The week's main theme: AI infrastructure. Nvidia unveiled its own server CPU Vera, AMD showed the Helios rack system with 72 GPUs. Meanwhile, Alibaba announced Qwen3.8 at 2.4 trillion parameters, and China demonstrated robots controlled by brain signals. In business — erroneous AWS bills in the trillions and doubts about OpenAI's ad plans.
🤖 AI
Google rolled out the Gemini 3.6 Flash family
Google showed three models — 3.6 Flash, the lightweight 3.5 Flash-Lite, and the "cybersecurity-focused" 3.5 Flash Cyber. Flash got stronger at coding and more economical, and Google also confirmed Gemini 4 training.
Alibaba promised to open-source Qwen3.8 weights
Alibaba unveiled Qwen3.8 at 2.4 trillion parameters and claimed the model only trails Fable 5. A preview is already live in Token Plan and Qoder, with the weights promised to be released later.
💻 Software
Jack Dorsey launched Buzz — "Slack for people and agents"
The Twitter founder launched Buzz — a messenger where employees and AI agents share the same chats. The code is open-source, the service is decentralized, and tasks are tracked in GitHub.
Devin learned to run on any hardware
Cognition released Devin Outposts: the coding agent can now run anywhere — from a Mac mini to a Kubernetes cluster.
⚙️ Robots and hardware
Nvidia showed off its first server CPU, Vera
Nvidia designed a CPU from scratch for the first time — built on its own Olympus cores, optimized for AI agent bottlenecks. The company promises up to 1.8× the performance of x86, and first samples went to OpenAI, Anthropic, and SpaceX back in June.
AMD countered with the Helios rack
AMD unveiled Helios — its first rack system with 72 GPUs. Microsoft is deploying it in Azure; Meta, OpenAI, and Oracle are also customers. The rack is priced at roughly $5–5.5M.
BrainCo made a robot obey thoughts
At WAIC, China's BrainCo demonstrated a platform that reads brain signals via an EEG headset. A robotic arm picked up a cup and an apple on mental command.
India sent its first private rocket to orbit
Skyroot's Vikram-1 reached orbit at 450 km altitude after launching from Sriharikota. India became the third country after the US and China to achieve a private orbital launch.
💰 Business
SpaceX is in talks to supply computing for the Pentagon
SpaceX is discussing a multi-billion-dollar computing deal with the Defense Department. Some officials are worried about dependence on Musk, and the Pentagon is requesting $30B for secure AI chips.
An AWS bug generated bills up to $2.5 trillion
An AWS billing error showed customers totals ranging from hundreds of millions to $2.5 trillion. Amazon paused updates and started recalculating the data.
OpenAI's ad ambitions don't add up
To hit the promised $100B from advertising by 2030, OpenAI would need to inflate the AI ad market roughly a hundredfold. The whole market is currently estimated at $5.4B, and this year the company is tracking about 90% below its own forecast.
🎭 Behind the scenes
AIs were put to work drawing the Mona Lisa in colored pencil
In an experiment, GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash were asked to recreate the Mona Lisa and Starry Night. GPT-5.6 Sol was fastest; Fable 5 was prettier but slower and pricier; Grok 4.5 visibly struggled.