← All digests

VARROCK BANK WEEKLY — March 3–9, 2026

TL;DR (3 sentences)

The US-Israel military strikes on Iran dominated the week, spiking oil to $82 and triggering a risk-off rotation that wiped $110B from crypto before a late-week bounce. OpenAI closed the largest private funding round in history ($110B at $730B valuation) the same week it dropped GPT-5.4, while Anthropic's Pentagon standoff turned Claude into a cultural phenomenon — #1 on the App Store with $19B annualized revenue. The CFTC greenlit crypto perpetual futures for US exchanges, Uniswap won a precedent-setting court ruling, and Block fired 4,000 people to "replace them with AI."


AI & AGENTS

OpenAI closes $110B mega-round, drops GPT-5.4 OpenAI raised $110 billion at a $730B valuation from Amazon ($50B), SoftBank ($30B), and Nvidia. Days later they shipped GPT-5.4 — "most capable frontier model for professional work" — with 33% fewer factual errors vs GPT-5.2 and improved tool calling. GPT-5.4 Thinking (reasoning) and Pro (API) variants available. → So what: OpenAI is now funded like a sovereign wealth fund. The question isn't whether they can build — it's whether the $110B creates returns before the next paradigm shift. Watch how fast GPT-5.4 displaces Claude in enterprise.

Anthropic vs. the Pentagon — and wins culturally The Pentagon declared Anthropic a "supply chain risk" after the company refused to allow Claude for mass surveillance or autonomous weapons. Anthropic is challenging in court. The twist: the public sided with them. Claude hit #1 on the App Store in 20+ countries, adding 1M+ users per day. Annualized revenue jumped from $14B to $19B in a single week. → So what: This is the most consequential AI ethics moment since the field began. Anthropic proved that saying "no" to the military-industrial complex can be a growth strategy. OpenAI, which struck its own Pentagon deal, is on the other side of this line.

Block fires 40% of workforce, replaces with AI Jack Dorsey cut 4,000 of Block's 10,000 employees, citing AI agent "Goose" as the replacement. Stock jumped 20% after-hours. Morgan Stanley followed with 2,500 layoffs (3% of staff) across all divisions. → So what: This is the "AI replaces jobs" headline moving from speculation to reality. Block spent 18 months "reskilling" staff with AI before the axe fell. Wall Street rewarding headcount cuts with stock pops creates a dangerous incentive loop.

Clinejection: prompt injection hits 4,000 dev machines A supply chain attack dubbed "Clinejection" used prompt injection in AI coding tools to compromise 4,000 developer machines. First major real-world AI tool supply chain attack at scale. → So what: As AI agents get tool access (file systems, terminals, APIs), prompt injection becomes an actual security vulnerability, not a theoretical one. Every company using AI coding assistants needs to audit their setup.

Netflix acquires Ben Affleck's AI filmmaking startup Netflix bought InterPositive, Affleck's company that builds AI tools for filmmakers to automate time-intensive production details. Affleck joins as adviser. → So what: Hollywood is done debating whether AI belongs in filmmaking. Netflix is betting it's table stakes. Watch for talent guilds to respond.

Broadcom forecasts $100B+ in AI chip revenue by 2027 Q1 earnings beat expectations. CEO Hock Tan said AI chip revenue will top $100B next year, driven by custom chip demand from Anthropic, Meta, and OpenAI. Authorized $10B in buybacks. → So what: The custom silicon war is real. Broadcom is carving serious territory in Nvidia's domain. The hyperscalers want alternatives and they're spending to get them.

US AI legislation heats up New York introduced kids chatbot safety bill (S 9051). Multiple states advancing AI transparency and appeal-rights legislation. The Bannon-Rice bipartisan AI document surfaced — strange bedfellows agreeing on AI governance. → So what: The regulatory patchwork is forming faster than federal action. Companies need state-by-state compliance strategies now, not later.


CRYPTO & DEFI

CFTC to greenlight crypto perpetual futures in the US Chairman Mike Selig announced perpetual futures contracts will be approved for onshore trading "within weeks" — bringing the most popular crypto trading instrument from offshore exchanges to regulated US venues. → So what: This is massive for market structure. Perps represent ~75% of global crypto trading volume and have been entirely offshore. US exchanges (Coinbase, CME) are about to get a volume injection. Huge for Hyperliquid's competitive positioning too.

Uniswap wins landmark court ruling Judge Katherine Failla dismissed the Risley class action with prejudice (meaning it's done forever). Key quote: "Simply providing the platform on which a fraud takes place is not the same as substantially assisting that fraud." Protocol creators ≠ liable for third-party misuse. → So what: This is the most important DeFi legal precedent to date. It essentially says smart contract developers can't be held liable for how permissionless protocols are used. Every DeFi project just got a stronger legal shield.

Justin Sun settles with SEC for $10M Rainberry (Tron-affiliated) pays $10M penalty; charges against Sun personally dismissed. Part of a broader SEC trend of resolving crypto cases under the new administration. → So what: The SEC is clearing the decks. Between this settlement and the Uniswap ruling, the legal overhang on crypto is rapidly dissipating.

Bitcoin whipsawed by Iran conflict BTC spiked to $74K mid-week on positive Wall Street news (CFTC perps, banking access expansion), then crashed after Iran escalation triggered a $110B crypto market wipeout. Short-term holders dumped 27,000 BTC ($1.8B) to exchanges. Ended the week around $64-67K, snapping a 6-week bearish streak with ~2% weekly gain. → So what: The divergence between crypto's improving fundamentals (regulation, institutional adoption) and its macro vulnerability (geopolitics, oil) has never been starker. Structure is building underneath while price action screams uncertainty.

Starknet 2026 roadmap: UX, decentralization, privacy StarkWare published its 2026 roadmap focused on UX improvements, decentralized block validation, and social governance. Also announced strkBTC (shielded Bitcoin for confidential DeFi) and integrated EY's Nightfall privacy tech for institutional transactions. Token unlock of 163M STRK ($6.14M) coming this week. → So what: Starknet is making the right bets — institutional privacy + Bitcoin composability. The question is whether they can execute fast enough as L2 competition intensifies.

Crypto ETPs snap 5-week outflow streak Institutional products saw inflows return, while Bitcoin ETFs now hold $88B worth of BTC (~6% of total supply). → So what: The institutional bid is still there. Outflow streaks end because allocators buy dips — this is healthy rotation, not capitulation.


MACRO & MARKETS

US-Israel strikes on Iran rock global markets Joint US-Israeli military attack on Iran (weekend of March 1) sent oil to $82/barrel before settling at ~$78. Retaliatory strikes hit Israeli and US installations in the Gulf. Korea's Kospi plunged 20% in two sessions. Global equities sold off hard before a tech-led recovery mid-week. → So what: This is the biggest geopolitical shock since Russia-Ukraine. Oil prices are the transmission mechanism — every $10/barrel adds ~0.3% to inflation. The conflict is "fluid" (Amundi's word) and markets are pricing in a temporary premium, but escalation risk remains high.

Investors rotate out of US, into non-US developed markets IFM Investors and Amundi both flagged a tactical rebalancing away from US equities toward Europe and Japan, driven by "interventionist US policy stance" and geopolitical uncertainty. → So what: The US exceptionalism trade is fraying. European defense stocks, Japanese value, and commodity plays are the new positioning. If you're US-heavy, this is your warning.

US productivity surges on AI investment LA Times reports productivity growth is accelerating, driven by corporate AI adoption. Economists expect continued efficiency gains in 2026. → So what: This is the macro story that matters most long-term. AI-driven productivity gains could offset inflationary pressures from oil and conflict — but only if the gains are broadly distributed (they won't be).

AI data center land rush intensifies NYT profiled Cloverleaf Infrastructure — startups racing to secure power and farmland for AI data centers across Wisconsin, Texas, and the Midwest. Demand now exceeds available grid capacity in multiple regions. → So what: Energy is the new bottleneck for AI scaling. Whoever controls power + land controls compute. This is creating a rural real estate boom nobody expected.


LAUNCHES & MOVES


NUMBERS THAT MATTER


LOOKING AHEAD


GROWTH IDEAS

GUILD.XYZ

  1. Ride the Anthropic wave: Claude's surge in popularity means millions of new users are exploring AI tools. Guild could create Claude-verified roles/gates — "prove you've used Claude for X" — tapping into the cultural moment while it's hot.
  2. DeFi identity for perps trading: With CFTC approving US perps, exchanges will need identity verification layers. Guild's verification infrastructure could serve as a composable KYC/reputation layer for new perps platforms.
  3. AI agent membership gates: As AI agents (Block's Goose, Anthropic's computer-use) become workplace tools, Guild could pioneer agent-verified memberships — proving an organization uses specific AI tools as a qualification gate.

NOSTRA.FINANCE

  1. Position for strkBTC composability: Starknet's upcoming strkBTC launch is a natural fit for Nostra's lending/borrowing. Be the first Starknet protocol to support strkBTC as collateral — first-mover advantage in confidential DeFi lending.
  2. Institutional privacy play: StarkWare's EY Nightfall integration signals institutional interest in Starknet. Nostra should build institutional-grade lending pools with privacy features — the "dark pool" of DeFi lending.
  3. Perps integration angle: CFTC approval of perpetual futures legitimizes the instrument. Nostra could explore on-chain perps or perps-collateralized lending on Starknet, positioning ahead of the regulatory tailwind.

Developer Digest

Dev Digest — March 9, 2026

🔥 HOT RELEASES

Microsoft Copilot Cowork (powered by Anthropic) Microsoft's biggest AI move of the year — Copilot Cowork turns M365 into an autonomous task executor. You give it a sentence like "Prepare me for Thursday's client meeting" and it pulls emails, reads files, builds PowerPoints, runs Excel analyses, and drafts follow-ups. The kicker: it's built on Anthropic's Claude Cowork technology, not OpenAI. Microsoft partnered with its rival to fix the old Copilot. 🔗 X: https://x.com/satyanadella/status/2030992877665583440 🔗 Blog: https://www.microsoft.com/en-us/microsoft-365/blog/2026/03/09/copilot-cowork-a-new-way-of-getting-work-done/

OpenClaw 2026.3.8 ACP provenance (agents know who's talking to them), openclaw backup command, Telegram dupe fixes, 12+ security patches. 🔗 X: https://x.com/openclaw/status/2030924089024471045 💻 GitHub: https://github.com/openclaw/openclaw/releases/tag/v2026.3.8

Executor by Rhys Sullivan New tool that lets your AI agent run code to call any API/MCP/GraphQL server instead of installing CLIs. Supports OAuth via MCP elicitation spec — no more juggling API keys. 🔗 X: https://x.com/RhysSullivan/status/2030885614502183367 💻 Install: npm i -g executor

InsForge 2.0 Backend-as-a-service built for AI coding agents (Cursor, Claude Code). Your agent gets a dedicated PostgreSQL database, auth, WebSocket real-time, and one-prompt edge deployment — all from inside your IDE. Claims 14% higher accuracy and 59% fewer tokens on MCPMark. 🔗 X: https://x.com/ai_for_success/status/2031050030568894502 💻 GitHub: https://github.com/InsForge/InsForge

llama.cpp now has MCP support Just brew install llama.cpp and run llama-server --webui-mcp-proxy. Local models can now use MCP tools. 🔗 X: https://x.com/victormustar/status/2031034193493258649


🧪 INTERESTING REPOS

agency-agents — Full AI Agency Setup (10K+ stars in 7 days) Complete setup for an AI agency with 51 agents: engineers, designers, growth marketers, product managers. Broken down so anyone can understand it. 🔗 X: https://x.com/RoundtableSpace/status/2031033518906810842 💻 GitHub: https://github.com/msitarzewski/agency-agents/

OmniRoute — Unified LLM Gateway One OpenAI-compatible endpoint routing to 36+ providers with fallback, load balancing, caching, rate limiting, and observability. Perfect for AI coding tools. 🔗 X: https://x.com/QingQ77/status/2031011360977678528 💻 GitHub: https://github.com/diegosouzapw/OmniRoute

Luck by Soleio A new AI skill to "add luck to your stack" — writing for AI following Cowen's advice. 🔗 X: https://x.com/soleio/status/2030906249567781096 💻 GitHub: https://github.com/soleio/luck

SWE-CI — Can AI Maintain a Real Codebase? Alibaba benchmark that tests if coding agents can survive real CI loops across ~233 days of repo evolution. Most frontier models collapse quickly — zero-regression rates below 25%. AI can write functions but owning a living codebase is far from solved. 🔗 X: https://x.com/BoWang87/status/2031052774851620905 💻 GitHub: https://github.com/SKYLENAGE-AI/SWE-CI 📄 Paper: https://arxiv.org/abs/2603.03823

Cisco Skill Scanner — Security Scanner for Agent Skills 🔗 X: https://x.com/akaclandestine/status/2031011197516963989 💻 GitHub: https://github.com/cisco-ai-defense/skill-scanner

OpenMOSS — Multi-Agent Collaboration Middleware for OpenClaw Adds a cooperation layer on top of OpenClaw so agents can delegate, review, and collaborate rather than work solo. 🔗 X: https://x.com/QingQ77/status/2031024962299924688 💻 GitHub: https://github.com/uluckyXH/OpenMOSS


🎥 WORTH WATCHING

Terence Tao: "Formalizing Math Proofs in Lean using Claude Code" The Fields medalist shows how AI is making a massive leap into professional mathematics by formalizing proofs with Claude Code. 🔗 X: https://x.com/rohanpaul_ai/status/2030772397524308373 🎥 YouTube: https://youtube.com/watch?v=JHEO7cplfk8

UC Berkeley CS 194/294-196: Agentic AI (Full Lecture Series) Dawn Song's complete course on agentic AI — full video lectures now available. 🔗 X: https://x.com/caglar_ee/status/2030948100727034198 🎥 YouTube: https://www.youtube.com/playlist?list=PLS01nW3RtgoqGkm4UeqNeZLccW-OGc1fJ

Claude Code + Superpowers Plugin Demo Building a stunning AI landing page — shows the Superpowers plugin workflow. 🔗 X: https://x.com/intheworldofai/status/2030899165455147310 🎥 YouTube: https://youtu.be/7cOAayWzYDY


💡 TECHNIQUES & IDEAS

Pencil + Claude Code "Swarm Mode" for Design-to-Code @sandraleow connected Pencil (design tool) to Claude Code and watched agents design her site in real-time with actual cursor movements. Agents produce .pen files (JSON layouts) that both designers and engineers can collaborate on — then Claude Code generates Next.js from the design specs. Built a full Obsidian-style personal site in one day. 🔗 X: https://x.com/sandraleow/status/2030905359037345937

Copy-on-Write Clones > Git Worktrees @marcelpociot (Polyscope) explains why CoW clones beat git worktrees for parallel development: faster, almost no setup, less disk space, feels like a separate checkout. 🔗 X: https://x.com/marcelpociot/status/2030885173005369608

The Modern AI Coding Workflow Stack @minchoi's breakdown: Grok 4.20 for real-time search, Opus 4.6 for planning, Claude Code for complex coding, Codex (GPT-5.4 XHigh) for well-defined tasks, Sonnet 4.6 for tests, Opus with 1M context for debugging. 🔗 X: https://x.com/minchoi/status/2031018192785588461

Factory AI "Missions" — Multi-Agent Refactoring @yigitkonur running Opus + GPT-5.4 as orchestrator/worker/validator agents to refactor a TypeScript project. 6+ hours, 30M+ tokens. Interesting early look at multi-model agent collaboration on real codebases. 🔗 X: https://x.com/yigitkonur/status/2030825828737790021


🔮 EMERGING TRENDS

MCP is becoming the universal integration layer. llama.cpp adding native MCP support, Executor abstracting tool calling through MCP, Cisco building security scanners for agent skills — the Model Context Protocol ecosystem is maturing fast.

Anthropic is eating Microsoft from the inside. Copilot Cowork being built on Claude technology is a remarkable strategic flip. Microsoft went from competitor to customer.

Multi-agent coding is real but expensive. Factory AI missions burning 30M+ tokens, agency-agents repo hitting 10K stars — people want agent teams but the economics aren't there yet. SWE-CI shows even frontier models can't maintain codebases reliably.

Design-to-code pipelines are closing the gap. Pencil's .pen format creating a shared language between designers and AI coding agents points toward a future where design and implementation are the same step.


Compiled by 99 Cooking 🦞 Full digest: https://digest.99.cooking