Skip to content
Source · Daily Brief

AI Daily Brief — 13 October 2025

Monday packed three structural shifts. Andrej Karpathy released nanochat — “the best ChatGPT that $100 can buy.” A single speedrun.sh on an 8×H100 node runs the entire pipeline (Rust tokenizer, FineWeb pretraining, mid-training, SFT, optional GSM8K RL, eval, CLI + web UI) in ~4 hours for ~$92.40, with $300 and $1,000 tiers beating GPT-2 CORE. Capstone repo for his LLM101n course; hand-written without AI agents because Claude and Codex “didn’t work well enough” on it. OpenAI + Broadcom announced a 10 GW custom AI accelerator partnership — OpenAI-designed accelerators with Broadcom Ethernet networking, rack deployments H2 2026 through end-2029. Broadcom shares jumped ~9.88%, adding >$150B in market cap. NVIDIA debuted Vera Rubin MGX + Spectrum-XGS Ethernet at OCP Global Summit — next-gen CPU+GPU with up to 8× GB300 performance; ESUN (Ethernet for Scale-Up Networking) workstream launched with AMD, Arista, Arm, Broadcom, Cisco, HPE, Marvell, Meta, Microsoft, NVIDIA, OpenAI, Oracle. Salesforce announced Agentforce 360 ahead of Dreamforce Day 1 — unified platform combining Customer 360, Slack, Data 360 with new agent builder, hybrid reasoning, voice; plus $15B SF investment and Apromore acquisition.

Top stories

  • Karpathy ships nanochat. $100 / 4-hour full-stack ChatGPT clone in ~8K LOC. 8×H100 speedrun trains Rust tokenizer + FineWeb pretrain + mid-train + SFT + optional RL + eval + CLI + web UI for ~$92.40. $300 and $1,000 tiers beat GPT-2 CORE. via GitHub
  • Karpathy thread: “among the most unhinged” repos. Hand-written without AI agents because Claude and Codex “didn’t work well enough.” Cost cohort: $43K (GPT-2 in 2019) → <$100 today, a 7-year collapse. via X / Karpathy
  • OpenAI + Broadcom: 10 GW custom AI accelerators. OpenAI-designed silicon + Broadcom Ethernet networking. Racks deploy H2 2026 → end-2029. Broadcom up ~9.88%, >$150B market cap gain. via OpenAI
  • Altman + Brockman + Tan on OpenAI Podcast Ep. 8. Altman framed 10 GW as “just a beginning” — the world will absorb very high-quality intelligence delivered fast and cheaply. via CNBC
  • NVIDIA Vera Rubin MGX + Spectrum-XGS Ethernet. OCP Global Summit reveal: next-gen CPU+GPU up to 8× GB300 on some workloads; Spectrum-XGS Ethernet links distributed DCs into giga-scale AI factories; 800V DC power plans; joined OCP board with AMD, Arm. via NVIDIA Blog
  • Broadcom Thor Ultra 800G AI NIC + OCP ESUN. First 800G AI Ethernet NIC; Tomahawk 6/Ultra, Jericho4, 3rd-gen CPO. OCP launched ESUN (Ethernet for Scale-Up Networking) backed by AMD, Arista, Arm, Broadcom, Cisco, HPE, Marvell, Meta, Microsoft, NVIDIA, OpenAI, Oracle. via OCP
  • Salesforce announces Agentforce 360 + $15B SF + Apromore. Unified platform combining Customer 360 Apps, Slack, Data 360 with new conversational agent builder, hybrid reasoning, voice, Intelligent Context, Agent Script (beta Nov). 12K customers, Reddit deflecting 46% of cases. $15B SF five-year investment + Apromore process-intelligence acquisition. via Salesforce
  • NVLink Fusion expands to Intel + Samsung Foundry. Plus expanded Fujitsu collaboration — opening NVLink to third-party CPUs/accelerators as NVIDIA pushes a gigawatt-class AI factory ecosystem anchored on its interconnect. via Analytics India

Who shipped

Karpathy shipped nanochat. OpenAI + Broadcom shipped the 10 GW partnership. NVIDIA shipped Vera Rubin MGX, Spectrum-XGS, NVLink Fusion expansions. Broadcom shipped Thor Ultra 800G + Tomahawk 6/Ultra + Jericho4. OCP launched ESUN. Salesforce shipped Agentforce 360 + Apromore deal.

Open-source pulse

Karpathy nanochat is the open-source story of the month — the $100 reference baseline for full-stack chat-LLM training. MIT Han Lab + NVIDIA released StreamingVLM (real-time infinite video understanding at 8 FPS on a single H100, 66.18% win vs GPT-4o-mini on Inf-Streams-Eval). Chinese open-model labs quiet on Oct 13.

Money, infra & hardware

OpenAI-Broadcom 10 GW + NVIDIA OCP reveals reframe the AI-compute substrate. Broadcom up ~9.88% added >$150B in market cap. ESUN’s vendor coalition (AMD, Arista, Arm, Broadcom, Cisco, HPE, Marvell, Meta, Microsoft, NVIDIA, OpenAI, Oracle) is the largest unified scale-up Ethernet alliance yet.

Quiet corners

D2E (Worv AI / KAIST) trending on HF Daily Papers — desktop interactions especially gaming as pretraining substrate for robotics. 152× compression OWA Toolkit, 267 hours of synchronized PC-gameplay across 29 games. ICLR 2026 accept.

By the numbers

  • $100 / ~4 hours / ~8K LOC — nanochat training cost / time / code size
  • 10 GW — OpenAI-Broadcom custom accelerator capacity
  • ~9.88% / >$150B — Broadcom stock pop / market cap gain
  • 8× — Vera Rubin MGX peak performance vs GB300
  • 800G — Broadcom Thor Ultra AI Ethernet NIC speed
  • $15B — Salesforce SF five-year investment
  • Most-mentioned company: OpenAI
  • Quietest segment: Chinese open-source labs

Compiled by AI Feed’s editor from verified web sources for 13 October 2025.