🗞️ AI Daily Briefing — April 8, 2026

🔥 Top Story

Anthropic triples revenue and locks in 3.5GW of TPUs. Anthropic announced its largest compute deal yet — a multi-gigawatt expansion with Google and Broadcom for 3.5 GW of next-gen TPU capacity coming online in 2027 — while disclosing that ARR has jumped from $9B at end of 2025 to over $30B. The number of customers spending >$1M/year on Claude doubled from 500 to 1,000+ in under two months. (TechCrunch, Bloomberg)

🚀 Model & Research News

  • OpenAI, Anthropic & Google form anti-distillation pact: The three labs are now actively sharing telemetry through the Frontier Model Forum to detect “adversarial distillation” — Chinese rivals scraping outputs to train cheap clones. Rare cooperation among bitter rivals. (Bloomberg, The Decoder)
  • NVIDIA open-sources DreamDojo: Jim Fan’s GEAR lab dropped the full stack — weights, code, post-training data, evals — plus a live VR teleop demo connecting a Unitree G1 to a “dream simulation.” Fan’s framing: 2026 is “the year of World Models for physical AI.” (NVIDIA Blog)
  • Sam Altman’s superalignment numbers exposed: New Yorker follow-up reporting confirms the original 20% compute pledge actually delivered 1–2% — and on the oldest clusters with the worst chips. Better hardware reportedly went to revenue products. (TechBrew)
  • dfs-mini1 launches: depthfirst quietly shipped its first in-house security model — an RL-post-trained open base that beats frontier general-purpose models on vuln-finding at 10–30× lower inference cost. (BusinessWire)

🛠️ Tools & Developer Updates

  • Genspark Workspace 4.0 demo tonight: Livestream at 8 PM PT (April 7) / noon JST (April 8) — first live demo of the next major version after AI Workspace 2.0 launched in February. (Genspark)
  • MCP at 97M+ installs: The Model Context Protocol is now de facto AI infrastructure — every major lab ships MCP-compatible tooling. If you’re not exposing tools/data via MCP, you’re the outlier. (crescendo.ai)
  • LangGraph 2.0 is the agent runtime: Production patterns are converging — LlamaIndex for retrieval, LangGraph for orchestration. The “LangChain vs LlamaIndex” framing is officially dead. (aimultiple)

💰 Funding & Business

  • Eclipse Ventures raises $1.3B for “physical AI”: The Cerebras backer’s biggest fund ever — $720M early stage + $591M later stage — explicitly targeting robotics, manufacturing, defense, and AI infra. (TechCrunch, Bloomberg)
  • depthfirst closes $80M Series B: Meritech-led, with Forerunner and The House Fund joining Accel and Box Group. Total raised now $120M, less than 90 days after exiting stealth. 80% fix-acceptance rate from devs. (SecurityWeek)
  • Aria Networks raises $125M Series A: London-based AI-data-center networking startup — hardware-agnostic, “AI-native” gear targeting the gigawatt-class buildouts. (Crescendo)
  • Hermeus closes $350M: $200M equity led by Khosla, with Founders Fund, In-Q-Tel, RTX Ventures, and Cox in. Autonomous hypersonic aircraft — defense-AI is officially uncapped. (TechCrunch)

🐦 Notable from the Timeline

  • @DrJimFan: Open-sourced DreamDojo and is hammering the message that “WAMs (World Action Models) learn best from diverse data, not repeated demonstrations.” Pixels as the universal robot bridge.
  • @fchollet: Still pushing ARC-AGI-3 — humans 100%, frontier AI 0.26%. The interactive-environments benchmark is the new AGI scoreboard now that v2 is being saturated.
  • @sama: Day two of silence on the New Yorker piece. OpenAI’s response so far: announce a “Safety Fellowship” program. Read into that what you will.
  • @hwchase17: Continuing to position LangGraph as the agent runtime, not a competitor to DSPy on optimization. Clearer narrative than 6 months ago.
  • @arena: New leaderboard adds today — glm-5.1 and dola-seed-2.0-pro joined the text arena. Top six models now within 20 Elo points of each other. (Arena)

📊 Benchmark Watch

  • LMArena (text): Claude Opus 4.6 Thinking holds #1 at 1504 Elo, with Gemini 3.1 Pro Preview (1493) and GPT-5.4 High (1484) in close pursuit. Six models within a 20-Elo band. (LMArena)
  • Total votes: 5,754,368 votes across 338 models as of April 7. The pace of new model additions is accelerating.
  • ARC-AGI-3: Still wide open. No frontier model has cracked the interactive-learning regime. (ARC Prize)

🎙️ Podcast Highlights

  • TBPN: Latest episode — first under OpenAI ownership — leans into editorial-independence framing while running guests Marc Lore, Adam Meyers, Jeremy Allaire, and Justin Levine. Slate published a sharp counter-take calling the deal “sleazy.” (TBPN, Slate)
  • Stratechery: Ben Thompson’s “OpenAI Buys TBPN” piece is still the cleanest analysis — read it as a token-distribution play, not a media play. (Stratechery)

🔗 Worth Reading