Tuesday, October 13, 2026

On Oct 12, China's CXMT reported 'key progress' on a 4F² DRAM design it plans to launch by end-2026, sending SK Hynix down about 4.5% and Samsung more than 4% in Seoul.

Broadcom CEO Hock Tan said over Oct 12–13 that chip and memory shortages are easing but capital availability is becoming the harder data-center constraint, warning many AI spenders cannot yet show revenue from the investment. Nvidia's finance team said Oct 13 demand would exceed supply even at 70–100% 2027 growth, naming land, power and rack shells as co-equal limits.

AMD CEO Lisa Su unveiled the sixth-generation EPYC 'Verano' server processor at the OCP Global Summit on Oct 12, projecting the server CPU market above $200 billion by 2030 at roughly 50% annual growth, a forecast raised three times in nine months. Supply-chain checks on Oct 12–13 put 2026 Rubin VR200 NVL72 rack shipments near 5,000, disputing that Morgan Stanley's sub-10,000 figure reflects a cut.

Artificial Analysis on Oct 13 scored Muse Wonder Full at 56 and Muse Wonder at 52 on its Intelligence Index, tying Sonnet 5.5 and trailing Opus 5.5 at 58, following the Oct 12 release of the model family Meta announced yesterday. Both top Muse Spark 1.3's 48, putting Meta Superintelligence Labs at frontier parity after 14 months; agentic-task performance remains untested.

Microsoft on Oct 13 reorganized its Azure cloud leadership and asked Azure president Girish Bablani's roughly 11,500-person group to identify at least 20% cost reductions, even as Azure revenue growth has accelerated. The move follows the Oct 12 Maia 300 launch, which marketed performance-per-watt rather than raw speed; whether the cuts reach infrastructure spending or only operating overhead is the open question.

• AMD CEO Lisa Su at OCP Global Summit 2026 unveiled the 6th-Gen EPYC 'Verano' CPU and projected server CPU TAM growing ~50% CAGR to over $200B by 2030, raising the forecast three times in nine months. (Oct 12)?
• Nvidia's Rubin VR200 NVL72 rack shipments in 2026 are now estimated at ~5,000 units; Morgan Stanley had cited a below-10,000 figure vs. an original 15,000 target, but checks suggest the original order was always ~5,000, not a cut. (Oct 12–13)?
• Intel began customer sampling of its D270 data-center GPU at OCP, pitching inference capacity and power efficiency over training speed; Intel CEO Lip-Bu Tan said AI compute shortages could last through 2029. (Oct 12)?
• AMD landed megakernel-decode support in upstream vLLM on ROCm before Nvidia's CUDA, marking one of the first major vLLM features exclusive to ROCm, according to SemiAnalysis. (Oct 12)?

• Nanya Technology posted Q3 2026 gross margin of 85.7%—a record for Taiwan wafer makers—on revenue of NT$133.6B, up 61.9% quarter-over-quarter, with AI infrastructure now above 20% of sales; Nanya's general manager said DRAM prices will keep rising in Q4 and buyers are seeking three-year supply deals. (Oct 12)?
• Micron's Sanjay Mehrotra-led board authorized a $35.16B share buyback—the largest in company history—set to begin December 9, following Micron's record fiscal quarter of $54.2B revenue and $33.42 non-GAAP EPS. (Sep 30 results; buyback effective Dec 9)?
• CXMT disclosed 'key progress' on a 4F² DRAM cell structure without EUV, triggering a two-session selloff: SK Hynix fell ~10%, Samsung ~8%, and Micron ~4–5%; the KOSPI dropped 9.7% on heavy foreign selling concentrated in Korean memory names. (Oct 12–13)?
• HBM4 specs remain contested: the market digested reports of an 8-high despec from 12-high, 12-high stack defect issues, and competitive divergence between Samsung and SK Hynix; SK Hynix's upcoming earnings are seen as the next key data point. (Oct 13)?

• Microsoft simultaneously reorganized its Azure cloud leadership and asked Azure president Girish Bablani's group of ~11,500 employees to identify at least 20% cost reductions, even as Azure revenue growth has accelerated. (Oct 13)?
• Power-efficiency has become the primary marketing axis for AI chips: Microsoft's Maia 300 launch emphasized performance-per-watt rather than raw speed, a marked shift from the H100 era. (Oct 12)?

• Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, scoring 52 and 56 on the Artificial Analysis Intelligence Index v4.3—each a material gain over Muse Spark 1.3's 48—completing MSL's 14-month climb to frontier status. (Oct 12)?
• Artificial Analysis benchmarks scored Muse Wonder Full at 56 on its Intelligence Index, tying Sonnet 5.5 and trailing Opus 5.5 at 58, with Astra, Gemini 4 Argon, and Fable 5.1 at 53. (Oct 13)?
• Global inference token volume reached 39 quadrillion in September 2026, 7.9× the volume from September 2025 and up 16% month-over-month, underscoring surging compute demand across frontier labs. (Sep 2026)?

• AI-hardware and optical stocks fell sharply across two sessions: Ciena −11.9%, Lumentum −11.8%, Credo −10.6%, Coherent −10.2%, and memory names down 4–10%; the selloff was partly triggered by a misread of Broadcom CEO Hock Tan's 'casino gambler' comment, which in context described investors who sell on a single short-term data point, not a bearish industry call. (Oct 12–13)?
• Broadcom's Hock Tan said wafer and memory shortages are easing but capital availability—not chips—is becoming the harder constraint for data-center buildout, and warned that many AI spenders cannot yet show revenue or profit from the investment. (Oct 12–13)?
• Retail sentiment as measured by Google Trends for 'Stocks to Buy' has fallen ~50% since June, coinciding with the AI supply-chain sell-off, while institutional investors have rotated from Korean tech to US names. (Oct 13)?
• J.P. Morgan said AI accelerators and CPUs remain the primary infrastructure bottleneck for at least 18 months; a 3Q26 CIO survey showed hardware spending expected to accelerate further into 2027. (Oct 13)?

• Bank of America raised its TSMC price target from $590 to $620 with a buy rating, citing 2nm ramp contribution and 5nm-and-below utilization pinned near 90–100% through 2028 on AI accelerator and custom chip demand. (Oct 13)?
• ASML unveiled a Hyper NA EUV system targeting 5nm single-exposure patterning ahead of its October 14 earnings; ASML's H1 2026 revenue was €18.1B, requiring €24.9–26.9B in H2 to hit its full-year guide of €43–45B. (Oct 13)?
• Applied Materials officially opened its $5B EPIC Center R&D facility in Sunnyvale, California—a 600,000 sq ft campus with nearly four acres of cleanroom—with TSMC, Intel, SK Hynix, Micron, and Samsung as founding partners. (Oct 12)?
• Teradyne agreed to make a $300M strategic equity investment in FormFactor and entered a broad IC test technology collaboration, targeting accelerated development of probe-card and test solutions for advanced packaging. (Oct 13)?

• Optical supply chain contacts at OCP confirmed demand has never been this high: Lumentum's CEO said active optical products are fully sold out through early 2029 and some customers have locked in 2030 orders, with supply gaps of 30–70% for some products through 2027–2028. (Oct 12–13)?
• Co-packaged optics yield math shows a whole switch yields ~19% at 95% per engine across 32 engines; raising per-engine yield to 99% before bonding pushes whole-switch yield to 72%, making pre-bond wafer-level test the critical value point. (Oct 12)?
• Credo introduced a 1.6T ZeroFlat active light cable drawing 12W per end—about 30% less than its own copper equivalent—while NVIDIA and Cisco jointly proposed USFP, a 16-lane 400G pluggable in roughly an OSFP footprint. (Oct 12–13)?
• Macquarie is forecasting an 89% CAGR for the transceiver market over 2025–2028, with supply/demand conditions remaining tight through 2027. (Oct 13)?

• Nvidia's share of China's data-center AI chip market fell to 34% in the first half of 2026, with 678,000 accelerators shipped—down 44% year-over-year—while Huawei shipped 572,000 chips for a 28% share, and non-GPU chips from Chinese suppliers surpassed GPUs by volume. (H1 2026)?
• Chinese AI startup Kimi has shifted 35–40% of its inference compute to domestic chips, up from ~15%, gradually replacing Nvidia hardware; lower-priced models carry 20–25% gross margins vs. 35–40% for flagships. (Oct 13)?
• Nvidia's CFO and investor relations team said demand continues to run well above supply and even a 70–100% 2027 growth scenario would not pull ahead of demand; management frames land, power, and rack shells—not just GPUs—as co-equal constraints. (Oct 13)?

• Anthropic disclosed that new defenses—including 'preserved thinking' in Claude Fable 5.1 and Opus 5.5—have reduced the data harvested by distillation attackers by roughly 90–99%; preserved thinking blocks model reuse of reasoning traces by unauthorized models. (Oct 12)?
• JPMorgan Chase CEO Jamie Dimon said the bank will not allow any outside AI agent access to internal systems due to client risk, while JPMorgan CFO Jeremy Barnum flagged that customers who give credentials to third-party agents bear responsibility for those agents' actions. (Oct 13)?
• OpenAI is scrambling to address safety issues with its next unreleased flagship model, which uses a loop-transformer or recurrent-depth architecture that makes monitoring internal reasoning harder; Microsoft accidentally disclosed the technique's use in GPT-6.1 Sol before editing the post. (Oct 13)?
• Microsoft CEO Satya Nadella argued that agentic AI systems need deterministic controls separate from frontier models, saying organizations cannot treat superintelligence as nested black boxes. (Oct 11, published Oct 12)?