Monday, October 12, 2026

Nvidia's most recent 10-Q, surfaced Oct 10, shows $530 billion of gross off-balance-sheet guarantees across six line items, up from $184 billion the prior quarter, driven mostly by commitments to buy memory. Separate Oct 11–12 reports place Nvidia in early talks to acquire, acqui-hire, or invest further in frontier lab Reflection AI; exposure if memory demand or pricing turns is the open question.

ASML on Oct 10 sent Samsung and SK Hynix a flat 10% increase on all extreme- and deep-ultraviolet replacement parts from January 2027, which both accepted; industry contacts say it gives Applied Materials, Tokyo Electron, and Advanced Energy cover to follow. By Oct 12 Intel, Qualcomm, and AMD had raised or were set to raise CPU prices, Panther Lake by 15–20%, partly on TSMC foundry costs; customer absorption is the uncertainty.

A Super Micro contractor pleaded guilty, per Oct 10–11 reports, to smuggling $2.5 billion of Nvidia chips to China, apparently the first criminal chip-smuggling conviction, while China invested further in Hua Hong Semiconductor, its second-largest foundry. A digest of the Oct 4–10 week adds DeepSeek raising more than $12 billion and Manus $500 million; enforcement follow-through and domestic foundry output bear watching.

DayOne's Nasdaq prospectus (Oct 9–11) discloses 2.3 GW of bookings, $4.4 billion of long-term debt, and one customer at 69% of revenue, while Firmus's withdrawn ASX IPO gains a detail: power-equipment supplier Maas Group had invested more cash into Firmus than Firmus had paid it. UBS on Oct 11 put 2027 compute shipments at 50 GW against 38 GW of powered shell; whether DayOne's offering prices is the test.

Memory spot prices spiked over Oct 11–12—DDR4 4Gb up 274%, DDR5 16Gb up 100%, 256Gb NAND wafers up 167%—and Barclays sees Micron DRAM supply constrained through calendar 2028, on top of Samsung's previously reported 2027 long-term agreements. Goldman Sachs on Oct 9 modeled Samsung HBM revenue at $81 billion in 2027 on blended prices above $4/GB; a NAND-based High Bandwidth Flash tier sized at $40–150 billion by 2028–2030 is the wildcard.

• Goldman Sachs projects Samsung Electronics' HBM revenue to surge 321% YoY to $81 billion in 2027, with blended HBM average selling prices exceeding $4/GB, a 159% YoY increase driven by severe supply shortages. (Oct 9)?
• Memory spot prices spiked sharply: DDR4 4Gb +274%, DDR5 16Gb +100%, and 256Gb NAND wafers +167%, while Barclays forecasts Micron DRAM supply constrained through CY28 with DDR5 spot pricing up 8% MoM. (Oct 11–12)?
• High Bandwidth Flash (HBF), a NAND-based near-compute memory tier using TSVs and UCIe interconnects, could reach a $40–150 billion TAM by 2028–2030 depending on adoption, potentially priced well below HBM4 on a per-GB basis. (Oct 11–12)?

• Google's AI infrastructure chief Amin Vahdat said Google must roughly double serving capacity every six months and that agentic computing will drive a 10x+ increase in compute demand. (Oct 10)?
• UBS forecasts 50 GW of compute shipments in 2027 against only 38 GW of powered data-center shell, a gap that widens further in 2028. (Oct 11)?
• DayOne filed a Nasdaq prospectus showing 2.3 GW of bookings, $4.4 billion of long-term debt, and a single customer accounting for 69% of revenue; Firmus pulled its ASX IPO the same week after its power-equipment supplier Maas Group had invested more cash into Firmus than Firmus had paid it. (Oct 9–11)?

• ASML sent a flat 10% spare-parts price hike to Samsung and SK Hynix covering all EUV and DUV replacement parts from January 2027; both accepted, and industry contacts say the move gives cover for Applied Materials, Tokyo Electron, and Advanced Energy to raise prices as well. (Oct 10)?
• Morgan Stanley projects TSMC optical engine shipments reaching 73.87 million units in 2028, up from 0.39 million in 2026, driven by yield improvements. (Oct 10)?
• Intel, Qualcomm, and AMD all raised or are expected to raise CPU prices—Intel's high-end Panther Lake up 15–20%—driven in part by TSMC's rising foundry costs; Vicor filed an ITC complaint naming roughly 20 AI server suppliers including Foxconn and Luxshare over power-system patents. (Oct 12)?

• Nvidia's most recent 10-Q disclosed $530 billion of gross off-balance-sheet guarantees across six line items, up from $184 billion the prior quarter, primarily driven by supply commitments to buy memory. (Oct 10)?
• UBS forecasts Nvidia revenue growing from $216 billion in FY26 to approximately $985 billion in FY29, before contracting in FY30; Goldman Sachs estimates a Nvidia B300 AI data center can generate a 43% gross ROIC per 1 GW, versus 7% for the Huawei 910C. (Oct 10)?
• Nvidia is in early talks to acquire, acqui-hire, or invest further in Reflection AI, a frontier model lab; Nvidia has already bought Hugging Face and acqui-hired Poolside and Essential AI, assembling open-source model capabilities alongside its hardware stack. (Oct 11–12)?
• Mizuho estimates Nvidia's server CPU revenue could reach approximately $60 billion in 2028, with Nvidia becoming a market leader in server CPUs by 2027 through its Vera CPU product. (Oct 11)?

• Anthropic's rumored $2 trillion IPO valuation drew skepticism as Meta's market cap sits at $1.9 trillion; Meta's Mark Zuckerberg announced a new family of frontier models called Muse Wonder for personal and professional agentic work on Oct 12. (Oct 12)?
• OpenAI's dots personal agent can now be built on a phone, sends tasks to Codex autonomously, and Codex has begun predicting users' next messages, positioning OpenAI as a goal-first layer before any app is chosen. (Oct 10)?

• Elon Musk described Terafab's goal of completing 10 chip designs per year versus the industry norm of one every two years, framing rapid iteration speed as the core technology lever; Musk was quoted back by Elon Musk himself after the post went viral, calling on world-class chip engineers to join SpaceX or Tesla. (Oct 11)?
• Analysis shows Terafab will need to spend more than $600 million per 1,000 wafers-per-month of advanced-node expansion; equipment lead times put the initial production ramp at risk through 2029 and SpaceX's orbital AI deployment targets of 10 GW in 2028 and 100 GW in 2029 depend heavily on Terafab execution. (Oct 12)?
• Musk's Google AI infrastructure counterpart Amin Vahdat noted that sun-synchronous orbits offer 40% more available power due to reduced atmospheric attenuation, supporting the orbital compute thesis. (Oct 10)?

• At OCP 2026 (Oct 12–15), Meta will present a 102.4T switch keeping lasers in 16 external pluggable modules, claiming 60% optics power reduction, while Nvidia will demonstrate co-packaged optics switches; memory makers appear on eight times as many OCP sessions as laser makers. (Oct 10–11)?
• Nomura estimates optical lasers/chips, DSP, TIA/driver chips, and PD chips collectively account for 66–72% of 800G/1.6T transceiver BOM costs; DSP/MCU/CDR leads 1.6T EML BOM at 27% and 200G EML/LD chip at 26%. (Oct 10)?
• Marvell's David Lazovsky stated that three years ago one-third of data-center infrastructure was focused on inference; today over 60% is, with that share expected to approach 90% within two to three years. (Oct 10)?
• Tower Semiconductor is expected to maintain 40%+ silicon photonics foundry market share through 2029; InP substrate supply remains the tightest bottleneck, with Sumitomo Electric and AXT controlling approximately 85% of that market, and Mizuho sees tightness lasting through 2028. (Oct 11)?

• DeepSeek raised more than $12 billion, Manus secured $500 million after splitting from Meta, and Kling is eyeing a $1 billion IPO, according to China AI Weekly Digest covering the week of Oct 4–10. (Oct 11)?
• China invested further in Hua Hong Semiconductor, its second-largest foundry after SMIC, as part of an all-in AI compute race; a Super Micro contractor pleaded guilty to smuggling $2.5 billion worth of Nvidia chips to China in what appears to be the first criminally convicted chip-smuggling case. (Oct 10–11)?
• Chey Tae-won, chairperson of SK Group, urged Korea to accelerate chip plant expansion, warning that slower action versus Chinese companies or Micron would shrink Korea's market share; Goldman Sachs modeled Samsung's operating profit at 373.5 trillion won in 2026, more than eight times 2025, with conventional DRAM the dominant driver. (Oct 11)?

• Anthropic shipped a 1,000-agent swarm configuration that caught 66 of 70 bugs in parallel; OpenAI's Codex gained a next-message prediction feature; Microsoft released Decision-1, a single-pass decision-scoring model post-trained from Qwen3.5-9B already deployed across incident response and quality control internally. (Oct 12; Oct 9)?
• Cloudflare released Clef-omni adding audio and video input to its open-weight decision model family, while also cutting Clef-flash pricing below Jev; SpaceXAI (xAI) published Grok 4.6 and Grok 4.5 with 500,000-token context windows. (event dates unknown, published Oct 9)?
• OpenAI produced 722 mathematical manuscripts grouped into 372 research families, prompting mathematicians to examine whether the proofs resolve long-standing problems, drawing comparisons to how Go AI transformed that field. (Oct 6 event, covered Oct 10–12)?
• A forward safety note: Meta's Muse drew concerns about model behavior at capability frontiers, and OpenAI researchers surfaced internal disputes over safety governance; SoftBank committed a $100 billion AI fund alongside renewed discussion of catastrophic AI scenario planning. (Oct 12)?