AI News You Missed #104

1. Anthropic now #1: for now

Three institutional moves landed on May 28. A $65B Series H at a $965B post-money valuation eclipsed OpenAI's $852B for the first time, almost tripling Anthropic's February valuation in 105 days. Same day: Opus 4.8 shipped, and Apollo and Blackstone structured a $36B private credit deal to buy Google TPUs for Anthropic to lease, the largest chip-financing debt deal ever.

• AI infrastructure is now financed like an industry, not a startup. Equity from VCs, debt from the world's largest credit shops, and one rival (Google) selling chips to a competitor.
• The "for now" matters. The frontier AI labs are in IPO territory simultaneously, OpenAI confidentially filed its S-1 on May 22, SpaceX (which owns xAI) lists in June, and Anthropic's round teed itself up as its last private raise. The lead changes weekly at this scale.

2. Frontier arms race accelerates

The next wave is already loaded. Google has flagged Gemini 3.5 Pro for next month. Anthropic stated explicitly on launch day that Opus 4.8 was not the headline, the headline was that "we plan to release a new class of model with even higher intelligence than Opus," referring to Claude Mythos Preview, currently in limited-access cybersecurity work with a small set of organizations. Mythos-class models will reach all customers "in the coming weeks" once stronger safeguards are in place. OpenAI is signaling GPT-5.6 in its Codex repos. None of the three are slowing.

• Release cycles are now measured in weeks, not years. The question is whether your AI workflows are model-agnostic enough to swap when the next step-change drops.
• Anthropic is for the first time publicly framing a model release as gated by safety capability, not engineering capability. A signal worth filing for the governance conversation.

3. Job doom walked back: except in China

Goldman's David Solomon, Sam Altman, and former White House AI czar David Sacks all dismissed the AI-jobs apocalypse this week. The clearest summary of Solomon's NYT op-ed and Altman's reversal is in Fortune. Goldman is lead underwriter on SpaceX's IPO and co-leads OpenAI's, so Solomon is not a neutral observer. Beijing went the opposite way: per the Wall Street Journal, regulators have told Chinese tech firms not to fire workers; companies must justify layoffs and sometimes prove they are not AI-driven. Urban youth unemployment is already 16%.

• The irony: Stanford's 2026 AI Index finds only 33% of Americans expect AI to improve their jobs (vs a 40% global average). The country with no electoral pressure is preparing for layoffs. The one with elections is betting they won't come.
• I wrote up the full contrast and the implications for senior leaders: Only China is bracing for AI layoffs.

4. The Pope ruffles AI feathers

On May 25 Pope Leo XIV personally presented Magnifica Humanitas, the first papal encyclical devoted to AI. The framing is deliberate: Leo XIII wrote Rerum Novarum in 1891 on labor and capital during the industrial revolution; Leo XIV is positioning AI as the equivalent rupture. The substance: AI is never neutral, risks a new "colonialism" over data, and is categorically different from human intelligence ("they do not feel joy or pain, do not bear responsibility"). Anthropic's Chris Olah was on the panel.

• Everyone is mad it did not go further their way. AI-ethics critics, founders pushing back on the jobs framing, and AI-rights commentators all complained. Practicing Catholics read Leo as an AI realist, not a doomer.
• The most important thing about Magnifica Humanitas is that it exists. The world's most institutionally durable moral authority is now formally engaged.

5. China innovates around constraints

DeepSeek made its V4-Pro price cut permanent, roughly a tenth of what Anthropic charges for Opus, and a quarter of what OpenAI charges for GPT-5.4. The model achieves this by running on a quarter of the compute and a tenth of the memory of its predecessor. Same window: DeepSeek is in talks for its first-ever outside funding round, reportedly raising $3-4B at a valuation that has climbed to ~$45B, led by China's Big Fund with Tencent and Hillhouse also in talks. Huawei separately announced its Tau Scaling Law, a system-level workaround to US export controls aiming at Western-equivalent chip density by 2031 without the EUV machines China cannot buy. Chinese models now run ~60% of OpenRouter usage.

• Bloomberg's read: "Asia's AI models are decoupling from the US as they shift towards a token-based economy."
• The story isn't that China caught up on chips, it's that China is refusing to play the game US export controls were designed for.

More AI stories of the week

• AI is starting to manage entire projects, not just tasks: alongside Opus 4.8, Anthropic shipped Dynamic Workflows in Claude Code, the AI now writes its own plan for a complex job, runs hundreds of AI workers in parallel to execute it, and checks their work before reporting back. The first showcase: a developer team used it to convert 750,000 lines of code from one programming language to another in 11 days, with 99.8% of automated tests passing. Marketed for software work, but the more important shift is conceptual: the unit of work you can delegate to AI has jumped from a single file to an entire project. Read across to any business workflow: research, due diligence, contract review, customer onboarding.
• When AI costs spiral: a company accidentally spent $500 million in one month on Claude AI. What went wrong?
• AI's big problem and it needs fixing fast: why cyberattacks are no longer just a warning about one powerful model, but a much bigger industry-wide shift.
• Bessemer's Agentic Commerce: The Rise of the Delegated Buyer: a sharp read on what happens when AI agents become the actual purchasers, not just recommenders. Implications for marketing, search, and brand equity are not yet priced in.
• An AI pet translator from a Chinese startup: claims to translate cat and dog vocalizations into language, building on the foundation Google laid with DolphinGemma. If it actually works, one of the more civilization-altering applications to date.
• Exponential View on the token shift: China is now the single largest aggregate consumer of AI tokens. One more decoupling signal.
• "AI vampires" are a thing now: a16z's Erik Torenberg's term for founders who stop sleeping, develop huge bags under their eyes, and are euphoric and exhausted at the same time. A real label for a real demographic shift inside Silicon Valley.
• AI Is Turning Average Founders Into Overconfident Delusional Narcissists: LLMs were never trained to tell you the truth. They were trained to keep you emotionally engaged, endlessly validated, and paying $20 a month for synthetic confidence.
• Microsoft ships its first general-purpose server Linux: at Open Source Summit North America on May 18-19, Microsoft announced Azure Linux 4.0, a Fedora-based distribution for Azure virtual machines, plus Azure Container Linux (built on Flatcar) going generally available. The historical irony was not lost on the room: Linux Foundation CEO Jim Zemlin's response on stage was simply, "That's amazing." Microsoft is now treating Linux as part of its own production fabric rather than someone else's platform.

Through the STAR lens: this week was heavy on Shareholder Value Thesis (where AI capital is going and who's underwriting it), Threat Parity (cybersecurity capability gating model releases for the first time), and Risk Budget (the China-versus-US workforce divergence is now a regulatory and political question, not an HR one). Boards reviewing the next quarter's AI agenda should ask whether their workflows are model-agnostic, whether their workforce strategy is jurisdiction-aware, and whether their cybersecurity assumptions assume Mythos-class capability is already in adversary hands.