Hello world, it’s Monday, September 7th.
AMD used its IFA keynote to answer Nvidia’s DGX Station with the Threadripper Halo Station, a liquid-cooled tower pairing a 96-core Threadripper Pro with up to four Instinct MI350P accelerators for local AI development. Kioxia is moving to sample its XL1 CXL module, a NAND-based product designed to replace a portion of DRAM in AI servers, with ties to Nvidia’s CMX platform confirmed. Meanwhile, Foxconn posted a record August with $29.14 billion in sales, up 52% year-over-year on AI server demand, and Anthropic disclosed $517 billion in compute commitments ahead of a mid-October IPO.
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! ❤️
AMD Answers Nvidia’s DGX Station With Threadripper Halo Station
AMD announced the Threadripper Halo Station at its IFA keynote, a tower workstation pairing a 96-core Threadripper Pro 9995WX with two to four Instinct MI350P accelerators for high-end local AI development. Each PCIe MI350P carries 144GB of HBM3E and 2.3 PFLOPS of MXFP8 compute, putting a four-card build at 576GB of accelerator memory alongside up to 2TB of DDR5. The MI350P cards are liquid-cooled — a first for a part that launched air-cooled for PCIe servers — and the demo unit leaned on off-the-shelf components, suggesting AMD can validate and ship quickly. The target is unmistakably Nvidia’s DGX Station, right down to the naming, though AMD spreads the work across multiple PCIe GPUs where Nvidia uses a single B300 plus Grace.
Kioxia XL1 CXL Module Pitches Flash as Viable DRAM Substitute in AI Servers
Kioxia is preparing to sample its XL1, a NAND-based CXL memory module built explicitly to displace a portion of DRAM in AI server racks, with the company reaffirming firm ties to Nvidia’s CMX platform as it readies customer units. The three-tier architecture positions flash-backed CXL capacity beneath fast DRAM rather than beside it, targeting cost-sensitive memory expansion slots where bandwidth requirements are tolerable but capacity density matters. The XL1 is the first flash-for-DRAM product from a major NAND maker aimed at production AI workloads, not just storage tiering. A reported long-term NAND supply agreement with Apple, cited by TrendForce, gives Kioxia the volume backstop to sustain the fab utilization the module strategy requires. Synopsys has separately updated its CXL 4.0 IP portfolio, a sign that the interconnect ecosystem is building toward exactly the memory-expansion configurations Kioxia is targeting.
Anthropic Discloses $517 Billion in Compute Deals Before Mid-October IPO
Across nine compute corridors signed in eleven months, Anthropic committed $517 billion in total agreements, with more than $100 billion flowing to AWS alone. FluidStack is leading a $50 billion slice of the build-out, while Nvidia sits in the mix as both hardware supplier and, in the ninth corridor, something closer to landlord. Anthropic has also expanded its revolving credit facility to $15 billion ahead of the public filing. The IPO itself, according to sources cited by Reuters, is now tracking toward a mid-October launch, putting a firm calendar date on what will be the first public market pricing of AI-native compute demand at this scale.
Huawei Kirin 9050 Pro Debuts in Mate XT 2 Tri-Fold After Six-Year Gap
Huawei launched the Mate XT 2 tri-fold smartphone in Guangzhou on Monday, pairing it with the Kirin 9050 Pro, its first flagship SoC in six years, built on the company’s proprietary Tau Scaling Law rather than conventional transistor-shrink approaches. Richard Yu Chengdong, chairman of Huawei’s consumer business group, called it the company’s “most powerful Kirin chip ever” at the event. The Kirin 9050 Pro’s Tau Scaling Law architecture addresses overheating problems that have dogged prior domestic chips, according to a Huawei-published paper cited by SCMP. The device ships with HarmonyOS 7 and a 5600mAh battery. Malaysia has separately indicated openness to Huawei AI silicon despite US warnings, Bloomberg reported, widening the geographic reach of Huawei’s revived chip ambitions.
Foxconn August Revenue Hits Record as AI Servers Carry the Quarter
Hon Hai Precision, the company most know as Foxconn, reported August sales of $29.14 billion, a 52% year-over-year jump that set an all-time monthly record and pushed the company to raise its third-quarter outlook above prior market expectations. The driver is almost entirely Nvidia AI server assembly, with hyperscaler orders filling production lines at a pace that has made the consumer electronics business look like a quiet side project by comparison. Foxconn now expects Q3 to outperform consensus estimates, citing AI demand strength that shows no sign of easing through the end of the quarter. Europe has also begun placing its own AI server orders with the company, broadening the customer base beyond North American cloud giants.
TSMC 3nm Nears NT$400 Billion as Revenue Crown Shifts
TrendForce reports that TSMC’s 3nm output value could exceed NT$400 billion in the second half of 2026, enough to displace 5nm as the company’s single largest revenue-generating node. That would be a first for any sub-5nm process. Demand is being driven by a surge in AI accelerator tape-outs from customers including Apple, Nvidia, and AMD, whose orders have grown large enough to move the aggregate revenue mix on their own. TSMC’s overall advanced-node share already exceeds 70 percent of the global foundry market, and capacity at its Taiwan fabs is running tight as the AI boom strains output. The 2nm node, currently in early ramp, is the next process in line to absorb that demand.
Nscale Seeks $3.5 Billion Pre-IPO as GPU Cloud Valuations Climb
Nscale, the Nvidia-backed, Aker-supported GPU cloud operator, is seeking $3.5 billion in pre-IPO financing ahead of a planned NYSE listing, according to Bloomberg. The raise follows a Wall Street Journal report that Nscale secured a $45 billion contract, which sent Aker shares climbing and put the company’s cumulative revenue commitments into a range that few pure-play compute providers have publicly disclosed. Nvidia holds a stake in the company, lending the capital raise a supplier-alignment angle that has become standard collateral in this corner of the market. A September NYSE debut would give investors their first clean public comparable for pricing CoreWeave and Crusoe, both of which operate similar GPU-rental models without yet trading on an exchange.
Qualcomm Snapdragon 8 Elite Gen 6 Launches September 22 With Stacked DRAM and Higher Prices
Die leaks show the Snapdragon 8 Elite Gen 6 Pro integrating side-mounted stacked DRAM alongside a heatsink, a configuration that pushes on-device memory bandwidth well beyond what discrete LPDDR arrangements can offer. Qualcomm is shipping two variants of the chip, with the confirmed September 22 launch date applying to both. Reference hardware carrying the chip has already surfaced at Notebookcheck, and Samsung is testing the Galaxy S27 across three regions with the unreleased Snapdragon part, per a Notebookcheck exclusive, with both top-end S27 models expected to use it.
OpenAI Publishes Internal View on Research Pace and AGI Governance
OpenAI released a research publication on September 6, 2026, framing the acceleration of its work as inseparable from questions of democratic oversight and public understanding. The company argues that AGI’s benefits depend on informed public debate about capabilities, risks, and safeguards, a position it presents as foundational rather than supplementary to its technical agenda. The publication contends that people everywhere need to understand the likely future trajectory of frontier AI to have any meaningful voice in how it is governed. Short on disclosure, long on framing, the document reads less like a technical report and more like an opening bid in a much larger conversation about who gets to decide what comes next.
Wafer AI Publishes GPU Performance Engineering Resource Repository
Wafer AI posted a curated GitHub repository of AI performance engineering resources. The project targets engineers working on GPU and AI workload optimization, aggregating learning materials into a single reference point rather than scattering them across documentation sites and research papers.
Sector Watch
Data Centers
CoreWeave will report its first quarter 2025 financial results on a date to be confirmed, marking its first earnings release as a public company. (CoreWeave)
Wistron launches a $1.5 billion global depositary share sale to fund raw-material purchases as AI server order books swell. (Reuters)
TCS commits up to $7.4 billion to build a one-gigawatt AI data-center campus in Hyderabad, one of India’s largest single infrastructure pledges, per Reuters. (Reuters)
Compute
Cerebras launches the CS-4 system claiming 750 petaflops and 30x performance over its predecessor, targeting large-model AI training and inference workloads.
AMD unveils the Threadripper Halo Station at IFA 2026, a high-density AI developer workstation pairing Threadripper Pro CPUs with purpose-built inference acceleration. (ServeTheHome)
Cisco expands its Secure AI Factory architecture with a Supermicro partnership, adding validated rack-scale AI infrastructure designs for enterprise deployments. (SecurityBrief Australia)
Foundry & Packaging
SKC will invest approximately 281 billion won in glass-substrate subsidiary Absolics as part of a 400 billion won round, accelerating mass-production readiness for advanced packaging. (The Elec)
Hanmi Semiconductor 2.5D thermo-compression bonders pass Taiwan foundry qualification, opening outsourced advanced-packaging volume to non-Samsung customers. (digitimes.com)
Kools pinpoints through-glass via metallization as the critical bottleneck to glass-substrate mass production and is developing targeted solutions to unblock volume ramp, per The Elec. (The Elec)
Power
Bloom Energy confirmed for S&P 500 inclusion ahead of the September 21 open, with Q2 revenue up 166% year-over-year, as the company progresses an Oracle data-center fuel-cell project in New Mexico. (barrons.com)
Infineon launches dual-phase smart power stages engineered for AI accelerator and data-center processor voltage-regulation, setting a 2 A/mm² current-density benchmark. (TimesTech)
GE Vernova launches a medium-voltage UPS system targeting continuous power protection for AI data centers and other high-density compute facilities. (Energies Media)
Policy & Trade
Belgium detains a Chinese national on suspicion of semiconductor espionage in what Reuters reports as one of the first such arrests in Europe tied to chip-technology theft. (Reuters)
The White House announces a national Foundry School with Arizona State University as a founding partner, targeting workforce development for domestic semiconductor manufacturing. (PR Newswire)
Taiwan is deploying chip diplomacy to distribute AI semiconductor access to allied nations as it faces mounting pressure to share its technology wealth, per Reuters. (Reuters)
Edge
MediaTek launches the MT8875, a 5G IoT chip with integrated edge generative-AI acceleration targeting industrial and smart-device applications. (Ubergizmo)
ASUS and Qualcomm deploy Snapdragon-powered AI pharmacy assistants to community pharmacists across Taiwan in a Copilot-style on-device inference rollout. (Qualcomm)
SambaNova unveils a new inference architecture designed to overcome the AI memory wall, enabling larger model deployment without proportional memory-bandwidth scaling. (Jon Peddie Research)
Optics & Networking
Vocus reroutes internet traffic after a break in its Australia-Singapore Cable in Indonesian waters disrupts a key trans-oceanic data route, per Light Reading. (Light Reading)
IQE CEO warns of an emerging China supply risk for compound semiconductors used in photonics and chip interconnects, flagging a gap in Western supply-chain resilience, per Bloomberg. (Bloomberg.com)
Google is establishing an AI chip design center in Israel by recruiting talent from local accelerator startup Hailo, per Haaretz. (Haaretz)
Tell us how to make Semi Doped better! Give us your feedback! ❤️