AI Industry News Brief — October 13, 2026

Collection window: 2026-10-11 20:47 UTC to 2026-10-12 20:47 UTC (last 24 hours). Publication times were checked on each article page and converted to UTC. Stories already covered in the previous brief (Nadella’s emergency brake, Amazon dropping data center NDAs, voice AI) were excluded. Getty Images going private, AMD’s Helion rack and IonQ’s acquisition of Elisity were left out for limited AI relevance or detail.

Meta releases its first frontier model, Muse Wonder

Meta Superintelligence Labs released Muse Wonder and a larger Muse Wonder Full. According to Meta’s model card, both score comparably to Fable 5.1 and GPT-6 Astra, and Meta claims they beat rivals on MemoryBench, which measures “active memory” — retaining and synthesizing user-provided information. Muse Wonder costs $4 per million input tokens and $20 per million output tokens via the Meta Model API, while the Full variant is limited to Muse Code subscribers. The models were trained at Meta’s Prometheus data center in Ohio and are the first state-of-the-art release since Alexandr Wang restructured Meta’s AI efforts in August 2025.

Reference: TechCrunch (2026-10-12 17:56 UTC) · Meta Research

Microsoft reportedly on track to pay Anthropic $2B+ a year for Copilot; HCLTech launches a dedicated Anthropic unit

The Information reports, citing a single source, that Microsoft is on track to pay Anthropic more than $2 billion a year to power Copilot. Office chief Charles Lamanna reportedly convinced Satya Nadella despite cost concerns, citing stronger performance on advanced tasks such as Excel financial functions and presentation generation; Microsoft now blends Claude and OpenAI models across Word, Excel, Outlook and PowerPoint. On the same day, HCLTech signed a Global Premier Partnership with Anthropic, setting up a dedicated Anthropic business unit and a Center of Excellence, and embedding Claude and Claude Code into its AI Force offering.

Reference: AI Weekly / The Information (2026-10-12 15:38 UTC) · PR Newswire – HCLTech (2026-10-12 17:08 UTC)

Microsoft details Maia 300, its third-generation AI accelerator

In an Azure blog deep dive, Microsoft described Maia 300: TSMC N2P compute chiplets paired with eight HBM4 stacks (288 GB, 18 TB/s). A 167 kW liquid-cooled rack holds 96 accelerators, double Maia 200’s density, delivering up to 7x rack-scale token throughput and more than 2.5x performance per watt. An optical second tier can link 12,288 accelerators across 128 racks into one superpod. Target workloads are Copilot, GitHub Copilot and Microsoft Foundry.

Reference: Azure Blog · AI Weekly (2026-10-12 14:01 UTC)

Nvidia contributes Vera Rubin rack designs to the Open Compute Project

At the OCP Global Summit, Nvidia contributed its Vera Rubin rack and compute tray designs through the MGX 1.2 open architecture, adding rack-scale blind-mate liquid cooling and 250 kW configurations. It co-published the Open Data Center Specification v0.7 with AMD, Google and Meta, and open-sourced four DSX components for configuration, cluster readiness, rack management and bare-metal provisioning. Meta committed to Spectrum-6 Ethernet switches for its 100,000+ GPU AI factories, and Nvidia and Cisco introduced USFP pluggable optics at 400 Gb/s per lane.

Reference: NVIDIA Blog · AI Weekly (2026-10-12 20:02 UTC)

Linus Torvalds pushes a Linux page-cache branch written with Claude Code

Linus Torvalds pushed a branch named noref-filemap-read-with-llm to his Linux tree; the cover letter says Claude Code wrote the code, tests and first draft. The series tackles cache-line bouncing when many threads do tiny reads from the same file by deferring folio frees until after an RCU grace period instead of taking references. On a 4-core laptop with 8 threads, throughput rose from about 10 million to 30+ million reads per second, but deferred frees hold memory longer and the branch is not merged.

Reference: Hardware Busters (2026-10-12 06:00 UTC)

Microsoft’s Brad Smith says the UK is best placed to lead AI regulation

On a BBC podcast, Microsoft President Brad Smith called AI “a weapon and a tool” and said the time has come for global safety rules. He pointed to the UK’s AI Security Institute, which tests U.S. frontier models before deployment, as the reason the UK — not the U.S. or EU — should lead. Calling last month’s voluntary White House pledge insufficient, he backed a legal duty to report breaches by rogue AI agents, outside evaluator approval before release, and kill switches.

Reference: TechCrunch (2026-10-12 15:31 UTC)

DeepSeek leads seed round for ex-Tencent scientist’s robotics startup Winverse AI

According to Bloomberg, DeepSeek is leading a seed round for Winverse AI, a stealth robotics startup in Hangzhou, at a post-money valuation of about $400 million, with Meituan’s Long-Z also participating. The move follows reports on October 6 of DeepSeek’s first external raise of more than $12 billion at a valuation above $50 billion, and its August investment in Unitree Robotics’ Shanghai IPO. It is seen as DeepSeek’s clearest step from language models into physical AI.

Reference: Startup Fortune (2026-10-12 12:20 UTC)

Z.ai opens a gated beta of GLM-Legal, a 320B-parameter legal agent

Z.ai unveiled GLM-Legal, a 320B-parameter mixture-of-experts model (18B active per token, 1M-token context) trained with reinforcement learning across tens of thousands of realistic legal environments. It reports a 19.2% all-pass score on Harvey’s LAB benchmark across 24 practice areas, 12.5 points above its base model. There is no public API — access requires an application — and the model is bundled into Z.ai’s AutoClaw desktop agent. No third party has replicated the results yet.

Reference: Z.ai Blog · AI Weekly (2026-10-12 17:08 UTC)

Today’s Summary

Meta joins the frontier-model race with Muse Wonder, while Microsoft leans further on Anthropic models and strengthens its own silicon with Maia 300.
AI infrastructure competition is shifting toward open rack standards (Nvidia at OCP), and the UK’s AI Security Institute model is drawing attention in the regulation debate.
From Torvalds’ kernel patch to a legal-specialist agent, AI coding and domain agents are moving deep into expert work.