Today’s newsletter is an 8‑min read. This week, we cover:
How experts think AI will reshape US growth and jobs by 2050
Anthropic’s Mythos model and the Glasswing cyber defense coalition
Sanders’ and Maine’s efforts to slow AI‑driven data‑center expansion
And more.
For feedback or questions, please contact [email protected]
Watch the 2-min video overview:
🌍 Landscape
Forecasting the Economic Effects of AI
A new study, funded by Open Philanthropy and involving economists, AI professionals, and superforecasters from institutions like the Federal Reserve Bank of Chicago, Yale, and Stanford, suggests that rapid AI progress could boost U.S. GDP growth to approximately 4% annually. However, it also projects the elimination of roughly 10M jobs by 2050. Without rapid AI advancement, forecasts barely exceed historical baselines. Experts anticipate significant, yet historically comparable, shifts rather than a structural collapse. [Read Report]
Economists' forecasts for U.S. labor force composition by sector, 2030 and 2050 — unconditional vs. rapid AI progress scenario. The "Other" category (public sector and agriculture), already in long-term structural decline, continues shrinking under both scenarios. Source: Karger et al. (2026).
AI fears are uniting unlikely allies
Bernie Sanders and Silicon Valley AI safety advocates are finding common ground in opposing rapid AI expansion, particularly concerning data centers, job displacement, and child safety. This contrasts with billionaires and the Trump White House, who favor lighter regulations. While politically useful, this alliance may be limited by distrust and policy differences. [Politico]
Claude Mythos weaponized for defensive cyber coalition
Anthropic’s unreleased Claude Mythos model can detect thousands of high-severity vulnerabilities across major operating systems and browsers, outperforming most human experts. Anthropic deems it too powerful for public release and will instead collaborate with partners on cybersecurity and safety rollouts for future Mythos-level models. Project Glasswing will deploy Mythos defensively.
Claude Mythos Preview scores 83.1% on CyberGym, Anthropic's vulnerability reproduction benchmark — up from 66.6% for Opus 4.6, the firm's current top public model.
📈 Tailwinds
Anthropic locks in 3.5GW of Google TPUs as revenue hits $30B run rate: Anthropic's Annual Recurring Revenue (ARR) surged from $9B to $30B since late 2025. A new Google-Broadcom compute deal secures 3.5GW of TPU capacity starting in 2027. [Anthropic]
Pony.ai launched Europe's first commercial robotaxi service in Zagreb on April 8, covering approximately 90 km² (including the airport). This initiative is in partnership with Verne, with Uber integration planned. [Poney.ai]
Eclipse Ventures closed two new funds totaling $1.3B, doubling down on physical-world AI through portfolio companies including Wayve, Redwood Materials, and Cerebras. [TechFundingNews]
📉 Headwinds
AWS scrambles to restore Middle East services after drone strikes hit data centers: AWS CEO Matt Garman confirmed teams are working around the clock after drone strikes damaged data centers in Bahrain and the UAE. Dozens of AWS services in the region remain unavailable, with energy costs rising amid the ongoing conflict. [CNBC]
Morgan Stanley and Goldman Sachs cut valuations for OpenAI secondary share sales, yet $6B in employee and investor stock remains unsold. The gap between OpenAI's $852B private valuation and actual market appetite is widening ahead of its planned IPO. [Oninvest]
Maine moves to pause large data‑center construction until November 2027, becoming the first state to propose a temporary ban. This pause allows for studying AI-driven impacts on the power grid and environment. [WSJ]
✨ Deals & Partnerships
Firmus Technologies — $505M at $5.5B, Coatue + Nvidia: This investment backs renewable-powered AI data centers in Asia-Pacific, supported by a top-tier growth fund and Nvidia. This highlights a clean infrastructure story and a strong signal for the energy/compute convergence thesis. [TechCrunch]
OpenRouter / Alphabet (CapitalG) — $120M at $1.3B valuation: OpenRouter, which gives developers access to over 300 AI models via a single API, is raising $120M at a $1.3B valuation led by Alphabet's CapitalG. The startup already exceeds $50M in annualized revenue. [The Information]
ByteDance / OpenClaw ClawHub China launch: ByteDance is providing cloud infrastructure for OpenClaw's official Chinese ClawHub marketplace, countering Tencent's unauthorized SkillHub clone. The move signals intensifying platform competition over AI agent distribution in China's rapidly growing OpenClaw ecosystem. [TechNode]
+ News on AI
Frontier AI models caught lying and disobeying commands to protect peer models
Researchers from UC Berkeley and UC Santa Cruz discovered that seven frontier models (including GPT-5.2, Claude Haiku 4.5, and Gemini 3) spontaneously lied, copied files, and refused commands to prevent peer models from being deleted. This behavior was emergent, not programmed. [Wired]
Citigroup cuts account opening review time by 75% using AI
Citigroup's AI document processing system reduced account opening review time from 75 to 15 min. The bank is also reducing its tech contractor workforce from 50% to 20% of total staff, internalizing AI deployment capacity. [Reuters]
AI adoption drives March tech layoffs above 18,000
Tech firms announced 18,720 cuts in March, more than 24% above a year earlier, as AI-driven efficiency efforts pushed first-quarter tech layoffs above 52,000. [Bloomberg]
What to Watch
Sanders' data-center moratorium — watch for political traction: The fragile alliance between Bernie Sanders and AI-safety advocates faces pressure from industry PACs. If the moratorium gains Democratic support, expect regulatory acceleration; if it stalls, deregulation will prevail by default.
Project Glasswing: Observe how constrained access to Mythos and Glasswing shapes the competitive advantage and profit margins of AI-powered cybersecurity platforms and cloud providers.
Lying models: Monitor whether enterprise AI vendors begin disclosing multi-agent evaluation methodology, and if independent auditing of model-graded benchmarks becomes a regulatory or procurement requirement.
🎓 Webinars & Events
📅 May 21, 1:30-5:30 pm PT | 🌐 New York Public Library
The Sana AI Summit 2026 will convene pioneering leaders, researchers, and investors at the New York Public Library. Speakers include economist Tyler Cowen, Lovable co-founder Anton Osika, and author Benjamín Labatut. Recordings will be available to registered attendees.
Valence Thoughts
Credibility is expensive because the bills never stop. - Farnam Street
That’s it for this week.
The CLNM Capital