Friday, October 9, 2026

Lumentum CEO Michael Hurlston told Bloomberg TV on Oct 9 that the company's AI optical parts are effectively sold out through nearly 2029, with some products facing 70% unmet demand in 2027 and laser chips the tightest bottleneck, extending the laser shortage flagged earlier. P Equity Research checks put 2026 optical-engine output near 60 million units but shippable volume at 17–18 million; assembly yield is the variable.

Samsung has locked 75–80% of its 2027 memory supply on long-term agreements with Nvidia, Google, and Microsoft, per Oct 9 reports, a day after its record quarter and smartphone output cut. Apple meanwhile cut iPhone 18 Pro component orders at least 15–20% in October on memory-driven price resistance, and Bernstein pins ~80% of semis' 119% year-to-date growth on memory pricing; consumer demand destruction is the counterweight.

Firmus, an Nvidia-backed AI cloud firm, pulled its planned $5 billion IPO on Oct 8–9 after finding insufficient demand, having built 42 MW of roughly 1 GW planned, while a same-day 30-year Treasury auction at 5.618% was heavily oversubscribed. A day after Broadcom's reported $50 billion-plus OpenAI chip-debt talks, SemiAnalysis flags $7 trillion of AI debt by 2029 and Electron Economics says CME GPU futures alone will not satisfy lenders; the terms of the next neocloud raise are the test.

Bloomberg reported on Oct 9 that Oracle is trucking compressed natural gas to AI data centers in Utah and Texas to bypass pipeline delays, a day after reports it sought financing for chip purchases. Powering just 100 MW of the 2.45-GW Project Jupiter site this way would require a large trailer every 40 minutes at roughly four times pipeline rates, so pipeline interconnection pace determines whether this stays a bridge or becomes a structural cost.

• Bloomberg separately reported OpenAI projects reaching $70B ARR by year-end, implying ~40% growth from September's $50B figure, driven by enterprise expansion. (Oct 9)?

• Lumentum CEO Michael Hurlston told Bloomberg TV the company's AI optical parts are effectively sold out through nearly 2029; some products face 70% unmet demand in 2027 and 30% in 2028, with laser chips the biggest bottleneck. (Oct 9)?
• Morgan Stanley expects Corning, Coherent, and TFC to play larger CPO roles than FOCI in Nvidia's Rubin NVL576 generation; Coherent is also being considered as a key fiber-kit supplier. (Oct 9)?
• Taiwan's upstream laser chain reported explosive September revenues: LandMark Optoelectronics up 190% YoY; P Equity Research supply chain checks show optical engine production could reach ~60 million units in 2026, but downstream assembly yield limits available shipments to 17–18 million. (Oct 8–9)?
• Bank of America forecasts CPO and NPO compound annual growth of ~200% from 2026 to 2030; Bernstein projects total networking TAM reaching $254B by 2030, with cloud/data-center the fastest-growing segment at ~39% CAGR (2026-28). (Oct 9)?

• Samsung has locked 75–80% of its 2027 memory supply on long-term agreements with Nvidia, Google, and Microsoft; the LTA model is spreading from HBM into MLCCs and power infrastructure. (Oct 9)?
• According to Bernstein, memory pricing drove ~80% of the semiconductor industry's 119% YTD revenue growth through August 2026. (Oct 9)?
• According to Photon Capital, Nvidia's NVHBM architecture moves the memory controller to the base die, reportedly prompting Micron and Samsung to outsource base-die production to TSMC; SK hynix leadership signaled confidence in sustained strong chip demand. (Oct 9)?

• According to Bloomberg, Oracle is transporting compressed natural gas by truck to power its AI data centers in Utah and Texas, bypassing pipeline delays; powering just 100 MW of its 2.45-GW Project Jupiter site would require a large trailer every 40 minutes and cost roughly four times pipeline rates. (Oct 9)?
• Firmus, an Nvidia-backed AI cloud firm, pulled its planned $5B IPO after finding insufficient demand; the company had only 42 MW built of roughly 1 GW planned, and the 30-year Treasury auctioned the same day at 5.618% was heavily oversubscribed. (Oct 8–9)?
• Electron Economics says CME GPU futures could provide a public H100 rental curve but lenders need more than a single benchmark; SemiAnalysis flags $7T of AI debt coming by 2029. (Oct 9)?

• Nvidia plans to invest in d-Matrix, a seven-year-old AI server chip developer that aims to compete with Nvidia, as part of a strategy to make its technology compatible with rival chip designers and draw them into its ecosystem. (Oct 8)?
• SemiAnalysis's InferenceX analysis estimates Nvidia's Vera Rubin NVL72 delivers 3.2× more profit per gigawatt and up to 10× better performance per dollar than the GB300 NVL72, running MiniMax M3 on vLLM. (Oct 8–9)?
• The Information reports Eliyan, a chip interconnect startup valued at $1B in July, has drawn takeover interest and is targeting a $3B valuation; The Information says SoftBank-owned Arm held early talks but did not make a bid. (Oct 8)?
• Synopsys and OpenAI are jointly building a GPT-Synopsys model post-trained on Synopsys tools; Synopsys guided FY27 revenue to $11.15B, up 15% YoY, with debate now focused on who captures the AI agentic-design layer. (Oct 8–9)?

• TSMC's CoWoS capacity could reach 260,000 wafers per month by 2028; Morgan Stanley takes the 2027 global semiconductor market to $2.5T, with AI semis contributing ~one-third, and bottlenecks in CoWoS, HBM, and ABF substrates. (Oct 9)?
• The semiconductor industry is transitioning from transistor-level scaling to system-level scaling, with advanced packaging, chiplets, and co-integrated optics becoming the primary performance levers as sub-2nm gains diminish. (Oct 9)?
• ASML CEO Christophe Fouquet stated that the AI infrastructure buildout will take 2–5 years and is now broadly recognized as real; next-generation EUV technology may still be 10 years away. (Oct 8–9)?

• A SemiAnalysis report found that among China's nine largest AI developers, only 3.6% of 857 model releases since 2021 included a published safety evaluation; Chinese AI regulation targets outputs rather than training compute, so no frontier obligation applies. (Oct 8)?
• Game maker MiHoYo, smartphone maker Xiaomi, and food-delivery giant Meituan have all entered China's already-crowded large-language-model race, mirroring a pattern where non-AI companies build frontier labs from scratch. (Oct 9)?
• Nomura estimates China's wafer-fabrication-equipment market will grow 27% to $55.4B in 2026 and reach $103.5B by 2030, with domestic foundries and IDMs accounting for ~97% of cumulative 2026–30 spending. (Oct 8–9)?

• The White House froze PERM green-card filings for Microsoft, Adobe, Cognizant, Infosys, TCS, Wipro, HCL, and Capgemini, citing alleged fraud; H-1B visas were not suspended, but PERM is the primary pathway from work visa to permanent residency. (Oct 8)?
• The same week, the Trump administration awarded National Medals of Science to Sergey Brin, Jensen Huang, Elon Musk, and Lisa Su — all immigrants — creating a widely noted policy contradiction. (Oct 9)?
• Microsoft responded that roughly 80% of its H-1B filings were extensions for existing employees, not new foreign hires; the freeze has no announced end date or off-ramp. (Oct 8–9)?

• OpenAI released 722 AI-generated mathematics manuscripts on GitHub, including claimed solutions to 90 of the top 500 open problems in mathematics; the model averaged three hours of compute per solution across ~4,000 problems attempted. (Oct 9)?
• Researchers subsequently noted that the Lean verification in at least one proof — the Navier-Stokes result — checks a formalization that differs from what the English-language proof actually claims, raising concerns about automated verification reliability. (Oct 9)?
• Three OpenAI safety researchers — Tomek Korbak, Mikita Balesni, and Jasmine Wang — say they were fired the prior week and published a letter to leadership alleging they were dismissed for prioritizing safety, linked to the METR/Hugging Face incident. (Oct 8)?
• Anthropic is preparing a usage-policy update prohibiting sustained abusive or cruel behavior toward its Claude models; Google Cloud launched Gemini as a single universal enterprise agent at its Gemini at Work event. (Oct 8–9)?