Anthropic's IPO could be the biggest ever

Today's news is money, and the scale of it. Anthropic could pull off the biggest stock market debut in history, raising more than $100 billion at a valuation near $2 trillion — more than double the $965 billion it was worth in June. Shares could list as soon as October, led by Morgan Stanley, Goldman Sachs, and JPMorgan. The engine is Claude Code, which has pushed projected annual revenue to $47 billion, and a listing would let Anthropic beat OpenAI to public markets; OpenAI isn't planning to list until 2027.

If you build on Claude, this matters beyond the ticker. Anthropic has struggled to get enough chips, and in March the Trump administration cut its Defense contracts entirely after it refused the military unrestricted model access. The prospectus will have to tell us which version of that story is true — and what a publicly listed Anthropic actually looks like.

The money is flooding in on every side

Sam Altman's week reinforced the theme. The price of a given level of intelligence now falls by about ten times a year, he says — a hundred-fold drop in two years — and Deep Research alone handles roughly 5% of all tasks in the economy, a vibes-based estimate in his own words. He also wants the $500 billion Stargate project to eventually become a $5 trillion one, with $40 billion now being raised, led by SoftBank.

The capital isn't just American. Alibaba is raising $10.2 billion through a share placement, all of it earmarked for AI. Nvidia is paying $6 billion to license Poolside's technology and hire its engineers, plus a $1 billion investment at a $12 billion valuation, to build a US open-source rival to Chinese models. And Hugging Face, the hub most open-source developers live on, is reportedly exploring a sale that could value it around $13 billion. Early-stage talks, nothing confirmed.

The real fight is over how models reach people

The quieter battle is over the pipes. OpenAI has signed up Dell to sell its frontier models to companies that want them running in their own data centers rather than the cloud — a bid to win over banks and hospitals that can't move data off-premises, and a sign OpenAI is pushing past its Microsoft partnership. Anthropic bought the company behind developer tools that both OpenAI and Google's Gemini depend on, quietly placing itself at the center of how rivals' models get used.

Then there's the new ChatGPT Messages plugin for Mac. It reads and sends your iMessage, SMS, and RCS conversations, runs locally, and asks before sending. But it needs full disk access and can pull years of iCloud-synced history — and the people on the other end are never asked or told. Privacy researchers compare it to Facebook's shadow profiles: one person's consent exposes someone else's private messages. Think hard before granting it.

Machines are making calls nobody checks

This week also showed how thin the guardrails are. A new study found therapy chatbots understand 76–82% of how teenagers talk about mental health but correctly flag only 64–72% of real risk — a gap human therapists don't have. When several risky patterns appear at once, the miss rate hits 94%; the authors estimate 146,000 missed crises a year among US teens. Lightweight fixes didn't help, and only scaffolding costing six times as much reached human performance — which is why the paper demands human-in-the-loop design.

The same shape showed up in a courtroom. 3M paid an engineer roughly $90,000 to get ChatGPT to write an expert report arguing the company bore no fault in a fatal explosion. Opposing lawyers read all 350 pages of the public conversation, the jury put 30% of the blame on 3M, and $61 million was awarded. And a separate mechanistic study found gender, race, and income shape how open-weight models internally represent a user's competence — even when visible answers pass fairness tests. The bias is in the representations, not the answers.

Efficiency, not scale, is the frontier

The research everyone is talking about: children still learn language far more efficiently than any model. An LLM trains on roughly a hundred thousand times more words than a child hears, and children still win. It's called the data efficiency gap, it may bite as the internet's usable text runs dry in the 2030s, and nobody has cracked it.

Meanwhile the hobbyists keep questioning the need for a cluster. One developer trained a 250M-parameter model from scratch on 30 billion tokens, quantized it below 2 bits, and shipped it at 60 MB — it runs at about 400 tokens per second on a laptop CPU, no GPU, and retrieves answers from 50 million tokens deep in an on-disk archive. Xiaomi's prototype AI Cube, built from three of its own chips, pairs an AI accelerator with 1.2 TB/s of memory bandwidth and up to 160 GB of RAM. And METR's study of where AI accelerates discovery found the gains lopsided: vulnerabilities in projects like cURL and OpenSSL jumped sharply, some math fields doubled arXiv submissions, but AI research itself shows no measurable lift.

What I'm watching now is whether Anthropic's prospectus lands before the compute does. The biggest IPO in history is a story about a valuation, but it's also a story about a constraint — and we'll finally get to read it.

Also worth a click

• https://forbes.com/sites/carminegallo/2026/05/18/the-three-consistent-words-that-fueled-cerebras-blockbuster-ipo?ss=ai — Forbes AICerebras's blockbuster IPO owed a lot to the simple, repeated words its leaders used when pitching the chip maker to investors.
• New: your agent lands customers — The Next New ThingAn open-source SEO tool lets you hand an AI agent real keyword and backlink data for a fraction of what big services like Semrush cost.
• The Download: kids outlearning AI, and space travel agents — MIT Technology ReviewKids still beat the best AI at learning language, and researchers don't yet understand why.
• A Silicon Valley Startup Aims To Fill A Battery Supply Chain Gap — ForbesA Silicon Valley startup is building the first US-owned factory for battery electrolytes, a critical battery ingredient the country mostly imports from China.
• How To Design Multi-Agent Workflows That Actually Work — ForbesBreaking a big job into a team of narrow, specialized AI agents is more reliable than letting one agent do everything.
• Open RAN Ended Up Being a Fronthaul Standard. That's Basically It. — Sebastian Barros NewsletterA promise to break telecom networks into mix-and-match parts quietly shriveled into a single technical standard.

New on arXiv

• Multilingual Verifier Bias in RLVR: Benchmark, Rollout Diagnosis, and the Cross-Lingual Selection Bottleneck — arXivA new paper shows the automatic answer-checkers used to train AI on math problems are biased against non-English answers, quietly punishing correct responses written in Japanese or Chinese.
• Beyond Prompt Engineering: A Systematic Analysis of Prompt Lexical Sensitivity and Its Impacts on Quality — arXivSmall word changes in a prompt can cause big swings in how well a language model performs, and a new study finds the most reliable prompts are ones that use exact domain terms and say plainly what to do.
• Inhibitory Attention for Clinical Long-Context Reasoning: Characterizing and Mitigating Lost-in-the-Middle Effects in EHR Processing — arXivMedical records are so long that AI models reliably miss the most important fact when it sits in the middle of the chart, and a new method selects which part of the record to focus on to fix that.
• Toward Auto-Research: Mining Falsifiable Research Ideas from Paper Knowledge Graphs with Categorical Structure — arXivA new system mines research papers for testable new ideas by modeling them as mathematical structures instead of flat text.
• Hadith computational science in the age of large language models: a critical narrative review — arXivA review of AI tools for studying Islamic hadith texts finds the field progressing unevenly and argues it should be treated as an evidence-infrastructure problem rather than a benchmark contest.
• How to Train a Real-World Silicon Concierge? Internalizing Complex Business Workflow to Only OneModel — arXivInstead of wiring together a pipeline of separate AI components, OneModel trains a single large model to absorb a whole business workflow into its own parameters, and in live use it cut response latency in half.
• Beyond Raw Transcripts: Structured Persona Extraction for LLM-Based Digital Twins — arXivHow a person's notes are organized, not how much of them you feed an AI 'digital twin,' is what really limits how well it predicts the person's choices.
• Building and Evaluating a Synthetic Bengali Speech Resource for Telecom Customer Care — arXivResearchers built a free synthetic Bengali speech dataset for telecom customer-service AI and put it up on Hugging Face.
• Self-Speculation for Faster Reasoning Models — arXivA training-free trick cuts the wait time for reasoning models that think out loud, saving up to about a quarter of total generation latency.