Skip to content
Source · Daily Brief

AI Daily Brief — 24 December 2025

Christmas Eve packed NVIDIA’s biggest-ever deal. NVIDIA buys Groq’s assets for ~$20B — largest NVIDIA deal in history (next-largest was Mellanox at ~$7B in 2019). Structured as non-exclusive license of Groq’s LPU inference technology + acqui-hire of founder Jonathan Ross + President Sunny Madra + senior engineers. Groq remains nominally independent under new CEO Simon Edwards; GroqCloud continues to operate. Analysts read structure as designed to keep appearance of competition + sidestep antitrust scrutiny while NVIDIA absorbs its most credible inference-chip rival. Groq newsroom: official non-exclusive inference-tech license “to accelerate AI inference at global scale” with Ross + Madra + leadership joining NVIDIA; agreement does not transfer Groq’s corporate entity — only the IP license + the people. NVDA pops past $190 → $4.6T market cap on the news. Cantor: NVDA top pick with $300 PT — “strategic masterstroke”. Truist’s William Stein: $20B headline modest vs NVIDIA’s cash + FCF. Skeptics — Hedgeye + DA Davidson’s Alex Platt — pushed back that Groq’s LPU is still “unproven” on large models because of limited memory capacity and suited only to narrow inference workloads. Tom’s Hardware: NVIDIA absorbs Groq’s full hardware stack + Jonathan Ross — who was foundational engineer on Google’s original TPU before founding Groq. NVIDIA paid ~2.9× Groq’s last private valuation to neutralize the most architecturally distinct LPU competitor. Fujitsu launches Kozuchi Physical AI 1.0 with NVIDIA NIM — first product from Oct 2025 collaboration. Integrates NIM microservices with Fujitsu’s Takane LLM. Multi-AI-agent framework + specialised procurement agents. Internal PoC: order-confirmation workload ~50% cut + NIM compatibility ~50% inference speed boost. Self-learning agentic foundation by end FY2025. Cornell + Berkeley Science study (390:6779, 1240) on LLMs in scientific publishing covered Dec 24: analysed ~2.1M preprint abstracts (Jan 2018 – Jun 2024). Researchers adopting LLMs posted significantly more papers: +59.8% social sciences/humanities, +52.9% biology/life sciences, +36.2% physics/math. Authors warn boom harder on reviewers/funders/policymakers separating substantive work from AI-polished filler. OpenAI + Anthropic doubled developer usage limits over Christmas window — temporarily doubling Codex/API call limits for paid ChatGPT users + Claude rate limits as “gift to developers”; runs through Jan 1, 2026. Part of OpenAI Shipmas cadence teeing up GPT-5.2-Codex-XMas edition Dec 25. Marvell-Celestial AI $3.25B from Dec 2 sets optical-interconnect backdrop — together with NVIDIA-Groq signals next round of AI consolidation centred on inference economics + rack-scale interconnect, not raw training FLOPs.

Top stories

  • NVIDIA buys Groq’s assets for ~$20B — biggest NVIDIA deal ever. Non-exclusive LPU license + Ross + Madra + senior engineers acqui-hire. Groq independent under Simon Edwards. GroqCloud continues. Next-largest: Mellanox $7B 2019. via CNBC
  • Groq + NVIDIA confirm via joint inference-tech licensing announcement. “Accelerate AI inference at global scale.” Only IP + people transferred, not corporate entity. via Groq
  • NVDA past $190 → $4.6T market cap. Cantor reiterates top pick at $300 PT — “strategic masterstroke.” Truist: $20B modest vs NVIDIA cash/FCF. Hedgeye + DA Davidson skeptics: Groq LPU “unproven” on large models due to limited memory. via Seeking Alpha
  • Tom’s Hardware: deal hands NVIDIA full hardware stack + Ross. Ross was foundational engineer on Google’s original TPU before founding Groq. ~2.9× last private valuation to neutralize most architecturally distinct LPU competitor. via Tom’s Hardware
  • Fujitsu Kozuchi Physical AI 1.0 with NVIDIA NIM. Multi-AI-agent + specialised procurement. Internal PoC: ~50% workload cut + ~50% inference speed. Self-learning agentic foundation by end FY2025. Trial environments now offered. via Fujitsu
  • Cornell + Berkeley Science: LLM-using scientists +36-60% papers but quality slips. 2.1M preprints Jan 2018-Jun 2024. +59.8% SSH, +52.9% bio/life, +36.2% physics/math. Strain on peer review. via Cornell
  • OpenAI + Anthropic double developer holiday usage limits through Jan 1. Codex/API + Claude rate limits — “gift to developers” during slow week. Part of OpenAI Shipmas cadence teeing up GPT-5.2-Codex-XMas Dec 25. via Techloy
  • Marvell-Celestial AI $3.25B (Dec 2) sets optical-interconnect backdrop. Together with NVIDIA-Groq = next round of AI consolidation centered on inference economics + rack-scale interconnect, not raw training FLOPs. via Marvell
  • IBM CUGA on Hugging Face Spaces (week of Dec 21). Apache 2.0 enterprise agent framework. #1 on AppWorld (750 tasks, 457 APIs), top-tier WebArena. via HF / IBM
  • Trump Dec 11 AI preemption EO backdrop: DOJ AI Litigation Task Force due within 30 days. Goodwin + Gibson Dunn + Davis Wright Tremaine: EO cannot invalidate state law and expected to draw immediate state pushback. via White House

Who shipped

NVIDIA + Groq shipped the $20B deal. Fujitsu shipped Kozuchi Physical AI 1.0. OpenAI + Anthropic shipped holiday API boosts. Cornell + Berkeley shipped Science paper. Google, Meta, xAI, DeepSeek, Alibaba quiet on direct launches.

Open-source pulse

IBM CUGA on HF Spaces continues to spread + #1 AppWorld result. Anthropic Skills + Bloom from prior week continue community adoption.

Money, infra & hardware

NVIDIA-Groq $20B + Marvell-Celestial $3.25B + Anthropic-Google-Broadcom multi-GW (Dec 19) = Dec 2025 inference-economy consolidation defining moment. NVIDIA inference dominance now nearly complete after absorbing most architecturally distinct LPU competitor.

Quiet corners

Cornell + Berkeley Science study is the substantive research moment — first peer-reviewed quantification of LLM-driven publication surge + quality strain. +59.8% / +52.9% / +36.2% by domain reframes academic AI productivity debate.

By the numbers

  • ~$20B / $7B / 2019 — NVIDIA-Groq deal / next-largest (Mellanox) / year
  • $190 / $4.6T / $300 PT — NVDA pop / market cap / Cantor target
  • ~2.9× / Sep $6.9B — NVIDIA paid vs Groq’s last private valuation
  • 50% / 50% — Fujitsu Kozuchi workload cut / inference speed improvement
  • +59.8% / +52.9% / +36.2% — LLM-using scientists paper growth (SSH / bio-life / physics-math)
  • 2× / Jan 1 — OpenAI/Anthropic API limit multiplier / end date
  • Most-mentioned company: NVIDIA
  • Quietest segment: frontier-lab launches

Compiled by AI Feed’s editor from verified web sources for 24 December 2025.