An AI found three drugs in a 12,456-compound haystack. Three IPOs walked to the public markets in the same week.
NIH NCATS published in Nature Medicine yesterday: a deep-learning model trained on 20 years of failed trials, animal studies, and molecular-interaction data surfaced three repurposing candidates for Alzheimer’s and Parkinson’s at 98% preclinical-screening accuracy — finerenone, originally a heart-failure drug, dropped amyloid plaque 42% in a mouse Alzheimer’s model. Plus the week’s IPO trajectory hardened to a triple-wave: SpaceX public S-1 under SPCX (May 20, $1.75T, June 12 Nasdaq target), OpenAI confidential S-1 (May 22, $852B), Anthropic October window holds.
Saturday morning. The week’s cleanest AI-doing-science result lands in Nature Medicine on the way out. The NIH’s National Center for Advancing Translational Sciences trained a deep-learning model on twenty years of failed clinical trials, animal studies, and molecular-interaction data, then screened 12,456 compounds and surfaced three repurposing candidates for Alzheimer’s and Parkinson’s. Screening accuracy: 98%. The standout: finerenone, originally approved for heart failure with preserved ejection fraction, reduced amyloid plaque by 42% in a twelve-week mouse Alzheimer’s model. The model architecture is unremarkable; the corpus shape is the result. Training on failure surfaced what twenty years of trial-by-trial human iteration missed. 12
The week’s other story hardened to a triple wave by Friday evening. SpaceX filed publicly with the SEC on May 20 under the ticker SPCX — Nasdaq debut targeted for June 12 at a $1.75 trillion valuation and a $75 billion raise. Lead underwriters: Goldman Sachs, Morgan Stanley, BofA, Citi, JPMorgan. Inside the S-1, the AI-infrastructure exposure is direct — xAI absorbed in February, the $40B+ four-year Anthropic compute commitment surfaced as a related-party contract, Starlink at a roughly $20B revenue run rate. Two days later, OpenAI’s confidential S-1 filing hit the SEC at an $852B private valuation with Goldman and Morgan Stanley running the deal. And Anthropic’s October window — projected against the WSJ-surfaced Q2 $10.9B and $559M first operating profit — holds. Three of the largest AI-infrastructure names walked toward the public markets in the same eight days. The S-1 risk-factor sections this fall will be a literature of the curve. 3415
Underneath the capital story is a structural buy-side anchor that landed quietly mid-month. The Ramp May 2026 AI Index — Ramp processes corporate-card spend across more than 50,000 U.S. businesses — flipped: Anthropic at 34.4% of business adoption, OpenAI at 32.3%, first crossover since the index began. Anthropic has quadrupled adoption in a year while OpenAI added 0.3 percentage points. The engine is Claude Code, which Ramp’s commentary pegs at four percent of all public GitHub commits. The IPO discipline question now reframes: OpenAI lists into a market where the enterprise daily-driver is no longer obviously theirs. Anthropic’s October window has a buy-side proof point — the Q2 $10.9B is the financial expression of the Ramp curve. 679
Carry into the weekend. The June ship calendar holds: Gemini 3.5 Pro confirmed (Sundar’s I/O line was “give us until next month”), GPT-5.6 rumored, Sonnet 4.8 source-map-leak pending. If two of three ship in the same window, frontier release velocity reprices. Figure F.03’s continuous-operation livestream is past 48 hours with no logged failure — Adcock’s bet is endurance first, throughput-rate disclosure later. Watch the rate cells when they land. SP-Index +1 to 75 on AI-doing-science. Capital substrate flat at its new triple-wave high; everything else holds. 121013
The Saturday read: a primary scientific result and a primary capital event in the same week, with the buy-side anchor already in the ledger. The Karpathy hire makes more sense at the start of Q3 than at the end of Q2. The S-1 risk-factor language is the next narrative spine. By the time the unredacted prospectus surfaces in late July or August, the curve will have moved across at least one more frontier release window — and the disclosures will have to catch up. 136
Three filings, one paper, one duopoly. The frontier looks different on a Saturday morning when the science lands first.
