The AI Dispatch — March 23, 2026

Strategic Infrastructure

Musk Unveils “Terafab” — A $25 Billion Chip Factory Backed by Tesla, SpaceX, and xAI

The ambitious project targets 2-nanometer AI chips, with 80% of compute capacity destined for orbital data center satellites, marking the largest cross-company infrastructure bet in AI history.

Sources: Fortune • Bloomberg

Elon Musk announced “Terafab” on March 22, a joint chip fabrication facility in Austin, Texas backed by Tesla, SpaceX, and xAI. The plant targets 2-nanometer process technology and aims to produce 100–200 billion AI and memory chips per year, with a stated goal of eventually reaching 1 terawatt of annual compute output. The project represents the most ambitious vertical integration play in AI infrastructure since NVIDIA’s acquisition spree.

Most strikingly, Musk envisions putting 80% of Terafab’s compute capacity into orbital data center satellites, building on SpaceX’s January FCC application for one million satellite launches. The space-compute vision would create an entirely new tier of AI infrastructure — latency-insensitive batch processing jobs running above the atmosphere, powered by continuous solar energy and cooled by radiative dissipation into space.

The $25 billion price tag makes Terafab the largest single capital commitment any AI-adjacent company has made. The announcement comes as global chip supply remains constrained, with TSMC, Samsung, and Intel all expanding capacity but facing multi-year backlogs. For xAI specifically, in-house fabrication could break its dependence on external suppliers for Grok training clusters — a strategic advantage that no other AI lab currently possesses.

Legal Showdown

Anthropic Files New Declarations Ahead of Tuesday Pentagon Hearing

Anthropic submitted sworn declarations to a California federal court from Head of Policy Sarah Heck and Head of Public Sector Thiyagu Ramasamy, directly contradicting the Pentagon’s claims. Heck’s declaration states Anthropic never demanded approval authority over military operations — a core DOD argument — and says the Pentagon’s concern about Anthropic remotely disabling Claude “appeared for the first time in the government’s court filings.” Microsoft filed an amicus brief supporting Anthropic. A hearing before Judge Rita Lin in San Francisco is scheduled for March 24.

Sources: TechCrunch • Federal News Network

Enterprise Land Grab

OpenAI Offers Private Equity Firms 17.5% Guaranteed Return in $4B AI Joint Venture

In an exclusive report, OpenAI is pitching private equity firms including TPG, Bain Capital, and Brookfield a guaranteed minimum 17.5% return on a ~$4B joint venture that would deploy OpenAI engineers to customize models for PE portfolio companies. Anthropic is simultaneously courting rival firms with a competing — but non-guaranteed-return — offer.

Sources: US News • Investing.com

Labor Market Analysis

Anthropic’s “Observed Exposure” Study: AI Is Already Slowing Hiring of Young Tech Workers

Anthropic published a landmark labor market report introducing “observed exposure” — measuring which jobs Claude users are actually automating, not just which jobs AI theoretically could. While unemployment in exposed roles hasn’t risen overall, the study found suggestive evidence that hiring of younger workers has meaningfully slowed in software, data, and administrative fields.

Computer programmers rank highest, with 75% of their tasks already reachable by AI. The distinction between theoretical and observed exposure is crucial: previous studies estimated AI’s potential reach, while Anthropic’s metric tracks what users are actually doing. The report suggests that AI’s labor market impact is arriving not as mass layoffs but as a quiet slowdown in new hiring — particularly at the entry level.

Sources: Anthropic • Fortune

Policy & Partnerships

Science Infrastructure

Google DeepMind Joins DOE’s Genesis Mission, Opens AlphaEvolve to All 17 National Labs

Google DeepMind formalized a partnership with the U.S. Department of Energy’s Genesis Mission, giving scientists at all 17 DOE National Laboratories accelerated access to the AI co-scientist platform and AlphaEvolve — a Gemini-powered algorithm-design agent. The DOE also announced $293 million in funding for interdisciplinary teams to tackle advanced manufacturing, nuclear energy, and drug discovery.

Sources: DeepMind • DOE

Political Spending

AI Industry’s $100M Shadow Campaign Floods 2026 Midterms Without Mentioning AI

Pro-AI super PACs, including “Leading the Future” backed by $50M from Andreessen Horowitz and $50M from OpenAI co-founder Greg Brockman, are pouring money into midterm primaries in Texas and North Carolina — but running ads exclusively on immigration and crime. The strategy is designed to elect AI-friendly candidates without triggering backlash over AI policy as the actual motive. A rival Democratic-aligned PAC is running counter-ads supporting state-level AI regulation candidates.

Sources: NBC News • Popular Information

Research & Models

Open-Source Benchmark

MiroThinker-H1 Dethroned GPT-5.4 and Opus 4.6 on BrowseComp Deep-Research Benchmark

MiroMind’s MiroThinker-H1 posted state-of-the-art results on BrowseComp, BrowseComp-ZH, and FrontierScience, surpassing all flagship models. The system uses “Effective Interaction Scaling” with a dual-layer verification mechanism — on BrowseComp’s hard subset, the local verifier cut reasoning steps by ~82% while improving accuracy by 26.4 points. The companion open-weight MiroThinker-1.7 achieves 74.0; MiroThinker-H1 hits 88.2.

Sources: PR Newswire • GitHub

On-Device AI

Qwen 3.5 Small Matches 120B-Parameter Models on Expert Science Benchmark — Runs Offline on iPhones

Alibaba’s Qwen 3.5 Small has demonstrated benchmark parity with models ten times its size on GPQA Diamond, a hard expert-level science reasoning benchmark. Most strikingly, the 2B-parameter variant runs fully offline on recent iPhones with 4GB of RAM — no API key, no subscription, no network connection. Capable expert-level reasoning may soon be a default device feature rather than a cloud service.

Sources: Hugging Face • GitHub

Quick Dispatches

WHO Launches Global AI Health Consortium, Warns Against “Emotional Dependence”

The World Health Organization held a pre-convening at TU Delft to establish a new Consortium of Collaborating Centres on AI for Health. Experts from 30+ countries agreed on three priorities: treating generative AI use as a public mental health concern, requiring co-design with people with lived experience of mental illness, and monitoring “emotional dependence” on AI companions. WHO

Okta MCP Server Adds Human-in-the-Loop Confirmation via Elicitation API

Okta’s self-hosted MCP server now enforces mandatory human approval before executing destructive operations like deleting applications or deactivating users. It’s one of the first production MCP servers to formally adopt the elicitation spec as a security control layer. Okta

LTX 2.3: Open-Source 4K Video Generation at 50 FPS With Synced Audio

Lightricks released LTX 2.3, a 22B-parameter open-source Diffusion Transformer generating up to 20 seconds of 4K video at 50 FPS in a single forward pass — with synchronized audio. Fully open weights, licensed for commercial use at no cost. Apatero

Apple’s $1B Bet: Gemini-Powered Siri Targets iOS 26.5 Beta Debut

Apple confirmed its revamped AI-powered Siri — built on Google’s Gemini model in a partnership reportedly worth $1 billion — targets a late-March appearance in the iOS 26.5 beta. The new Siri promises deeper personal context, on-screen understanding, and per-app controls through Apple’s Private Cloud Compute. MacRumors

HBR: AI Cutting Job Postings 17% in Automation-Heavy Roles, Boosting Hybrid Roles 22%

A Harvard Business Review analysis of nearly all U.S. job vacancies from 2019 through early 2025 finds AI is simultaneously destroying and creating work — but the new roles demand a combination of human judgment and AI fluency. 94% of surveyed workers prefer AI as a collaborative assistant over full replacement. HBR

WEF: AI-Powered “Narrative Attacks” Now a Top-Tier Global Risk

The World Economic Forum places AI-driven cognitive manipulation in its top five short-term global risks for 2026, arguing precision psychological profiling and emotional targeting have graduated from electoral nuisance to systemic threat. “Faster, more convincing, and harder to detect.” WEF

GitHub Trending

Repository	Language	Stars	Description
openclaw/openclaw	TypeScript	210k+	Local-first personal AI assistant connecting any LLM to 50+ integrations including WhatsApp, Telegram, Slack, and Signal
666ghj/MiroFish	Python	39.6k	Swarm intelligence engine simulating up to 1M agents for financial market and sentiment prediction
microsoft/BitNet	C++	27k+	Official inference framework for 1-bit LLMs — 6x speedup and 82% energy reduction on CPU
TauricResearch/TradingAgents	Python	29.9k	Multi-agent LLM financial trading framework mirroring real-world trading firm dynamics
nearai/ironclaw	Rust	10.3k	Privacy-first OpenClaw reimplementation with local encryption and WebAssembly sandboxing
lightpanda-io/browser	Zig	23.4k	Headless browser built from scratch for AI agents — 9x less memory, 11x faster than Chrome headless
ruvnet/RuView	Python	37.4k	WiFi DensePose: real-time human pose estimation and vital sign monitoring using commodity WiFi signals