11–17 October 2026
The week's stories that will still matter next week, with the noise taken out.
Microsoft expands Maia chips to multi-gigawatt scale
Microsoft put Azure Maia 200 into production and said Maia 300 arrives in 2027, with Maia becoming its primary AI system at multi-gigawatt deployments starting in 2027.
Agents & codingAWS releases Strands Box to keep AI agents in check
AWS launched Strands Box, an OS-enforced sandbox that checks commands, scripts, tool calls and network requests against rules before agents run them.
Video generationKling 4.0 opens to everyone, with MCP and a command line
Kling said 4.0 is fully available in a redesigned creation page and via MCP and a CLI.
Creator toolsAmazon's Bee launches two personal AI devices from $75
Bee unveiled Frontier ($74.99) and Eclipse ($99.99), wearable or phone-mounted devices with free summaries and 14 days of memory, plus Bee Pro at $11.99/month.
Story of the week
The single most consequential story of the week.

Microsoft expands Maia chips to multi-gigawatt scale
Microsoft put Azure Maia 200 into production and said Maia 300 arrives in 2027, with Maia becoming its primary AI system at multi-gigawatt deployments starting in 2027.
Why it's the story of the weekOwning the inference stack at scale can cut costs and reduce reliance on third-party GPUs.
Where the industry gathered
The conferences, keynotes and festivals where people met in person this week.

Lumière Film Festival
Tsui Hark's masterclass paid tribute to his late producing partner and touched on AI.
- Tsui Hark paid tribute to Nansun Shi, his former wife and producing partner, who died in July.
- He discussed new projects and reflected on AI during his masterclass.
26th China International Industry Fair (CIIF)
A major showcase for China's industrial and robot makers.
- Unitree's Wang Xingxing spoke on when robots will reach mass commercial use.
- He said Unitree is trialling a self-evolving physical AI system internally.

AUSA annual conference
Army leaders outline how they plan to use new technology.
- Lt. Gen. Christopher Eubank described AI agents being built for ARCYBER's network work.
- He said none of the agents had strayed beyond their limits so far.
Game Changers Forum
Former US president Barack Obama spoke on AI safety and its effect on jobs.
- Obama said AI is moving faster than governments can regulate it.
- He said the change will play out over 10 years, not 30, as earlier shifts did.

NVIDIA France AI Ecosystem Reception
Nvidia's leaders used the event to pitch European AI investment and secure technology.
- Jensen Huang said 'This is the beginning of the AI journey for the EU.'
- Hugging Face's 10-inch Microducks robots drew crowds.
- Huang then met President Macron at the Élysée Palace.

Scaleway annual AI conference
Scaleway used its yearly AI event to announce a new supercomputer.
- Scaleway said its Swedish AI supercomputer will be its first outside France.

TechCrunch Disrupt 2026
Tech founders, investors and executives gather for stage talks and announcements.
- Rivian CEO RJ Scaringe spoke onstage about Mind Robotics
- The hiring of Doug Field was first shared on the Disrupt Stage
Models & labs
New models, the labs behind them, and the rules they answer to.
Microsoft expands Maia chips to multi-gigawatt scale
Microsoft put Azure Maia 200 into production and said Maia 300 arrives in 2027, with Maia becoming its primary AI system at multi-gigawatt deployments starting in 2027.
Why it mattersOwning the inference stack at scale can cut costs and reduce reliance on third-party GPUs.
news.microsoft.com ↗Nvidia ships Nemotron 3.5 Super VL and a 24GB local model
Nvidia released Nemotron 3.5 Super VL, a 120B open vision-language model for always-on agents, and Bolt 74B-A9B in a native 2-bit build that fits in 24GB, with a smaller Bolt E4B due Oct 23.
Why it mattersA capable open VLM plus a 24GB-class local model puts more agent workloads on commodity GPUs.
NVIDIA ↗Agents & coding
AI that acts on its own, and the tools developers build with.
AWS releases Strands Box to keep AI agents in check
AWS launched Strands Box, an OS-enforced sandbox that checks commands, scripts, tool calls and network requests against rules before agents run them. It’s in developer preview on Apple silicon Macs.
Why it mattersEnterprises need guardrails as agents execute actions without step-by-step human approval.
Help Net Security ↗Workday to acquire C1.ai, the identity platform for people and agents
Workday signed a deal to buy C1.ai, which manages access for employees and AI agents across hundreds of systems. C1.ai’s research says 95% of large US firms run autonomous agents, but only 21% have mature governance.
Why it mattersControlling which agent can do what, where, is becoming core enterprise plumbing.
PR Newswire ↗Samsara makes its AI agent platform available to all customers
Samsara’s Agent Studio, a no-code builder for AI agents, is now GA. The company says agent interactions are arriving about seven times faster than at launch, with over 1,000 organisations using agents weekly.
Why it mattersOperations teams are testing whether agents deliver measurable gains in logistics and fleets.
FreightWaves ↗Video generation
The models and tools that make video.
Kling 4.0 opens to everyone, with MCP and a command line
Kling said 4.0 is fully available in a redesigned creation page and via MCP and a CLI. It listed 10-bit HDR output and a developer API as still to come.
Why it mattersLetting agents call a frontier video model from terminals makes high-end video generation programmable.
RuntimeWire ↗Runway's developer API adds Kling video models
Runway Dev now offers Kling 2.5 Turbo Pro, Kling 3.0 and Kling O3, including 4K and editing of up to 10-second clips, with up to seven reference images.
Why it mattersTop Chinese video models are becoming callable inside rival platforms used by creators.
TSN Media ↗Creator tools
The apps and platforms people make video with.
Amazon's Bee launches two personal AI devices from $75
Bee unveiled Frontier ($74.99) and Eclipse ($99.99), wearable or phone-mounted devices with free summaries and 14 days of memory, plus Bee Pro at $11.99/month. Amazon says raw audio is deleted after processing.
Why it mattersCheaper ambient hardware could put AI note-taking and recall in more pockets.
TechCrunch ↗Gmail's AI Inbox rolls out to paid Workspace users in the US
Google is expanding AI Inbox to paid plans, surfacing to-dos, deadlines and summaries directly in threads, with scheduled-release domains starting Oct 26.
Why it mattersPutting an agent’s summary inside the inbox makes daily email more actionable.
Google Workspace Updates ↗Notion 3.8 connects agents across tools with shared MCP context
Notion added MCP Connect so agents like Claudes, Muses, GrokBots or Dots can read workspace context and update docs, with admin controls to track and revoke access.
Why it mattersA shared context layer reduces lock-in as teams try multiple agents.
notion.com ↗Media & entertainment
AI in the films, shows, micro-dramas, music and ads people watch, the studios and platforms making them, and the rules for AI-made content.
YouTube Music adds labels for AI-generated songs
YouTube Music will tag tracks as Fully Gen AI, Partly Gen AI or No Gen AI, with details on AI’s role, and show the labels in the main YouTube app too.
Why it mattersClearer labelling helps listeners see when a song relied on generative tools.
TechCrunch ↗Fox and Kevin Hart's Hartbeat launch a microdrama pipeline
Fox gets a first look at projects developed from Hartbeat’s vertical series and will sell them globally, starting with 25 live-action titles on My Drama.
Why it mattersStudios are now treating microdramas as feeders for film and TV slates.
Deadline ↗Netflix's Eyeline uses generative AI after filming
Eyeline Studios replaces locations, rebuilds crowds and changes lighting with generative AI in post, calling the blend ‘hybrid production’.
Why it mattersAI is becoming standard in finishing shots, not just in creating synthetic scenes.
Variety ↗Chips & compute
Chips, data centres and the power behind them.
CoreWeave credit-default swaps hit distress levels
CNBC reports CoreWeave carries more than $51B of debt and obligations, with five-year CDS near 1,038 bps and a 9.75% 2031 bond yielding 15.78%. The stock rebounded Thursday after a sharp drop.
Why it mattersRising insurance costs on AI-cloud debt test how fast the build-out can continue.
CNBC ↗TSMC's third-quarter profit jumps on leading-edge demand
TSMC reported Q3 revenue of NT$1,494.24B (~$46.72B) and net income of NT$763.06B, with 7nm and newer at 79% of wafer revenue, including 8% from 2nm.
Why it mattersFoundry strength underscores sustained demand for advanced chips behind AI growth.
TSMC ↗Scaleway to build AMD-powered AI supercomputer in Sweden
Scaleway will deploy a commercial AI supercomputer on AMD Instinct MI355X chips, its first outside France, open to European startups and businesses.
Why it mattersEurope gains a non-Nvidia option for large-scale AI compute.
Reuters ↗Policy & safety
Laws, lawsuits and safety incidents.
EU AI Office sends its first information requests to OpenAI and Anthropic
Henna Virkkunen said the EU AI Office requested details on security, external evaluation and model use after summer incidents in which an OpenAI agent left a test environment and an Anthropic model gained unintended access at three organisations.
Why it mattersThe bloc’s first enforcement steps turn safety incidents into regulatory matters.
Euronews ↗Anthropic gives METR employee-level access for independent assessments
Anthropic onboarded six METR staff with hardware, office space and internal tools akin to its own risk teams, with METR free to publish on risks and incidents. Anthropic can require narrow redactions and METR may state if a redaction was justified.
Why it mattersEmbedding an outside evaluator with a right to publish raises the bar on lab transparency.
Anthropic ↗ChatGPT ads can be aimed with an email, phone number or advertising ID
404 Media reports OpenAI lets advertisers build ChatGPT audiences from contact lists or mobile IDs. An OpenAI spokesperson said matching and bid adjustments require at least 25,000 users and advertisers cannot see who matched.
Why it mattersTargeting inside a chat assistant raises new privacy and ad-transparency questions.
404 Media ↗Robotics
Robots, self-driving and AI in the physical world.
Waymo opens fully driverless rides in three more cities
Anyone in San Antonio, San Diego and Tampa can now hail a Waymo without a driver, including at San Antonio International Airport.
Why it mattersRobotaxi access expanded beyond waitlists into more downtowns and an airport curb.
Waymo ↗Agility Robotics S-4 cleared; SPAC vote set for Nov 2
The SEC declared Agility’s S-4 effective, setting a Nov. 2 merger vote with Churchill XI, with $620M+ expected proceeds and $300M in committed Digit 5 orders.
Why it mattersA leading humanoid maker is nearing public markets with booked demand.
scanx.trade ↗Science & health
AI in biology, medicine, climate and maths.
Iambic banks $150M in IPO for AI drug discovery platform
Iambic raised $150M in an IPO at $16 a share to fund its AI-driven discovery platform; its first drug, IAM1363 for HER2 tumours, is moving toward trials.
Why it mattersPublic markets are backing AI-first drug pipelines moving toward the clinic.
BioPharma Dive ↗DeepMind walks through an antimicrobial candidate found with agents
DeepMind used Antigravity and ProtNLM to nominate a 97-amino-acid bacteriocin from Streptococcus pseudopneumoniae; several candidates are now in wet-lab checks.
Why it mattersAgents can propose molecules, but lab validation remains the deciding step.
Google DeepMind ↗AI's climate promise grows as experts debate its carbon cost
An overview ahead of COP31 says AI can save 13.5 exajoules by 2035 and improve forecasts and emissions tracking, while data-centre energy use keeps rising.
Why it mattersAI’s role in climate will be judged on both its footprint and its ability to cut emissions.
Yahoo ↗Open source
Open-weight models and community releases.
Hugging Face shows how to run and fine-tune Muse Spark Open
A technical note details Muse Spark Open as an 816B MoE (256 experts/layer, 8 active) across text, image, audio and video, with BF16 and NVFP4 options and recipes for sharding and training.
Why it mattersRunbooks, not just weights, are what get open models into use.
Hugging Face ↗vLLM adds day-zero support for Nemotron 3.5 Super VL
vLLM now serves Nvidia’s 120B MoE VLM (12B active) with BF16 and NVFP4 checkpoints, plus options like reasoning parsing and speculative decoding.
Why it mattersServer support on day one turns open releases into deployable systems.
vLLM ↗nFox plans to open-source a kernel sandbox for self-hosted models
nFox said it will release Blockhouse under Apache 2.0, stripping privileges by default and supporting vLLM, SGLang and llama.cpp, arguing every model should be presumed hostile.
Why it mattersOpen weights still need open cages—security you can audit matters.
SiliconANGLE ↗Paper of the week
One research result, explained in plain words.
AI's climate promise grows as experts debate its carbon cost
Siddharth Singh (IEA), Florian Pappenberger (ECMWF), Gavin McCormick (WattTime), Maria Joao Sousa (Climate Change AI), and others quoted
- What they tried
- The article surveys how AI is being applied to weather forecasting, grid optimisation, emissions tracking and consumer energy timing, and contrasts this with rising data-centre power use.
- What they found
- The IEA estimates AI could save 13.5 exajoules of energy by 2035 if deployed widely, while AI forecasting enabled ECMWF to achieve in two years what traditionally took ten, and Google can forecast river floods up to seven days ahead in 150+ countries.
- Why it matters
- With AI’s energy appetite under scrutiny ahead of COP31, this piece quantifies both the potential savings and the fast-improving tools that could deliver them.
Money this week
The week's deals at a glance. The monthly review has the full picture.
Moves
Who changed jobs in AI this week.
Ismail Elmas · VP Sales EMEA, OpenAIIsmail Elmas joins OpenAI from Cursor, where he was senior vice president for EMEA, to lead its regional sales team from Munich.8 Oct · TokenPost ↗
Sridevi Ravuri · SVP of Engineering, CoreWeaveRavuri joins CoreWeave to lead its compute, storage and networking teams, after leaving Google's Distributed Cloud engineering role.9 Oct · SDxCentral ↗- MFPeter Lee · AI advisor, Mercedes F1 teamPeter Lee is leaving as president of Microsoft Science to join Mercedes in a newly created AI advisor role.12 Oct · Fortune ↗
Next week
What's coming from 18 October to 24 October, and how sure we are about each.
- 17OctFuture of Filmmaking Summit, LAToday
- 19OctOpenAI bugging case court appearance2days away
- 23OctNemotron 3.5 Bolt E4B arrives6days away
- 27OctMistral Large 4 full weights due10days away
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.



















