Maia 300 Unveiled, Character.ai Bankruptcy, GlobalFoundries Japan Deal, and More

AI

Microsoft Unveils Maia 300, Its Third-Generation Custom AI Chip

Microsoft announced Maia 300 on October 12, a custom AI inference chip designed for Azure data centers. Built on TSMC’s N2P (2 nm) process with a multi-chiplet architecture, it packs eight HBM4 stacks delivering 288 GB of capacity and 18 TB/s of bandwidth. Peak compute reaches 13.8 PFLOPS at FP4 and 6.9 PFLOPS at FP8.

Each chip draws 950 W, and a full rack holds 96 accelerators at roughly 167 kW. Microsoft claims 40% more tokens per megawatt than leading GPU-based systems, though it did not name the GPU it compared against. These figures are vendor specifications, not independently verified.

Maia 200 is already operational, powering services including Microsoft 365 Copilot. Maia 300 deployments are planned for 2027 at gigawatt scale. Compared to its predecessor, memory bandwidth jumps roughly 2.5x while FP4 compute rises about 37%, making this a bandwidth-centric generational leap.

Character.ai Files for Chapter 11 — Disney Bids $20 Million

AI chatbot startup Character Technologies filed for Chapter 11 bankruptcy on October 12 in the U.S. Bankruptcy Court for the District of Delaware. Disney Worldwide Services has been named the stalking-horse bidder at $20 million for a non-exclusive, perpetual IP license and select assets. Competing bids are accepted through late November or early December.

The company estimated assets and liabilities at $50 million to $100 million each. Lawsuits alleging harm to minors’ mental health and state investigations in Kentucky, Texas, Florida, Pennsylvania, and New Mexico are cited as contributing factors. The platform will continue to operate through the sale process, but will shut down if no buyer commits.

Once valued in the billions, Character.ai’s decline to a $20 million bid illustrates how regulatory exposure and legal liability can erode value rapidly in the consumer AI space. Whether user chat data is included in the sale has not been disclosed.

Big Tech

GlobalFoundries Secures $21 Billion Japan-Backed Deal for Two New York Fabs

GlobalFoundries announced on October 14 that it will operate two new 300 mm fabs in Malta, New York under a U.S.-Japan semiconductor initiative. Japan provides up to $21 billion in investment while GlobalFoundries contributes technology and operational expertise. The first fab, backed by roughly $10 billion in committed funding, targets 2029 operations with about 900,000 wafers per year. A second fab would add equal capacity but its funding remains conditional.

According to Reuters, Apple and Japan’s Renesas are among the initial customers. The project is expected to create approximately 27,000 construction and ecosystem jobs plus over 1,000 manufacturing positions.

The deal underscores a dual trend: the U.S. push to onshore semiconductor manufacturing and Japan’s expanding overseas chip investments. The U.S. Department of Commerce reportedly takes an approximate 4% equity stake, though exact terms have not been confirmed in official filings.

Li Auto’s Chip Subsidiary Machxin Raises RMB 1.625 Billion in External Funding

According to a 6-K filing with the SEC, Li Auto’s chip and domain controller subsidiary Shanghai Mahe Xin Technology (MHX) signed equity subscription agreements totaling RMB 1.625 billion. Li Auto’s subsidiary contributes RMB 1 billion while external investors Hangzhou Jingtao, Lanchi Ventures, Huaye Tiancheng, and MiraclePlus put in RMB 625 million. Hangzhou Jingtao has a one-time option to subscribe additional shares by December 31, 2026.

Earlier reporting from LatePost indicated MHX was targeting a pre-money valuation of roughly RMB 15 billion. This follows Nio’s GeniTech completing a RMB 2.257 billion first external round in February. Chinese EV makers are accelerating chip self-sufficiency as an industry-wide strategic priority.

Startups

DeepSeek Leads Winverse AI Seed Round at $400 Million Valuation

Bloomberg reported that DeepSeek has agreed to lead a seed round for Winverse AI, a robotics AI startup founded in 2026 by former Tencent chief scientist Hu Han. The post-money valuation is approximately $400 million, with Meituan’s Long-Z also participating. The round is worth tens of millions of dollars, though final terms are not yet settled.

DeepSeek previously invested CNY 140.8 million in Unitree’s Shanghai IPO in August and agreed to jointly develop AI models for humanoid robots. The move from model training into robotics investment signals a broadening strategic scope.

Winverse remains in stealth with no public product details. Hu Han previously led the vision-language and embodied foundation model team at Tencent before founding the company.

MMI Closes $125 Million Series D for Robotic Microsurgery

Medical Microinstruments (MMI) closed a $125 million Series D led by BioStar Capital, with S3 Ventures, Angelini Ventures, and an undisclosed strategic corporate investor participating. This follows a $110 million Series C in 2024.

MMI’s Symani surgical system has been installed at roughly 50 sites worldwide and used in over 3,500 precision microsurgery procedures. It received FDA de novo classification for microsurgery in 2024 and CE mark certification in Europe, with an additional clearance for soft tissue dissection in 2025.

The funding will support U.S. commercial expansion and the REMIND clinical study, which evaluates whether reopening small lymphatic drainage pathways in the neck could help clear harmful proteins from the brain, exploring potential applications related to Alzheimer’s disease.

Open Source

Linus Torvalds Uses Claude Code to Achieve 3x Kernel Page Cache Speedup

Linus Torvalds published a noref-filemap-read-with-llm branch in his kernel repository. The eight-patch series, built on Linux 7.3-rc6, eliminates folio reference count contention when multiple threads read the same file from the page cache. The commit message explicitly credits Claude Code for the code, test programs, review, and a first draft of the text, all done under his direction.

On his laptop with some security protections disabled, eight threads reading a single file went from about 10 million reads per second to between 30 and 33 million. The tradeoff: freed memory stays idle briefly before reuse, and there is overhead on cache misses.

The patches have not been merged into mainline. At Open Source Summit in Prague, Torvalds said he “actually really likes using AI”, a marked shift from his 2024 characterization of the AI industry as “90 percent marketing.”

Open Source AI Week Kicks Off — PyTorch Conference, AGNTCon + MCPCon

Open Source AI Week runs October 16 through 25 across the San Francisco Bay Area. Hugging Face’s Open Together meetup launched the week on October 16 at The Midway, featuring 36 live community demos. Arcee AI and Reflection AI sponsored the event, with Allen AI, Qwen, Nous Research, and Unsloth as co-hosts.

PyTorch Conference NA takes place October 20-21 at the San Jose McEnery Convention Center. Sessions cover post-training, serving, and kernel work for AI engineers, researchers, and infrastructure builders.

AGNTCon + MCPCon NA follows on October 22-23 at the same venue. The Agentic AI Foundation’s flagship conference focuses on agent infrastructure, production agent systems, and the Model Context Protocol (MCP). PyTorch Conference attendees get a 20% discount on AGNTCon registration.

Gaming

Pokemon Pokopia Launches Its First Halloween Event

Nintendo Switch 2 exclusive Pokemon Pokopia begins its first Halloween event, ‘Pumpkaboo’s Sweet Surprises’, on October 17. Running until November 1, players befriend Pumpkaboo to collect pumpkin candy, which can be exchanged for nine types of Halloween decorations.

The core mechanic uses Camouflage, a skill learned from Zorua. Players disguise themselves as objects to surprise other Pokemon and earn rewards. Completing the event unlocks a dedicated Gourgeist habitat.

ID@Xbox Showcase Set for October 20 with 18-20 Indie Games

The ID@Xbox Showcase, co-hosted by IGN, airs October 20 at 10 AM PT. The event promises 18 to 20 indie game reveals including exclusive trailers and world premieres. It streams on IGN’s website, YouTube, and Twitch.

Among confirmed titles, The Wolf Among Us Remastered launches October 29 for PS5, Xbox Series X|S, and PC. The remaster of Telltale Games’ 2013 original is expected to receive additional details during the showcase.

References

Microsoft Maia 300 HBM4 N2P Chiplet 96-Accelerator Rack — StorageReview
Microsoft Maia 300 AI chip arrives with 288 GB HBM4 memory — Notebookcheck
Character AI Files for Bankruptcy and Signs $20 Million Disney Deal — PYMNTS
Disney Readies $20 Million Bid for Character.ai Assets — The Wrap
GlobalFoundries to Operate Two New Semiconductor Fabs Under $21B U.S.-Japan Deal — Construction Review Online
GlobalFoundries to build two chip factories with $21 billion Japanese backing — ROIC.AI (Reuters)
Li Auto 6-K Filing — MHX Technology Equity Financing — SEC
Li Auto chip unit to pursue first funding round — CnEVPost
DeepSeek to lead seed round in $400mn robot AI start-up — The Next Web
MMI raises $125M to scale up in robotic microsurgery — MedTech Dive
MMI Raises $125 Million Series D — Pulse 2.0
Linus Torvalds noref-filemap-read-with-llm commit — kernel.org
Linus Torvalds Says He Likes Using AI, While His New Kernel Branch Credits Claude Code — It’s FOSS
PyTorch Conference North America 2026 — Linux Foundation
AGNTCon + MCPCon North America 2026 — Linux Foundation
Pokemon Pokopia Halloween Event Adds Pumpkaboo and Gourgeist — Siliconera
Pokemon Pokopia Pumpkaboo’s Sweet Surprises — GamingProMax
ID@Xbox Showcase Announced for October 20th — GamingBolt
ID@Xbox Showcase 2026 Date, Livestream, Indie Games — Khel Now