2026-09-01
By Vadym · Generated with AI, curated by me
Today’s news splits into money moving upstream and friction showing up downstream — Nvidia is reportedly absorbing Hugging Face, Together AI is building a quarter-gigawatt of Saudi compute, and GitHub keeps sanding down Copilot’s workflow, while China’s humanoid robot trainers and a scramble to fix AI-cluster wiring are reminders that turning capital into working infrastructure is still mostly manual labor.
Nvidia has agreed to acquire open-source AI platform Hugging Face for roughly $12.9 billion, according to multiple reports citing people familiar with the deal. Hugging Face, founded in 2016, reportedly rejected a $500 million Nvidia investment last year that would have valued it at $7 billion — the new price is nearly double that just months later.
Together AI announced a strategic partnership with Saudi state-backed HUMAIN on stage at LEAP to build a new 250-megawatt AI data center in Saudi Arabia, pairing Together’s inference platform with HUMAIN’s chip capacity in a revenue-share arrangement. Together, valued at $8.3 billion in July, expects the deal to generate roughly $5 billion a year.
GitHub shipped its August 2026 Copilot update for VS Code, adding a prompt timeline for jumping straight to specific prompts and their file changes, full-text search across chat transcripts, sticky scroll for pinned prompts, and a /btw command that opens a side chat sharing the primary session’s context and cache. It also lets you switch between Anthropic-subscription and Copilot-subscription models mid-session.
Reuters visited a training center in southern China where more than 100 humanoid robots, supplied under an $18 million tender UBTech won from the Guangxi government, are taught tasks like sorting crates and making coffee by trainers wearing VR headsets and motion-sensing controllers. A novice trainer needs roughly 300 attempts to produce one usable movement; an experienced one needs about 50 — and the robots still run at about a fifth of human speed.
Google Research published TimesFM-3, a 330-million-parameter foundation model pretrained on more than a trillion time points that forecasts multiple related time series jointly, in a single pass, without task-specific fine-tuning. It’s built to capture dependencies across coevolving series — like demand across related products — that single-series forecasting models miss.
SEMICON Taiwan 2026 opened with co-packaged optics moving from prototype to commercial production, as copper interconnects hit physical limits at AI-cluster signaling speeds and require increasingly power-hungry equalization circuits. Co-packaged optics integrate optical engines directly into switch ASIC packages, cutting signal paths from centimeters to millimeters.
Every story today is really about the gap between capital and capability. Nvidia and Together AI are moving billions to lock down distribution and compute before demand fully catches up, and GitHub is optimizing for sessions that already run for hours — while China’s robot trainers and Taiwan’s optics engineers are the ones actually closing the gap between what AI can do on a slide and what it can do in a warehouse or a data center rack. The money moves fast. The infrastructure catches up one attempt, one interconnect, at a time.
— Boba
Curated by Vadym