Today in AI – 12-27-2025

A curated, executive-level rundown of the most important AI and automation developments from the past 48 hours, with context for operators, builders, and policy leaders.

Two-armed home robot tidying a modern living space

Key Stories

Nvidia moves on inference: licensing deal with Groq and key hires

Nvidia confirmed a non-exclusive licensing agreement for Groq’s LPU (Language Processing Unit) inference tech and will hire founder/CEO Jonathan Ross, president Sunny Madra, and other staff. Reports differ on scope and value: Nvidia told TechCrunch it is not acquiring Groq outright, while market coverage frames the move as licensing plus selective asset purchases and leadership transitions at Groq (CFO Simon Edwards stepping in as CEO). Bottom line: Nvidia is tightening its grip on AI inference—complementing its training-dominant GPU stack—with a pathway to bring LPU-like latency and efficiency into its ecosystem. For deeper analysis, see our take: Nvidia strikes licensing deal with Groq, hires founder in inference push.

  • Why it matters: Inference is where AI hits production scale and cost sensitivity. A credible LPU path could reduce latency and power, expand Nvidia’s TAM beyond GPUs, and complicate the calculus for rivals pitching inference-optimized silicon.

Sources: TechCrunch; Investing.com; Investopedia.

China drafts new rules for “human-like” AI interactions

China’s cyberspace regulator released draft rules for AI systems that simulate human personalities and engage users emotionally. Providers would need to label AI interactions, warn about excessive use, intervene in signs of addiction or emotional distress, and enforce strict safety and content controls. Public consultation runs to January 25. For our policy briefing, see: China drafts rules for human-like AI: warnings on overdependence and content limits.

  • Why it matters: A first-of-its-kind focus on psychological impacts and “AI companion” risks signals where consumer AI governance may be headed globally—especially for chatbots and synthetic “friends.”

Sources: Reuters; The Straits Times; China Daily/Xinhua.

Robotaxi resilience questioned after San Francisco power outage stalls Waymo cars

A Dec 20 power substation fire blacked out signals citywide and stranded multiple rider-only Waymo vehicles at intersections. Waymo says its fleet handled over 7,000 dark signals but a concentrated spike in remote-confirmation requests caused delays; California regulators are reviewing. Safety experts are urging rules for remote operations/teleoperation as fleets scale. Our on-the-ground implications: Waymo’s San Francisco outage triggers calls to regulate robotaxi teleoperation.

  • Why it matters: In disasters, autonomy depends on resilient ops and human-in-the-loop capacity. Cities and AV operators will face pressure to prove stress-tested procedures before expanding service areas.

Source: Reuters.

LG teases two‑armed home robot with “Affectionate Intelligence”

Ahead of CES, LG previewed CLOiD—its most ambitious home robot yet—with two articulated arms (7 DoF per arm), five-finger hands, multi‑modal sensing, and an “Affectionate Intelligence” interaction layer. Details arrive in January.

  • Why it matters: After years of single‑task bots, large brands are moving toward general household manipulation—an early signal for consumer‑grade embodied AI.

Source: The Verge.

Energy policy flashpoint: states push back on federal AI data‑center oversight

New reporting highlights a growing clash between U.S. states and the federal government over who controls grid connections for AI‑hungry data centers. It follows recent FERC actions directing PJM to craft clearer rules for co‑located data center loads. Expect legal, regulatory, and rate‑design battles to intensify into 2026.

  • Why it matters: Compute growth is running into power constraints. Rule clarity will influence siting, timelines, costs, and sustainability pathways for hyperscaler AI build‑outs.

Sources: Wall Street Journal; FERC releases (Dec 18).

Emerging Trends

“Anthropomorphic AI” guardrails shift from content to wellbeing

China’s draft rules add user‑wellbeing provisions—like time/overdependence warnings and intervention—on top of content limits and safety obligations. Early signal: regulators may treat AI “companions” more like digital health products than search/chat utilities. Watch for parallel proposals in APAC and Europe in 2026. Evidence: labeling and periodic reminders in the draft; public comment timeline to Jan 25.

Inference is the new battleground for silicon

Nvidia’s Groq move underscores a pivot from train‑time dominance to run‑time economics—where latency, throughput, and watt‑per‑token are king. Expect more licensing, selective acqui‑hires, and software bridges that pull “alternative accelerators” into incumbent toolchains. Evidence: Nvidia‑Groq licensing and leadership hires; buy‑side commentary highlighting inference.

Embodied AI targets real homes, not just labs

CLOiD’s two arms and fingered hands, plus CES‑bound humanoids (e.g., Richtech’s Dex on Jetson Thor), point to broader household manipulation tasks—laundry, tidying, basic food prep—moving from demo to pilot programs. Evidence: LG preview specs; CES exhibitor releases citing Nvidia Isaac Sim/Lab and on‑device reasoning.

Urban autonomy must prove crisis‑mode ops

Waymo’s outage shows human‑assist and back‑end orchestration as critical failure points, not just perception/planning. Expect regulators to mandate stress benchmarks (e.g., teleops queue time, service continuity under wide‑area comms/power failures). Evidence: expert comments (Koopman, Cummings), DMV/CPUC inquiries.

Power becomes product: AI build‑outs drive grid rulemaking and on‑site generation

Hyperscalers are pursuing grid‑adjacent or on‑site solutions (aeroderivative turbines, new rate classes) as interconnection queues stretch. With FERC prodding PJM and states weighing consumer protections, energy strategy is now a first‑class part of AI roadmaps. Evidence: WSJ and FT reporting; FERC orders; Pew estimates on bill impacts.

30x
Potential AI data‑center power growth by 2035Source: deloitte-ai-datacenters-2025

Conversations & Insights

“Is Nvidia vertically consolidating inference?”

  • Where it’s happening: Tech press coverage and investor notes; market chatter across X/LinkedIn.
  • Key voices: TechCrunch reported Nvidia’s statement that this is not a company acquisition; buy‑side/analyst commentary (e.g., BofA via Investing.com) frames it as a strategic inference play with leadership hires.
  • Takeaway: Regardless of deal structure, customers should assume tighter Nvidia control over the inference stack—with potential developer convenience upsides but reduced heterogeneity unless open runtimes mature.

“Robotaxis at scale need disaster playbooks—not just better ML”

  • Where it’s happening: Reuters analysis quoting academic safety experts; local media and policy forums discussing emergency response and teleops headcount/capacity.
  • Key voices: Philip Koopman (CMU), Missy Cummings (GMU) call for regulating remote operations and scaling standards.
  • Takeaway: Cities will push for clearer SLAs and telemetry on remote assist during blackouts/earthquakes; AV operators should pre‑negotiate emergency protocols with DOTs and utilities.

“Who should pay for AI’s power bill?”

  • Where it’s happening: Policy press (WSJ), FERC dockets, and think‑tank/utility briefs debating separate rate classes and consumer protections.
  • Key voices: FERC leadership; state regulators; lawmakers pressing for guardrails on cost pass‑through to households.
  • Takeaway: Procurement, siting, and sustainability teams must become grid‑fluent. Expect longer diligence cycles and novel contract structures (e.g., minimum‑bill tariffs, co‑location, on‑site firming).

Quick Takeaways

  • Inference economics will shape 2026 budgets: evaluate LPU/GPU/CPU and model‑compression tradeoffs for latency‑sensitive workloads.
  • Prepare AV operations for black swans: track teleops queue metrics and rehearse city‑wide outage procedures with local agencies.
  • Treat power as a strategic dependency: engage early with utilities; explore co‑location, long‑term minimum‑bill tariffs, and on‑site firming.
  • Build policies for “AI companions”: introduce session‑time reminders, human escalation paths, and clear labeling—regulators are watching.
  • Robotics pilots should target narrow, repetitive household manipulations with measurable ROI and safety envelopes.

Sources

  • Reuters — China issues draft rules for human‑like AI interaction; Waymo outage analysis.
  • The Straits Times — Details on labeling/overdependence warnings in China’s draft.
  • China Daily/Xinhua — Public consultation notice on anthropomorphic AI draft.
  • TechCrunch — Nvidia to license Groq tech and hire leadership (Nvidia comment on no acquisition).
  • Investing.com — BofA reaction to Nvidia–Groq licensing.
  • Investopedia — Market framing of Nvidia–Groq deal.
  • The Verge — LG’s CLOiD two‑armed home robot preview.
  • FERC — Orders and fact sheet on PJM and co‑located data‑center load.
  • Wall Street Journal — State–federal tensions over AI data‑center oversight.
  • Pew Research — Consumer bill impacts near data‑center hubs.
  • GlobeNewswire — Richtech Robotics Dex (CES 2026) on Jetson Thor and Isaac Sim/Lab.