$AMKR Customers are prepaying Amkor about $3.2B for capacity, more than the ~$2.5B it will spend on capex this year. The 10-Q puts dates on that money. Our call is Buy with a $48 target, plus what would make us drop it.

$SSNLF 005930 KS Samsung's chip division made ₩106.5T of operating profit in 3Q26. That's more than all of $SKHY earned in the same quarter (₩83.3T), and it's 99% of Samsung's group profit. 3Q26 final (K-IFRS): Revenue ₩195.3T, +14% QoQ, +127% YoY (prelim ₩195.0T) Operating profit ₩107.4T, +20% QoQ, margin 55.0% (2Q 52.2%) DS (chips) ₩106.5T on ₩147.1T of sales, a 72% margin (2Q 70%) Memory sales ₩140.7T, +16% QoQ Phones, TVs and appliances -₩0.8T, same as 2Q Net profit ₩75.1T, EPS ₩11,462 Net cash ₩197.6T The number that moved most is capex. Samsung spent ₩26.9T on property and equipment in the quarter, almost double 2Q's ₩14.1T and 2.5 times a year ago. Revenue growth slowed from 28% to 14% QoQ while spending stepped up. On 2027, the company expects AI demand to tighten memory supply further than this year, and says it'll build its sales base on long-term contracts. HBM4 sales "significantly scaled up" in 3Q, with prices up across all applications. Foundry's first Taylor fab starts operations, and Samsung says its 2nm HPC wins now include major hyperscalers. Shareholders get ₩30T in 3Q dividends, ₩4,601 a share. Of the ₩90–110T planned for 2026, ₩35T is done or committed. The other ₩55–75T will include a large buyback and cancellation, with the plan set at the board meeting in late January. Below operating profit, the finance line swung to a ₩9.5T loss from a ₩4.7T gain in 2Q. The materials don't say why. For the call (10am Seoul): C1 What drove the ₩9.5T finance loss C2 How much of 2027 HBM4 is under long-term contract, and on what pricing terms C3 Where the ₩26.9T of capex went, and what 2027 capex looks like C4 Whether memory costs keep phones and TVs in the red into 4Q

$MRVL $LITE Marvell lost about $18B of market value after hours on a Bloomberg report that it's in advanced talks to buy Lumentum. Lumentum gained about $9B. Add the two together and they were worth roughly $9B less than at the close. What's reported so far: Bloomberg, citing people familiar with the matter, says the talks are advanced and a deal could be announced as soon as next week. Neither company has confirmed it, and no price or structure has come out. So for now it's a report, not a deal. After hours, Lumentum went from $913.95 to about $1,019 (+11.5%) and Marvell from $251.65 to about $231 (-8.2%). Coherent was up 2.7% and Applied Optoelectronics 1.3%. The size is the first thing that stands out. At the close Lumentum was worth $82.0B, a bit more than a third of Marvell's $220.7B. And Lumentum's stock is up 4.6x over the past year vs 2.85x for Marvell, so Marvell would be paying for it near the top of that run. What Marvell would get: Marvell sells the DSPs and other electrical chips inside optical modules, and Lumentum makes the indium phosphide lasers that put light on the fiber. Together they'd cover both the electrical and the optical side of the transceiver. If it's announced, the numbers that matter are the price per share and the split between cash and stock. With Lumentum at more than a third of Marvell's size, that split decides how much of Marvell current holders keep.

$SIMO Silicon Motion guided Q4 revenue to $588–616M. Analysts had $550M for Q4 going in, so even the low end is 7% above that, and the Q3 it just reported was already $560M. Q3 revenue $560.2M, up 24% from Q2 and 131% from a year ago, against a $519–541M guide and $529M consensus. That's every quarter since Q2 2025 above the top of its own guide. Non-GAAP gross margin 53.7% (guide 50–51%), operating margin 32.1% (guide 27.5–28.5%), EPS $4.35 vs $3.33 expected. GAAP EPS fell to $1.91 from $3.99, while operating income rose to $162.6M from $101.1M. The gap is investments: a $78.0M loss this quarter after a $74.7M gain in Q2. By product, vs Q2: SSD controllers +20–25% eMMC/UFS controllers −10–15% Ferri and boot drives +75–80% Management says rising memory costs are squeezing some consumer devices, while enterprise and Ferri customers care more about getting supply than about price. The Q4 guide also has non-GAAP gross margin at 54–55% and operating margin at 36–37%. Q1–Q3 plus the Q4 range puts 2026 revenue at $1.94–1.97B, vs $886M in 2025 and a $1.87B full-year consensus. Cash is $953.6M, up from $181.8M, after the $1.15B zero-coupon convert in August. The stock was around $227 after hours, about 3% above the close, after a roughly 16% run in the two sessions before the report. For the call (8am ET): C1 How much MonTitan revenue is in the Q4 guide, and which customers are ramping. C2 What takes gross margin to 54–55%: product mix, or pricing on Ferri and boot drives. C3 Whether the eMMC/UFS drop is price or units, and whether it keeps falling in Q4. C4 What the $1.15B is for, and where the $78M investment loss came from.

$MSFT Microsoft guided December-quarter Azure growth to 49–50% in constant currency, faster than Q1's 48%, and it's doing that while some rented GPU capacity slips. Amy Hood said a few third-party GPU-as-a-service contracts were pushed into the second half, and engineering efficiency across its own fleet has to fill the gap. The same delay showed up in capex. Q1 capex including finance leases was $45.2B, short of the "over $50B" guide, because third-party capacity deliveries expected in the quarter came late. Q2 is guided over $50B again, and calendar-year spend is now about $170B instead of about $175B. Roughly three quarters of Q1 capex went to short-lived assets, mainly CPUs and GPUs, and Hood called out higher memory and storage costs inside that. Our four questions for the call: C1 December-quarter Azure guide: answered. 49–50% cc. Demand still exceeds supply. C2 Capex including finance leases: partly answered. $45.2B, finance leases $7.4B, cash paid for PP&E $39.4B. How much spending moved to operating leases wasn't quantified. C3 Why commercial RPO stopped growing: answered. RPO usually falls in Q1, and one large Azure commitment made this quarter's dip smaller. Commercial bookings fell 12% against last year's large Azure deals. C4 Operating margin: partly answered. Q1 came in near 50%, up slightly. Q2 is guided down slightly year over year. No full-year number. Two more from the call. Higher-than-expected token pricing helped Azure's Foundry revenue more than triple. And asked how Microsoft would fund the buildout, Hood listed leasing, debt markets it hasn't tapped yet, and other options "when they make sense", while saying FY27 free cash flow stays positive.

$GOOGL Google Cloud's margin held at 36.2% in Q3 because the rented compute barely showed up yet. On the call, management said the third-party "bridging" capacity came online only at the very end of the quarter, usage ramps in Q4, and that means margin headwinds in the near term. So for anyone asking what this means for the neoclouds: Google is renting more outside compute this quarter, not less. Our four questions for the call: C1 Cloud backlog and the TPU share: partly answered. Backlog rose to $561B from $514B in Q2, mostly typical GCP contracts, and just over 55% should turn into revenue within 24 months. No TPU split. TPU system sales keep scaling in Q4, with a meaningful revenue ramp in 2027. Google now also sells TPU systems through Oracle and a joint venture with Blackstone. C2 Capex: partly answered. The $195–205B full-year guide is unchanged, so Q4 has to come in at $63–73B after $51.4B in Q3. 2027 capex will "increase significantly", with numbers on the Q4 call. Nobody asked why Q3 came in light. C3 Third-party compute: answered. Online at the very end of Q3, ramping in Q4, a margin headwind near term. C4 ATM and new notes vs 2027 capex: not answered, and nobody asked. Management said ROIC stays well above the cost of capital while free cash flow stays under pressure from capex and TPU inventory. Cash and marketable securities ended at $265.1B, long-term debt at $125.1B. Two more things from the call. Search growth slowed to 12% from 17%, and FX was the single biggest reason, along with a World Cup-heavy June and AI Mode queries that still earn less. Google is now taking ads in AI Mode international and expects lapping last year's acceleration to cost a couple of points again in Q4. And the TPU inventory that dragged on Q3 cash flow is only built after a contract is signed, so it's already sold.

$AAOI $LITE $COHR Applied Optoelectronics guided Q3 cable TV revenue to $100–110M. Run that for four quarters and you get $400–440M a year, already above the "more than $325M a year" it called its longer-term target in August. The cable side is riding the DOCSIS 4.0 upgrade cycle. Q2 CATV revenue was a record $80.6M, up 44% from a year ago and 21% from Q1, led by 1.8 GHz amplifiers shipped to its largest cable customer. In May, AOI said Mediacom picked it as primary vendor for its DOCSIS 4.0 upgrade. Cable was 42% of Q2 revenue. Datacenter was $107.7M, up 140%, so the AI side is still the bigger engine. At the midpoint, the Q3 cable guide is about 39% of the $255–290M total guide. Cable has been lumpy here before. CATV fell 24% from Q3 to Q4 last year, from $70.6M to $54.0M, even though full-year 2025 cable revenue nearly tripled to $245M. AOI reports Q3 on Nov 5. Two numbers matter for cable: whether Q3 lands inside $100–110M, and whether the Q4 cable guide holds up better than last year's drop.

$GOOGL Google Cloud did $30.7B in the September quarter, double a year ago, and its operating margin went up, to 36.2% from 35.6% in Q2. In July Alphabet had warned that renting third-party compute would put mild pressure on that margin. Alphabet revenue $127.6B, up 25%, a little under the $128–132B range of consensus estimates. Operating income $40.6B, up 30%, margin 31.8%. EPS $2.12 vs $2.87 a year ago: other income swung to an $8.2B loss from a $12.8B gain, so operating income is the cleaner read. Search & other $63.4B, up 12%. Our five questions from yesterday: Q1 Capex vs the $57–62B a quarter the full-year guide implies: answered. $51.4B, below that range. Year to date is $132.0B, so hitting $195–205B now needs $63–73B in Q4. No 2027 number in the release. Q2 Cloud margin vs 35.6%: answered. 36.2%. Q3 Cloud backlog vs $514B: not in the release, left for the call. Q4 Search growth vs Q2's 17%: answered. 12%, against last year's tougher comparison. Alphabet says query growth is the fastest in five years. Q5 Free cash flow and the ATM: answered. FCF -$9.1B, after -$5.9B in Q2. Alphabet sold $12.4B of stock through the ATM in the quarter with $27.6B left, and raised $28.7B in new notes. For the call: C1 Cloud backlog, and how much of it is TPU systems. C2 Whether the capex shortfall is delivery timing or a lower full-year number, and the 2027 figure. C3 How long third-party compute stays in the cloud cost base. C4 How the ATM and the new notes fit the 2027 capex plan.

$MSFT Azure did $33.2B in the September quarter, up 48%. Microsoft guided 44–45% in July, so it came in 3–4 points above its own guide, in the first quarter it reports Azure in dollars. Revenue $92.3B, up 19%, vs $90.7B consensus and an $89.85–90.95B guide. Non-GAAP EPS $4.89 vs $4.72 consensus and $4.13 a year ago. Agents and Infra $76.8B, up 25%, above its $75.15–75.75B guide. Devices and Consumer $15.5B, down 3%, above its $14.7–15.2B guide. Operating margin 49.8%, up from 48.9% a year ago. Our five questions from yesterday: Q1 Azure vs 44–45% and $32.2B: answered. 48% and $33.2B. Q2 Capex vs "more than $50B": partly. Cash paid for property and equipment was $39.4B, up from $35.8B in June. The figure including finance leases comes with the CFO commentary. Q3 Commercial RPO: answered on size, not on mix. $677B, up 73% from a year ago, but flat with June's $678B. Who added and who didn't isn't in the release. Q4 Margin flat while capex grows 40%+: answered. Margin rose about 0.9 points. Q5 December-quarter Azure guide: on the call. For the 5:30pm ET call: C1 December-quarter Azure guide, and whether growth still accelerates through the first half as Microsoft said in July. C2 Capex including finance leases, and how much moved to operating leases under the new accounting. C3 Why commercial RPO stopped growing quarter on quarter. C4 Full-year operating margin with Azure running ahead of plan.

$GOOGL Google Cloud revenue doubled. $30.7B, up 103%, with a 36.2% operating margin, up from 35.6% in Q2. Alphabet revenue $127.6B, up 25%. Operating income $40.6B, up 30%. EPS $2.12, down from $2.87, after an $8.2B other-expense loss. Capex $51.4B, below the $57–62B a quarter the full-year guide implies. Free cash flow -$9.1B.

$MSFT Microsoft beat. Revenue $92.3B, up 19%, vs $90.7B consensus and a $89.85–90.95B guide. Non-GAAP EPS $4.89 vs $4.72. Azure $33.2B, up 48%, above the 44–45% guide. Commercial RPO $677B, flat with June's $678B. Cash paid for property and equipment $39.4B. Guidance comes on the 5:30pm ET call.

$SKHY $MU $CXMT The 2028 memory glut case has a real date. Micron's second Idaho fab, its Hiroshima expansion and its Singapore NAND fab all start output in the second half of 2028. What the case doesn't have yet is buyers pulling back. SK hynix said Wednesday that long-term contracts cover more than half of its 2027 and 2028 capacity, and that HBM customers are asking for more than twice what it can supply. What SK hynix said on its 3Q26 call, with the qualifiers: HBM requests are "more than twice" the volume it can supply. That's today's figure for HBM, no year attached, and it counts what customers asked for. Orders can be smaller. DRAM and HBM are "effectively" sold out for the next three years through 2029. That counts signed contracts plus volumes still under negotiation. Signed so far: over 50% of 2027–28 capacity, 14 customers, typically five-year terms. The contracts carry deposits, and prices move with the market inside a range agreed with each customer. It expects demand to stay above supply through 2030. That's a seller in the middle of contract talks, so we discount it. Who is making the 2028 argument, and with what: The FT in July: record expansion at Samsung, SK hynix and Micron could bring oversupply back from 2028–29 if AI demand falls short. Morgan Stanley, as reported by the FT: China takes about 30% of the world's net new DRAM wafer capacity through 2028. Kwon Seok-jun of Sungkyunkwan University: aggressive expansion could turn into oversupply around 2028 if AI demand slows. He rates an HBM glut as less likely and puts commodity DRAM's turn around 2029. Michael Burry, who has disclosed a short position in Micron, called the expansion a sign the upcycle is ending. He's positioned for that outcome. The timing in that case holds up. Micron's ID1 starts wafer output in mid-2027, and the three sites above follow in 2H28. SK hynix opens its Yongin Phase 1 cleanroom in early 2027 and says 2027 capex rises significantly. Samsung has said P5 runs at full scale from 2028. The FT reported CXMT is aiming for 500k wafers a month by end-2028; CXMT doesn't publish a wafer count. Citi's Lee Se-cheol said in August that meaningful volume from Yongin and P5 shows up in 2H28. The demand side looks different from past cycles because buyers are putting money down. Micron has 26 agreements with take-or-pay volumes, about $150B of remaining obligations counted at minimum volume and minimum price, and $32B of customer commitments, mostly cash deposits. Samsung said in July it aims to put 60–70% of planned capacity under five-year contracts. Amkor's 10-Q lists about $3.2B of customer advance payments for packaging through 2027. Each of the three also said on its own call that it builds cleanroom shells early and adds tools as demand shows up. So late 2028 is when the space opens. How many wafers go into it is a call each company makes later, fab by fab. The parts of the bear case we think are real: 1. "Sold out through 2029" includes deals that aren't signed. The signed figure is 50%+ of 2027–28. 2. The contracts protect volume better than price. SK hynix's prices float inside a range and Micron's have floors. If spot falls, contract prices drift toward those floors. 3. China isn't in any of these contracts, and CXMT's added output goes mostly into commodity DRAM, the segment Kwon worries about. 4. In a shortage customers over-ask, and the 2x HBM figure counts requests. Deposits put a cost on over-asking, which is why they're the number to watch. Our view: a broad DRAM glut in 2028 isn't supported by what's on the record today. Micron expects industry DRAM bits to grow in the low 20s percent in both 2027 and 2028 and stay supply-constrained, and SK hynix puts 2027 DRAM demand the industry can actually supply at mid-20s growth. A glut needs the 2H28 space fully equipped, demand growth well below that, and buyers walking away from deposits, all at once. The narrower risk is real: commodity DRAM and NAND, China-exposed, late 2028 into 2029. TrendForce already has NAND loosening in 2H27. That's where a first crack would show, well before HBM. What would change our view: equipment spending running ahead of construction in 2027 (Micron says construction grows faster this fiscal year), spot DDR4 and DDR5 falling while contract prices hold, contracts reopened or deposits handed back early, and a new CXMT fab showing up in a filing. Samsung reports full 3Q26 results at 10am Seoul time Thursday, and its contract coverage update is the next data point. SK hynix's US shares were up 13.6% at $164.98 around 1pm ET. Seoul closed up 10.9%.

$GOOGL Alphabet reports after the close today. To hit its own 2026 capex guide of $195–205B, it has to spend $114–124B in the second half, about $57–62B a quarter. Last quarter it spent $44.9B, against $39.1B of operating cash flow. What Alphabet said in July, and where the Street is: Capex guide raised to $195–205B from $180–190B, "primarily due to an acceleration in the delivery of capacity." 2027 capex is expected to "increase significantly" Cloud will lean on more third-party capacity in Q3 as a bridge, which brings "modest margin pressure in the near-term" TPU system sales are a small part of revenue this year and ramp as 2026 ends. The vast majority lands in 2027 Search starts lapping last year's acceleration in Q3 Consensus revenue is $128–132B depending on the provider ($132.0B on Finnhub), up from $102.3B a year ago. EPS around $3.0–3.1 The EPS line won't tell you much tonight. Q2 GAAP EPS was $9.11 because other income included a $98.0B net gain, mostly unrealized gains on equity holdings. Operating income is the cleaner read: $40.8B in Q2, $31.2B in Q3 last year. The spending is now partly funded with stock. In June Alphabet raised $49.6B net from common and mandatory convertible preferred shares, and set up an at-the-market program for up to $40.0B more. Free cash flow was -$5.9B in Q2. What we'll check tonight: Q1 Q3 capex vs the $57–62B average the full-year guide implies, and whether a 2027 number shows up Q2 Cloud operating margin vs 35.6% once third-party capacity costs land Q3 Cloud backlog vs $514B. It rose by more than $50B in Q2, and just over half is expected to become revenue within 24 months Q4 Search & other growth vs 17% in Q2, now that it laps last year's acceleration Q5 Free cash flow, and how much of the $40B ATM has been used How it could go: Bull: Cloud growth near 80% with margin holding around 35%, capex in the high $50Bs, and a 2027 capex figure that comes with backlog to match Base: revenue near consensus, Cloud margin a few points lower on third-party capacity, capex in the low-to-mid $50Bs Bear: Cloud margin back in the 20s, free cash flow deeper in the red with heavy ATM use, Search growth under the mid-teens Options were pricing about a 6% move either way through Friday, going by the Oct 30 at-the-money straddle at 11:15 AM ET. The stock was $323.6, 19.6% below its May 13 closing high of $402.63.

$INTC $IONQ $RGTI $QNT $IBM Intel is paying a reported $2.3B, more than 3x Classiq's last private valuation, for quantum software built to run on rivals' hardware. Classiq's SDK lists 15 places a program can run. Intel is one of them. IBM, Google and IonQ are on the same list. Intel's Oct 26 release doesn't give a price. The $2.3B in cash and stock comes from Calcalist and Bloomberg. Calcalist also puts Classiq at about 150 people and under $100M in annual revenue, and says its 2025 round valued it at $600-700M. What the release does say: Classiq turns a high-level quantum program into a circuit built for whichever machine runs it. Intel plans to pair that with its silicon spin qubits and cryogenic control chips, which it makes on standard chip equipment. After closing, Intel says it will keep supporting Classiq's existing multi-platform integrations. Closing is expected in the first half of 2027, pending regulatory approval. So the asset Intel wants is the layer that sits above the hardware, and it has promised to keep that layer open to rivals. The pure plays jumped on Oct 27: $RGTI +8.7%, $QNT +7.4%, $IONQ +7.3%, $QBTS +5.2%. By 10:33 AM ET today they had given back between 1.7% and 5.1%. $INTC rose 3.6% the same day, but it also announced an Oracle deal and chips rallied broadly, so the deal's share of that move can't be pulled out. $IBM fell 1.2%. If IBM and IonQ customers keep writing code in a tool Intel owns, how long does "hardware-agnostic" stay that way?

$MU $SKHY $CXMT CXMT's HBM3E reportedly yields 20% to 30%, against about 70% at SK hynix, and CXMT is going into production anyway. Its customers in China can't buy the 70% product without a US export license, so the gap lands mostly in CXMT's costs while the chips still ship. Reuters reported today, citing two unnamed people, that CXMT plans to start limited HBM3E production over the next two months (that timing came from one of them). Samples went to several Chinese AI chip designers, which Reuters didn't name. Some relays named Huawei and Cambricon as the sample customers, but the Reuters text doesn't. In testing the parts performed closer to SK hynix's previous-generation HBM3, and they're expected to cost more than imported equivalents. CXMT didn't comment, and its exchange filings say nothing about HBM. How big the yield gap is: at 70% vs 20–30%, SK hynix gets roughly 2.3 to 3.5 times as many good chips out of the same input. Micron was reported at about 75% on 8-high HBM3E and 70% on 12-high last year. On cost per good stack, CXMT starts well behind. Here's why that's not the whole story. Scale. CXMT's share of global DRAM revenue went from 7.6% in Q1 to 9.5% in Q2 (TrendForce), fourth behind Samsung, SK hynix and Micron. Capacity is reported at about 320k wafers a month, heading to 420k by 2027. A captive market. Since December 2024 US export rules have covered all HBM stacks in production. A Chinese accelerator maker that needs HBM3E-class memory has a short list of suppliers, and CXMT is on it. Speed. A year ago CXMT was reportedly sampling HBM3 and talking about HBM3E in 2027. If production starts by the end of this year, that's ahead of its own reported plan. Yield comes with volume, and a protected home market supplies the volume. What it doesn't change yet: HBM revenue in Q2 split 50/33/18 among SK hynix, Samsung and Micron (Counterpoint), and CXMT's first HBM3E goes to Chinese buyers at roughly last-generation performance. The number to watch is how fast that 20–30% climbs. Next markers are whether production actually starts by year end, any yield read on the second design CXMT sampled to fix yields, and whether Chinese accelerators ship with CXMT HBM in 2027.

$MSFT Microsoft reports after the close today, and it's the first earnings report with Azure revenue in dollars. Put Microsoft's own 44–45% Azure guide on its restated base and Azure comes to about $32.2–32.5B for the quarter. On Microsoft's restated numbers, the whole Intelligent Cloud segment made $24.1B two years ago. What Microsoft guided in July, and where the Street is: Revenue $89.85–90.95B (+16–17%). Consensus $90.74B, near the top Azure +44–45% in constant currency on the new definition; last quarter was 42% Operating margin roughly flat year on year Capex over $50B, vs $34.9B a year ago EPS consensus $4.72 (25 analysts), up from $4.61 90 days ago. A year ago it was $4.13 excluding OpenAI The capex line is bigger than it looks. From this fiscal year Microsoft depreciates data center buildings over 25 years instead of 15, which moves more future leases into operating leases that don't count as capex. That's why its calendar-2026 capex figure went to about $175B from about $190B. This quarter's guide is still over $50B after that change, so the spending keeps going up and only part of the label moved. It's also the first report on two segments, Agents and Infra ($75.15–75.75B guided) and Devices and Consumer ($14.7–15.2B), so the old three-segment comparisons won't line up. What we'll check tonight: Q1 Azure growth vs 44–45%, and whether the dollar figure clears $32.2B Q2 Capex vs "over $50B", and how much the lease change took out Q3 Commercial RPO was $678B in June, and all of last quarter's sequential growth came from customers outside the frontier labs. Does that hold? Q4 Whether operating margin stays flat with capex up 40%+ Q5 The December-quarter Azure guide. In July Microsoft said it expects first-half growth to accelerate How it could go: Bull: Azure above 45% and a December guide that accelerates again Base: Azure 44–45%, capex just over $50B, margin flat Bear: Azure at 43% or lower, or capex well above $50B with margin slipping Options were pricing about a 5% move either way, going by Tuesday's last trades. The stock was $529.36 premarket, 2.6% below its Oct 13 closing high.

$ORCL $CBRS $GOOGL $NVDA Oracle put three non-Nvidia chips on its cloud roadmap yesterday, and none of them comes with a volume, a price or a 2026 date. The chip customers can rent on OCI today is Nvidia's Vera Rubin, which Oracle plans to take to multi-gigawatt scale. What the four press releases from Oracle AI World say: Google TPUs arrive on OCI in calendar 2027 and ramp over several years. Oracle calls it the first time TPUs are offered on a hyperscale cloud outside Google, and Oracle supplies the power, cooling and networking, so the TPUs sit in Oracle's data centers. Cerebras inference on OCI is also planned for calendar 2027. The release says it isn't a commitment to deliver. SambaNova's SN50 has no date. SambaNova pitches it as a decode chip running next to Nvidia GPUs, with the GPUs doing prefill. Nvidia Vera Rubin NVL72 and Vera CPUs are generally available on OCI now. OpenAI trained and served GPT-6 Astra in Abilene on an OCI cluster with more than 100,000 Blackwell GPUs. The new chips mostly come in on the inference side. Training on OCI stays with Nvidia, and Google's TPU 8t becomes the second option for it from 2027. None of the four releases says anything about cutting reliance on Nvidia. That framing came from the coverage. $CBRS still closed up 10.1% on the day, against 3.6% for $ORCL, 1.2% for $GOOGL and 0.9% for $NVDA. The biggest move went to a plan dated 2027. Has anyone seen Oracle put a capacity number on the TPU side yet?

$SKHY 000660 KS SK hynix kept 80 won of every 100 won of sales as operating profit in 3Q26. Operating profit came in at ₩83.3T, 8% above FnGuide consensus and above the ₩81.5T analysts had three months ago, before the stronger won made them cut. 3Q26 vs what was expected (K-IFRS, preliminary): Revenue ₩104.6T vs consensus ₩98.2T, up 32% QoQ and 328% YoY, the first quarter above ₩100T Operating profit ₩83.3T vs consensus ₩77.0T, up 38% QoQ Operating margin 80%, last quarter 76%, a year ago 47% Net profit ₩39.8T, down from ₩93.9T in 2Q26. The release doesn't explain the gap below operating profit Cash ₩114.5T, borrowings ₩15.8T, net cash ₩98.7T The new disclosure is the LTA book. SK hynix has long-term agreements with 14 customers covering more than half of its 2027 and 2028 capacity. Prices move with the market inside a range agreed with each customer, and the contracts typically run five years. HBM4 should pass HBM3E in both shipments and revenue in 4Q, with M15X at full capacity by year end. Management expects demand to exceed supply through 2030, which is the supplier's own forecast. Shareholders get ₩20,000 a share for 3Q, half of it a special dividend from the Kioxia gains. Total 2026 returns come to about ₩100T, and the ₩40T buyback announced in August finished last week. For the call (9am Seoul): C1 What sits between ₩83.3T of operating profit and ₩39.8T of net profit C2 How wide the LTA price ranges are, and how much of 2027 HBM4 pricing is already set C3 How big the 2027 capex step-up is, with Yongin Phase 1 coming online

$TER beat the top of its own guide, and the CEO's first named demand driver was memory and hard drive test, ahead of compute. Q3 revenue $1.31B was up 70% YoY and non-GAAP EPS $2.29 came in 11% above the Street. Q3 2026 vs what was expected: Revenue $1.306B vs guide $1.20–1.30B and consensus ~$1.27B, up 70% from $769M a year ago (Q2 was $1.329B) Non-GAAP EPS $2.29 vs guide $1.85–2.15 and consensus ~$2.06 Non-GAAP gross margin 59.1% (Q2 59.8%), non-GAAP operating margin 31.6% Semiconductor Test $1,072M (+77% YoY), Product Test $138M (+57%), Robotics $96M (+28%) Q4 guide: revenue $1.25–1.35B, non-GAAP EPS $1.95–2.25, vs consensus ~$1.27B / $2.03 Greg Smith: "Demand extended across the AI value chain, from memory and hard disk drive test to high-speed interconnect and robotics." That fits the $STX print 25 minutes earlier: storage makers are running at record margins, and test demand is showing up in Teradyne's numbers. For tomorrow's 8:30 ET call: C1 Memory and HDD test as a share of Semi Test, and whether HBM test orders run into 2027 C2 Why revenue dipped from Q2 while the guide still points flat to up C3 Robotics at $96M: is the turnaround real or one quarter

$STX now keeps more than half of every revenue dollar as operating profit. Non-GAAP operating margin hit 50.7% in FQ1, up from 29.0% a year ago, and the guide for next quarter sits about 5% above the Street. FQ1 FY27 vs what was expected: Revenue $4.23B vs guide $4.0–4.2B and consensus ~$4.15B, up 61% from $2.63B a year ago Non-GAAP EPS $7.68 vs guide $7.10–7.50 and consensus ~$7.43 Non-GAAP gross margin 57.8%, last quarter 52.7%, a year ago 40.1% Free cash flow $1.5B, $1.2B of debt retired, dividend raised to $0.76 FQ2 guide: revenue $4.8B ± $0.1B, non-GAAP EPS $9.20 ± $0.20, vs consensus ~$4.56B / $8.61 Mosley pins it on "long term customer demand visibility" and on growing exabytes through HAMR areal density "while maintaining supply discipline." So the extra bits come from denser drives, and Seagate isn't adding much capacity to get them. For the 5pm ET call: C1 How much of 2027 exabyte demand is already locked in with the cloud customers C2 HAMR / Mozaic share of exabytes shipped, and the cost-per-TB gap vs PMR C3 Whether pricing keeps rising, or whether the margin gain is now mostly mix and cost