Thursday, October 15, 2026

TSMC's Oct 15 results showed Q3 revenue of NT$1,494.2B (US$46.72B), up 50.9% year over year, a 66.4% gross margin, net income up 68.7%, and 2nm at 8% of wafer revenue versus 3% in Q2, filling in margins behind last week's above-guidance revenue print. Management said growth will exceed its ~55% CAGR target and committed capex well above the prior three years through 2028; a self-operated Texas expansion is only being explored.

Anthropic disclosed Q3 revenue of $16.8B at an Oct 14 private investor meeting, up ~45% from the $11.5B Q2 in its leaked S-1 but below a ~$18B buy-side estimate, and semiconductor stocks fell after hours. Margins were 85% adjusted gross and 11%+ adjusted operating with positive free cash flow and a ~$110B year-end annualized target, but no ARR or valuation; the S-1 expected around Oct 26 and curtailed distillation revenue are the checkpoints.

SK Hynix on Oct 14 made its planned US NAND fab conditional on Washington approving more equipment exports to its Dalian, China plant, while Korean memory stocks extended the CXMT-triggered selloff to five sessions (SK Hynix ~-25%) on added 2029-oversupply fears. Pricing cuts the other way—Barclays shows spot DDR5 up 8% month over month, Micron quotes Q4 2026 SSD pricing up 40–45% versus 20% consensus, and J.P. Morgan models a 23-week HBM shortage by 2028—leaving the export decision as the swing variable.

CoreWeave shares fell 15.5% on Oct 14 as its 2032 bonds slid to 77 cents and five-year CDS broke through 1,000bp, beyond the intraday approach noted yesterday, while Oracle's five-year CDS widened 26bp to a record 288bp. Morgan Stanley's Q3 CIO survey found only 13% of CIOs report ROI above AI token spend versus 28% below, yet 83% are adjusting procurement, including pulling 2027 spending into 2026; whether pull-forward offsets financing stress is unresolved.

Google's first TPUv7 FP8 submission to SemiAnalysis's Agentic Inference Benchmark, reported Oct 14–15, showed 1.67x better performance per dollar than Nvidia's GB300 NVL72 at FP4—the first result to beat Nvidia's top system there—while a vLLM/Inferact/NVIDIA co-design called wideEP claimed 7.2x better performance per TCO on that rack versus other chips. At OCP, Meta said it needs over 100 million optical transceivers in 2027 and will pilot near-packaged optics; benchmark comparability and CPO manufacturing readiness remain uncertain.

• Anthropic disclosed Q3 revenue of $16.8B at a private San Francisco investor meeting, up ~45% from Q2's $11.5B but below a buy-side estimate of ~$18B, sending semiconductor stocks lower in after-hours trading. (Oct 14)?
• The company reported adjusted gross margin of 85%, adjusted operating margin above 11%, positive free cash flow, and told attendees it targets ~$110B annualized revenue by year-end, while withholding an ARR figure and valuation. (Oct 14)?
• Growth decelerated sharply from Q2's ~140% sequential rise; distillation revenue, estimated at 5–7% of total, was materially curtailed during Q3, accounting for roughly 5–6% of the revenue shortfall. (Oct 14)?
• Anthropic is preparing an S-1 filing expected around Oct 26 and is designing a custom chip for internal use; the pre-IPO S-1 draft devoted nearly a third of its pages to risk factors, including existential AI risk disclosures. (Oct 14)?

• TSMC reported Q3 revenue of NT$1,494.2B (US$46.72B), up 50.9% YoY, with gross margin of 66.4% and net income up 68.7%; 2nm rose to 8% of wafer revenue from 3% in Q2, and HPC reached 68% of revenue. (Oct 15)?
• TSMC's CEO said AI demand shows no sign of slowing and that growth will exceed the prior ~55% CAGR target; the CFO committed to CapEx significantly above the prior three years through 2028 and confirmed TSMC is exploring a Texas expansion it would insist on operating directly. (Oct 15)?

• Korean memory stocks fell sharply over five sessions — SK Hynix down ~25%, Samsung ~20%, Micron ~15% — driven by two fears: near-term CXMT competition after China's CXMT announced a next-generation DDR5 RDIMM using 4F² architecture, and a projected DRAM oversupply beginning around 2029 as capacity expansions accumulate. (Oct 14–15)?
• SK Hynix is conditioning its planned US NAND fab on Washington approving additional US chipmaking-equipment exports to its Dalian, China facility; its US subsidiary Solidigm wants to fund the American plant partly from expanded Chinese output. (Oct 14)?
• Despite the stock declines, spot DDR5 prices rose 8% month-over-month in October and 20% over three months per Barclays; Micron is quoting Q4 2026 SSD pricing up 40–45% QoQ versus consensus of 20%; J.P. Morgan estimates the HBM market grows from ~$64B in 2026 to ~$160B/$280B in 2027/2028 with a 23-week accumulated shortage by 2028. (Oct 14–15)?
• Nvidia's next-generation Rubin accelerators are expected to use predominantly 12-layer HBM4; Google is in discussions to extend memory long-term agreements beyond 2030 to 2035; Morgan Stanley's supply-chain checks show the AI semis supply chain strengthened over the prior 30 days with most conviction in DRAM, NAND, HDD, and GPU. (Oct 15)?

• CoreWeave's stock fell 15.5% on Oct 14 after bond prices declined to worryingly low levels; CoreWeave's 2032 bonds fell to 77 cents on the dollar and its 5Y CDS broke through 1,000bp; Oracle's 5Y CDS widened 26bp to a record 288bp; Gartner expects 80% of neocloud providers to cease independent operations by 2030. (Oct 14–15)?
• Digital Realty re-signed large leases at rates 66.7% higher as assumed debt costs rose to as much as 5.5%; AEP agreed to supply Google via a $4B fuel cell plant in Wyoming and is also in a deal for up to 1.6 GW of Mainspring linear generators, illustrating how hyperscaler contracts backstop large generation investments. (Oct 14–15)?
• FERC rejected most of PJM's data center curtailment plan but approved a large load registry; Oregon's governor effectively imposed a moratorium on new data center permits until July 2027; NERC reliability standards for AI data centers and crypto miners of 50 MW or more moved closer to adoption. (Oct 14–15)?

• At the OCP Global Summit in San Jose, NVIDIA disclosed that the Vera Rubin Ultra NVL576 will use co-packaged optics for scale-up; a vLLM/Inferact/NVIDIA co-design called wideEP achieved 7.2x better performance per TCO on GB300 NVL72 versus other AI chips by exploiting the copper backplane's 72-GPU scale-up domain. (Oct 14)?
• Google submitted its first TPUv7 FP8 vLLM results to SemiAnalysis's Agentic Inference Benchmark, demonstrating 1.67x better performance per dollar than the GB300 NVL72 FP4, the first result to beat NVIDIA's top system on that benchmark. (Oct 14–15)?
• NVIDIA launched Nemotron 3.5 Super VL (120B parameters, 5x faster task completion than comparable open models) and Nemotron 3.5 Bolt (74B, 24GB footprint for local RTX/DGX devices), and announced a $250M five-year compute commitment to UK agentic safety research with new investments in UK startups. (Oct 14–15)?
• Nvidia CEO Jensen Huang confirmed that SpaceX and Nvidia are co-engineering a new server type called Star Mind for space-based data centers, expected to reach volume production in 2026; Ian Buck, Nvidia's vice president of hyperscale and HPC, cited three reasons customers choose GPUs: performance, software ecosystem, and density, noting service providers can now fit 30% more GPUs per data center footprint. (Oct 14–15)?

• Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full (collectively Muse) on Oct 12 — the company's first frontier models since Llama 4 — trained on the Prometheus cluster with a 1M-token context window; Truist projects Muse will generate ~$28.5B in incremental revenue by FY2030 in its base case. (Oct 12, analyst note Oct 15)?
• Google made Gemini 4 Argon broadly available on Oct 13 with introductory API pricing of $2/$10 per 1M tokens; OpenAI's agent orchestration platform is used by twice as many companies as Anthropic's per a VentureBeat September survey; Anthropic's Claude Code added idle compaction to auto-refresh 1-hour cache TTLs and save user costs. (Oct 13–14)?
• Microsoft announced Maia 300 — its third-generation AI accelerator, 96 chips in a 167kW rack with 18 TB/s HBM4 bandwidth and a scale-up fabric for 12,288 accelerators on TSMC N2P — and plans to make it the primary inference system for Microsoft's own products and services starting next year. (Oct 14–15)?
• OpenAI named Uday Ruddarraju, formerly head of infrastructure engineering at xAI during the Colossus build, as CTO of Compute; the company aims to move from fully off-premises to self-build infrastructure; OpenAI is also accelerating its advertising business, with nearly half of its ~140 ad employees joining between July and September. (Oct 15)?

• The US is shifting CHIPS Act implementation from construction subsidies toward demand creation, urging chipmakers to use existing domestic capacity; separately, Tower Semiconductor sued Intel for fraud over an abandoned manufacturing deal following the collapse of their $5.4B merger, claiming over $450M in damages. (Oct 15)?
• TSMC said it sees no ceiling for its profitability and confirmed it is exploring a Texas expansion it will insist on operating itself; Chenming Hu, former TSMC CTO and Academia Sinica academician, attributed TSMC's advanced-node lead to early preparation and high yields. (Oct 15)?
• ASE Technology Holding's shares rose ~177% since end-2025 on AI and advanced packaging demand; ASE launched vertically integrated 3D SiP and PowerSiP solutions targeting AI data center power supply; MediaTek is targeting 300kW rack designs compatible with OCP and NVIDIA NVLink Fusion. (Oct 15)?
• ASML's China chip equipment sales rose 109% despite US export-curb pressure, with China expected to account for 20% of ASML's 2026 total sales; China's chip exports jumped 136% to US$45B; AMD announced price increases on adaptive embedded computing products effective December 27, 2026 citing supply-chain cost inflation. (Oct 14–15)?

• Meta told OCP attendees it needs over 100 million optical transceivers in 2027 (six per GPU) and will pilot near-packaged optics in its scale-out network; NVIDIA's Vera Rubin Ultra NVL576 CPO uses eight-wavelength DWDM, and co-packaged optics technology is moving from R&D to manufacturing readiness with Ayar Labs advancing toward production-scale deployment. (Oct 14–15)?
• OCP Day 2 was dominated by near-packaged optics, CPO/NPO format competition (six competing formats presented between pluggable and full CPO), CXL memory hierarchies extending beyond HBM, and power architectures progressing from ~100kW racks toward 1MW+ designs. (Oct 14)?
• Optics and testing stocks fell broadly on Oct 14–15 — AAOI -8%, COHR -8%, LITE -7%, AEHR -10%, FORM -10% — linked to rate-hike fears and AI infrastructure cost concerns; ASML is building a hybrid bonding die-to-wafer tool based on patent filings and roadmap evidence, which could pressure Besi. (Oct 14–15)?

• September PPI came in at 6.3% YoY versus 5.3% estimated and CPI at 3.7%, well above the Fed's 2% target, reinforcing expectations of another rate hike; AI infrastructure and neocloud stocks led the broader market decline on Oct 14–15. (Oct 14–15)?
• A Morgan Stanley Q3 2026 CIO survey found only 13% of CIOs reported ROI above AI token spend, versus 28% reporting ROI below token spend; 83% of CIOs are making at least some procurement adjustments including pulling 2027 spending forward into 2026 due to higher prices and longer lead times. (Oct 14–15)?
• Mizuho projects fast inference will grow from ~$1B in 2025 to $550B by 2030 (291% CAGR); Arm projects the overall CPU market reaches $300B by 2030 from ~$52B in 2026, driven by agentic AI workloads demanding dedicated CPU capacity for orchestration and persistent tasks. (Oct 15)?