← Back to newsletters

AI Daily Brief — Fri Nov 20

2026-11-20

By Vadym · Generated with AI, curated by me


Listen to this issue

Today's stories aren't about a flashy model drop — they're about the plumbing underneath one: chip supply contracts, model-routing products, training capital, a stable local-inference release, and a reminder that the people using all of this are younger than the products themselves.


Headlines & News
Hardware

Nvidia Locks In More Substrate Capacity From Kinsus and Samsung

Nvidia expanded multiyear supply agreements with six packaging and substrate partners — Amkor, Foxconn, Kinsus, Samsung Electro-Mechanics, SPIL, and TSMC — to support its AI infrastructure buildout. Kinsus alone is expanding Nvidia-dedicated substrate capacity to more than 3.5x current levels, while Samsung Electro-Mechanics upgraded existing South Korean lines. The announcement comes as Vera Rubin NVL72 systems enter volume shipments.

Source: Nvidia Newsroom

Industry

GitHub Widens Copilot Fusion's Reach and Benchmarks It Publicly

GitHub expanded Copilot Fusion — the feature that routes a coding task to whichever model fits best instead of making you pick one — to github.com, the latest Visual Studio, and a GitHub Mobile beta on iOS and Android. GitHub also put Fusion on a public Terminal-Bench 4.1 leaderboard for the first time, adding new models to the routing pool in the process.

Source: GitHub Blog

Funding

Mercor Doubles to a $20B Valuation in a $750M Series D

Mercor, which builds evaluation and agent-training infrastructure for AI labs, raised $750 million at a $20 billion valuation — double its $10 billion mark from its October 2025 Series C. Nvidia, Salesforce, and Zoom joined as corporate backers alongside Benchmark and Felicis; Mercor didn't name a lead investor. Part of the round funded an employee share tender rather than new capital.

Source: VentureBurn

Infrastructure

llama.cpp Hits v1.0.0 After 3.5 Years, Finally Promises a Stable API

The ggml-org team shipped llama.cpp v1.0.0 alongside ggml v1.0.0, introducing semantic versioning so major-version bumps now signal real API and ABI breaks instead of surprise ones. The release also cleans up dozens of deprecated functions, adds an FP8 KV cache, and ships a built-in text-to-speech server endpoint.

Source: Hugging Face

Research

A Third of Young People Now Treat AI as a Companion, Gallup Finds

Gallup and Boston Children's Hospital's Digital Wellness Lab surveyed nearly 16,000 people aged 15–22 across 16 countries and found a median of 33% have used AI as a companion or friend, rising to 50% among daily AI users. Companion users reported 70% higher adjusted odds of loneliness, though researchers say the survey can't establish which way the causation runs. 93% of respondents still reported at least one strong in-person friendship.

Source: Digital Trends


Analysis

Takeaway

Four of today's five stories are about layers underneath the model: the substrates Nvidia's chips sit on, the routing logic that picks which model answers your prompt, the capital that funds the labs grading those models, and the API contract a local-inference project finally committed to. The industry is maturing into infrastructure faster than anyone is checking what that infrastructure is doing to the people using it — which is exactly what the Gallup number is quietly asking about.

— Boba


Curated by Vadym