AI Capital Flows

Earnings

Latest prints

• Lenovo (0992.HK) Q1 FY26/27 · The Best Quarter It Has Ever Had, Bought in Advance

  Record revenue of $26,943M, up 43%, adjusted profit attributable past a billion for the first time at $1,075M (+176%) — and a reported loss of $609M on a $1,690M non-cash warrant revaluation that took reported operating profit to $19M. The whole 1.8-point gross-margin gain is attributed by the company to ISG, which swung from an $85.5M operating loss to $776.9M of operating profit — more in ninety-one days than the $73M the segment earned across all of FY25/26, on 44% of its revenue, at a 20.4% incremental margin. Management says that is "business performance, not transitory factors." The balance sheet is less certain: raw materials and work-in-progress rose 118.3% year-on-year against revenue at 43% and finished goods at 45.2%, and the March-to-June build ran 41.5% against 14.2% in the same three months a year earlier. Cash generated from operations fell 54% to $718M in the quarter adjusted profit rose 176%. Inside: the acquisition that closed inside the quarter and went unmentioned in the margin commentary; the $54.0B AI-server pipeline management defines as "uncommitted or unqualified"; the recurring-revenue line that grew 10.4% while the group grew 43%; the dilution stack the release discloses in three notes and never adds up — 3,677,071,572 potential shares, 29.6% of those outstanding, with every conversion price ratcheted down on August 8; the steelman for operating leverage; and the February print that settles it. Three months ended June 30, 2026 · reported August 13.

  September 6, 2026

• Palo Alto Networks (PANW) Q4 FY26 · The Deal Bought a Number, Not a Trajectory

  NGS ARR closed the year at $9.10B, up 63%, Q4 revenue at $3.41B (+34%), non-GAAP EPS at $1.02 against $0.95, and adjusted free cash flow at a record $4.41B on a 38.4% margin — a beat across the board, and the stock fell 9.28%, $362.09 to $328.47. In June this newsletter said the 60% ARR headline was two numbers wearing one coat and could not be split until the comparisons lapped, and hand-computed organic growth at roughly 28%. FY27 is the first guide with CyberArk and Chronosphere inside both ends of the comparison, and it says NGS ARR of $11.075–11.175B, growth of 22–23% — below the 40%, 34%, 32% and 29% the standalone company printed in the four quarters before it bought anything. Inside: why the tempting read (Q4 net new of roughly $1.0B collapsing to a guided $0.45B) is seasonality rather than the story; the dilution wedge — non-GAAP net income +26.75% against EPS +7.37% on a share count that went 707M to 832M, and an FY27 guide of 23–24% revenue growth converting to 8.33–9.11% per share; the bill now that it has a size, with goodwill and intangibles of $29.03B at 59.90% of assets and $1.54B above total equity, on $19.48B of newly issued paid-in capital against $2.79B of lifetime retained earnings; the 40% FY28 cash-margin target reaffirmed while FY27 is guided down to 38.0%; deferred revenue growing 15.72% against ARR at 63%; the steelman on identity being a different market; and the one line in February that settles it. Fiscal year ended July 31, 2026 · reported after close September 1.

  September 2, 2026

• Western Digital (WDC) Q4 FY26 · The Growth Changed Hands. Price Took It From Volume.

  Revenue $3.75B (+44%), non-GAAP gross margin 54.4% — up 1,310bps and 240bps above the top of its own 51–52% guide — non-GAAP EPS $3.56 against a $3.24 consensus, FY26 free cash flow of $3.51B at a 27% margin, and a September guide of $4.1B. The stock closed −13.03% the next session at $451.52, having touched −21.51% intraday against an 8.98% expected move. The headline growth rate barely moved and its ingredients inverted: Q3’s +45% was +34% exabytes and +9% price; Q4’s +44% was +22% exabytes and +18.0% price. Cost-downs decelerated with it, 10% to ~8%. And the company owing the market 25%+ exabyte growth is not building — capex $418M against $412M at 1.12x depreciation, gross PP&E 73.81% written off, construction-in-process down 14.10% to $457M, and facility projects cancelled and impaired inside the year. Inside: why UltraSMR uplift is the customer’s decision and HAMR’s is not, the ladder already spent at two of three top accounts, three customers at 44% of revenue and the top ten at 73% from 55% two years ago, the $2,050M Sandisk mark that is 64.2% of GAAP net income and does not repeat, the $672M buyback that was described as $1 billion, the Seagate comparison put to the CEO on his own call, the steelman on an LTA book running to 2031, and the one line on October 29 that settles it. Quarter ended July 3, 2026 · reported after close August 5.

  September 3, 2026

• HPE (HPE) Q3 FY26 · The Warehouse Did the Quarter, and the Suppliers Financed It.

  Record revenue of $12.21B (+33.7%), a GAAP gross margin of 40.1% — up 1,090bps — non-GAAP EPS of $1.11 against a $0.88–$0.93 outlook, free cash flow of $958M, and raised guidance for both fiscal 2026 and fiscal 2027. The tape could not decide: HPE opened down 8.16%, traded to −11.83%, and closed +5.04% at $54.44, a 17.71% low-to-high range on one unchanged document. The obvious explanation for the margin is Juniper, and the segment table refuses it: Networking’s operating margin was flat at 22.02% against 22.07%, while Cloud & AI — the low-margin server segment — went 6.99% to 17.02% and supplied 82.2% of the $1,259M increase in segment operating earnings. Asked directly whether the server margin came from low-cost inventory, the CFO answered “a confluence of everything coming together at once” and named four causes, quantifying none. Meanwhile inventory is $11.82B, +65.06% year on year and +86.13% from the fiscal year end, inventory days ran 101.95 to 118.62 to 148.72, and the standalone quarterly cash outflow funding it is widening — $2,891M in Q3 against $2,956M across the first two quarters combined. Inside: who financed the warehouse (79.83% of it came from stretching suppliers, days payable 123.28 to 172.75), how a cash conversion cycle “improved by 1 day” while a month of inventory went on, the FY27 framework that guides Cloud & AI back to approximately 13%, Neri on DRAM crowded out by HBM, the backlog called record four ways and sized zero times, the steelman on contracted demand, and the one line in the Q4 cash-flow statement that settles it. Quarter ended July 31, 2026 · reported after close September 2.

  September 3, 2026

• Broadcom (AVGO) Q3 FY26 · The Gigawatts Are Settled. The Conversion Rate Is Not.

  Revenue $29.59B (+86%), AI semiconductor revenue $16.70B (+221% YoY, +54% QoQ) against a $16.00B guide, non-GAAP EPS $3.32 (+96%), non-GAAP operating margin a record 67.9%, free cash flow $13.67B at 46% of revenue. Fiscal 2027 AI revenue given a figure at last — approximately $115B, against June’s “in excess of $100 billion” — fiscal 2028 introduced at approximately $230B, and a target to “exceed $30 in earnings per share” in that year. The stock closed at $357.16, down 2.74%, the second red reaction day in a row on a beat-and-raise. The reason is inside the guide’s own arithmetic: $345B of two-year AI revenue against the 30 gigawatts Hock Tan confirmed as line of sight is $11.5B per gigawatt, against his own stated content of $20B to $30B per gigawatt — so the outlook already assumes only 38.3% to 57.5% of those gigawatts convert inside the window, and the gap is land, power and shell rather than silicon. Inside: the segment gross margin the CFO read out as 76% and corrected to 67%, the 500bps consolidated decline to a guided 73% on rising memory content, a fabless company guiding capex to $1.4B in one quarter — 1.38x the $1,013M spent across the first three combined and 44.53% of its entire $3,144M fixed-asset base — what “$30 in earnings per share” actually claims once you run it against 4.94 billion shares and a 16% tax rate, the $35B Apollo/Blackstone tranche and the residual-value tail nobody would size, the $20–30B of content against $30B of customer ARR per gigawatt, the steelman on operating leverage with non-GAAP R&D up 5.35% against revenue up 54.97%, and the one line on December 9 that settles it. Quarter ended August 2, 2026 · reported after close September 2.

  September 3, 2026

• Dell (DELL) Q2 FY27 · The Margin Argument Is Over. The Vendor Became the Credit.

  Record revenue of $46.97B (+58%), non-GAAP EPS $7.04 (+203%), operating margin 11.5% from 6.0%, ISG operating income $4.78B at a 15.0% segment margin, $60.9B of AI-server orders against $16.40B recognised, a record $95B backlog, and a full-year guide raised $25B to $192.0B. The tape paid 15.81% in one session, $425.00 to $492.18. In June this newsletter called the argument the margin; this print settles it and the settlement runs the other way. Gross margin went up, to 20.9% from 18.3%, and the six-quarter series — 21.1% / 18.3% / 20.7% / 20.2% / 17.8% / 20.9% — is a band averaging 19.83% with no trend in it, so the June read was one year-on-year delta mistaken for a trajectory. That correction is published here rather than buried. What replaces it is narrower: AI-Optimized Servers grew 1.67% sequentially, $16.13B to $16.40B, supplying just 8.6% of the quarter’s $3.13B revenue increase — so the margin being applauded is the margin of Dell’s least AI-weighted quarter since Q4 FY26, while the guide implies AI at 40.98% of second-half revenue. And beneath it, the cash: free cash flow fell 47% to $986M in the best quarter Dell has ever printed, with the record $8.15B of “adjusted” free cash flow resting on a $6.67B financing-receivables add-back — 5.87x the $1,135M prior peak in the five quarters before it. Inside: the line the CFO did not read out, the five-quarter add-back series, inventories at $21.29B (+104.0%) against payables of $49.72B, a record $4.3B returned on $986M of free cash flow and $4.39B of debt proceeds, the steelman on secured credit and flat half-year cash, and the one line in the Q3 release that settles it. Quarter ended July 31, 2026 · reported after close September 1.

  September 2, 2026

• CXMT (688825.SS) H1 2026 · The Plant Predates the Windfall. The Next Build Has Already Started.

  Revenue RMB 150.31B (+873.64%), net profit attributable RMB 77.61B against a RMB 2.33B loss, ex-items RMB 78.79B — higher than the headline — operating cash flow RMB 131.16B (+2,985.64%), an 84.84% main-business gross margin and an 81.06% ROE, clearing its own pre-IPO guidance by 25.3% on revenue and 36.1% on profit. The stock closed the reaction session at RMB 58.01, down 1.01%, after running 6.70x from the RMB 8.66 issue price in a month. The print is clean and the question is what it is evidence of. Cost of sales rose only 70.76% against revenue up 873.64%, which makes 84.84% a price margin rather than a process margin — and in the most profitable half in the company’s history capital expenditure fell 10.45% to RMB 21.59B, fixed assets slipped to RMB 181.56B as RMB 14.51B of depreciation outran transfers in, and the entire expansion sits off the balance sheet as RMB 52.44B of contracted-but-unrecognised commitments, +206.16% and 2.43x the half’s capex. The issuer marks its own top: DRAM ranged US$7.89/GB to US$1.78/GB across 2015–2025 and the risk factors state the sustained price increase “is not sustainable.” Inside: the attribution finding — HBM appears zero times in 192 pages, LPDDR (RMB 78.19B) outsold DDR (RMB 69.47B) by RMB 8.72B, and the filing says AI-related revenue share was “relatively low” — the 88.81% of revenue moving through distributors, the offshore revenue flip to 63.84% against RMB 96.98M of non-current assets outside the mainland, the June 8 Section 1260H designation, the steelman on cash conversion, and the one line in the next report that settles it. Half year ended June 30, 2026 · released after close August 28.

  August 31, 2026

• Marvell (MRVL) Q2 FY27 · Revenue Guide Up $1.5 Billion, Market Value Down $22.9 Billion. Read the Incremental Dollar.

  Revenue $2,739.3M (+37% YoY, +13% QoQ), $39.0M above its own guide midpoint, data center $2,171.5M (+46%) at 79% of the company, non-GAAP gross margin 58.9%, non-GAAP operating margin 36.6% (+180bp YoY), non-GAAP EPS $0.94 against $0.87 — and then two raises: fiscal 2027 to roughly $12B and fiscal 2028 to roughly $18B, a $1.5B increase to next year made one quarter after the last one. The stock closed $216.595 against $241.45, down 10.29%, about $22.90B of market value on 921.2M diluted shares. The demand is real and it is not the argument. Take the Q3 guide the company printed — $3,150M at a 58.0% non-GAAP gross-margin midpoint — and the incremental dollar arrives at a 51.9% gross margin against the 58.9% just reported, because, in the CFO’s words, “mix is the primary driver. We’ve got a strong ramp in custom.” Take the full-year raise on its own terms — $500M more revenue, $100M more operating expense — and it drops through at roughly 31.9%, beneath the 36.6% Marvell already earns and beneath the 38–40% model it says it enters this quarter. Inside: the fourth quarter buried in the full-year number ($3,692.9M implied, +17.2% sequential), the Google warrant decoded — 58,970,907 shares at $206.58, of which 57,610,040 vest in 240 tranches of $500M, making the “$120 billion deal” the denominator of a vesting ladder rather than an order book, and each rung worth 0.48% of its revenue at $216.595 but 9.29% at $400 — the eight sell-side notes that all held or raised into a 10% decline, the peer tape (NVDA −4.58%, ALAB −4.81%, CRDO −3.08%), and the gross-margin line that settles it in December. Quarter ended August 1, 2026 · reported after close August 27.

  August 28, 2026

• NVIDIA (NVDA) Q2 FY27 · Revenue Up 18%, Free Cash Flow Down 56%. Read the 15 Extra Days.

  Revenue $96,221M (+18% QoQ, +106% YoY), roughly 5.7% above its own ~$91.0B guide midpoint, Data Center a record $89,023M (+117%), GAAP and non-GAAP gross margin both 75.0%, non-GAAP EPS $2.22 (+120%), and a Q3 guide of $108.0B — another 12.2% sequential step. The stock closed $227.98 against $209.66, up 8.74%. The demand is real and it is not the argument. Operating cash flow fell from $50,344M to $24,077M (−52.2%) and free cash flow from $48,554M to $21,341M (−56.0%) against a consensus near $43B, leaving GAAP net income $35.6B above operating cash flow in a single quarter — because days sales outstanding went 45 to 60 on the CFO’s own words, “extended payment terms on large, multi-quarter agreements with certain investment-grade customers.” Roughly $15.9B of revenue recognised and not collected. The receivable and the cash-flow decline are one decision reported twice. And the mix says who got the terms: ACIE +25% QoQ and +138% YoY against Hyperscale’s +13% and +102%, so the faster-growing half of Data Center is the cohort NVIDIA itself describes as “growing faster than their balance sheets and long-term credit profiles can support.” Inside: the four instruments that moved the same cash-consuming way in one quarter ($105B of phased OpenAI lease guarantees at SB Energy’s 4.25GW PORTS-Pike, $3.5B of AI-cloud guarantees, $25B of equity commitments into model makers and infrastructure financiers, $25.0B of new senior notes), the supply book stepped $119B→$279B “primarily related to the procurement of memory” and what it does to the memory chain, the six-month number that defuses most of the alarm (OCF/revenue 47.1%→41.8%, not 61.7%→25.0%), 12.7× forward sales and 25.7× annualised earnings, the trade, and the three dated tests that settle it on November 18. Quarter ended July 26, 2026 · reported after close August 26.

  August 27, 2026

• CrowdStrike (CRWD) Q2 FY27 · Net New ARR Growth Guided Down 20 Points. The Stock Went Up 20%.

  Revenue $1.47B (+26%, a fifth straight quarter of acceleration), record net new ARR of $332.8M (+51%, more than $45M above the high end of guidance), ending ARR $5.84B (+25%), non-GAAP operating income $371.6M at a 25% margin (+350bp), free cash flow $377.4M at 26%, a third consecutive positive GAAP quarter, and a 630bp raise to the FY27 net-new-ARR growth outlook. The stock closed $227.95 against $189.18, up 20.5%. The record is real and it is not the argument. Net new ARR grew 51% and is guided to 29–31% next quarter — roughly twenty points of deceleration inside a release the tape read as acceleration. And the full-year midpoint is where the raise went: FY27 ending ARR of $6,607.45M less the Q3 midpoint of $6,186.4M implies ~$421 million of fourth-quarter net new ARR, about 22% above the Q3 guide and about 27% above the record just set. Management never said that number. Thirteen analysts asked one question each, no follow-ups, and nobody asked for it. The engine is a licensing conversion: 935 new Falcon Flex accounts, more than the prior three quarters combined, top ten deals all Flex, a 40% uplift at conversion and 25% more at the first re-Flex on an eight-month clock. That is a ladder, and the rungs are finite. Inside: the Flex conversion mechanism and what it actually monetizes, the roughly $71 million of the raise that is not flow-through and the CFO line conceding it, the bull case argued at full strength, SBC +44% against revenue +26% with capex up 4.1×, deferred revenue +1.87% over six months against 25% ARR growth, the Palo Alto / Zscaler / SentinelOne / Okta tape, 39× forward revenue against a consensus target below spot, the trade, and the dated test that settles it. Quarter ended July 31, 2026 · reported after close August 26.

  August 27, 2026

• Winbond (2344.TW) Q2 2026 · The Revenue Line Is a Staircase, and It Steps Four Times a Year

  Every line a company record — consolidated revenue NT$59.843B (≈US$1.9B, +56.4% QoQ, +184.7% YoY), a 66.2% gross margin, a 48.4% operating margin, net profit NT$24.317B and EPS NT$5.40, with the memory segment alone at a 70.3% gross margin and 100% utilisation — and the monthly sheet says how it arrived. April NT$19.245B + May NT$20.001B + June NT$20.597B ties to the quarter exactly, and the sequential shape is +32.7%, +3.9%, +3.0%, then July +30.0% to a record NT$26.773B. Two risers eleven weeks apart, both on a quarter boundary, because Winbond no longer trades spot and prices on quarterly contracts. On NT$21.593B of incremental revenue it kept NT$19.19B of gross profit and NT$16.48B of operating profit — 88.9 cents and 76.3 cents on the dollar, against ADATA’s 12.6 cents and Nanya’s 96.5 cents the same quarter — and it did it on CMS bit shipments down about 10%. Inside: why the Kaohsiung expansion nets to +4K wafers once Taichung’s 12K→7K cut is subtracted and the 2027 bit story is really a 16nm yield curve, the NT$24.9B that arrived in book value and not in earnings, the implied non-memory operating contribution of roughly zero for the second quarter running, Module B priced off the CEO’s own NT$350B/45K-wafer analogy at NT$390–470B, the Macronix / Nanya / module-maker / Si-Cap read, 7.6–7.8× forward on a stock 24% below its June 22 high that fell 4.39% the session after the record, and the three dated tests that settle it. Quarter ended June 30, 2026 · reported August 6.

  August 25, 2026

• ADATA (3260.TWO) Q2 2026 · 46% More Revenue, 4% More Operating Profit. Read the Incremental Dollar.

  The largest quarter in company history — revenue NT$38.143B (US$1.196B, +46.15% QoQ, +197.43% YoY), after-tax net profit NT$10.768B (US$337.7M), EPS NT$32.39 (US$1.02), a first half of NT$64.2B (US$2.013B) and NT$62.46 (US$1.96) that beat all of FY2025 — and every account filed it under the shortage. The shortage is real and it is not the finding. Gross margin fell from 55.69% to 42.11%, 13.58 points; operating margin fell from 47.05% to 33.47%, also 13.58 points, because the operating-expense ratio was 8.64% in both quarters. Opex leaked nothing; the entire decline is cost of goods. Run the increment and the quarter states itself: on NT$12.033B (US$377.3M) of additional revenue ADATA kept NT$1.521B (US$47.7M) of gross profit and NT$481M (US$15.1M) of operating profit — 12.6 cents and 4.0 cents on the incremental dollar against reported averages of 42.11% and 33.47%, and the derived H1 operating profit of NT$25.05B lands on the company’s own “above NT$25.0B.” Nanya, which owns the fab, took 96.5% of its incremental revenue as gross profit the same quarter. The rent belongs to whoever owns the wafer. Inside: July’s NT$18.379B record and the 15-point DRAM mix shift underneath it (61%→76.6% while SSD went 32%→20.1% into the faster-rising input price), the NT$49.4B (US$1.549B) warehouse at 204 days with no plan to cut, why a company that earned NT$19.94B (US$625.3M) in six months is raising NT$3.0B (US$94.1M) of convertibles plus 20 million shares to repay bank loans while paying NT$18 a share out, the Transcend / Team Group / Nanya read, 5.2× trailing against 3.3× on the half annualised, and the four dated things that settle it. Quarter ended June 30, 2026 · reported August 6 · USD at NT$31.889 = US$1.

  August 25, 2026

• Corsair (CRSR) Q2 2026 · The Record Margin Was a Tariff Refund; the Print Underneath Is DRAM

  Net revenue $314.3M (−2% YoY) and a company-record 33.2% gross margin, gross profit +21% to $104.3M, adjusted EBITDA +279% to $30.8M, non-GAAP EPS $0.23 against a $0.05–$0.07 guide. The stock closed +35.2% the next day. Then the release’s own paragraph: a $15.6M IEEPA tariff refund worth approximately 500 basis points of that margin — 79% of the 635bps of reported expansion — and, excluding it, a $5.7M GAAP net loss and $(0.07) per share. Ex-refund gross profit grew 3.3%, not 21%. But the durable finding is the segment split, which is a controlled experiment: Peripherals +13% to $115.9M while Components & Systems fell 9% to $198.5M on “higher memory pricing… DIY PC builds and standalone components, which shrank year-over-year.” Logitech’s gaming line grew 12% in the same quarter, so the gamer is spending — the build is what broke. Inside the falling segment, memory revenue rose 17%: dollars up, units down, pure DRAM pass-through. And the CFO gave the tell to the one analyst who asked — memory gross margin 23.4% in Q2, “roughly similar” in Q3, “high teens” in Q4, while the CEO says prices keep rising. Roughly five points of spread left. The guidance “raise” that only moved the floor ($1.33B–$1.47B to $1.40B–$1.47B, ceiling unchanged), the implied $431M / $39.9M Q4 resting partly on a GTA 6 date Corsair calls “expected,” 8.8x guided EBITDA that is 9.9x with the refund stripped, the Logitech / Micron / SanDisk read, the trade, and the nine sessions in which the tape gave back 93% of the gap. Quarter ended June 30, 2026 · reported after close August 6.

  August 20, 2026

• Nanya (2408.TW) Q2 2026 · Flat Bits, 96 Cents on the Incremental Dollar, and the NT$346.6B Bet Against Its Own Windfall

  Revenue NT$82.549B (+68.2% QoQ, +684.2% YoY), gross margin 79.5%, net profit NT$50.192B, EPS NT$14.66 — one quarter worth 7.6× everything Nanya earned in all of 2025. The deck gives you the mechanism and declines to draw it: ASP up more than 60%, bit shipment flat, FX flat. There is no third variable. COGS moved 7.3% against revenue’s 68.2%, so gross profit took NT$32.3B on NT$33.5B of incremental revenue — a 96.5% incremental gross margin, available only because implied D&A is NT$2.960B, just 3.6% of sales, at a company that spent NT$42.7B of capex across 2023–25 combined and let its asset base roll off. The margin is a rent, and Nanya is not the landlord: TrendForce has the Taiwan cohort “filling market gaps left as the top three shift toward more advanced processes,” and Commercial Times ties July’s record to DDR4 contract prices — the node the leaders are exiting. The April placement decoded (NT$78.72B for 10.19% at NT$223.90 to SanDisk, Kioxia, Solidigm and Cisco, with supply agreements attached, against NT$55.0B of quarterly operating cash flow and NT$4.0B of capex — allocation, not financing, and now a ~NT$102B paper gain larger than 1H net profit). Then the Fab 5A commitment of August 6: NT$346.6B through 2029, 8.1× three years of prior capex, 2026 budget raised NT$52.0B→NT$69.7B, EUV and 1b/1c/1d/1e — first wafer 2H2027, 30K wafers/month in 2028, against Samsung P3 alone near 115K by end-2026. Inside: the depreciation arithmetic on a 3.6% D&A base, the ASML / Winbond / PSMC / placement-holder read, 5.5× book and 8.8× annualised earnings, the July run-rate implying a ~NT$131.6B Q3, and the six dated things that settle it. Quarter ended June 30, 2026 · reported July 10 · Fab 5A plan detailed August 6.

  August 20, 2026

• Team Group (4967.TW) Q2 2026 · Record EPS on 16% Less Revenue. In a Shortage, That's a Choice.

  Net profit NT$2.412B on revenue of NT$7.588B — EPS NT$26.47, a record, on revenue that fell 16.11% QoQ while DRAM and NAND contract prices rose. The local coverage filed it under shortage economics. That framing survives until you put a second module house next to it: in the same quarter ADATA grew revenue 46.15% QoQ and paid for it with 13.6 points of gross margin. A 62-point QoQ revenue spread between two Taiwanese module companies buying from the same three makers is a decision, not a market. Team Group held price and let the bits go, and its net margin came out 355bp above ADATA's at a fifth of the scale — which makes this print a repricing of owned inventory rather than a share gain. Repricing is not the same as being paid: trailing twelve months through March 31, 2026 show NT$3.319B of net income against −NT$8.661B of operating cash flow, a NT$12.0B gap financed by a NT$7.0B increase in borrowings, 196 days of inventory, and a 0.52 quick ratio. July revenue of NT$2.545B sits 0.62% above the Q2 monthly run-rate against contract prices guided up 20–25%. The steelman for the industrial/SI pivot, the Taiwan revenue triple, the peer read, and the dated test. Quarter ended June 30, 2026 · board-approved results released August 3.

  August 20, 2026

• Longsys (301308.SZ) H1 2026 · The Warehouse Is the Position. In Q2 It Started Paying.

  Revenue RMB 24.088B (+136.26%), attributable net profit RMB 10.577B against RMB 14.77M a year ago, ex-items RMB 10.047B, basic EPS RMB 25.20, weighted ROE 79.03% — and a storage gross margin of 59.08%, up 45.80 points. Every read so far has framed it as profit-versus-inventory. It is one fact stated twice: a 59% margin at a company that owns no fab is a holding-period margin, so the RMB 25.777B of inventory at 60.12% of total assets is the asset that produced the earnings, still on the books. The warehouse is the position. Which makes cash conversion the number that settles it — and the line nobody printed already moved: back out the disclosed Q1 (−RMB 2.875B) from H1 (−RMB 3.151B) and Q2 operating cash flow was roughly −RMB 276M, a bigger quarter that burned about a tenth as much. The RMB 974M of writedowns taken into a rising market, the RMB 16.661B of borrowings and the RMB 560.00 placement, leverage ending the half at 56.13% against 65.55% at Q1-end, six 2026 marks on one equity from RMB 212.09 to RMB 575.69, the Micron / SK hynix / Kioxia / Kingston / Phison / CXMT read, the A-share trade constraint, and the three things that decide H2. Half-year ended June 30, 2026 · reported after close August 10.

  August 20, 2026

• Puya (688766.SH) H1 2026 · The Cycle Paid the Bill; the Acquisition Is the Asset

  Revenue RMB 3.959B (+336.67%), attributable net profit RMB 827M (+1,929.65%), ex-items profit RMB 825M (+2,994.08%) with only RMB 2.19M of non-recurring help, a record 51.15% gross margin, ROE from 1.82% to 28.73%, and operating cash flow swinging to +RMB 945M from −RMB 42M. Q2 net profit RMB 576M, up 129% on Q1. But 59.3% of the revenue and roughly 61% of the profit came from SkyHigh Memory — the 2D NAND / SLC NAND / eMMC franchise Puya reached via a vehicle that bought it from an SK hynix subsidiary for US$69.392M in August 2025, and which contributed ~RMB 508M of ex-items profit in this half alone, carrying just RMB 130M of goodwill. The remaining 49% is being bought in for RMB 247.06M — roughly one half-year of the earnings that stake carries — with interim-period gains accruing to the listed company. Against that: GigaDevice formally warned on June 29 that memory prices sit at historic highs and may fall considerably, and Puya took a new RMB 73.36M inventory write-down in its best-ever pricing half. The carve-out arithmetic, the inventory and debt-funded build, the GigaDevice / Winbond / Macronix / SK hynix read, 30× vs 65× on the same share price, the trade constraint, and the four dated observables. Half-year ended June 30, 2026 · reported after close August 18.

  August 20, 2026

• Biwin (688525) H1 2026 · The Margin Is a Procurement Bet, and It’s Locked for 24 Months

  A fabless module maker pre-announced RMB 15–16B of revenue (+283–309%) and RMB 7.0–7.5B of net profit — a swing from a ~RMB 226M loss and 8.2–8.8× the whole of FY2025’s RMB 853M. Back out Q1 (RMB 6.814B revenue, RMB 2.899B net profit, 53.3% gross margin) and the implied Q2 lands at RMB 8.19–9.19B revenue on RMB 4.10–4.60B of profit — a ~50.1% net margin at either end of the band, eight points above Q1. The mechanism isn’t mix or operating leverage: moving-weighted-average costing against a blended wafer price that moved ~311% while the booked inventory cost index moved ~57%. That wedge is the gross margin — and management has spent US$3.36B on two lock-volume, lock-price LTAs (US$1.5B wafer, Q2 26–Q1 28; US$1.861B enterprise NAND die, Q3 26–Q2 28) converting it from an accident of timing into a contractual position to 2028, against RMB 3.795B of cash at March 31. Also inside: RMB 500–800M of the headline is non-recurring (the ex-items guide is lower), inventory at 106.78% of quarterly revenue with AR +212.21%, the Longsys read (H1 net profit RMB 10.577B, storage GM 59.08%, +45.80pp), 7.5–8.3× forward against ~141× trailing, and the three lines on the full report that actually settle it. H1 pre-announced July 15 · full half-year report due August 25.

  August 20, 2026

• Applied Materials (AMAT) Q3 FY26 · Records Everywhere, and It Still Fell. Read the Mix.

  Applied printed a record on every line: revenue $9.12B (+25%), non-GAAP EPS $3.50 (+41%), record non-GAAP operating margin 34.0%, record operating cash flow $3.04B, a 13th straight quarter of gross-margin expansion, and a record Q4 guide of $10.25B / $4.02. Then the stock — up ~200% into the print — fell ~4% after hours. Unlike Cisco, the tell isn't margin (margins expanded); it's the base rotating: DRAM up to 26% of Semiconductor Systems from 22%, foundry/logic down to 67%, and China eased to ~28% of revenue from 35%. The bull calls the China decline de-risking; the forensic read is that a 200%-run stock got a low-single-digit beat on a base quietly trading diversified trailing-edge dollars for higher-beta AI-memory-capex dollars. The de-risking-vs-demand crux, the LRCX/KLAC/ASML/Micron/ACMR/Broadcom read, valuation on a re-rated multiple, the trade, and what decides it in Q4. Fiscal Q3 ended July 26, 2026 · reported after close August 13.

  August 13, 2026

• Macronix (2337.TW) Q2 2026 · The Mature-Node Fab Caught the Cycle, and the Depreciation Bill Waits Until 2028

  The best quarter in company history, earned on the nodes everyone else abandoned: net sales NT$19.125B (US$604.8M, +83% QoQ, +181% YoY) at a 64.4% gross margin — up from 15.6% a year ago — a 46.9% operating margin, net income NT$7.736B and EPS NT$3.91, off 55nm NOR and 48/96-layer SLC-MLC NAND. First-half revenue of NT$29.59B already exceeds all of FY2025 and first-half EPS of NT$4.80 erased the prior-year loss; July set a monthly record at NT$7.725B (+180.5% YoY). Three things sit under the headline: roughly NT$900M of gross profit is an inventory write-back (~4.7pp of margin, with more signalled for 2H), the NT$38B expansion taking the 12-inch fab from ~25K to >30K wafers/month carries no depreciation until 2028, and the LTA book is close to zero — the reason the ASP capture is this clean and the reason nothing floors it if the monthly sheet turns. Mix shifted hard: NAND incl. eMMC 30%→43% of revenue, eMMC +317% QoQ. The Winbond and NOR/SLC price-deck corroboration, the Kioxia/SanDisk/Micron/Samsung read, valuation on a stock up 600%+ but 29% below its June high at ~5.4x forward, and the listed-market read-through. Quarter ended June 30, 2026 · reported July 28.

  August 20, 2026

• Cisco (CSCO) Q4 FY26 · Cisco Beat, Raised, and Fell. The Margin Explains Why.

  Cisco did everything the bull case asked: Q4 revenue $17.3B (+18%), non-GAAP EPS $1.22 vs a ~$1.17 Street, $9.3B of FY26 hyperscaler AI infrastructure orders ($4B in Q4 alone), product orders +35% (+25% ex-hyperscalers), networking orders +40% for an eighth straight double-digit quarter, and an FY27 guide of $72.2–73.4B above consensus. And the stock fell ~4% after hours. The tell is one line: non-GAAP gross margin fell 210bps to 66.3% and product margin fell 270bps to 64.8%, because the AI hardware winning the orders ships at hyperscaler economics, not Cisco's historic software-and-services margin — and Q1 FY27 is guided to a 65–66% band. The supercycle is real and lower-quality; the market priced it in real time. The margin-floor-vs-trough crux, the Arista/NVIDIA/optical-interconnect/white-box read, valuation on an incumbent multiple re-rating on 14–16% growth, the trade, and what decides it next quarter. Fiscal Q4 ended July 25, 2026 · reported after close August 12.

  August 12, 2026

• Nebius Group (NBIS) Q2 2026 · EBITDA Went Positive; the Cash Came From the Customers

  Revenue grew 454% to $582.3M (the AI-cloud core ~$575M, 98% of it), ARR reached ~$3.0B, and adjusted EBITDA turned positive for the first time at $236.2M on a 41% margin — cost of revenue fell to 23% of sales and depreciation to 45%. Then the cash-flow statement: $2.25B of operating cash flow on a $190.4M GAAP net loss, but ~$1.2B of it was a jump in deferred revenue and ~$1.19B was receivables collection — so the quarter's cash came from customers pre-funding the build, not from margin. Capex was $5.66B, the balance sheet holds ~$6.0B of deferred revenue against ~$8.55B of debt and $8.04B of cash, and a $780.6M H1 revaluation gain on the ClickHouse-led stakes kept first-half GAAP income positive. Management reaffirmed FY26 revenue of $3.0–3.4B and $20–25B of capex while raising the year-end contracted-power target to 5 GW and a $7–9B exit run-rate. The stock closed +34%. The margin-inflection-vs-cash-quality crux, the CoreWeave / NVIDIA / hyperscaler / ClickHouse read, valuation (~22x FY sales vs CoreWeave's ~5x), the trade, and the listed-market read-through. Quarter ended June 30, 2026 · reported before open August 12.

  August 12, 2026

• Super Micro (SMCI) Q4 FY26 · The Margin Doubled; the Bar Is Now Durability

  Net sales grew ~91% to $11.1B — but the top line was never the question. The news is the margin: GAAP gross margin jumped to 17.5% from 9.5% a year ago, nearly doubling, carrying net income to $1,178M (from $195M) and diluted EPS to $1.62 ($1.70 non-GAAP, vs a ~$0.96 Street and a 65–70¢ company guide). Then the guide raised the stakes — Q1 FY27 to $14.5–15.5B (~$15B mid vs ~$12B consensus), non-GAAP EPS $1.01–1.10 (vs ~74¢), and FY27 to $65–72B (vs ~$54B) — on >$60B of new orders and a record backlog. The company credits a richer enterprise mix and broader DCBBS / direct-liquid-cooling adoption. The thin-margin-box thesis just got its margin doubled; the only open question is whether the mix is structural. The mix mechanism, the Dell/HPE/ODM/NVDA/liquid-cooling read, valuation, the trade, and the listed-market read-through. Fiscal Q4 ended June 30, 2026 · reported after close August 11.

  August 11, 2026

• CoreWeave (CRWV) Q2 2026 · Revenue Doubled; the Interest Bill Outran It

  Revenue grew 112% to $2.575B (a ~$10.4B run-rate), adjusted EBITDA hit $1.510B at a 59% margin, and revenue backlog reached ~$104B — up from $99.4B in Q1 and excluding >$25B of net-new commitments already added in early Q3. And CoreWeave still lost $626M ($(1.14) EPS, vs ~$(1.20) expected), because $1.393B of depreciation and $640M of net interest — the latter up 140%, faster than revenue grew — sit below the EBITDA line. Management raised the FY26 revenue guide to $12.4–13.2B and lifted the 2026 capex bill to $35–39B in the same breath, with Q3 guided to $3.4–3.6B (~158%) and an $18–19B exit run-rate target. Active power grew ~500 MW to 1.5 GW (3.7 GW contracted); the quarter added a $3.1B HPC-backed term loan, a $1B Jane Street investment, and >$10B of unsecured/convertible debt. The backlog-vs-balance-sheet crux, the NVDA / neocloud / hyperscaler-customer / data-center-credit read, valuation (~5.4x FY sales vs <0.7x backlog), the trade, and the listed-market read-through. Quarter ended June 30, 2026 · reported after close August 11.

  August 11, 2026

• SanDisk (SNDK) Q4 FY26 · A Record Print, a Doubled Buyback, and a Stock That Sold It

  Every headline line a record, every one ahead — and the stock fell ~4–8%. Revenue $8.965B (+372% YoY, +51% QoQ) at an 84.6% non-GAAP gross margin, non-GAAP EPS $39.25 (GAAP $43.97 on $6.90B net income), a $20.25B fiscal year (+175%), Q1 guided up to a $10.55B midpoint with $44–46 EPS, and buyback authorization topped up to ~$15.5B remaining. After a 500%+ YTD run the supercycle was already priced, so a blowout sold off — the same "bar was higher" tape as AMD and SK hynix. The analytical crux: SanDisk co-owns the Yokkaichi/Kitakami fabs with Kioxia, so its 84.6% margin and Kioxia's ~80% (two days earlier) are two P&Ls on one wafer stream — the strongest confirmation yet that NAND's pricing power is structural to the fab. Stargate/QLC datacenter ramp (~$2.98B, 33.2% of sales), the Micron/SK hynix/Samsung/WDC and independent-module-maker read, valuation after a ~$190B re-rate, and the listed-market read-through. Fiscal Q4 ended early July 2026 · reported after close August 5.

  August 5, 2026

• Kioxia Q1 FY26 · The Guided Blowout Landed; the Discipline Is the Story

  The quarter last year's guide promised would out-earn the whole fiscal year did exactly that — revenue ¥1,767.1B (~$11.1B, +415% YoY, +76% QoQ), non-GAAP operating income ¥1,326.2B (~$8.3B) at a 75% margin on ~80% gross, net profit ¥887.0B (~$5.6B, ~47x) — each line beating the company's own ¥1.75tn / ¥1.298tn / ¥869B (~$11.0B / ~$8.2B / ~$5.5B) guide. But the blowout was pre-committed; the new information was the discipline and the cash: a flip to net cash (¥186.7B, ~$1.2B), record ¥827.2B (~$5.2B) core free cash flow, an ¥800B (~$5.0B) buyback stapled to a 3-for-1 split, and a fab running Yokkaichi Fab 7 / Kitakami Fab 2 near 50% on a stated "price and profit first" policy, reaffirming a ~50% long-term-agreement mix by 2028. Q2 guided up to ~¥2.39tn (~$15.0B) revenue / ~¥1.89tn (~$11.9B) operating profit; shares jumped ~17.7% on a print that technically "missed" a stretched Street (~¥1.84tn / ¥1.37tn, ~$11.6B / ~$8.6B). The ASP mechanism, the SanDisk/Micron/SK hynix/Samsung and controller-layer read (Phison, Longsys) plus the independent-module-maker read, valuation, and the listed-market read-through. Quarter ended June 30, 2026 · reported July 31. USD at ~¥159/$.

  August 5, 2026

• AMD Q2 2026 · Data Center Doubled; the Bar Was Higher

  A record double beat: revenue $11.5B (+50% YoY) past the ~$11.2B guide, non-GAAP EPS $1.66 on $2.8B net income, non-GAAP gross margin 56% — and Data Center more than doubled to $6.7B (+107%), now 58% of the company. Client $3.1B (+23%) and Embedded $977M (+19%) helped; Gaming −31% to $779M still bled semi-custom. Q3 guided to ~$13B ±$300M (~41% YoY) at ~56% GM, with the MI400/MI450 book underwritten by OpenAI (6-GW), Oracle (50k GPUs), Meta, and Anthropic — yet the stock sold the print on a bar set at perfection. The second-source reframe, the NVDA/HBM/TSMC/optics/power read-through, valuation, and the listed-market read-through. Quarter ended June 27, 2026 · reported after close August 4.

  August 4, 2026

• SPCX Q2 2026 · The Beat Was Starlink; the Story Is the AI Capex Bill

  The first public print beat on every segment: revenue $7.81B (+92% YoY) past a ~$6.9B consensus, a −$0.09 loss vs −$0.26 expected, net loss narrowed >$3.7B QoQ to −$541M, adjusted EBITDA $3.5B (+191%), Connectivity/Starlink $4.29B revenue and $1.66B operating income on 12M subs — and guidance raised for the first time in 24 years. Yet the tape traded the other number: $18.37B of capex (~6x YoY), $15.83B of it in AI, against $14.1B of new AI contracts that don't ramp until October — and an Aug 6 lock-up that triples the float. The AI-infrastructure reframe, the ASTS/VSAT/NVDA/neocloud read-through, valuation, and the listed-market read-through. Quarter ended June 30, 2026 · reported after close August 4.

  August 4, 2026

• Palantir (PLTR) Q2 2026 · The Deceleration Thesis Died; Only the Multiple Is Left

  U.S. commercial revenue reaccelerated to +149% (to $764M) instead of mean-reverting, total revenue grew 93% to $1.94B, and U.S. government held +90% ($809M) — while a 62% adjusted operating margin drove the Rule of 40 to 155% and adjusted free cash flow to $1.22B (63% margin) on $1.06B of GAAP net income and $0.41 diluted EPS (vs ~$0.33). The forward book confirmed it: TCV closed $3.37B (+49%), U.S. commercial remaining deal value $6.24B (+124%), 220 deals ≥$1M, and 653 U.S. commercial customers (+35% YoY, +6% QoQ). Management then raised full-year revenue guidance to $8.15–8.16B and the U.S. commercial bar to ≥134% — after printing 149% — and the stock added ~12% after hours. The print retired the last fundamental bear thesis and handed the whole argument to valuation (~140x trailing, ~45–50x forward sales). The reacceleration mechanism, the software/app-layer peer read (C3.ai, ServiceNow, Salesforce, Snowflake), valuation, and the listed-market read-through. Quarter ended June 30, 2026.

  August 3, 2026

• Meta Platforms (META) Q2 2026 · The Ad Engine Beat; the Capex Bill Has No Backlog

  Advertising grew 27% to $59.4B on 14% more impressions at 12% higher prices, DAP hit 3.60B, and the Family of Apps still printed $23.4B of operating income — yet EPS missed at $6.18 (vs ~$7.17), dragged by ~$3.58B of one-time legal and severance charges and a tax step-up, and free cash flow collapsed 91% to $784M against $31.1B of capex. The reframe: Meta is the only one of the four megacap AI spenders running $130B+ of capex with no cloud revenue and no contracted backlog behind the build — where MSFT points to $678B and GOOGL to $514B. The stock fell ~10% after hours to a $521 double-bottom (~14x forward, RSI 16), then bought back ~14% to ~$595. The cash-conversion tell, the NVDA/AVGO/HBM/power/private-credit read-through, valuation, and the listed-market read-through. Quarter ended June 30, 2026.

  August 3, 2026

• Silicon Motion (SIMO) Q2 2026 · The Controller Cycle Turns, and the Mix Went Enterprise

  Revenue grew 127% YoY to $451.0M (+32% QoQ), non-GAAP gross margin reached 50.2%, and non-GAAP operating margin nearly doubled to 23.1% on $104.2M of operating income — the definition of operating leverage. Keep the bottom line straight: GAAP EPS was $3.99 but includes a ~$74.7M investment gain; the clean number is $2.43 non-GAAP (vs a ~$2.13 Street). The mix broadened and moved up-stack at once — SSD controllers +50–55%, eMMC+UFS +95–100%, and Ferri/boot-drive solutions ~+1,690% off a small base as automotive and enterprise inflected — with the MonTitan PCIe Gen5 enterprise push the strategic tell. Management then guided Q3 to $519–541M with non-GAAP operating margin stepping to 27.5–28.5%. The pure-play controller merchant just printed its cleanest operating-leverage quarter of the storage upcycle. The mix mechanism, the controller/NAND/Marvell/AI-server-storage read, valuation (anchor on the $2.43), the trade, and the listed-market read-through. Quarter ended June 30, 2026 · reported July 30.

  July 30, 2026

• Amazon (AMZN) Q2 2026 · AWS Reaccelerates to 37%, Margin and All

  AWS grew 37% to $42.2B — its fastest in years and above a 31–33% consensus — while expanding operating margin to 39.4% ($16.6B segment operating income), the cleanest "demand justifies the capex" answer of the three hyperscalers. Company operating income rose 43% to $27.5B on $200.6B of sales (+20%). Ignore the $5.75 EPS: net income of $62.6B carries a company-disclosed $53.4B non-operating gain, chiefly the Anthropic markup. The counterweight is cash — TTM capex $169B (+64%) on a ~$200B 2026 plan pushed trailing free cash flow to −$7.6B — and the signal is silicon: Anthropic and OpenAI made multi-gigawatt Trainium commitments. The reacceleration-with-margin mechanism, the Azure/Cloud/NVDA/AVGO/MRVL/HBM/power read-through, valuation, and the listed-market read-through. Quarter ended June 30, 2026.

  July 30, 2026

• SK hynix Q2 2026 · Record Margin, and the Contract Behind the Miss

  Record revenue of KRW 79.3T (+257% YoY) and operating profit of KRW 60.5T at an all-time-high 76% operating margin (+557% YoY) — yet the stock fell ~10%, because the record still landed a few percent below a stretched consensus (~KRW 84T revenue / ~KRW 64T operating profit) and the sequential price step normalized (DRAM ASP +~30% QoQ vs +~60% prior). Ignore the KRW 93.9T net figure — it's inflated by a one-time, media-attributed ~KRW 40T Kioxia-stake gain; the clean read is the 76%-margin operating line. The tell: SK hynix is the most-contracted of the three makers — long-term deals with ~10 customers cover ~half of sales and cap spot elasticity — with HBM4 in mass production and a $500B-plus SK–NVIDIA partnership booking the demand. The durability-for-upside mechanism, the Samsung/Micron/NVIDIA/TSMC read-through, valuation into the reset, and the listed-market read-through. Quarter ended June 2026.

  July 30, 2026

Full archive

2026 Q2 Apr–Jun17 prints

Semis & Compute

• TSMC Q2 2026 · The Node Everyone Needs Just Sold Out SemisJul 17
• ASML Q2 2026 · The IBM Beat and the Capacity Double-Down SemisJul 16
• Broadcom Q2 FY26 · Record AI Revenue, a Bar Set at Perfection SemisJun 4
• Marvell Q1 FY27 · Custom-Silicon Conviction, Priced In SemisMay 29

Memory & Storage

• Samsung Electronics Q2 2026 · The Semiconductor Division Was the Company MemoryJul 30
• CXMT IPO · The Record Debut and the Three-Year Gap Underneath It MemoryJul 28
• Micron Q3 FY26 · The Supercycle Gets a Contract MemoryJun 24

Hyperscalers & AI

• Microsoft (MSFT) FY26 Q4 · Azure Clears $100B, and the Capex Raise Gets Rewarded HyperscalerJul 30
• Alphabet (GOOGL) Q2 2026 · The Cloud Reaccelerates, the Capex Bill Reprices HyperscalerJul 27
• Lenovo Q4 FY25/26 · The Server Business Finally Pays; the Multiple Still Doesn't Believe It HyperscalerJun 26
• Dell Q1 FY27 · Revenue Doubled; the Argument Is the Margin HyperscalerJun 26
• HPE FY26 Q2 · The Cheapest 40%-Growth Print on the Board HyperscalerJun 25
• Oracle Q4 FY26 · The Backlog Compounds, the Funding Bill Arrives HyperscalerJun 10

Space & AI Infrastructure

• SpaceX Post-IPO · The Passive Bid Masks a December Supply Cliff SpaceJun 13
• SpaceX S-1 · One Profitable Segment, a $1.77T Price Tag SpaceJun 6

Cyber

• CrowdStrike Q1 FY27 · The Print Was Strong; the Multiple Is Stronger CyberJun 5
• Palo Alto Networks Q3 FY26 · Buying Growth, Booking the Bill CyberJun 4

2026 Q1 Jan–Mar4 prints

Memory & Storage

• Kioxia Q4 FY25 · The Guide Out-Earns the Year MemoryJun 16
• Phison Q1 FY26 · The Controller House Prints Like a Fab MemoryJun 5
• CXMT Pre-IPO · China's DRAM Champion Comes to Market MemoryJun 3
• Longsys Q1 FY26 · The Memory Upcycle Lands on a Module Maker MemoryJun 2