 2026
Basis: Verbatim management statement on a public earnings call by chief executive Jim Anderson: 'Indium Phosphide capacity continues to be our primary constraint...we are not constrained in the assembly and test capacity right now.' It locates the binding physical stage as compound-semiconductor wafer fabrication rather than module assembly.
Limits: A third-party transcript, not a company-published document. It is one supplier's view of its own factories and may not generalise to every transceiver maker, and it does not quantify how far short of demand the company is.
- FactFrontier · Sep 2026Primary sourceF30
In the June 2026 quarter, GE Vernova signed twenty gigawatts of new gas equipment contracts and shipped three.
Source: GE Vernova Inc. Form 8-K, Exhibit 99, second-quarter 2026 financial results (opens sec.gov)
Event 30 Jun 2026 · Published 22 Jul 2026
Basis: Company-disclosed Power segment operating metrics for the three months ended 30 June 2026, in gigawatts of heavy-duty gas turbine electrical capacity. The 20 GW signed comprises 18 GW of slot reservation agreements plus 2 GW of firm orders; separately 10 GW of existing slot reservations converted to orders, and 3 GW of physical equipment shipped.
Limits: One quarter is not a run rate: shipments were 4 GW in Q1 2026, and GE Vernova states it reaches a 20 GW annualised output rate only in Q3 2026, so the quarterly shipment figure understates full-year capability. A slot reservation agreement is a capacity reservation, not a firm order, and is cancellable on terms not disclosed.
- FactFrontier · Sep 2026Primary sourceF31
GE Vernova says it reaches a twenty-gigawatt annual output rate in the third quarter of 2026, and it has a hundred and sixteen gigawatts under contract or reserved.
Source: GE Vernova Inc. Form 8-K, Exhibit 99, second-quarter 2026 financial results (opens sec.gov)
Event 30 Jun 2026 · Published 22 Jul 2026
Basis: Chief executive Scott Strazik in the earnings release: 'we remain on track to deliver 20 GW of annual gas turbine output in the third quarter of 2026, with 24 GW in 2028, and we are implementing actions to produce 30 GW in 2030.' Combined gas equipment backlog and slot reservation agreements grew from 83 GW at end-2025 to 100 GW at end-March 2026 and 116 GW at end-June 2026.
Limits: These are company targets, not achieved output; only the 20 GW rate is described as imminent. The figure covers gas turbines only, not the transformers, balance of plant or grid connection needed to turn a turbine into delivered electricity. Slot reservations are cancellable.
- FactFrontier · Sep 2026Primary sourceF32
Lumentum, which owns the laser wafer fabs, took its gross margin from thirty-seven point eight per cent to fifty point four while its revenue doubled. Fabrinet, which assembles the modules, went from twelve point two to twelve point zero. Arista, which buys the finished parts and sells the switch, went from sixty-five point two to sixty-two point nine in a year its revenue grew thirty-eight per cent.
Event 30 Jun 2026 · Published 17 Aug 2026
Basis: Lumentum quarter ended 27 June 2026: revenue $1,006.3M against $480.7M (+109.3%); non-GAAP gross margin 37.8% to 50.4%; GAAP 33.3% to 47.4%. Fabrinet quarter ended 26 June 2026: revenue $1,315.8M against $909.7M (+44.6%); GAAP gross margin 12.0% against 12.2%. Arista quarter ended 30 June 2026: revenue $3,035.7M against $2,204.8M (+37.7%); GAAP gross margin 62.9% against 65.2%.
Limits: Three different businesses with different models: Fabrinet is a cost-plus contract manufacturer, so a stable low gross margin is its design, not evidence of powerlessness, and its return on capital can be high. Lumentum's margin expansion mixes price, mix shift and fixed-cost absorption on doubled volume, and the same quarter carried a large non-cash GAAP loss. Arista's margin also moves with customer mix and tariffs, both cited in its own filing. The comparison is a pattern, not a controlled experiment. Lumentum's non-GAAP figures are used because the company leads with them; the GAAP pair is stated alongside.
- FactFrontier · Sep 2026Primary sourceF33
In one season of filings: the memory maker kept eighty-four point six cents of every dollar; the foundry sixty-seven point seven; the packaging company sixteen point eight; the server builder seventeen point five.
Event 30 Jun 2026 · Published 27 Aug 2026
Basis: Gross profit divided by revenue, most recently reported fiscal quarter for each: Micron Q3 FY2026 ended 2026-05-28, revenue $41,456M, gross profit $35,056M = 84.56%. TSMC Q2 2026 ended 2026-06-30, 67.7% as stated verbatim in its own earnings release. Super Micro Q4 FY2026 ended 2026-06-30, net sales $11,119.777M, gross profit $1,942.631M = 17.47%. Amkor Q2 2026 ended 2026-06-30, net sales $1,897.965M, gross profit $318.590M = 16.79%. Also compiled on the same basis: NVIDIA 74.98% (quarter ended 2026-07-26), Broadcom 69.48% (2026-05-03), Samsung group 69.56% (2026-06-30), Vertiv 37.71% (2026-06-30), Dell 20.93% (2026-07-31).
Limits: FISCAL CALENDARS DO NOT ALIGN and the chart says so: the period-end dates span nearly three months, from Broadcom's 3 May to Dell's 31 July, in a market where memory prices moved materially inside that window. Gross margin is not comparable across business models — a fabless designer, a foundry, a packaging subcontractor and a server assembler capitalise and allocate costs differently, and Vertiv publishes no gross profit line so its figure is derived. Gross margin is not profit: it sits before research, selling and administrative costs.
Research note
This is the scene's whole argument: being the bottleneck STEP is not the same as owning the bottleneck CAPACITY.
- FactFrontier · Sep 2026Primary sourceF34
SUMCO, one of the two largest silicon wafer makers in the world, reported an operating loss for the first half of 2026 — in the same period it described shipments of the wafers everything else is built on as increasing significantly on AI demand.
Event 30 Jun 2026 · Published 6 Aug 2026
Basis: Six months to 30 June 2026 against the same period of 2025, Japanese GAAP consolidated: net sales ¥215,016M against ¥205,372M (+4.7%); operating LOSS ¥6,363M against an operating profit of ¥7,457M; loss attributable to owners ¥12,895M against a profit of ¥3,081M.
Limits: The loss is driven substantially by depreciation from newly ramped plants and by rising energy and labour costs, not solely by wafer prices — the filing does not itemise a price effect, so this must NOT be presented as proof that wafer prices fell. SUMCO reports on a January-December fiscal year.
- FactFrontier · Sep 2026Primary sourceF35
In the same season Shin-Etsu's electronics materials business grew operating income twenty-three per cent, and said in its own results that it had worked on increasing the volume and revising the prices.
Event 30 Jun 2026 · Published 24 Jul 2026
Basis: Three months ended 30 June 2026, Japanese GAAP. Electronics Materials segment net sales ¥279.2 billion against ¥240.2 (+16%); operating income ¥101.9 billion against ¥83.1 (+23%). Company narrative states it 'worked on increasing the volume and revising the prices' of semiconductor materials including silicon wafers, photoresist and photomask blanks.
Limits: 'Revising the prices' does not quantify the change or state direction product by product, though the context implies increases. Electronics Materials includes photoresist, photomask blanks, rare-earth magnets and other products alongside silicon wafers, so this is not a wafer-price signal on its own. Shin-Etsu and SUMCO are not identical businesses and the contrast is illustrative, not controlled.
- FactFrontier · Sep 2026Primary sourceF36
The insulating film inside a great many AI accelerator packages is made by Ajinomoto — the company that sells seasoning. The segment that contains it grew business profit sixty-three per cent in the June 2026 quarter.
Event 30 Jun 2026 · Published 6 Aug 2026
Basis: Three months ended 30 June 2026, IFRS as filed. The Healthcare and Others segment, which contains the Functional Materials business including the Ajinomoto Build-up Film interlayer insulating material used in semiconductor packages, grew business profit 63.3% year over year to ¥25.1 billion. The company raised full-year guidance citing electronic materials for AI, server and network high-performance substrates.
Limits: The Healthcare and Others segment also contains amino acids and pharmaceutical businesses, so ¥25.1 billion is NOT an ABF-only figure and overstates the film business. Ajinomoto does not disclose ABF revenue or profit separately. 'Business profit' is the company's own subtotal, not operating profit under IFRS.
- FactPrimary sourceF37
Micron's gross margin for the full year was twenty point two per cent in 2016, fifty-eight point nine in 2018, minus nine point one in 2023 and thirty-nine point eight in 2025.
Event 28 Aug 2025 · Published 1 Oct 2025
Basis: Annual gross profit divided by annual revenue, both taken as full-year duration facts from the same 10-K: FY2016 $2,505M on $12,399M = 20.20%; FY2017 41.51%; FY2018 $17,891M on $30,391M = 58.87%; FY2019 45.72%; FY2020 30.57%; FY2021 37.62%; FY2022 45.18%; FY2023 gross profit NEGATIVE $1,416M on $15,540M = −9.11%; FY2024 22.35%; FY2025 $14,873M on $37,378M = 39.79%.
Limits: Computed ratio, not a tagged fact — no XBRL concept for gross margin percentage exists. GAAP basis only; Micron's non-GAAP gross margin excludes stock-based compensation and runs a point or two higher. Micron's fiscal years end in late August or early September and do not align with calendar years. FY2020 was a 53-week year.
- FactPrimary sourceF38
Micron's quarterly gross margin peaked at sixty-one per cent in the summer of 2018 and was twenty-eight point six per cent four quarters later — thirty-two points of margin gone in a year.
Event 29 Aug 2019 · Published 1 Oct 2019
Basis: Quarterly GAAP gross margin computed from Micron's own XBRL: FY2018 Q4 61.0%, FY2019 Q1 58.3%, Q2 49.1%, Q3 38.2%, Q4 28.6%. The same series shows the 2016 trough at 17.2%, the 2023 trough at −32.7% (FY2023 Q2), and 84.6% in FY2026 Q3. Where a quarter is not separately tagged it is derived from the full-year figure less the nine-month year-to-date figure, and those four derived points are flagged in the data module.
Limits: Computed ratio, not a tagged fact. Four of the 43 quarterly points are derived by subtraction rather than read directly. Fiscal quarters do not align with calendar quarters.
- FactPrimary sourceF39
Micron's capital spending peaked in the 2019 financial year, at nine point seven eight billion dollars — one year after its margin peaked, and in the year its revenue fell twenty-three per cent.
Event 29 Aug 2019 · Published 1 Oct 2019
Basis: Annual figures as tagged in Micron's audited 10-K filings. Revenue: FY2018 $30.39bn, FY2019 $23.41bn (−23%). Capex: FY2017 $4.73bn, FY2018 $8.88bn, FY2019 $9.78bn, FY2020 $8.22bn. Full-year gross margin peaked at 58.87% in FY2018.
Limits: Cash paid for property, plant and equipment, not 'net capex' — Micron guides net of partner contributions and government incentives, so these gross figures run above its own headline numbers. One company is not the industry: Samsung and SK hynix capex would be needed to characterise the total supply response. Capex is spent across DRAM, NAND, technology transitions and construction, so the attribution to the shortage is directional.
- FactPrimary sourceF40
SK hynix earned a fifty-two per cent operating margin in 2018 and a ten per cent operating margin in 2019 — operating profit fell eighty-seven per cent in a single year.
Source: SK hynix annual results releases for 2018 and 2019 (opens news.skhynix.com)
Event 31 Dec 2019 · Published 30 Jan 2020
Basis: Company-reported K-IFRS annual results: 2018 operating profit KRW 20.84 trillion, 2019 operating profit KRW 2.71 trillion, a decline of 87%, with operating margin falling from 52% to 10%.
Limits: K-IFRS, not US GAAP. Operating margin sits after research and administrative costs and is not comparable to a gross margin. Won-denominated figures embed exchange-rate movement.
- FactPrimary sourceF41
SK hynix said exactly what had ended its own boom, in its January 2019 results: as memory demand slowed in the second half and the supply shortage was resolved, the memory market environment rapidly changed.
Source: SK hynix results release, January 2019 (opens news.skhynix.com)
Event 31 Dec 2018 · Published 24 Jan 2019
Basis: Verbatim from the company's own results release announcing a record year: 'as memory demand slowed down in the second half and the supply shortage was resolved, the memory market environment rapidly changed.'
Limits: A company's own retrospective characterisation of its market, written at the moment of a turn it did not forecast a year earlier.
- FactPrimary sourceF42
Micron's own 2019 annual report blames part of its thirty per cent memory price collapse on a shortage of processors — a shortage somewhere else in the chain destroyed the memory shortage by starving it of demand.
Source: Micron Technology Form 10-K for fiscal 2019 (opens sec.gov)
Event 29 Aug 2019 · Published 1 Oct 2019
Basis: Micron's fiscal 2019 annual report attributes the DRAM price decline partly to CPU shortages, alongside customer inventory adjustments and macroeconomic conditions. DRAM average selling prices fell approximately 30% in fiscal 2019 and a further 34% in fiscal 2020.
Limits: One of several causes the company names, not the sole cause, and the relative weight is not quantified. A company explaining a bad year has an interest in external attributions.
- Historical analogyResearchF43
Between 1996 and 2000 the amount of data a single strand of optical fibre could carry rose from about two and a half gigabits a second to about a hundred — a fortyfold increase, with no new glass laid at all.
Event 31 Dec 2000 · Published 1 Jan 2003
Basis: The article identifies three sources of long-haul fibre overcapacity, the second being the dramatic increase in the capacity of installed fibre through wavelength division multiplexing: capacity per strand rose from roughly 2.5 Gbps in 1996 to roughly 100 Gbps by 2000. Corning's own optical fibre and cable sales fell 70%, or $2.03 billion, in 2002, and the company permanently closed two international fibre plants, mothballed its Concord, North Carolina facility, and announced approximately 7,100 job cuts.
Limits: The authors write in their own names, not for the Federal Reserve Bank. Capacity per strand is not the same as deployed lit capacity, and the glut also had two other causes the same article names. This is an analogy about a mechanism, not a prediction about memory.
Research note
This is the episode's most important historical mechanism: a bottleneck can die because the INSTALLED BASE gets better, with no new capacity built at all.
- FactPrimary sourceF44
In 2021 the chips that actually stopped the world's car factories were not the newest ones. They were microcontrollers and analog parts on processes going back to eight hundred nanometres — technology from the early 1990s.
Event 31 Dec 2021 · Published 25 Jan 2022
Basis: The Commerce Department survey of more than 150 companies identified the most acutely constrained parts as microcontrollers at 40 to 250 nanometres and analog devices at 40 to 800 nanometres. The same survey found the median buyer in key sectors held under five days of chip inventory in 2021, against a median of 40 days in 2019.
Limits: A voluntary survey of self-selected respondents at one moment, not a census of the market. The node ranges are categories, not a single part. Inventory days are a median across dissimilar buyers.
- FactPrimary sourceF45
Microchip named the killer in its own annual report: among the causes of a forty-two per cent revenue decline it lists shorter product lead times, which left customers holding inventory and delaying orders.
Source: Microchip Technology Form 10-K (opens sec.gov)
Event 31 Mar 2025 · Published 1 May 2025
Basis: Microchip's annual report lists among the causes of a 42.3% revenue decline 'shorter product lead times, which factors resulted in many customers having higher levels of inventory and delaying or reducing orders.' Operating margin fell from 36.9% in the year to March 2023 to 6.7% in the year to March 2025. The company closed Fab 2 in Tempe, Arizona, announced 2 December 2024 and shut in May 2025. Over a comparable window onsemi's operating margin fell from 30.8% in 2023 to 1.4% in 2025.
Limits: One of several causes named in the filing. Lead times shortening is itself a symptom of supply catching up, so this is a description of the mechanism rather than an independent cause.
- Historical analogyPrimary sourceF46
Maersk earned thirty point nine billion dollars of operating profit in 2022 and three point nine billion in 2023 — and it published the number that would do it, inside its own record-breaking annual report: three hundred and sixty-two ships scheduled for delivery, about ten per cent of the world fleet.
Source: A.P. Moller - Maersk A/S Annual Report 2022 and Annual Report 2023 (opens investor.maersk.com)
Event 31 Dec 2023 · Published 8 Feb 2024
Basis: Maersk reported operating profit of $30.9 billion for 2022 and $3.9 billion for 2023, an 87% decline in one year; by the fourth quarter of 2025 its Ocean business reported an EBIT of minus $153 million. The 2022 annual report states that '362 vessels with a nominal capacity of 2,482k TEU are scheduled for delivery in 2023. This corresponds to around 10% of the current nominal fleet.' UNCTAD attributes the 2024 wave of deliveries directly to capacity ordered during the post-COVID cash-flow boom.
Limits: Freight is not semiconductors, and the container market was also moved by Red Sea rerouting, which added an estimated 17% to global ton-miles in 2024 and recreated scarcity without any change in physical capacity. The point is the mechanism — capacity ordered at the top arrives into the trough — not a forecast about memory.
- FactFrontier · Sep 2026Primary sourceF47
Micron's largest customer agreements generally carry a ceiling price for existing products at approximately the market price in the second calendar quarter of 2026 — the most constrained supplier on earth wrote its own upside cap into a contract.
Event 28 May 2026 · Published 25 Jun 2026
Basis: Verbatim: 'The largest agreements generally have a ceiling price for existing products that approximates the market price in the second calendar quarter of 2026, and a floor price through the term of the agreement.' Micron adds that it expects gross margins from the agreements with price bands, even at floor pricing, to yield gross margins well above its peak quarterly margins in any past cycle.
Limits: The floor-price margin expectation is a forward-looking management statement, not a result, and Micron flags it as such. A minority of the agreements have no fixed pricing or price bands. The ceiling applies to 'existing products', leaving future product generations outside the band.
- FactFrontier · Sep 2026Primary sourceF48
NVIDIA took a four-hundred-million-dollar charge for excess inventory and purchase obligations on the H200 — the chip that was rationed and waitlisted two years earlier — because demand for it diminished.
Source: NVIDIA Corporation Form 10-Q for the quarter ended July 26, 2026 (opens sec.gov)
Event 26 Jul 2026 · Published 26 Aug 2026
Basis: NVIDIA recorded a $0.4 billion charge in the first half of fiscal 2027 for excess H200 inventory and purchase obligations 'as the demand for H200 products diminished', after a $4.5 billion charge on H20 in the prior-year first half. Over the same six months NVIDIA's raw-material inventory rose from $3.807 billion to $11.341 billion while finished goods fell from $8.774 billion to $6.857 billion.
Limits: H200 demand was also affected by export-control policy and by the arrival of newer products, so 'demand diminished' is not purely a scarcity-ends story. The inventory shift is consistent with both stockpiling scarce inputs and with slowing finished-goods sell-through, and the filing does not resolve which.
- FactFrontier · Sep 2026Primary sourceF49
SK hynix reported the best quarter in the history of the memory business, and its shares fell about ten per cent on the day it reported.
Source: SK hynix Q2 2026 results (Form 6-K) and Korea Exchange price data for 28 July 2026 (opens sec.gov)
Event 29 Jul 2026 · Published 29 Jul 2026
Basis: SK hynix reported a 76% operating margin for the second quarter of 2026 on 29 July 2026 (Seoul), consolidated K-IFRS, revenue KRW 79.3187 trillion and operating profit KRW 60.5426 trillion. Its KRX-listed shares fell approximately 10% on that session.
Limits: A single day's price move has many causes — guidance, positioning, the wider market — and the episode does not assert the results alone caused it. It is narrated as an observation about expectations, never as a causal claim and never as investment advice. AN EARLIER VERSION OF THIS FACT SAID ~15% ON 28 JULY AND WAS WRONG ON BOTH THE SIZE AND THE DATE; it was corrected by two independent checkers.
- FactFrontier · Sep 2026ResearchF50
China's CXMT is estimated to end 2026 with roughly three hundred and fifty thousand memory wafer starts a month, against Micron's three hundred and seventy-five thousand.
Event 30 Jun 2026 · Published 31 Aug 2026
Basis: CXMT is estimated to reach roughly 350,000 DRAM wafer starts per month by end-2026 against Micron's approximately 375,000, and held 7.6% of the global DRAM market in the first quarter of 2026. In its first interim report since listing, CXMT reported first-half 2026 revenue up 873.64% year over year.
Limits: Wafer-start estimates are research-firm figures, not company disclosures, and CXMT's technology node and yields are not independently verified. Wafer starts are not bits and say nothing about product qualification with Western customers. CXMT's HBM position specifically is NOT established and the episode makes no claim about it.
- Counter-argumentFrontier · Sep 2026JournalismF51
A class action was filed in a California federal court in June 2026 alleging the three memory makers coordinated a shift to high-bandwidth memory as cover for curtailing ordinary memory output. The allegations are unproven.
Event 25 Jun 2026 · Published 3 Jul 2026
Basis: Filed 25 June 2026 by indirect purchasers against Samsung, Micron and SK hynix, alleging the defendants deliberately shifted toward manufacturing high-bandwidth memory for AI hyperscalers and thereby choked production of consumer memory. Micron's own risk factors separately name government or regulatory focus on elevated memory pricing as a risk.
Limits: Every figure in the complaint is an unadjudicated plaintiff allegation and must be labelled as such on screen. Filing a complaint requires no proof. The lawful explanation — that HBM is more profitable per wafer and capacity followed the money — is fully consistent with the same facts, and the episode says so.
- HypothesisFrontier · Sep 2026JournalismF52
Google published a technique in March 2026 that it says cuts the memory a model needs to hold a conversation by at least six times, without retraining.
Event 25 Mar 2026 · Published 26 Mar 2026
Basis: Reported that Google's TurboQuant cuts key-value cache memory use by at least six times, quantising to three bits without training or fine-tuning and, per Google, without compromising model quality.
Limits: Trade reporting of a company research blog post, not a peer-reviewed result or a deployed product, and the claim is Google's own. It addresses the key-value cache used during inference; it does not touch model weights or training, which is where high-bandwidth memory demand principally sits. Narrated as a hypothesis about how a constraint could ease, never as an accomplished fact.
- FactFrontier · Sep 2026Primary sourceF53
NVIDIA's own filing attributes slower consumer PC sales to elevated memory and systems prices.
Source: NVIDIA Corporation Form 10-Q for the quarter ended July 26, 2026 (opens sec.gov)
Event 26 Jul 2026 · Published 26 Aug 2026
Basis: MD&A: Edge Computing revenue of $7.2 billion, up 27% year over year, driven by strong sales of Blackwell workstations 'but partially offset by slower consumer PC sales that were tempered by elevated memory and systems prices.'
Limits: One line in one segment discussion; NVIDIA does not quantify how much of the slowdown is memory. Consumer PC demand is also affected by replacement cycles and macro conditions.
- FactFrontier · Sep 2026ResearchF54
Counterpoint Research reported in August 2026 that memory has become the most expensive component in a premium smartphone, passing the main processor.
Source: Counterpoint Research (opens counterpointresearch.com)
Event 7 Aug 2026 · Published 7 Aug 2026
Basis: Counterpoint reported on 7 August 2026 that DRAM has become the most expensive component in premium smartphones, surpassing the system-on-chip, and that low-end smartphone bill-of-materials costs rose 70% year over year as of the second quarter of 2026.
Limits: A commercial research firm's bill-of-materials estimate, not a manufacturer disclosure. Teardown cost estimates are modelled, not audited, and vary between firms. 'Premium smartphone' is the firm's own category definition.
- FactFrontier · Sep 2026ResearchF55
Measured on closing prices from 29 August 2025 to 1 September 2026, the one-year price returns ran: SanDisk up two thousand eight hundred and twenty-nine per cent, Micron up six hundred and eighty-four, SK hynix up five hundred and twenty-nine in won, Coherent up two hundred and one, Amkor up eighty-nine, GE Vernova up forty-seven, and NVIDIA up twenty-five. The most famous name in the chain delivered the smallest price return of the seven.
Event 1 Sep 2026 · Published 2 Sep 2026
Basis: PRICE return only, dividends excluded. Base 29 August 2025 (the last US session before Labor Day) against the close of 1 September 2026. SNDK $52.47 to $1,536.87; MU $119.01 to $933.44; SK hynix KRW 269,000 to KRW 1,693,000; COHR $90.47 to $272.03; AMKR $24.19 to $45.73; GEV $612.97 to $898.53; NVDA $174.18 to $217.44.
Limits: Price return is not total return: including dividends changes Micron (about +685.6%), Amkor (about +90.7%) and GE Vernova (about +46.9%). SK hynix is quoted in won; converted at ECB reference rates the same move is about +537%. These are two specific closing days and the figures move every session — SK hynix was already about 4.2% lower the next session, which alone takes its figure to roughly +503%. Seven hand-picked companies are not an index. Nothing here is a recommendation and no forward statement is made.
Research note
Falsifier: any subsequent session. This is narrated inside a dated frontier frame for exactly that reason. AN EARLIER VERSION PUT SK HYNIX AT +538.67% AND WAS CORRECTED TO +529.37% BY AN INDEPENDENT CHECKER.
- FactPrimary sourceF56
Micron traded on about seven times earnings at the end of its 2022 financial year. The following year it lost five point eight billion dollars.
Event 31 Aug 2023 · Published 1 Oct 2023
Basis: Micron's fiscal 2022 ended 1 September 2022; at that year-end the shares traded on roughly seven times trailing earnings. Fiscal 2023 produced a net loss of $5.833 billion, so there was no trailing multiple at all. Over the same span its price-to-book went from about 1.27 to about 1.74.
Limits: A trailing multiple compares a price to earnings that have already happened, and in a cyclical business those earnings are the thing that moves. That is precisely the point being taught, and the episode states it as a warning about interpretation rather than as a rule about any current price. A cross-company forward-multiple comparison was compiled and DROPPED: forward multiples come from one provider's non-GAAP consensus and are not struck on the same forecast year for every company, which is not a standard this course can put on a chart.
- FactFrontier · Sep 2026ResearchF57
Micron closed on 1 September 2026 at nine hundred and thirty-three dollars — about twenty-six per cent below the intraday high it reached on 25 June, the session after it reported the best gross margin in its history.
Event 1 Sep 2026 · Published 2 Sep 2026
Basis: Close of $933.44 on 1 September 2026, market capitalisation about $1.05 trillion on about 1.13 billion shares. The 52-week high of $1,255.00 is an INTRADAY high, reached on 25 June 2026 — the session after Micron reported an 84.6% gross margin.
Limits: The high is intraday, not a closing high, and the two are not the same measure. One company's share price over one window is not evidence about any other company or any future period.
Research note
Falsifier: any subsequent session.
6 further audited series are recorded but not drawn
Micron gross margin by fiscal quarter, FY2016 Q1 to FY2026 Q3
Source: SEC XBRL company facts, CIK 0000723125, us-gaap GrossProfit / revenue as tagged in Micron's own filings · Audit: Computed ratio, not a tagged fact. Four of 43 points are derived as full-year less nine-month year-to-date and are flagged in the data module. Zero is drawn heavier than the gridlines because the series crosses it and zero is the line between selling above cost and below it.
Micron capital spending by fiscal year against gross margin, FY2016 to FY2025
Source: SEC XBRL, us-gaap PaymentsToAcquirePropertyPlantAndEquipment and computed gross margin · Audit: Cash paid for PP&E, not net capex — Micron guides net of partner contributions and government incentives, so these run above its headline numbers, and the chart says so.
Gross margin by layer of the chain, most recently reported fiscal quarter
Source: SEC XBRL and filed 10-Q/10-K documents for Micron, NVIDIA, Broadcom, Vertiv, Dell, Super Micro and Amkor; TSMC and Samsung earnings releases · Audit: FISCAL CALENDARS DO NOT ALIGN — period ends span 3 May to 31 July 2026 and the chart prints that. Vertiv publishes no gross profit line and its figure is derived. Gross margin is not comparable across business models and is not profit.
HBM share of the top three suppliers' DRAM wafer input against its share of their DRAM bits, 2026
Source: TrendForce, 2 June 2026 · Audit: Research-firm estimate covering the top three suppliers only. Wafer input is measured at wafer start and leads shipped bits, so the two shares in one year are an approximation, not an identity — stated on the frame.
SEC filings containing the phrase 'memory costs', four consecutive twelve-month windows to 1 September 2026
Source: SEC EDGAR full-text search · Audit: One exact phrase only, so 19 filers is a floor and not a census — HP writes 'memory and storage costs' and is not counted. Control phrase 'supply chain disruptions' returned 6,991-9,144 in every window, proving the filters behave.
One-year price return by layer of the chain, 29 August 2025 to 1 September 2026
Source: Exchange closing prices, cross-checked at more than one venue per line · Audit: Price return only, dividends excluded, and the frame says so. SK hynix in won. The axis is linear and unkind on purpose — SanDisk's bar dwarfs the rest and NVIDIA's is a sliver, which is the argument. Every figure moves every session, so the scene is a dated frontier frame.
The strongest case against this
The long-haul fibre glut, 1996-2003
Carried at full strength, before the conclusion — not as a footnote.
It is the case that teaches the way a bottleneck dies that almost nobody models. Not new capacity — better use of the capacity already in the ground.
Arc: Between 1996 and 2000 the data a single fibre strand could carry rose from about 2.5 Gbps to about 100 Gbps through wavelength division multiplexing. Every strand already buried silently became forty strands. Corning's optical fibre and cable sales fell 70%, or $2.03 billion, in 2002 alone; it closed two plants, mothballed Concord, North Carolina, and announced about 7,100 job cuts.
Transferable lesson: A constraint can be relieved by a technology that multiplies the output of the installed base, with no new build at all — which is why a supply-side thesis that only watches capacity announcements is watching the wrong variable. The 2026 parallel is not asserted; it is left as the question the learner is handed.
History, under test
The strongest historical case
What the past licenses — and, stated just as plainly, what it does not.
The 2017-2019 memory cycle, told through Micron's own filed gross margin
It is the same industry, the same three suppliers and the same product as the present shortage, so it is not an analogy the learner has to accept — it is the identical machine, running one cycle earlier. And every number is in a filing.
Arc: Micron's full-year gross margin went 20.2% (FY2016) → 58.9% (FY2018) → −9.1% (FY2023) → 39.8% (FY2025), and its quarterly margin peaked at 61.0% in the summer of 2018 and was 28.6% four quarters later. SK hynix's operating margin went 52% (2018) to 10% (2019). Micron's capital spending peaked in FY2019 — a year AFTER its margin peaked, in the year revenue fell 23%.
What ended it: Micron's own fiscal 2019 annual report attributes part of the DRAM price collapse to CPU shortages: a constraint somewhere else in the chain killed the memory constraint by starving it of demand. SK hynix's January 2019 release names the other half — demand slowed and the supply shortage was resolved.
Transferable lesson: The capacity ordered at the top arrives in the trough, because the same lead time that creates the pricing power guarantees the response lands late.
What remains uncertain
What we still do not know
Stated by the research team, in full, rather than smoothed over.
- How much of Micron's 84.6% gross margin is scarcity rent and how much is previously written-down inventory being sold at high prices. Cost of goods sold is not a unit-cost measure and the filing does not separate the two.
- Whether the deceleration in contract price increases from roughly 90% to a projected 13-18% is the top of the cycle or a pause. Only later quarters can say.
- Whether CXMT's commodity capacity actually converts into qualified supply for Western customers, and on what timescale. Its wafer starts are estimated; its qualification is not disclosed.
- Whether architectural efficiency gains — quantisation, key-value cache compression, higher-capacity stacks — reduce memory intensity enough to matter economically. The published techniques address inference, not training.
- How much of the consumer price move is a genuine pass-through and how much is measured shrinkflation: CPI applies hedonic quality adjustment, so a device that costs the same but ships with less memory registers as a price increase.
- Whether packaging capacity is still binding by the same measure it was in 2023, or whether TSMC's characterisation reflects a different and larger demand base. No capacity number has ever been published.
- Whether the price returns in S11B are already stale by the time the episode is watched. They are stamped with both endpoint dates on the frame rather than 'as of', and the narration frames them as what has already happened, but a viewer six months later is looking at history and the frame has to make that obvious.
What we refused to publish
12 claims we would not say — and why
The do-not-narrate list. Some are popular; some are true but unproven; some are simply not this note’s to make. Each refusal is enforced in production, not just recorded.
- UnverifiableX01
“That the memory shortage caused Nintendo to raise the price of the Switch 2.”
Verdict: UNVERIFIABLE
Why: Nintendo raised the US price from $449.99 to $499.99 effective 1 September 2026 and stated the reason as 'various changes in market conditions, which are expected to extend over the medium to long term'. It does NOT name memory, DRAM, NAND or component costs. Attributing the rise to memory is journalist and analyst inference. Tempting, vivid, and not sourceable — so it stays out of the film.
- Partly unverifiableX02
“That HBM consumes exactly three times the wafers per bit of conventional DRAM.”
Verdict: PARTIALLY UNVERIFIABLE
Why: Micron's filing states the relationship qualitatively and gives no number. The 'roughly three times' figure comes from a spoken answer at a conference, reported by trade press, and compares wafer AREA for the same capacity between HBM3E and DDR5 — which is not the same as wafers consumed per bit shipped, since that also depends on yield and stack height. It may be narrated only as 'roughly three times, per Micron', never as a precise ratio and never on a chart axis.
- UnresolvableX03
“That the memory makers colluded, or that price-fixing has been established.”
Verdict: UNRESOLVABLE — MUST BE STATED AS UNKNOWN
Why: A class action alleging coordination was filed on 25 June 2026 and is unadjudicated. Micron has denied the allegations. Every figure in the complaint is a plaintiff construction — including a ~700% price claim that matches no supplier-disclosed series. The episode may report that the suit exists and that the allegations are unproven, and must state the lawful alternative explanation in the same breath.
- UnverifiableX04
“That memory prices will fall, or that the current margins will not last, as a prediction with a date.”
Verdict: UNVERIFIABLE
Why: The episode's argument is that every previous bottleneck ended and that the mechanisms which end them are already visible and dated. That is a structural claim about mechanisms, not a forecast. No date, no direction of travel for any price, and no implication about any share price is permitted.
- RefutedX05
“Any statement that a named company is attractive, cheap, expensive, or worth buying or selling — including 'look one layer up' framed as where to invest.”
Verdict: REFUTED BY POLICY
Why: The reasoning move the episode teaches is that the economics of a shortage do not necessarily land on the famous name, and the SUMCO-versus-Shin-Etsu pair and the packaging margin ladder are there to show that they often do not land upstream either. It is a way to ask a question, never a recommendation. Buy/sell/hold language is prohibited by the course standard.
- RefutedX06
“That the constraint has MIGRATED from packaging to memory.”
Verdict: REFUTED
Why: The tidy relay story the episode originally planned is not what the evidence shows. TSMC said on 16 July 2026 that packaging capacity is still so tight it is limiting customers' growth. The honest and more useful finding is that constraints MULTIPLIED — packaging still binds while memory became acute and substrates, optics and power joined. The episode's instrument was rebuilt to show accumulation rather than a single moving light.
Show the remaining 6 refused claims
- RefutedX07
“Any grid, siting, water, emissions or interconnection-politics teaching.”
Verdict: REFUTED BY SCOPE
Why: Episode 6's subject. Power appears here only as the newest place a queue formed, quantified just far enough to prove that constraints accumulate — GE Vernova's twenty gigawatts signed against three shipped — and is then handed forward explicitly.
- RefutedX08
“Any bubble thesis, capital-cycle verdict, or statement about whether AI capital spending is wise.”
Verdict: REFUTED BY SCOPE
Why: Episode 12's subject. This episode says the cycle has always turned and shows the mechanisms by which it turns. It says nothing about whether this one is a bubble or when it ends.
- RefutedX09
“Microsoft's component-price attribution, Amazon's memory-cost capex increase, and the SK hynix margin-versus-2023-loss contrast as scene-carrying evidence.”
Verdict: REFUTED BY SCOPE
Why: All three were narrated in Episode 4. The SK hynix contrast may be called back in one clause in the cycle scene because the episode is explicitly about the cycle, but none of the three may carry a scene, and Microsoft's and Amazon's capex sentences do not appear at all.
- UnverifiableX10
“That CXMT can make competitive high-bandwidth memory at volume.”
Verdict: UNVERIFIABLE
Why: CXMT's commodity DDR5 and LPDDR position is credible on the evidence and its wafer-start trajectory is sourceable as an estimate. Its HBM capability and its access to advanced lithography are not established at volume, and the adversarial lens explicitly recommended not building an on-screen claim on them. The episode narrates the commodity capacity only.
- RefutedX11
“That memory prices rose ~700% over four years.”
Verdict: REFUTED
Why: This figure comes from the plaintiffs' complaint and matches no supplier-disclosed average-selling-price series. Micron's own filed disclosures give banded ASP changes, and those are what the episode uses.
- Partly unverifiableX12
“A cross-company forward price/earnings comparison across the chain.”
Verdict: PARTIALLY UNVERIFIABLE
Why: A forward-multiple ladder (SK hynix ~4, Micron ~6.5, SanDisk ~7, Amkor ~17, NVIDIA ~18, Coherent ~29, GE Vernova ~43) was compiled and is genuinely interesting — the market pays least for the constraint that binds now and most for the one that has not bound yet. It was dropped because the checkers established the figures come from one provider's non-GAAP consensus and are not struck on the same forecast year for every company. That is not a basis a chart in this course can carry. The one valuation fact narrated instead is Micron's own filing-grade history.
What to watch next
Dated material, and what would make it stale
Micron's fiscal fourth-quarter and full-year FY2026 results
expected late September 2026Status: pending
What changes: The gross-margin series, the ~$100 billion remaining-performance-obligation figure, and whether cost of goods sold has begun to rise in proportion to revenue — the single test of whether the rent is durable.
Server memory contract prices projected to rise 13-18% in Q3 2026, after roughly 90% in Q1 2026
Aug 2026Status: current
What changes: The deceleration is the clearest live signal that the price shock is passing its peak; a re-acceleration would falsify it. Research-firm estimate, not a filed figure.
NVIDIA's next quarterly filing and the $279 billion supply-commitment line
expected late November 2026Status: pending
What changes: Whether the largest single expression of confidence in the memory constraint grows, holds or is exercised down — the filing itself says the agreements may be cancelable or rescheduled.
CXMT capacity additions and the antitrust class action
25 Jun 2026Status: open
What changes: Two of the five classic thesis-killers — new capacity and policy — are already in motion and dated. Neither has resolved.
GE Vernova gas equipment: 116 GW contracted or reserved against a 20 GW annual output rate
22 Jul 2026Status: current
What changes: The newest queue. It is in the episode to prove constraints accumulate, and it is handed to Episode 6.
One-year price returns across the chain, and Micron about 26% below its June intraday high on record margins
1 Sep 2026Status: current
What changes: Every session. SK hynix alone moved about 4.2% the next day, taking its figure from +529% to roughly +503%. This is the fastest-rotting frame in the episode and is dated on screen for that reason.
Ideas we borrowed, and tested
Thinkers, taken seriously enough to argue with
Claim → author → evidence → counter-argument → historical test → current relevance. Never doctrine.
Chris Miller
Chip War (2022)
That the semiconductor industry's chokepoints are physical, few, and slow to duplicate — and that this is what makes them strategically and economically decisive, rather than the frontier node everyone watches.
- Evidence:
- The 2021 automotive shortage bore this out at the opposite end of the technology scale: the parts that stopped the car factories were microcontrollers and analog devices on processes down to 800 nanometres, per the US Commerce Department's own survey.
- Counter-argument:
- Chokepoints are not permanent: the fibre case shows a constraint dissolving through better use of installed capacity, and CXMT's growth shows one dissolving through new entrants.
- Historical test:
- Every case in this packet's history lens ended within six to twelve quarters of extraordinary economics.
- Current relevance:
- Used as a framing idea for why a boring component can be decisive. Never as authority for a fast-changing factual claim.
Andrew Odlyzko
Internet traffic measurement work, cited in the Federal Reserve Bank of Richmond Economic Quarterly
That the telecom boom's governing demand forecast — internet traffic doubling every three to four months — was arithmetically impossible, and that measured growth was roughly a doubling per year.
- Evidence:
- A doubling every three months implies internet use growing seventeen-million-fold between 1994 and 2000; measured growth was roughly sixty-fourfold.
- Counter-argument:
- Being wrong about a forecast is not the same as being wrong about the technology: the fibre was real and is still carrying traffic today.
- Historical test:
- The capacity built against that forecast produced the glut that destroyed the industry that built it.
- Current relevance:
- The reason the episode treats confident demand extrapolation as the most fragile input in any bottleneck thesis. Not narrated as a claim about AI demand.
Evidence & sources
53 sources, by tier
Tier 1 is primary and authoritative — filings, regulators, official statistics. Journalism and books are attributed ingredients, never proof by reputation.
- Primary sourceSEC EDGAR full-text search (efts.sec.gov), root forms 10-K, 10-Q and 8-K, exact phrase (opens efts.sec.gov)supports F01
- Primary sourceSirius XM Holdings Form 8-K Exhibit 99.1, second-quarter 2026 earnings release (opens sec.gov)supports F02
- Primary sourceMotorola Solutions Form 10-Q for the quarterly period ended July 4, 2026 (opens sec.gov)supports F03
- Primary sourceDiebold Nixdorf Form 10-Q for the quarterly period ended June 30, 2026 (opens sec.gov)supports F04
- Primary sourceRaspberry Pi Ltd company announcements by chief executive Eben Upton (opens raspberrypi.com)supports F05
- Primary sourceNETGEAR Form 10-Q for the quarterly period ended June 28, 2026 (opens sec.gov)supports F06
- Primary sourceMicron Technology Form 10-Q for the quarterly period ended May 28, 2026, Item 1A Risk Factors (opens sec.gov)supports F07
- JournalismTom's Hardware, reporting Micron's Hot Chips 2026 presentation by HBM Design Architecture Fellow Raghu Sreeramaneni (opens tomshardware.com)supports F08
- ResearchTrendForce, 'Tight DRAM Supply Gives Suppliers Greater Pricing Power in HBM' (opens trendforce.com)supports F09
- Primary sourceSK hynix Inc. SEC-filed US listing prospectus (opens sec.gov)supports F10
- Primary sourceSK hynix Newsroom, 'SK hynix Announces Q3 2025 Financial Results' (opens news.skhynix.com)supports F11
- Primary sourceMicron Technology fiscal Q3 2026 earnings call prepared remarks, chief executive Sanjay Mehrotra (opens s25.q4cdn.com)supports F12
- Primary sourceMicron Technology Form 10-Q for the quarterly period ended May 28, 2026 (opens sec.gov)supports F13, F25
- Primary sourceNVIDIA Corporation Form 10-Q for the quarter ended July 26, 2026 (opens sec.gov)supports F14, F48, F53
- Primary sourceMicron Technology Form 8-K Exhibit 99.1 (fiscal Q3 2026 earnings release) and Form 10-Q for the quarter ended May 28, 2026 (opens sec.gov)supports F15
- Primary sourceMicron Technology Form 10-Q for the quarterly period ended May 28, 2026, Management's Discussion and Analysis (opens sec.gov)supports F16
- Primary sourceSEC XBRL company facts, Micron Technology (CIK 0000723125): us-gaap GrossProfit and Revenues, as tagged in the company's own 10-Q and 10-K filings (opens data.sec.gov)supports F17
- Primary sourceMicron Technology Form 10-Q for the quarter ended May 28, 2026, operating income by business unit (opens sec.gov)supports F18
- Primary sourceMicron Technology news release, 'Micron Announces Exit from Crucial Consumer Business' (opens investors.micron.com)supports F19
- Primary sourceZebra Technologies Form 10-Q for the quarterly period ended July 4, 2026 (opens sec.gov)supports F20
- Primary sourceHP Inc. Form 10-Q for the quarter ended July 31, 2026, segment information (opens sec.gov)supports F21
- Primary sourceHP Inc. Form 10-Q for the quarter ended July 31, 2026 (opens sec.gov)supports F22
- Primary sourceSamsung Electronics second-quarter 2026 results announcement (opens news.samsung.com)supports F23
- Primary sourceSK hynix Inc. Form 6-K filed with the SEC, preliminary results for the second quarter of 2026 (opens sec.gov)supports F24
- Primary sourceASML Holding N.V. 2025 Annual Report (US GAAP), audited (opens ourbrand.asml.com)supports F26
- ResearchSEMI, 300mm Fab Outlook, 2Q26 edition (opens prnewswire.com)supports F27
- Primary sourceTSMC Q2 2023 earnings call transcript (hosted on TSMC Investor Relations), compared with the Q2 2026 earnings call (opens investor.tsmc.com)supports F28
- JournalismCoherent Corp. fiscal Q4 2026 earnings call transcript (The Motley Fool) (opens fool.com)supports F29
- Primary sourceGE Vernova Inc. Form 8-K, Exhibit 99, second-quarter 2026 financial results (opens sec.gov)supports F30, F31
- Primary sourceLumentum Holdings Form 8-K Exhibit 99.1 (fiscal Q4 2026); Fabrinet Form 8-K Exhibit 99.1 (fiscal Q4 2026); Arista Networks Form 10-Q for the quarter ended June 30, 2026 (opens sec.gov)supports F32
- Primary sourceSEC XBRL company facts and filed 10-Q/10-K documents for Micron, Amkor and Super Micro; TSMC quarterly earnings release (opens data.sec.gov)supports F33
- Primary sourceSUMCO Corporation, consolidated financial results for the six months ended June 30, 2026 (Tokyo Stock Exchange filing) (opens finance-frontend-pc-dist.west.edge.storage-yahoo.jp)supports F34
- Primary sourceShin-Etsu Chemical Co., Ltd., consolidated financial results for the first quarter ended June 30, 2026 (opens shinetsu.co.jp)supports F35
- Primary sourceAjinomoto Co., Inc., consolidated financial results for the first quarter of the fiscal year ending March 2027 (opens finance-frontend-pc-dist.west.edge.storage-yahoo.jp)supports F36
- Primary sourceSEC XBRL company facts, Micron Technology (CIK 0000723125): annual us-gaap GrossProfit and revenue as tagged in each Form 10-K (opens data.sec.gov)supports F37
- Primary sourceSEC XBRL company facts, Micron Technology (CIK 0000723125), quarterly us-gaap GrossProfit and revenue (opens data.sec.gov)supports F38
- Primary sourceSEC XBRL company concept API, Micron Technology (CIK 0000723125): us-gaap PaymentsToAcquirePropertyPlantAndEquipment, as reported in Forms 10-K (opens data.sec.gov)supports F39
- Primary sourceSK hynix annual results releases for 2018 and 2019 (opens news.skhynix.com)supports F40
- Primary sourceSK hynix results release, January 2019 (opens news.skhynix.com)supports F41
- Primary sourceMicron Technology Form 10-K for fiscal 2019 (opens sec.gov)supports F42
- ResearchFederal Reserve Bank of Richmond Economic Quarterly, Couper, Hejkal and Wolman, on the sources of long-haul fibre overcapacity (opens richmondfed.org)supports F43
- Primary sourceUS Department of Commerce semiconductor supply-chain request-for-information results (survey of 150+ companies) (opens commerce.gov)supports F44
- Primary sourceMicrochip Technology Form 10-K (opens sec.gov)supports F45
- Primary sourceA.P. Moller - Maersk A/S Annual Report 2022 and Annual Report 2023 (opens investor.maersk.com)supports F46
- Primary sourceMicron Technology Form 10-Q for the quarter ended May 28, 2026, Industry Conditions — Strategic Customer Agreements (opens sec.gov)supports F47
- Primary sourceSK hynix Q2 2026 results (Form 6-K) and Korea Exchange price data for 28 July 2026 (opens sec.gov)supports F49
- ResearchResearch-firm estimates of DRAM wafer starts and market share, with CXMT's own Shanghai Stock Exchange STAR Market interim report (opens trendforce.com)supports F50
- JournalismPutative antitrust class action filed 25 June 2026, US District Court for the Northern District of California (opens notebookcheck.net)supports F51
- JournalismTrendForce, summarising Tom's Hardware's report of a Google research blog post on TurboQuant (opens trendforce.com)supports F52
- ResearchCounterpoint Research (opens counterpointresearch.com)supports F54
- ResearchExchange closing prices — Nasdaq, NYSE and the Korea Exchange, cross-checked at more than one venue per line (opens nasdaq.com)supports F55
- Primary sourceMicron Technology Form 10-K for fiscal 2023 and market pricing at the fiscal 2022 year-end (1 September 2022) (opens data.sec.gov)supports F56
- ResearchNasdaq closing and intraday price data for Micron Technology, cross-checked at more than one venue (opens nasdaq.com)supports F57
Complete evidence depth
The research desk
Everything the research evaluated before it was distilled — including what it rejected, and why.
This note predates DowJo’s research-preservation standard, so the pre-collapse shortlist was not kept. Everything the film was allowed to say — and refused to say — is above.
Community
Talk it through
Discuss this research
Members only · in 📡 Intelligence Watch
Challenge the argument, bring evidence, or ask what would change our mind. This note is discussed in 📡 Intelligence Watch, a DowJo room where members compare reads. Filings, news events, and what they actually mean — the discussion room for the intelligence feed.
Read and reply, then keep following the story — you come straight back to this note.
Education-first ground rules apply in every room: mechanisms and evidence, never calls.
Now train it
Take it further
Reading is where judgment starts. Practice is where DowJo measures it — and remembers what to train next.
Train it
Can you spot what the market already priced in?
Five reps on today's tape. Commit a call before the market grades it.
Train itTrain it
Need the concept first?
Learn it visually in the Academy, then come back to the evidence.
Learn itLearn the concepts underneath
Join the discussion
Challenge the argument, bring evidence, or ask what would change our mind — in the open, with other members.
MentorAsk Jo what could invalidate this thesis.
Jo reads this note before answering. Sign in to ask, with this evidence attached — you come straight back here.
Ask Jo with this evidence attachedProvenance — where this note comes from
This page is a deterministic projection of canonical research artifacts. It adds presentation and discovery; it never adds, removes or softens a finding. Question taken from the descriptor; thesis from the script. Projected 21 Sep 2026 by dowjo-research-projector v1.1.0 from origin commit c83c8797ab79.
| Artifact | Path (origin-relative) | Identity | Version |
|---|---|---|---|
| Episode registry | course/episodes.json | sha256 09032826ac14… | — |
| Evidence packet | course/research/AI05-evidence-packet.json | sha256 344fd6aee7b1… | compiled 2026-09-01 · rev 1 |
| Approved script | course/scripts/AI05-the-boring-component.json | sha256 b160cd7f539a… | v 1 |
| Episode manifest | course/manifest/AI05.manifest.json | sha256 fa038a31a421… | — |
| Approval record | renders/approved/AI05-FINAL.json | sha256 9ed26d784c83… | compiled 2026-09-20 |
| Poster still | qa/stills-AI05/RESEARCH-POSTER.png | sha256 a6a2829c3c2b… | — |
Approved master (AI05-final02-corrected.mp4): sha256 0a36f65e6499100391cc9d4cc0cd4658469d16723a860de90d948f2c38178dcb
Commercial disclosure: none. No affiliate relationship, review copy, sponsorship or publisher relationship applies to this note.
For education only. Not financial advice. No buy, sell or hold recommendations, ever.
Related research
That looks interesting too
Investing in the AI Era · Episode 1
Everyone Knows AI Is the Future. So What Do You Actually Invest In?
You can be completely right about AI and completely wrong about the investment.
Shares Artificial Intelligence, Semiconductors, AI Infrastructure, Markets
Investing in the AI Era · Episode 4
AI Is Not a Chip Trade
A dollar of AI spending does not stop at a chip.
Shares Artificial Intelligence, AI Infrastructure, Semiconductors, Corporate Margins
Investing in the AI Era · Episode 6
AI Eats Electricity — and Then the Grid Fights Back
"Electricity" is not one constraint.
Shares AI Infrastructure, Artificial Intelligence, Markets