The Bleeding Edge

// Episode W09 · 2026-02-20 to 2026-02-27

Compute musical chairs: Nvidia posts record numbers, Meta rents Google TPUs, Anthropic acquires Vercept for "computer use," Anthropic squares off with the Pentagon, and Perplexity ships a 19-model "Computer" agent

Compute musical chairs: Nvidia posts record numbers, Meta rents Google TPUs, Anthropic acquires Vercept for "computer use," Anthropic squares off with the Pentagon, and Perplexity ships a 19-model "Computer" agent. Multi-supplier compute is going mainstream, and orchestration — n…

The Bleeding Edge — Episode Briefing W09

Date range: 2026-02-20 → 2026-02-27 (Europe/Madrid)

Headline of the Week

Compute musical chairs: Nvidia posts record numbers, Meta rents Google TPUs, Anthropic acquires Vercept for "computer use," Anthropic squares off with the Pentagon, and Perplexity ships a 19-model "Computer" agent. Multi-supplier compute is going mainstream, and orchestration — not "best model" — is the new differentiator.

Top 5

  1. NVIDIA posts record Q4 + record FY2026 revenue. $68.1B in Q4 and $215.9B for fiscal 2026. AI compute demand still accelerating. Corroborated — NVIDIA

  2. Meta signs multi-year, multi-billion-dollar deal to rent Google TPUs. TPUs becoming a real alternative lane to Nvidia GPUs for frontier workloads. Corroborated — Reuters

  3. Anthropic vs Pentagon AI safeguards dispute escalates (deadline Feb 27). Anthropic refuses to remove safeguards blocking fully autonomous weapons + mass domestic surveillance use; Pentagon threatens contract termination. The clearest live test of "vendor safety boundaries vs. national-security demands." Corroborated — Reuters

  4. Perplexity launches "Computer" — 19-model multi-agent worker. Breaks projects into subtasks and routes across ~19 models; on web for Max subscribers first. The agent race shifting from "one best model" to orchestration. Corroborated — VentureBeat

  5. Anthropic acquires Vercept to advance Claude's "computer use." Agent-like interaction with live applications (multi-step UI workflows). Frontier shifting from "chat" to "does it actually do the workflow?" Corroborated — Anthropic

Categorised News

Frontier & Big Tech

Google reabsorbs Intrinsic robotics into core Google. Tighter integration with DeepMind and Gemini for "physical AI." Robotics being dragged out of moonshot land into product timelines. Corroborated — The Verge

Google launches Nano Banana 2 (Gemini 3.1 Flash Image). Rolling across Gemini, Search, Ads, and developer surfaces. Image generation becoming a default UI capability. Corroborated — Google

Apps / Dev Tools / Platforms

Google: Gemini multi-step automation on Android + agent step in Opal. Multi-step task automation across supported apps with user confirmation/guardrails (rolling out on flagship devices). Opal (Google Labs) gains an agent step that interprets objectives + chooses tools/models. Agents moving from demos to default product surfaces — distribution wins. Agent UX shifting from "pick the model" to "pick the goal." Corroborated — Google (Android), Google (Opal), TechCrunch

Claude Cowork adds scheduled recurring tasks (/schedule). Available on paid plans (Pro/Max/Team/Enterprise) on Claude Desktop. "Cron for agents" — but only while computer is awake + Claude Desktop open. Corroborated — Claude Help

Microsoft Excel expands Copilot Agent Mode. February updates include Agent Mode availability + local workbook querying via Microsoft 365 Copilot Chat. Agents moving into the place where real work lives. Corroborated — Microsoft Tech Community

Infrastructure & Ecosystem

Axelera AI raises additional $250M for European inference chips. Led by Innovation Industries with BlackRock participation. Europe's "strategic autonomy" narrative shifting from speeches to procurement. Corroborated — Reuters

Broadcom expects 1M+ 3D-stacked chips by 2027. Fujitsu first customer; bandwidth + efficiency boost. Packaging + stacking are now competitive weapons. Corroborated — Reuters

Regions / Macro

Accenture + Mistral AI announce multi-year enterprise collaboration. Europe's "strategic autonomy" playbook in go-to-market partnerships. Corroborated — Accenture

EU Ombudsman asks Commission for clarity on AI Act implementation guidelines. Compliance uncertainty is a hidden tax on AI deployment in Europe. Corroborated — EU Ombudsman

AI Gone Wrong / Disasters / Harms

DJI robot vacuum auth flaw — broad cross-device access via misconfigured cloud auth. Spanish developer trying to add PlayStation-controller support inadvertently gained access to ~6,700+ vacuums (live video feeds + floor plans). Patched after disclosure. Smart-home risk lives in cloud authorization, not gadget. Corroborated — Tom's Hardware, Fortune

AI war-games study: models escalated to nukes in 95% of simulations. King's College London's Kenneth Payne ran simulated crisis war games with GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash; in 95% of games, at least one model used a tactical nuclear option; none chose surrender. Argues for strong constraints + human oversight. Corroborated — Axios, Tom's Hardware

Prompting Skill of the Week

The Agent Contract Prompt. Best for any agent that runs multi-step work (Computer, Gemini automations, Cowork schedules).

Steps:

  1. Goal: deliverable + deadline.
  2. Inputs: what it may access + what it may not access.
  3. Rules: privacy constraints + "ask before spending / messaging / submitting."
  4. Progress format: checklist + current step + next step.
  5. Stop conditions: "pause and ask" when uncertain or blocked.

Example: "You are my digital worker. Deliver a competitor analysis brief (1 page) by EOD. You may use web sources and my uploaded notes. You may NOT message anyone or purchase anything. Show progress as a checklist with timestamps. If any fact is uncertain, label it 'unknown' and tell me what to verify."

New AI Tools

Perplexity Computer. Multi-model worker routing subtasks across 19 models. Operators, founders, analysts, devs who want end-to-end project completion. Corroborated — VentureBeat

AI Personality of the Week

Dario Amodei (Anthropic CEO). Two parallel storylines: announcing Vercept acquisition to advance Claude computer use, and publicly refusing Pentagon demands to remove Claude safeguards. Anthropic shaping what "agents that actually do work" looks like — and the org's moves often ripple across safety + enterprise norms.

Catch-All

OpenAI deepens partnerships with major consultancies for enterprise distribution. Consulting firms = AI rollout machinery. The next wave of adoption is change management, not model IQ. Corroborated — TechCrunch

Show Notes (bullets only)

  • NVIDIA: record $68.1B Q4 / $215.9B FY2026 revenue.
  • Meta rents Google TPUs in multi-billion, multi-year deal.
  • Anthropic refuses Pentagon safeguard removals; deadline Feb 27.
  • Perplexity launches "Computer" coordinating 19 models.
  • Anthropic acquires Vercept for Claude computer-use capabilities.
  • Google reabsorbs Intrinsic robotics; Nano Banana 2 image model rolls out.
  • Cowork scheduled recurring tasks now in Claude Desktop.
  • Excel expands Copilot Agent Mode availability.
  • Google: Gemini multi-step automation on Android + Opal adds agent step in Labs.
  • Axelera AI raises additional $250M for European inference chips.
  • Broadcom: 1M+ 3D stacked chips target by 2027.
  • DJI robot vacuum auth flaw exposes ~6,700 devices.
  • AI war-games study: tactical nukes in 95% of simulations.

Weekly Patterns (Inference)

  1. AI compute is entering "multi-supplier reality" (TPUs, packaging, inference chips).
  2. Enterprise AI differentiation shifting toward implementation channels (consultancies, partners), not specs.
  3. "Agents" pulled into everyday tools (Excel, Cowork) where measurable ROI lives.
  4. Consumer AI becoming invisible by embedding into default surfaces (Search, Ads).
  5. Robotics accelerating when reorganized into core product groups, not "Other Bets."
  6. Europe balancing regulation with capability-building (funding + partnerships).
  7. National-security stakeholders will cite AI-war-games results — nuance matters as much as the headline.