Hello world, it’s Tuesday, September 22nd.
Alibaba unveiled what it’s calling China’s most powerful AI chip alongside a new 10-trillion-parameter model at its Apsara Conference in Hangzhou. Texas froze all state-issued data-center permits pending a grid audit, California signed a new oversight package, and TSMC pulled its A14 pilot run forward to Q1 2027 as Qualcomm eyes Samsung for 2nm.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Tell us how to make Semi Doped better! Give us your feedback! ❤️
Alibaba Unveils AI Chip and 10-Trillion-Parameter Model at Apsara
At its annual Apsara Conference in Hangzhou on Tuesday, Alibaba Group Holding introduced what it calls China’s most powerful AI chip, a new accelerator positioned to compete directly with Nvidia and support a major expansion of the company’s data center capacity. Group chairman Joe Tsai told attendees that Alibaba is building “full-stack AI” end-to-end capabilities, a commitment that came paired with the company’s first public mention of plans to train a model reaching up to 10 trillion parameters. Alibaba’s shares rose as much as 5% on the news, according to Reuters. The parameter count would place the forthcoming model well beyond most publicly disclosed architectures, though no release date was given for either the chip or the model at scale.
Vik: Lots of juicy details in the linked SCMP post.
Qwen 5 will be a 5-10T parameter model. Qwen Intelligence is a full-stack agent platform for AI on smartphones.
Zhenwu V900 processor is the most power AI chip in the market today. Clusters of 500,000 GPUs will comprise of these chips. Next gen: Zhenwu J900
216 GB of HBM, 1.2 TB link to other chips. GA in Q1 2027
Plans for next-generation proprietary CPUs tailored for agentic AI tasks – the Yitian 720 and Yitian 730, both due in 2027.
Texas Freezes Data-Center Permits, California Signs Oversight Package
Governor Abbott halted all state-issued data-center permits in Texas pending completion of a grid audit, a sweeping pause that hits the largest data-center market in the US before a single new facility can break ground. California’s governor signed a broad oversight bill package in the same news cycle, adding formal state-level review requirements to what had previously been a tangle of local opposition.
Two states, two different political parties, one shared conclusion: hyperscale buildout needs a leash. Bloomberg notes the permit freeze has put Republicans in an awkward spot, caught between grid-reliability constituents and the industry investment they courted. Neither state has disclosed a timeline for resolving its process, leaving capacity schedules for major cloud and colocation operators in open-ended limbo.
Vik: Its not Dario saying that he will pace the frontier of AI. It’s ultimately NIMBY-ism that will pace it for us. We will humbly take a victory lap for saying this will happen.
Vicor Lifts Q3 Outlook More Than 20% as AI Power Royalties Accelerate
Vicor raised its Q3 2026 revenue guidance to reflect over 20% growth, citing royalties from its Vicor Power-Delivery architecture licensed into high-density AI server racks. The VPD licensing stream means Vicor collects a fee regardless of which GPU vendor occupies a given server slot, a position the company has been building toward as rack power densities climb beyond what conventional voltage-regulator designs can handle. Shares VICR 0.00%↑ have now gained 316% in 2026 heading into the updated guidance, reflecting investor appetite for the royalty model over pure component sales. The guidance revision covers the quarter ending September 2026.
Vik: This 20% bump is a licensing play, not a demand surprise. Vinciarelli basically says unlicensed hyperscalers only called after their systems got banned from import over infringing power modules. VPD is vertical power delivery, feeding the GPU from underneath instead of across the board - that’s how you survive these rack densities.
TSMC Pulls A14 Pilot to Early 2027 as Qualcomm Eyes Samsung for 2nm
TSMC has moved its A14 pilot run forward to Q1 2027, according to TrendForce, compressing what had been a comfortable lead over Intel’s competing 14A node. The schedule shift is a direct response to Intel’s accelerating foundry roadmap. Qualcomm, watching this tighten from the customer side, is reportedly evaluating Samsung as a co-source for 2nm production alongside TSMC, a move that would give the chip designer supply-chain insurance while keeping both foundries under competitive pressure. TSMC pulling A14 to Q1 2027 puts its most advanced node on a collision course with Intel 14A within the same calendar window. Dual-sourcing at 2nm isn’t a simple task; Samsung’s yield track record at leading edge has historically trailed TSMC’s, and Qualcomm’s evaluation is described only as ongoing.
Vik: Qualcomm dual-sourcing 2nm is really about TSMC capacity being locked up by Apple, not love for Samsung. And we’ve seen this one: Samsung built the Snapdragon 8 Gen 1 on 4nm in 2021, it ran hot, and Qualcomm went straight back to TSMC. Yields still trail. but it is still “Insurance”, reportedly.
Google Launches $899 Googlebook as MediaTek Challenges Qualcomm for AI-PC Silicon
Google opened pre-orders for Googlebook laptops starting at $899, a five-model lineup built on a hybrid of Android and ChromeOS with Gemini AI at the center and tighter integration with Android phones. Qualcomm’s Snapdragon X Elite is the launch silicon, giving the chip its first major Google platform win in the premium laptop tier. MediaTek isn’t sitting out: the company debuted the Dimensity CX C10 Max, a 3nm chip with an 8-core Arm CPU, 11-core GPU, and 55 TOPS NPU, built specifically for Googlebook devices, with Lenovo already signed as a design-win partner. Android Central confirmed the Snapdragon X Elite as the processor powering the first wave of Googlebook hardware. Google’s Sameer Samat, president of the Android Ecosystem, framed the platform as a direct challenge to Apple and Microsoft in the AI PC race.
Vik: MediaTek pitches the C10 Max as built for Googlebook, but it’s basically the Kompanio Ultra 910 Chromebook chip rebranded - same X925 core, same Immortalis GPU, just clocked a little higher. Qualcomm still grabs the premium launch slot with Dell and HP.
Sector Watch
Memory
YMTC wins two German patent injunctions against Micron on NAND technology, marking a significant IP offensive by China’s leading flash maker. (IAM Patent)
CXMT boosts dies-per-wafer by 50% with fifth-generation 12nm-class DRAM using quadruple-patterning, advancing without EUV. (chosun.com)
SK hynix CEO Kwak Noh-Jung publicly commits to sustaining HBM and next-generation memory technology leadership as AI demand accelerates. (The Elec)
Foundry & Packaging
Hanmi Semiconductor ships its first FC Bonder 75, entering the 2.5D AI chip packaging market with global delivery of the new flip-chip bonding tool. (Chosunbiz)
JNTC will invest 347 billion won to build a through-glass via (TGV) glass substrate plant in Gimcheon, South Korea, targeting AI packaging demand. (The Elec)
EST files for Kosdaq IPO targeting glass substrate and OLEDoS component markets, joining a wave of Korean advanced-packaging material plays. (The Elec)
Compute & Edge
Cadence expands its ChipStack AI Super Agent with a new RTL-generation and early PPA optimization agent, deepening AI-driven EDA automation. (Business Wire)
Rebellions NPU now powers four live AI services at SK Telecom, marking a commercial deployment milestone for the Korean AI chip startup. (The Elec)
d-Matrix joins Nvidia NVLink Fusion ecosystem with its Raptor inference chip delivering 100 TB/s bandwidth, expanding the in-network compute tier. (shattered.io)
Optics & Networking
Flag selects Ciena’s Waveserver platform to upgrade the Echo subsea cable, extending Ciena’s coherent-optics footprint on a key Pacific route. (Light Reading)
Semtech joins the XPO Multi-Source Agreement with 200G-per-lane modulator driver and TIA ICs, adding merchant silicon to the extra-dense pluggable optics effort. (EE News Europe)
Lightera announces a major capital investment to expand submarine optical fiber manufacturing capacity as hyperscale cable demand surges. (PR Newswire)
Power
Nvidia launches DSX Ready, a qualification program for power and cooling products certified for AI factory infrastructure deployments. (Nvidia News)
Infineon launches the PSOC Control C3 MCU family for real-time power and motor control, adding post-quantum cryptography security to the platform. (Bisinfotech)
Mitsubishi Electric introduces a chip-to-grid power blueprint for Nvidia-powered AI factories, covering conversion from utility supply to chip-level delivery. (Business Wire)
Data Centers
Crusoe closes an initial $3.9 billion Series F backed by Nvidia, pivoting to modular AI data center builds and reaching a $30.9 billion valuation. (HPCwire)
CoreWeave-tied data center project launches a Goldman-led $1.1 billion junk-bond offering, the latest in a string of project-finance deals for leased AI capacity. (Bloomberg.com)
BDx breaks ground on a 640 MW AI data center in West Java, Indonesia, backed by 845 MVA of secured grid power. (PR Newswire)
Policy & Trade
ASML executive warns Europe risks falling behind in chipmaking after revealing no lithography machines have been sold to European fabs. (The Edge Malaysia)
Australia state bans new data centers from residential zones, a first-of-kind zoning restriction with implications for hyperscale site selection across the region. (Bloomberg.com)
China and US hold a second day of trade talks covering AI governance and investment access, with the Pentagon entity blacklist remaining a sticking point. (Bloomberg.com)
Tell us how to make Semi Doped better! Give us your feedback! ❤️