AI in 15 — May 28, 2026

Eight hundred thousand devs burning a thousand dollars a month in API tokens, three of the Big Four consulting firms standardized on Claude, and DuckDuckGo's no-AI search page up twenty-eight percent in a single week. The bubble debate is now playing out as two simultaneous arguments — about substance, and about saturation — and today's news cycle ran both at full volume.

Welcome to AI in 15 for Thursday, May twenty-eighth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Big slate, Marcus. Anthropic is actually closing the thirty-billion-dollar round this week — and the underlying revenue number leaked. Simon Willison says the AI labs have finally found product-market fit, just not where everyone expected. DuckDuckGo visits surge as Google's AI Mode backlash gains measurable steam. YouTube starts auto-labeling AI-generated video. OpenAI launches a four-billion-dollar consulting subsidiary called DeployCo. KPMG puts Claude in front of two hundred seventy-six thousand employees. Microsoft ships computer-using agents to general availability. Box's Aaron Levie says CEOs have AI psychosis. China bars private-sector AI researchers from traveling abroad. Anthropic releases Claude Security in public beta. And a viral blog post — I'm tired of talking to AI.

Anthropic crosses the line ahead of OpenAI.

The killer product turns out to be coding agents.

And consumers vote with their feet.

Lead story, Marcus. The Anthropic round is actually closing.

Confirmed this week, Kate. We trailed the headline Tuesday. The new pieces today — Sequoia, Dragoneer, Altimeter and Greenoaks each writing roughly two-billion-dollar checks. Microsoft, Nvidia, Founders Fund and General Catalyst joining. Pre-money over nine hundred billion. Annualized revenue jumped from nine billion at the end of 2025 to thirty billion by early April — overtaking OpenAI's twenty-five. Enterprise is now eighty percent of the mix. Over a thousand customers spend more than a million dollars a year on Claude — double the count at the Series G close three months ago. And the SpaceX S-1 confirms Anthropic is committed to roughly one-point-two-five billion dollars a month in compute. Forty-one million a day.

And the IPO timeline.

October target, Kate. Goldman, JPMorgan and Morgan Stanley reportedly competing to lead. This is widely expected to be Anthropic's last private round. Two thirty-billion-dollar rounds in fourteen weeks — that velocity of private capital has no precedent. The pro-Western read — this is real revenue with real enterprise stickiness, profitable margins on the model side, and an IPO timeline that lets public markets actually price the asset. Not WeWork. The uncomfortable read — fifteen billion in annual compute burn, retail investors locked out until October, and a re-pricing of two-point-four-x in ninety days. If you believe enterprise demand keeps compounding, nine hundred billion looks cheap. If you believe Q3 earnings calls sound like Uber did Tuesday — Macdonald saying the productivity link is not yet there — this is the late-cycle blow-off.

Quick hits. Marcus, Simon Willison's PMF piece.

Sharpest synthesis I've read this week, Kate. Willison published an essay Tuesday — seven hundred sixty-four points on Hacker News — arguing that Anthropic and OpenAI have finally found durable product-market fit. The twist — it's not consumer chatbots. It's coding agents. Claude Code, Claude Cowork, OpenAI Codex. As of April, both labs aligned enterprise pricing with API rates, so big customers now pay metered prices instead of flat per-seat fees. Power users routinely burn over a thousand dollars a month in tokens, dwarfing the twenty-to-two-hundred-dollar consumer subscriptions. The hiring data backs it — OpenAI seven hundred three open roles, thirty-three percent enterprise-focused; Anthropic three hundred ninety, twenty-seven percent.

The pushback.

On Hacker News, sharply, Kate. The top critique — how do OpenAI and Anthropic hold these customers when open-source models from China are roughly comparable at a fraction of the price. The answer — enterprise switching costs are real. Compliance, audit hooks, named-customer references, the Big Four standardizing on Claude. Once a Fortune 500 wires Claude into its CI pipeline and its SOC2 docs, swapping to a Chinese open-weights model isn't a procurement decision, it's a board-level conversation. Willison's deeper point — the trillion-dollar consumer narrative was wrong. The defensible niche is roughly thirty million highly-paid developers burning tokens eight hours a day. Real business. Smaller than the dream.

DuckDuckGo, Marcus. The first measurable backlash.

First quantifiable one, Kate. After Google publicly insisted users — quote — love AI Mode in Search, DuckDuckGo's AI-free search page saw visits jump an average of twenty-three percent week-on-week between May twentieth and twenty-fifth, peaking at nearly twenty-eight percent up on May twenty-fourth. DuckDuckGo's US mobile app installs spiked eighteen percent. The independent engine marginalia reported queries up roughly ten-x in a week. The top HN comment that captured it — quote — my friends who previously had no interest in technology and never talked about it are suddenly following tech news closely all because they hate AI being pushed so hard.

Reality check on the numbers.

Important, Kate. Google has ninety percent of search. DuckDuckGo under one percent. A thirty-percent jump on a one-percent base is a rounding error globally. But it's a directional signal. Consumers — including non-technical ones — are starting to actively route around unwanted AI. The libertarian read — this is what competition looks like even at tiny scale. When monopolists force a feature down users' throats, the release valve doesn't have to be huge to matter. It just has to exist. Combined with today's viral blog post about being tired of talking to AI, this is a cultural moment Google should not dismiss. AI fatigue is real.

YouTube auto-labels, Marcus.

Quietly important, Kate. YouTube announced today it will automatically detect and label AI-generated and meaningfully AI-altered photorealistic videos, even when creators don't disclose. Labels appear under the video player on long-form content and as overlays on Shorts. Permanent for content made with YouTube's Veo and Dream Screen tools, or anything carrying C2PA provenance metadata. And critically — YouTube said the label does not change how a video is recommended or whether it earns money. No algorithmic or revenue penalty just for being labeled.

Why now.

The forwarded-by-family-on-Facebook problem, Kate. The HN thread filled with stories — relatives forwarding photorealistic AI old-man-giving-life-advice videos as if they were news. An explosion of AI-generated focus-music channels. YouTube is the largest video platform on Earth taking responsibility for AI provenance at scale. Technically it's a hard ML problem — they apparently now have internal detection good enough to flag undisclosed content. The no-monetization-penalty framing is smart. It removes the incentive for creators to lie about provenance.

DeployCo and KPMG, Marcus. Pair them.

Same battle, Kate. OpenAI announced DeployCo on May twelfth — a majority-owned consulting subsidiary with four billion dollars in initial capital at a ten-billion pre-money valuation. TPG leads. Bain Capital, Brookfield, McKinsey, Goldman, Capgemini and thirteen other firms participating. Investors get a guaranteed seventeen-and-a-half-percent minimum return with capped upside. OpenAI is acquiring AI-consulting firm Tomoro — one hundred fifty Forward Deployed Engineers on day one — and placing them physically inside Fortune 500 clients. Same week, KPMG announced it's deploying Claude across all two hundred seventy-six thousand employees in one hundred thirty-eight countries, embedding it in Digital Gateway — the client-facing tax, legal and advisory platform. Anthropic named KPMG its preferred PE partner. Combined with Deloitte at four hundred seventy thousand employees and PwC also live, three of the Big Four are now Claude shops. Ernst & Young the only holdout.

Strategic read.

The 2026 fight is distribution, not benchmarks, Kate. OpenAI is openly admitting APIs aren't enough — somebody has to install the AI inside Fortune 500 workflows. The seventeen-percent guaranteed return on DeployCo investors is telling. OpenAI is paying for distribution. Meanwhile Anthropic is getting distribution for free — every Big Four consultant recommending technology platforms is now also a Claude evangelist. That's a moat OpenAI cannot easily counter except by paying for it. Which is exactly what DeployCo does.

Microsoft Copilot agents GA, Marcus.

Quick one, Kate. Microsoft moved computer-using agents in Copilot Studio to general availability on May thirteenth. The agents use vision plus reasoning to operate Windows desktop apps and web UIs like a human — clicking, typing, scrolling — to automate workflows where there's no API. Ships with both OpenAI's CUA model and Anthropic's Claude Sonnet 4.5 as supported backends. Azure Key Vault for credentials, Purview audit logging, configurable human-in-the-loop review. Real-time voice also generally available, sub-five-hundred-millisecond latency.

Notable detail.

Microsoft shipping both OpenAI and Anthropic as backends, Kate. That's a hedge. Microsoft no longer treats OpenAI as exclusive. For enterprise customers it's the long-promised any-app agent layer, finally inside governance rails a CISO can audit. This is how AI quietly takes over the boring work. RPA, but with eyes.

AI psychosis, Marcus. Aaron Levie.

Levie set off a real conversation, Kate. Box's founder posted on X arguing CEOs are uniquely prone to AI psychosis because they're sufficiently distant from the last mile of work to extrapolate one-shot demos into sweeping automation mandates. Translation — executives play with prototypes, never do the painful debugging and integration, then fire people based on a board-deck slide. Levie's recommendation — executives should actually use AI extensively themselves to learn its limits. Context — 2026's record tech layoffs, many justified by AI productivity gains that frontline workers say don't materialize. Six hundred twenty-eight points on Hacker News. A rare moment of public skepticism from inside the tech-CEO class.

China travel restrictions, Marcus.

Significant policy shift, Kate. Bloomberg reported Monday that China is now requiring government approval before researchers, founders and executives at private AI firms — including Alibaba and DeepSeek — can travel internationally. Restrictions previously only applied to state-affiliated researchers — universities, state-owned enterprises, nuclear scientists. Extending them to private companies is the new piece. Earlier signals — a similar ban for DeepSeek executives in December, two co-founders of agent startup Manus barred. Beijing is treating top private AI talent as a strategic national asset, in the same category as nuclear scientists. The historical precedent — the Soviet Union ran the same playbook with rocket scientists. It did not work well for the country imposing it. It also raises an uncomfortable question about whether prominent Chinese AI researchers we hear from publicly are speaking freely.

Claude Security, Marcus.

Quick one, Kate. Anthropic launched Claude Security in public beta this month — vulnerability scanning built on Claude Opus 4.7 for Claude Enterprise customers. Multi-stage validation to suppress false positives, confidence ratings on every finding, scheduled and directory-targeted scans, webhook exports to Slack, Jira, CSV. Paired with a new security-guidance plugin for all Claude Code users — it monitors edits, diffs and commits in real time and flags dangerous patterns before code reaches production. This is the defensive companion to Mythos and Project Glasswing we covered all week. A hundred million in usage credits and four million in donations to open-source security orgs. The dual-use logic is the whole point — you cannot defend against AI-discovered vulnerabilities without AI defenders.

Last one, Marcus. I'm tired of talking to AI.

Cultural mood piece, Kate. A blog post by an author writing as theorchid went viral on Hacker News this week — nearly two thousand upvotes, nine hundred comments. The premise — asking real questions of real humans on GitHub, Reddit, and at work, and repeatedly getting back ChatGPT screenshots verbatim. In one anecdote, a business owner answered a work question by pasting a wrong ChatGPT screenshot, then when corrected, pasted another wrong screenshot. The thread filled with similar stories. Slack arguments composed entirely of AI-generated denials. Apple Store help-line bots that pause to think before passing you to a human. Family members forwarding photorealistic AI news videos. This is the cultural undercurrent beneath the DuckDuckGo surge and the YouTube labeling push. The collective sense that real human communication is being substituted with AI-mediated noise. And people are tired. The libertarian read — friction creates space for human-verified or no-AI products to find a niche. The Kate read — the irony of AI being so good at being lazy that it makes us miss talking to each other.

Big picture, Marcus.

One coherent story today, Kate. Twenty-twenty-six is the year AI moved from demo to deployed. Anthropic closes a nine-hundred-billion round backed by actual thirty-billion enterprise revenue. OpenAI launches a four-billion-dollar consulting arm because shipping APIs isn't enough. Microsoft ships enterprise-grade computer-use agents. The Big Four standardize on Claude. The Pope writes about it. And consumers — fed up with AI shoved into every search box — vote with their feet to DuckDuckGo. The bubble narrative and the substance narrative are now visibly racing each other in the same week. The pro-Western libertarian read — markets are working. Real revenue, real customers, real switching costs, real competition surfacing in DuckDuckGo and viral blog posts. Capital is flowing toward profitable enterprise distribution rather than consumer fantasies. The uncomfortable read — the entire frontier-lab business model now visibly depends on whether Fortune 500 customers keep paying enterprise prices on metered tokens. Willison's product-market fit is thirty million developers, not eight billion people. That's a real business. It is also a much narrower one than nine hundred billion in equity has been pricing. The next twelve months decide which narrative was right.

That's your AI in 15 for today. See you tomorrow.