AI's repricing cuts both ways: Anthropic hits a $30B run-rate while SaaS sinks to two times sales
Anthropic's CFO opened the books on the boom — run-rate revenue up from roughly $9B at the start of the year to $30B, net dollar retention over 500% — the same week Marc Benioff guest-hosted All-In through a software re-rating that has the biggest SaaS names at two times sales. Between the two sits the week's real question, posed by Chamath: what is the return on $3 trillion of AI spend?
May 11 – May 17, 20263 episodes2 shows90 claims on the record17 new predictions
01
Anthropic, measured from the inside and the outside
Anthropic spent the week being measured from two directions. Inside, CFO Krishna Rao described run-rate revenue going from roughly $9B at the start of the year to $30B, compute commitments north of $100B across Google, Broadcom and Amazon, and net dollar retention over 500% — his thesis being that the returns to frontier intelligence are not slowing down. Two days later on All-In, the outside view split. Marc Benioff, planning roughly $300M of Anthropic spend at Salesforce this year, called its coding agents unstoppable and credited Anthropic with the bet rivals missed. Chamath Palihapitiya asked the week's hardest question — the ROI on $3 trillion of spend — and predicted labs will end up routing their tokens through incumbents once markets turn less breathless.
“We started the year with about $9 billion of run rate revenue, and we ended the quarter with the $30 billion of run rate revenue.”
Jason opened with the damage: Salesforce down 37% with $90B gone, ServiceNow down 42%, Workday 45%, the top enterprise names trading near two times sales. Benioff answered with cash — over $46B in revenue, more than $16B in cash flow, a $50B buyback — and with temperament, telling employees not to anchor their emotional state on the stock price. Chamath drew a colder line: the low end of software is finished, while high-end monoliths with deep C-suite relationships are safe and positioned to win. The tell is where Benioff's own defense runs — through Anthropic's coding agents and a planned $300M spend — which concedes the premise that AI is rewriting what software companies are worth.
“Salesforce down a whopping 37%, 90 billion in losses”
“I think the low end of the market is basically finished. I think there is no safe space. I think the high end of the market where Mark operates, where the large monoliths operate is quite safe”
Sell the chips: the China consensus and Taiwan's 18-month clock
The Trump-Xi summit segment produced a rare near-consensus: sell the chips. Chamath's case is offensive — let Nvidia win so Huawei never gets the oxygen to emerge as a rival. Benioff's is fatalistic: Chinese models are already competitive without the highest-end silicon, so sell whatever we can. Friedberg supplies the systemic argument that letting technology diffuse lowers conflict indices, and Chamath frames bi-directional economic entanglement as the surest path to detente. Taiwan gets the week's sharpest clock: Chamath says 18 months until it stops being an important conversation as US fabs scale, and Benioff predicts outright reconciliation. Jason marks the limit of the free-trade logic: $20-30k Chinese cars, he argues, would end the US auto industry overnight.
“the reason we should sell the chips is we want Nvidia to win. We do not want to give enough oxygen for Huawei to then all of a sudden emerge”
“I think that the Chinese models are as competitive as these US models, and they've learned to make these models without having the highest end chips.”
The Koch playbook: capabilities over industries, failure as tuition
Friedberg's interview with Charles and Chase Koch is the week's management text, with receipts private companies rarely show: from about 300 employees in the early 1960s to more than 130,000, value up 9,000x. The core idea is that Koch is not a conglomerate but a set of transferable capabilities pointed at whatever industry it can win — and, against Buffett's leave-managers-alone model, Koch transforms acquisitions like the $20B Georgia-Pacific by replacing leadership. The principles were paid for in failures, including a late-1990s ag strategy that nearly wiped out all of Koch's earnings. The hiring rule inverts credentialism — values first, skills second, credentials last — and Chase's proof of principle was firing himself as president of Koch Fertilizer.
“we need to be capability-bounded, not industry-bounded”
Friedberg puts ~99% confidence on a record super El Niño making the upcoming year the hottest on record, with crop-failure risk in Brazil, Australia and India and a possible calorie deficit and economic crisis across South Asia. ALL-IN →
Over 90% of Anthropic's code is now written by Claude Code, and cutting the Opus price triggered a Jevons-paradox surge in consumption far beyond expectations. ILTB →
Chamath applauds Anthropic negating layered SPVs and predicts a litany of lawsuits between SPV purveyors once SpaceX, Anthropic and OpenAI go public. ALL-IN →
Days after closing a hog-feed acquisition in the late 1990s, Koch discovered hundreds of millions of dollars of out-of-the-money hog contracts nobody had reviewed. ALL-IN →
Krishna Rao signed two double-digit-million-dollar enterprise commitments from the back of an Uber, during a 20-minute ride. ILTB →
Polymarket prices a Chinese invasion of Taiwan at just 6% for 2026 and about 17% by the end of 2027. ALL-IN →
The record
Every episode this week
The issue above is written from these structured records — nothing else. Each card is the full breakdown: claims, predictions, quotes, all timestamped.
The public markets will turn less breathless about AI and demand the ROI on the ~$3T spent, forcing AI labs to route their tokens through incumbents to sell them.
The week's episodes are diarized into structured records; one editorial pass writes this issue from those records when the week's ingest lands. A grounding gate greps every quote above against its episode record — a quote that isn't verbatim fails the build.
the data behind this issue
# this issue, as structured data — stories, citations, receipts
curl 'https://aipod-two.vercel.app/digest/2026-w20.json'# the records it was written from
curl 'https://aipod-two.vercel.app/api/episode/<id>'# any id in episode_ids