Weekly Intel - 2026-05-10

The theme this week is ambition outrunning its container: AI models diagnosing patients better than doctors, meme-stock companies making $55B acquisition bids, coding agents scaling across entire fields. Everyone’s acting like the constraints that used to govern their domain no longer apply, and in some cases they might be right.

Markets

GameStop makes $55.5B takeover offer for eBay Ryan Cohen is pitching a $55.5 billion acquisition of eBay using an $11.9 billion company and $20 billion in debt, a structure that would immediately saddle the target with the acquirer’s leverage. The strategic thesis is that eBay’s brand is underleveraged and $2 billion in cost cuts (mostly sales and marketing) would unlock value, positioning it as a real Amazon competitor. This is less a conventional M&A play and more a bet that Cohen’s meme-stock credibility can will a deal into existence despite the math working against it. What eBay’s board does with it, treat it as a serious proposal or a pressure campaign to surface other buyers, is what determines where this goes.

AI Industry Moves

Higher usage limits for Claude and a compute deal with SpaceX Anthropic just locked up the entire compute capacity of SpaceX’s Colossus 1 data center, and the immediate payoff is doubled rate limits for Claude Code and significantly higher API limits for Opus models. AI companies are now competing on infrastructure access as aggressively as they compete on model quality, because the bottleneck has shifted from capability to availability. If you’re building on Claude’s API or have teams using Claude Code heavily, the capacity constraints that forced workarounds are loosening. Whether this signals a stable new baseline or just the opening move in a compute arms race is what’s worth watching.

AI & Software

OpenAI’s o1 correctly diagnosed 67% of ER patients vs. 50-55% by triage doctors A Harvard trial published in Science gave OpenAI’s o1 and pairs of ER doctors the same electronic health records for 76 emergency patients. The AI diagnosed correctly 67% of the time versus 50-55% for the physicians. The setting matters: real patient data in the highest-stakes, most time-compressed clinical environment there is, not a synthetic benchmark. AI is pulling ahead fastest where decisions rely on integrating structured data under pressure, which describes a lot more than medicine. If a 12-to-17-percentage-point accuracy gap holds up in larger trials, the liability calculus for not deploying these systems starts to flip, and every health system, insurer, and employer with self-funded plans will need a position on that.

AlphaEvolve: Gemini-powered coding agent scaling impact across fields Google DeepMind’s AlphaEvolve has moved past research prototype into production infrastructure. It’s now designing circuits for next-gen TPUs and optimizing core systems like Google Spanner, producing solutions so counterintuitive that human engineers wouldn’t have proposed them. AI systems that improve the substrate on which AI itself runs create a compounding loop that’s difficult for competitors without equivalent scale to replicate. Nobody knows yet whether this recursive self-improvement in infrastructure represents a durable moat or an accelerant that eventually commoditizes, but the compounding dynamic is already visible.

Agents for financial services and insurance Anthropic is shipping ten pre-built agent templates targeting the grunt work of finance (pitchbooks, KYC screening, month-end close), designed to deploy in days through Claude Cowork, Claude Code, or their Managed Agents platform. The more consequential move is the Microsoft 365 integration, where context persists across Excel, PowerPoint, Word, and Outlook without re-prompting. That’s the difference between a chatbot and a workflow layer. AI vendors are racing to become the connective tissue between enterprise data sources and the applications people already live in, and the firm that locks in those integrations first owns the switching costs.

Agents can now create Cloudflare accounts, buy domains, and deploy Cloudflare now lets AI coding agents provision accounts, start paid subscriptions, register domains, and obtain API tokens on behalf of users: no dashboard visits, no copy-pasting credentials, no manual payment entry. The human stays in the loop only for permission grants and terms of service. This is Cloudflare positioning itself as the default infrastructure layer for agentic workflows, and the real signal here is what it implies about distribution: if agents increasingly choose where to deploy code, the competitive moat shifts from developer experience to agent experience. Watch whether your cloud vendors are building for the human buyer or the autonomous agent that will increasingly act on their behalf. Those are different products.

A recent experience with ChatGPT 5.5 Pro Timothy Gowers, a Fields Medalist and one of the sharpest mathematical minds alive, reports that ChatGPT 5.5 Pro produced a piece of PhD-level mathematical research in about an hour, with no serious intellectual input from him. Gowers is describing something specific: novel research at doctoral level, not textbook problem-solving or literature retrieval. Each time the AI community crosses a threshold experts said was years away, the goalposts move quietly and the implications land loudly. If a model can now do in an hour what takes a doctoral student months, for any organization investing in R&D, technical hiring, or knowledge work: assume the reshape is coming, plan around the timeline.

Privacy & Governance

US healthcare marketplaces shared citizenship and race data with ad tech giants Bloomberg found that nearly all 20 state-run health insurance marketplaces were leaking applicant data (citizenship status, race, incarceration history of family members) to Google, Meta, LinkedIn, Snap, and TikTok via misconfigured tracking pixels. No hack, no breach in the traditional sense. The predictable outcome of dropping ad tech instrumentation onto pages that collect sensitive information without anyone auditing what those pixels actually capture. If you run any organization that embeds third-party trackers on authenticated or form-heavy pages, assume you have this problem until you’ve actually verified otherwise.

Financial Markets

Grand Theft Oil Futures: Insider traders keep making a killing at our expense Krugman documents what’s becoming a pattern impossible to ignore: massive, precisely timed bets in the oil futures market, in this case $920 million in shorts placed 70 minutes before an Axios report on a US-Iran deal, netting roughly $125 million as crude dropped 12%. Sophisticated analysis doesn’t produce positions like that, and neither does lucky timing at 3:40 AM. The signature points to advance knowledge of policy announcements. For anyone with energy exposure or supply chain sensitivity, oil prices are looking increasingly like a policy insider’s playground. That changes how much you can trust any forward pricing signal you’re using to plan.

Legal & Regulation

Zuckerberg ‘Personally Authorized and Encouraged’ Meta’s Copyright Infringement Five publishers and author Scott Turow are suing Meta and Zuckerberg personally, alleging he directly authorized the mass torrenting of copyrighted books and articles from pirate sites to train Llama. The legal theory is notable: it’s not just a corporate liability claim but an attempt to pierce the CEO shield by naming Zuckerberg as a defendant who encouraged the infringement. If that framing gains traction, it reshapes the risk calculus for every executive greenlighting AI training pipelines, and whether “move fast and break things” turns into personal exposure at the top.

That’s what I’m watching. What caught your attention this week?

-Eric