Tech Digest – March 20, 2026

AI Designs Its Own Silicon

An AI Agent Designed a Working CPU in 12 Hours. The Usual Timeline: One Year.

Verkor’s Design Conductor, an autonomous AI agent, took a 219-word requirements document and produced a complete, tape-out-ready RISC-V processor — from Verilog through synthesis and place-and-route to verified GDSII layout — in 12 hours, with no human intervention. The resulting chip, VerCore, meets timing at 1.48 GHz on the ASAP7 7nm process and scores 3261 on CoreMark, roughly equivalent to a 2011 Intel Celeron. According to the team, this is the first known instance of an autonomous agent building a complete CPU from specification to fabrication-ready layout.

Note: The chip itself isn’t impressive by 2026 standards — the compression of time is. A quarterly engineering cycle completed over a lunch break. For institutions planning digital infrastructure with hardware dependencies, lead times just became a variable, not a constant.

Sources: arXiv (Verkor), Verkor

Agents Outgrow Their Sandboxes

Coding Agents Now Work While You Sleep — and Their Makers Are Already Watching for Misalignment

Anthropic added asynchronous event channels to Claude Code, allowing MCP servers to push CI results, chat messages, and system alerts so the agent can react while the developer is away. Separately, OpenAI published its framework for monitoring its own internal coding agents for signs of misalignment — the first public acknowledgment that autonomous systems operating inside company codebases now require their own oversight layer. The two developments bracket the same reality: agents are persistent enough to act autonomously and consequential enough to require surveillance.

Note: If the companies building these agents consider them risky enough to watch continuously, the question for any IT department deploying similar tools isn’t whether to monitor — it’s whether “we’ll review the logs” is still a credible answer.

Sources: Anthropic (Claude Code Channels), OpenAI

G42 Opens Formal Recruitment for AI Agents — One Billion by Year’s End

Abu Dhabi-based technology group G42 launched a structured recruitment process for AI agents in enterprise roles, complete with technical validation, performance testing, and a probationary period before scaled deployment. This isn’t symbolic: CEO Peng Xiao stated the company targets one billion AI agents deployed by the end of 2026, consuming close to one gigawatt of AI infrastructure. Developers of accepted agents will receive performance-linked compensation — a contractor-style model where the “contractor” is software.

Note: Structured reviews, probation periods, value-linked pay. G42 is building an HR framework for non-human workers. When the first European institution asks “should we do this too?”, the template will already exist in Abu Dhabi.

Sources: G42, Khaleej Times

Expertise Gets Automated

PwC to Partners: Adopt AI or Leave. Anthropic Builds a Deployment Arm with Blackstone.

PwC’s US leader warned that partners who resist AI integration “will have no place at the firm,” signaling a conversion of tax and advisory work into subscription AI products designed to operate without a human in the loop. From the other direction, Anthropic is in talks with Blackstone and Hellman & Friedman to form a Palantir-style AI consulting joint venture — embedding Claude directly into portfolio companies across holdings worth over $1 trillion. In a survey of 81,000 people across 159 countries, Anthropic found professional excellence was the top desire people expressed for AI, at 18.8%. The demand signal and the supply response are converging fast.

Note: Public institutions procure billions in consulting annually. When the Big Four replace billable hours with AI subscriptions — and AI labs build their own deployment arms to compete — the delivery model changes fundamentally. The question isn’t whether this reaches institutional procurement. It’s whether procurement frameworks can evaluate what they’re buying when the product is a model, not a person.

Sources: Financial Times, The Information, Anthropic

Google Launches “Vibe Design” with Stitch. Figma Loses 8% in a Day.

Google released a major update to Stitch, an AI-native design canvas that turns natural language and voice input into high-fidelity UI prototypes. Instead of starting with wireframes, users describe a business objective or a desired feeling, and the tool generates multiple design directions with instant clickable prototypes. Designs export to Figma, HTML/CSS, and multiple development frameworks. The tool is free. Figma’s stock dropped roughly 8% on the announcement, extending an already steep decline from its August 2025 IPO.

Note: Free, voice-driven, code-exporting design — from the company that owns the browser, the cloud, and the ad ecosystem. The cost and timeline for professional interface design just shifted materially. This doesn’t replace a design team, but it compresses the exploratory phase from weeks to hours.

Sources: Google Blog, John Wang (X)

Platform Governance Meets AI

Meta Replaces Human Moderators with AI and Abandons the Metaverse. The UK Responds with Labeling Rules.

Meta is cutting third-party content moderators in favor of AI moderation systems, shifting billions of daily trust-and-safety decisions to automated tools. The move is part of a complete strategic pivot: the company is simultaneously walking away from its metaverse investment, leaving VR on life support as it redirects fully toward intelligence products. On the regulatory side, the UK government announced it will examine mandatory labeling of AI-generated content as part of wider copyright reforms — among the first binding content-provenance requirements proposed by a major Western economy outside the EU’s AI Act framework.

Note: Two sides of a widening gap. Platforms are automating the judgment calls that shape what citizens see online, while governments are only beginning to require labels on AI output. For any institution whose public communications run through these platforms — or that regulates those who do — the governance architecture just changed hands, and the rule book hasn’t caught up.

Sources: Bloomberg, New York Times, Reuters

The Physical World Reorganises

Waymo Hits 170 Million Miles with 92% Fewer Serious Injuries. Uber Commits $1.25 Billion to Follow.

Waymo’s autonomous fleet has logged 170 million miles with 92% fewer serious-injury crashes compared to human drivers. Uber responded by investing $1.25 billion in Rivian to deploy 50,000 robotaxis, adding capital-backed scale to the safety data. The convergence of demonstrated safety and fleet-scale investment signals that autonomous urban transport is moving from pilot phase to deployment — on a timeline measured in years, not decades.

Note: The safety data is accumulating faster than the regulatory frameworks. The longer that gap persists, the more pressure builds — from industry pushing for access and from citizens asking why a demonstrably safer option isn’t available. Cities that engage early will set the terms. Those that wait will inherit them.

Sources: The Verge (Waymo), The Verge (Uber/Rivian)

Bezos Raising $100 Billion to Automate Factories. Alphabet Spins Out Anori to Fix Permitting.

Jeff Bezos is raising $100 billion to acquire manufacturing companies and retrofit them with AI-driven automation — one of the largest single bets on industrial transformation to date. In parallel, Alphabet’s X moonshot factory spun out Anori with $26 million to streamline the building approval process, targeting the 2–4 year gap between a developer’s decision to build and breaking ground. Anori’s first partnership is with Rio de Janeiro, where it is digitizing the city’s urban licensing process. The round was led by Prologis, one of the world’s largest real estate owners.

Note: Anori deserves close attention. Permitting is among the least glamorous and most consequential bottlenecks in public infrastructure — and it’s a function that sits squarely inside municipal government. If the platform delivers even partial automation of the approvals chain, the institutions that issue permits face pressure to match that pace — or become the bottleneck themselves.

Sources: WSJ, TechCrunch

The AI Scaling Race Needs Energy. Not Everyone Has It.

Behind the trillion-dollar capital flows into AI infrastructure sits a question that gets less attention than it deserves: energy. G42’s billion-agent plan would consume close to one gigawatt of power. Bezos’s factory automation push adds industrial-scale electricity demand on top of existing loads. Meanwhile, China is building solar parks in Cuba — now supplying roughly 10% of the island’s electricity — as a US oil blockade deepens what may be Cuba’s worst energy crisis in decades. The episode is a reminder of how quickly energy dependency becomes a geopolitical lever, and how renewable alternatives can shift the balance.

Note: Every AI initiative — institutional or industrial — ultimately runs on electrons. The institutions proactively diversifying their energy sourcing and investing in renewables aren’t just meeting climate targets. They’re building operational resilience against the kind of supply shocks that can paralyze entire economies overnight.

Sources: Washington Post, Khaleej Times