Thursday, October 8, 2026

Samsung Electronics posted preliminary Q3 2026 operating profit of 107.4 trillion won (~$80 billion) on Oct 8, up 783% year-over-year and a fourth consecutive record on AI memory demand, even as shares fell 2.42% on institutional selling. Its mobile unit is cutting Q4 smartphone output up to 30% because memory now costs more than handset margins, while HBM4E cleared validation for Nvidia's Rubin Ultra; the planned 1c capacity shift toward HBM by 2027 bears watching.

TSMC reported Q3 2026 revenue of NT$1.494 trillion (~$46.7 billion) on Oct 8, up 50% year-over-year and ~2% above guidance, with September's $16.1 billion extending its run of monthly records to seven. GlobalFoundries the same day signed a $2 billion multi-year agreement to supply silicon interposers from Malta, New York—the first U.S.-based source of this CoWoS component—and the pace of that capacity add is the variable.

The Wall Street Journal reports Broadcom is in early talks to arrange more than $50 billion of debt for OpenAI's custom-chip program with Apollo and Blackstone among lenders, after readying up to $60 billion for Anthropic, while Oracle separately is talking to Apollo and Goldman Sachs about financing chip purchases. Applied Digital's $77.4 million interest bill exceeded its $65.8 million of base rent in its Oct 7 quarter, so debt service against contracted revenue is the variable.

OpenAI's annualized revenue was approaching $50 billion at end-September, the Financial Times reported on Oct 8—about $20 billion below a circulated $70 billion figure, reflecting gross-versus-net accounting that excludes Microsoft and Amazon Bedrock revenue shares. Usage is rising regardless: J.P. Morgan found OpenRouter token volume up 71% month-over-month in September, while Anthropic's Claude Haiku 5.5 (Oct 7) matched GPT-6 Luna's $0.10 per million input tokens, keeping price pressure on; net-revenue disclosure is the open variable.

Optical circuit switch revenue has risen from under $10 million a quarter a year ago to Lumentum projecting over $100 million for its September quarter, with P Equity Research seeing more than $2.5 billion in 2027. Its Oct 7 channel checks show AMD testing a next-generation Helios rack with Marvell's 6.4T NPO engines at an early stage, while Nomura forecasts co-packaged optics engine shipments jumping from 2 million in 2027 to 22.5 million in 2028, where UBS says testing remains the bottleneck.

• Samsung Electronics reported Q3 2026 preliminary operating profit of 107.4 trillion won (~$80 billion), up 783% year-over-year — its fourth consecutive record quarter — driven by AI-fueled memory demand. (Oct 8)?
• Despite record earnings, Samsung stock fell 2.42% on Oct 8 as institutions sold; its mobile division is cutting Q4 smartphone production by up to 30% because memory inside many devices now costs more than the margin they generate. (Oct 8)?
• Samsung's HBM4E memory completed testing validation for NVIDIA's next-generation Rubin Ultra GPU, and Samsung is shifting 10+ percentage points of its 1c node capacity from LPDDR5X to HBM by 2027. (Oct 7–8)?

• TSMC reported Q3 2026 revenue of NT$1.494 trillion (~$46.7 billion), up 50% year-over-year and ~2% above guidance, with September revenue of $16.1 billion marking a seventh consecutive all-time high. (Oct 8)?
• TSMC now holds 72.5% of global foundry revenue, up from 53.6% in early 2022; Taiwan's share of global chip sales by customs value fell to 16.9% as value-add concentrates in finished systems assembled elsewhere. (Oct 8)?
• TSMC signed a $2 billion, multi-year agreement for GlobalFoundries to supply silicon interposers from its Malta, New York fab — the first U.S.-based source for this CoWoS component. (Oct 8)?
• Taiwan shipped $15.1 billion of computer units and modules to Mexico in January–August 2026, ~17× the 2024 pace and mostly AI server sub-assemblies rather than finished computers, according to Schulz Research's reading of Taiwan customs data. (Oct 8)?

• UBS calls co-packaged optics (CPO) testing a key bottleneck to commercialization, creating new equipment revenue beginning in 2027; Nomura forecasts CPO optical-engine shipments rising from 2 million in 2027 to 22.5 million in 2028. (Oct 8)?
• According to P Equity Research, optical circuit switch revenue surged from under $10 million per quarter a year ago to Lumentum projecting over $100 million for its September quarter, with total OCS revenue seen exceeding $2.5 billion in 2027. (Oct 8)?
• Channel checks from P Equity Research indicate Marvell and AMD are collaborating on scale-up, with AMD testing a next-generation Helios rack using Marvell's 6.4T NPO engines in both compute trays and scale-up switches, work at an early stage but ongoing. (Oct 7)?
• Bernstein projects AI accelerator chiplet-adjusted shipments to more than double from 2025 to 2028; Broadcom announced it will present rack-based networking innovations for scale-up, scale-out, and scale-across AI clusters at the 2026 OCP Global Summit (Oct 12-15). (Oct 8)?

• Anthropic released Claude Haiku 5.5, priced at $0.10 per million input tokens — matching GPT-6 Luna exactly — and scoring 43 on the Artificial Analysis Intelligence Index, 26 points above the prior Haiku. (Oct 7)?
• OpenAI's annualized revenue was approaching $50 billion at end of September per a Financial Times report, roughly $20 billion below a $70 billion figure previously circulated; the gap reflects gross-vs-net accounting, with Microsoft and Amazon Bedrock revenue shares excluded from the net figure. (Oct 8)?
• J.P. Morgan found AI token volume on OpenRouter surged 71% month-over-month and 30× year-over-year in September 2026, with open-weight model volumes up 83% month-over-month and 101× year-over-year. (Oct 8)?

• WSJ reports Broadcom is in early talks to arrange more than $50 billion in debt financing for OpenAI's custom AI chip program, with Apollo and Blackstone among lenders; WSJ reports Oracle is separately in talks with Apollo and Goldman Sachs to finance chip purchases. (Oct 8)?
• AWS EC2 capacity-block listed prices rose 15%, following Nebius's earlier H100 price increase of 17% and B300 increase of 21%; Morgan Stanley estimates AWS Bedrock revenue at approximately $10 billion, growing ~100% sequentially. (Oct 8)?
• William Blair initiated coverage of CoreWeave and Nebius, projecting global neocloud revenues growing from ~$26 billion in 2026 to $200–300 billion by 2031. (Oct 8)?
• Applied Digital reported Q1 revenue of $341.9 million — 2.7× consensus — but paid $77.4 million in interest against $65.8 million in base rent from AI data centers, and targets 3.5–4 GW of operating capacity by end of 2030. (Oct 7)?

• An expert call with a process integration engineer at Micron's Taiwan fabs revealed that over 88% of the Taoyuan union voted to strike; compensation concerns were said to span multiple major Taiwan sites, not just one facility. (Sep 30 interview; Oct 8 publication)?
• YMTC's enterprise SSDs are being trialed by some domestic Chinese customers due to tight capacity at leading overseas suppliers, but data center buyers still prioritize Samsung, SK Hynix, and Micron. (Oct 8)?
• Wolfspeed received a conditional $1.5 billion, 30-year U.S. government loan commitment; the first $600 million tranche refinances existing debt, extending maturity from 2030 to 2056 and reducing interest payments. (Oct 7)?
• Teradyne's Titan HP system enters high-power AI chip burn-in, creating direct competition with Aehr Test Systems' Sonoma at the package level; Aehr simultaneously announced a ~$6 million follow-on Sonoma order as a second AI processor moves onto its platform. (Oct 7)?

• Jensen Huang, founder and CEO of NVIDIA, received the National Medal of Science; fellow honorees receiving the Medal of Science or Medal of Technology and Innovation included Lisa Su (AMD CEO), Elon Musk, Sergey Brin, Satya Nadella, and Michael Dell. (Oct 8)?
• Nvidia and Microsoft unveiled RTX Spark PCs integrating NVIDIA's AI stack into Windows laptops and compact desktops with up to 128 GB unified memory and 1 petaflop FP4 performance; Jensen Huang credited the Windows partnership as foundational to NVIDIA's history. (Oct 8)?
• AMD CEO Lisa Su visited Korea on Oct 7 for an AI Semiconductor & Startup Roundtable co-hosted with South Korea's Ministry of Science and ICT, and is seeking multi-year supply agreements with SK Hynix for the next AI cycle. (Oct 7)?
• AMD's Taiwan capacity negotiations are shifting from one-to-two-year procurement cycles to three-to-five-year planning horizons, pulling TSMC, advanced packaging, substrates, and server ODMs deeper into AMD's AI infrastructure expansion. (Oct 8)?

• The FOMC's September minutes listed 'surging AI-related investments' alongside oil as a source of inflation pressure; the Fed hiked 12-0 in September with 12 of 18 officials projecting one more hike in 2026. (Oct 7)?
• ERCOT's interconnection queue reached 474 GW of new large-load requests by June 2026, roughly 90% data centers; a provisional Texas grid study sorted 66 GW as base load, ~127 GW as studied, and excluded 302 GW. (Oct 8–9)?
• Transformer prices have risen from $7/kVA in 2014–2022 to $13/kVA in 2026, with new 50–100 MVA units in Europe quoted at $30–35/kVA; WoodMac finds 4-hour battery storage now cheaper than gas peakers across global markets. (Oct 8)?
• Voltus raised $225 million; Fervo's first enhanced geothermal system plant went commercial in September; Virginia's lieutenant governor opposed the NextEra-Dominion merger as public hearings began. (Oct 8)?

• Meta and Sierra, joined by Genesys, Shopify, Stripe, and Walmart, launched the Personal Agent Protocol to govern how personal AI agents interact with business services and handle payments. (Oct 8)?
• Google Cloud launched a Gemini enterprise agent at Gemini at Work 2026, positioning it as a competitive entry into cross-system agentic workflows for enterprises. (Oct 8)?
• The Every team built a company-wide agent on Anthropic's Claude Managed Agents deployed in Slack, migrating from a self-hosted fleet to reduce infrastructure overhead, then released it to subscribers. (Oct 8)?
• AI labs are purchasing companies' operational records — emails, contracts, accounting ledgers — to train systems capable of autonomously running business functions; Synopsys and OpenAI announced a specialized GPT-Synopsys model for chip design trained on Synopsys tools and workflows. (Oct 8)?