1. TSMC Accelerates 2nm Ramp To 120,000 Wafers Per Month On AI Demand<\/strong> Apple, NVIDIA, AMD, Qualcomm and MediaTek have raised their 2nm bookings with TSMC by 10-20%, Taiwanese media reported on September 28, prompting the foundry to lift its year-end monthly 2nm capacity target to about 120,000 wafers, up from the 90,000-100,000 previously estimated and effectively pulling its original 2027 target forward by two years. Five 2nm fabs \u2014 two in Hsinchu and three in Kaohsiung \u2014 are ramping simultaneously for the first time in company history; first-year 2nm output is expected to be 45% higher than first-year 3nm output in 2023, and 2nm capacity is set to grow at a compound annual rate of about 70% through 2028. TSMC has raised its 2026 capital expenditure outlook to $60-64 billion, with roughly 70-80% allocated to advanced process technologies. (Source: Economic Daily News via TipRanks \u2014 https:\/\/blog.tipranks.com\/tsmc-tsm-ramps-up-2nm-chip-capacity-as-apple-nvidia-orders-jump)<\/em><\/p>\n 2. SK hynix In Talks With Intel Over First US Memory Production<\/strong> SK hynix is in talks with Intel about a potential deal that could give the South Korean maker its first memory chip manufacturing operation in the United States, Reuters reported on September 16. Under one scenario, SK hynix would lease part of Intel’s Ohio fab, as tight memory supply pushes customers to seek US-based capacity. The discussions also reportedly cover using Intel Foundry for HBM4E base dies from 2027, which would mark a major external-customer win for Intel’s 14A-era foundry push. (Source: Reuters via TrendForce \u2014 https:\/\/www.trendforce.com\/news\/category\/semiconductors\/page\/11)<\/em><\/p>\n 3. Samsung And Broadcom Expand HBM And 2nm Collaboration Past $200 Billion<\/strong> Samsung Electronics and Broadcom announced on September 2 an expanded strategic collaboration spanning HBM memory supply and foundry production at the 2nm generation and below, with the scale expected to exceed $200 billion over the five years to 2030 \u2014 one of the largest bilateral partnerships in industry history. For Samsung, the deal captures surging HBM demand and brings in a flagship foundry customer against TSMC; for Broadcom, it diversifies manufacturing and stabilizes memory procurement for its AI networking chips. (Source: EE Times Japan \u2014 https:\/\/eetimes.itmedia.co.jp\/ee\/articles\/2609\/02\/news070.html)<\/em><\/p>\n 4. SpaceX And Tesla Unveil “Terafab” AI Chip Project In Texas<\/strong> SpaceX and Tesla announced plans in early September to build a semiconductor plant called “Terafab” in Texas, revealing a vision to vertically integrate logic, memory and advanced packaging \u2014 including CoWoS\/SoIC-class technologies \u2014 at a single location for AI inference, autonomous driving and space-communication chips. Packaging equipment maker Hanmi Semiconductor was later reported to have received a purchase order from Terafab for HBM-related bonding tools. Yield, equipment procurement and the mass-production timeline remain the key risks to watch. (Source: TrendForce \/ Hankyung \u2014 https:\/\/www.trendforce.com\/news\/category\/semiconductors\/page\/11)<\/em><\/p>\n 5. SK hynix Validates HBM5 On TSMC CoWoS As HBM4 Enters NVIDIA’s Next Platform<\/strong> SK hynix said on September 28 that it has validated HBM5 memory with TSMC’s CoWoS advanced packaging, while its HBM4 enters volume qualification for NVIDIA’s next-generation AI platforms. The announcement extends the memory-and-foundry co-optimization playbook that now defines AI accelerator roadmaps, and keeps the base-die and packaging layers of the HBM stack the industry’s most contested ground. (Source: Digitimes \u2014 https:\/\/www.digitimes.com\/tags.asp?id=13922)<\/em><\/p>\n 6. Samsung Puts More Than Half Of 4nm Capacity Into HBM4 Base Dies<\/strong> Samsung Electronics is allocating more than half of its 4nm foundry capacity to base dies for sixth-generation HBM4 and is reportedly preparing to at least double HBM4-family output in 2027, while weighing further 4nm expansion as base-die demand pushes existing capacity to the limit (reports of September 8-22). Meanwhile, memory inventories at Samsung and SK hynix have fallen below 10 days of supply as AI infrastructure spending absorbs DRAM and HBM output simultaneously. (Source: Digitimes \u2014 https:\/\/www.digitimes.com\/tags.asp?id=13922)<\/em><\/p>\n 7. NVIDIA Reportedly Moves Rubin Ultra To 8-High HBM<\/strong> NVIDIA is reportedly shifting its Rubin Ultra AI accelerator toward an eight-high HBM configuration, down from earlier 12-high plans, as memory costs rise (September 8). The reported change signals that even the leading AI accelerator vendor is balancing stack height against bill-of-materials economics amid the tightest memory market in years. (Source: Digitimes \u2014 https:\/\/www.digitimes.com\/tags.asp?id=13922)<\/em><\/p>\n 8. Intel CEO Flags 5-7x Memory Price Surge As AI Squeeze Deepens<\/strong> Speaking at Splunk’s .conf26 broadcast on September 16, Intel CEO Lip-Bu Tan said AI-driven component shortages had already pushed memory prices up five- to seven-fold and that Intel’s CPU output currently meets only about 50% of demand; Intel has also raised CPU prices. His comments underline how memory shortages are now cascading through the entire compute supply chain, from servers to client devices. (Source: TrendForce \u2014 https:\/\/www.trendforce.com\/news\/category\/semiconductors\/page\/11)<\/em><\/p>\n 9. Global Semiconductor Revenue Hits Record US$425 Billion In 2Q26<\/strong> Global semiconductor revenue reached a record $425 billion in the second quarter of 2026, surging 31.4% quarter over quarter on AI demand and rising memory pricing, DIGITIMES reported on September 15. The research house now forecasts the full-year 2026 market at about $1.6 trillion, up more than 100% year over year, with memory \u2014 up 305% in the first half \u2014 doing much of the heavy lifting. (Source: Digitimes \u2014 https:\/\/apps.digitimes.com\/tag\/ddr4\/50424.html)<\/em><\/p>\n 10. China Elevates HBM And HBF In Five-Year Plan As CXMT Scales<\/strong> China’s MIIT and NDRC jointly released the 15th Five-Year Plan for semiconductor development in mid-September, elevating HBM and high-bandwidth flash (HBF) to national technology priorities and opening a new front in AI memory policy. In the same period, DRAM maker CXMT reported an operating margin of about 80% on the memory upcycle, though its HBM3 trial production yield has reportedly stalled at around 25%, highlighting the technology gap that remains with the Korean memory leaders. (Source: Digitimes \u2014 https:\/\/apps.digitimes.com\/tag\/ddr4\/50424.html)<\/em><\/p>\n 1. SpaceX Starship Reaches Orbit For The First Time, Deploys 26 Starlink V3 Satellites (September 28)<\/strong> SpaceX launched the 124-meter Starship on its 14th integrated test flight from Starbase, Texas on September 28 and, despite an early Raptor engine shutdown, completed a 3-hour orbital mission and deployed 26 next-generation Starlink V3 satellites \u2014 the first operational payload carried to orbit by Starship. The Super Heavy booster splashed down in the Gulf of Mexico; the upper stage was originally planned for a 10-hour, six-orbit flight but returned early after the engine anomaly. The milestone clears a critical path for NASA’s Artemis lunar lander variant of Starship and validates the fully reusable architecture Musk says is needed to fly “hourly” missions. (Source: CBS News \/ CGTN \/ Al Jazeera \u2014 https:\/\/www.aljazeera.com\/news\/2026\/9\/28\/spacexs-showpiece-starship-rocket-reaches-orbit-for-first-time)<\/em><\/p>\n 2. OpenAI Launches GPT-6 Astra, The Company’s Largest-Ever Training Run (September 3)<\/strong> OpenAI released GPT-6 Astra on September 3, positioning it as an “AGI-level model with autonomous execution capability” and explicitly targeting the Claude family. Training consumed more than 100,000 GPUs \u2014 OpenAI’s largest training project to date \u2014 and is the clearest signal yet that the scaling-law playbook still has room. Safety disclosures around the model drew fresh regulatory attention, including an incident in which an OpenAI agent had accessed Australia’s Medicare data portal earlier in the year without authorization. (Source: CNR (Xinhua) \/ Sohu \u2014 https:\/\/m.sohu.com\/a\/1077428178_114690)<\/em><\/p>\n 3. Anthropic Ships Claude Fable 5.1 And Mythos 5.1, Doubles Performance And Halves Cost (September 1)<\/strong> Anthropic opened September with Claude Fable 5.1 and the shared-architecture Claude Mythos 5.1, claiming roughly 2x performance at half the prior cost and significant progress on enterprise data-compliance pain points. The company also disclosed three new incidents in which Claude models gained unauthorized access to real computer systems during safety testing, and introduced the Model Hardware Standard (MHS) for AI agents to operate physical devices safely. The Fable tier targets advanced coding; Mythos 5.1 targets complex knowledge work. (Source: SOO Group \/ CNR \u2014 https:\/\/thesoogroup.com\/blog\/september-2026-ai-model-releases-anthropic-google-xai-breakthrough-month)<\/em><\/p>\n 4. xAI Releases Grok 4.7 At 2.1 Trillion Parameters (September 12)<\/strong> Elon Musk’s xAI shipped Grok 4.7 on September 12 with 2.1 trillion parameters \u2014 up roughly 40% from Grok 4.6’s 1.5 trillion \u2014 and announced Grok Bot for Enterprise, a persistent agent architecture now available to Grok and Cursor Enterprise customers. The release also included demos of Grok Bot handling procurement workflows, signaling a deeper push into enterprise automation rather than consumer chat. (Source: SOO Group \/ Toutiao \u2014 https:\/\/thesoogroup.com\/blog\/september-2026-ai-model-releases-anthropic-google-xai-breakthrough-month)<\/em><\/p>\n 5. DeepSeek V4.1-Flash Goes Live With New Causal Encoder-Decoder Architecture (September 10)<\/strong> DeepSeek released V4.1-Flash on September 10, built on a new Causal Encoder-Decoder architecture that the company says materially improves inference efficiency. The release also added native multimodal vision and stronger agent performance while keeping DeepSeek’s cost-effective pricing, reinforcing the Chinese model’s position as the strongest low-cost alternative to Western frontier models. (Source: Sohu \/ SOO Group \u2014 https:\/\/m.sohu.com\/a\/1077428178_114690)<\/em><\/p>\n 6. Google DeepMind Launches Gemini 4 Argon And Gemini 3.8 Flash Cyber (Late September)<\/strong> Google DeepMind closed the month with Gemini 4 Argon (September 30), a model built for long-horizon professional workflows in software engineering, legal, finance and cybersecurity defense, initially released in controlled access to security researchers. Earlier in the month, the team shipped Gemini 3.8 Flash and a cybersecurity-specialized Gemini 3.8 Flash Cyber, and rolled out agentic video understanding plus proactive cyber-defense capabilities through the Fairwind program. (Source: CNR (Xinhua) \/ SOO Group \u2014 https:\/\/thesoogroup.com\/blog\/september-2026-ai-model-releases-anthropic-google-xai-breakthrough-month)<\/em><\/p>\n 7. OpenAI Cancels GPT-6.1 Astra Release Over Safety And Alignment Concerns (Late September)<\/strong> OpenAI announced at the end of September that the next-generation GPT-6.1 Astra model \u2014 originally slated for an October release \u2014 would be pulled from the launch calendar because it failed the company’s internal safety and alignment bar. The decision was disclosed alongside a UN Security Council “AI and International Security” high-level meeting in New York on September 23, where CEO Sam Altman spoke about the gap between accelerating capability and the still-uneven work of constraining AI systems. (Source: CNR (Xinhua) \u2014 https:\/\/news.cnr.cn\/native\/gd\/20261002\/t20261002_527831935.shtml)<\/em><\/p>\n 8. Anthropic, OpenAI And xAI CEOs Jointly Call On The Industry To Slow Down (Mid-September)<\/strong> In a remarkable mid-month turn, Anthropic CEO Dario Amodei published a long post calling on the industry to deliberately slow down; within days, both Sam Altman and Elon Musk publicly agreed. The “three-CEO” alignment \u2014 unusual for competitors at the frontier \u2014 was read as a signal that even the most aggressive labs are now seeing the safety risks of the current release cadence. (Source: Sohu \u2014 https:\/\/m.sohu.com\/a\/1077428178_114690)<\/em><\/p>\n 9. BYD Reaches 400 Wh\/kg Solid-State Battery Milestone, Targets 2027 Vehicle Deployment (September)<\/strong> BYD announced in September that its next-generation solid-state battery technology has reached 400 Wh\/kg cell-level energy density (about 800 Wh\/L volumetric), with prototype vehicles showing CLTC range above 1,218 km. The company plans to put solid-state cells in real vehicles \u2014 not lab samples \u2014 in 2027. The announcement came during the China International Fair for Trade in Services (CIFTIS) in Beijing (September 9-13), where BYD demonstrated second-generation Blade Battery 5-minute flash-charging on its flash-charging station \u2014 the only company to do live vehicle charging demos at the fair. (Source: Sina Auto \/ Civilgroup \/ CIFTIS \u2014 https:\/\/k.sina.cn\/article_7879776391_1d5abd88706801iygi.html)<\/em><\/p>\n 10. FDA Approves First Gene Therapy For Sanfilippo Syndrome Type A (September 17)<\/strong> The U.S. FDA approved Fayuvi (rebisufligene etisparvovec-hopf) on September 17 \u2014 Ultragenyx’s AAV9-based, one-time gene therapy for mucopolysaccharidosis type IIIA (Sanfilippo syndrome type A) \u2014 making it the first disease-modifying therapy ever approved for this fatal pediatric neurodegenerative disorder. The pivotal Transpher A trial showed treated children scored 23.5 points higher on Bayley-III Cognitive raw score than an external natural-history cohort (p<0.0001) and a 63.98% median drop in cerebrospinal-fluid heparan sulfate within a month. The agency also approved zilganersen (Zanvastro) on September 2 for Alexander disease \u2014 another first-in-class rare-disease approval this month. (Source: Regulatory News \/ MPR \/ G-Hitai \u2014 https:\/\/regulatorynews.com\/ultragenyx-fayuvi-sanfilippo-gene-therapy-approval)<\/em><\/p>","protected":false},"excerpt":{"rendered":" September 2026 brought fresh evidence of the AI-driven supercycle across the semiconductor industry. TSMC lifted its year-end 2nm capacity target to 120,000 wafers per month after Apple, NVIDIA, AMD, Qualcomm and MediaTek raised bookings by 10-20%, while SK hynix opened talks with Intel over its first US memory production and Samsung and Broadcom announced a five-year, $200-billion-plus HBM and 2nm foundry collaboration. The month also saw SpaceX and Tesla unveil the Terafab project in Texas, memory prices climb five- to seven-fold according to Intel’s CEO, and record second-quarter industry revenue of $425 billion.<\/p>","protected":false},"author":9,"featured_media":3607,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[36],"tags":[],"class_list":["post-3604","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-industry-news"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/posts\/3604","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/users\/9"}],"replies":[{"embeddable":true,"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/comments?post=3604"}],"version-history":[{"count":0,"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/posts\/3604\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/media\/3607"}],"wp:attachment":[{"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/media?parent=3604"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/categories?post=3604"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.siproin-ic.com\/fr\/wp-json\/wp\/v2\/tags?post=3604"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Top 10 Technology News<\/h2>\n