The Five Year Desert to Product Market Fit & a $5.3BN Valuation with Shiv Rao, Founder @ Abridge

Summary

• The company’s five-year desert (founded 2018, three months after the Transformer paper; hot only since 2023) distills to one survival rule: hold the thesis, pivot everything else. Shiv Rao was willing to pivot product, go-to-market, and business model, “but I wasn’t willing to move on the thesis” — that clinical conversations are “the most human signal” powering healthcare. Until the market opens, “you just need to not die” — be standing when the sky opens.
• On the existential question every vertical AI founder gets: “If you are fighting against them, you’ve already lost.” If foundation-model tailwinds aren’t yours to leverage, you’re screwed. The company’s defense is depth — a regulated, $5.3T industry (18-19% of US GDP) where post-training on daily user edits across doctor types, settings, and languages is the product: “people in 2023 thought that was last mile. That’s actually most of it. The model piece is much less.”
• ~40% of the company’s in-product model outputs come from in-house models — possibly 60% next month as distilled open-source models replace frontier calls. The rule: binary, ring-the-bell tasks go in-house (faster, cheaper, set-and-forget); never-perfect tasks where the market rewards continuous improvement “ride the frontier wave.” Owning the lower stack gives “agency on the P&L” — a lever he doesn’t need now but “my future self might care more.”
• The healthcare GTM trap: it’s not one $5.3T market, and founders who take the “start down-market” advice literally never time their “YOLO shot” up-market — 800K practicing doctors are concentrated in large delivery networks. The company’s opening came in 2023 when burnout (40-50% of doctors; a JAMA study saying 30% of nurses wanted out within 12 months) met LLMs: “when the sky opened, we ran right in.” Now doctors reportedly refuse to join hospitals that don’t have “a bridge.”
• The note was never the product — the note is a bill. Clinicians “are compensated for the care that they documented that they deliver,” so the company sequenced notes → orders → billing along jobs nobody wanted, threading CMIO, CIO, and CFO. Result: Microsoft/Nuance (a $20B+ acquisition) and likely Nabler “we don’t see either of those companies much anymore” — not through bundling defense but category creation; Emory doctors now use “a bridge” as a verb.
• OpenAI and Anthropic hiring forward-deployed engineers and partnering with PE firms is, to Rao, a clear bull signal for vertical AI: “If that wasn’t a sign that there is absolutely an incredible opportunity… I don’t know what is.” The mess is the moat — “it’s not SOC 2, it’s HITRUST,” behavioral-health data handled differently, enterprise data cleanup that takes real machinery to scale.
• Offered frontier-model access six months early versus the best talent pool for six months, Rao takes talent, “no question about it” — great people build your own primitives and the frontier arbitrage “is not the be-all and end-all.” On Goldman’s estimate that agents drive 24X token consumption in five years: directionally right — jobs you’re never good enough at are “bottomless pits,” and Jevons means “a tsunami of healthcare needs” the system can’t deliver.
• Founder mode, via investor Jensen (who cold-called Rao at midnight replying to a two-line email): not micromanagement but “tours of duty” — and “your job is to fall in love with whatever the job is,” CRISPR-ing new DNA that loves living on airplanes. Rao is unapologetically a wartime CEO (“who’s a peacetime CEO?”), envisions flattening his 450-person org around “super ICs,” and hires for Josh Wolfe’s “chips on shoulders makes chips in pockets.”

Deep dive

1. Five years in the desert: die on the thesis, pivot everything else

• Founded 2018 — “we started 3 months after the Transformer paper” — the company was hot for only the last 2-3 years; Harry calls it “a 5-year wilderness walk.” Rao’s survival mechanics: a true north you feel in your bones — that conversations are “the most human signal” in healthcare, on which “you can build a whole new set of rails” — plus brute persistence: “You just need to stay standing. You just need to not die. You just need to be there when it’s happening.”
• Would he have died on that hill? Yes. He was willing to pivot on product, feature order, go-to-market, and business model — “but I wasn’t willing to move on the thesis.” If the thesis failed, “we’d just shut it down and start something new.”
• The self-diagnosed mistake: pre-LLM they were fine-tuning BERT, BioBERT, Longformer, Pegasus, and T5 (publishing a 2021 paper on doing today’s jobs with pre-LLM models) while building a direct-to-consumer patient app (“Hey, can I record?”). Consumer healthcare companies “paint themselves into a corner” — business models that “would require you to take multiple showers every single day cuz you’re selling really private sensitive information.” “We realized that too late”; he’d have spent a couple more cycles in the “research caves” instead.

2. How USV happened: music taste as diligence, in both directions

• Rao stalked Union Square Ventures for years before the first meeting. What convinced him: their ritual of talking about music — “country western on Monday, indie hip-hop on Wednesday, Swedish death metal on Friday.” If they could pattern-match across genres, they could think about healthcare in a new way. He got in through an angel via an MIT friend and “knew at the end of that meeting that we were going to work together.” The seed: $5M on a $15M pre, 2018-19.
• Andy Weissman’s tell, as Rao tells it: ask the world’s biggest coffee snob about his ritual and he says “coffee’s coffee, man”; ask about music and it’s “I like all the sounds.”
• A company value written in 2021-22, before the taste discourse: “you have to taste good things to have good taste.” In practice: read the latest arXiv papers, absorb the newest UI/UX primitives, be “living at the edge of culture” — because the best companies are also creating it.

3. Healthcare GTM: it’s not one $5.3T market — time your YOLO shot

• The standard advice — start down-market, swim upstream — is a trap here. Of roughly a million US doctors, maybe 800,000 practice, and they’re concentrated in integrated delivery networks, payer-providers, and academic medical centers (the Emorys, Yales, UCSFs). “The trap a lot of healthcare founders fall into is that they stay down market… they don’t time their YOLO shot to go up market at the right moment.”
• The 2023 opening: 40-50% of doctors reporting burnout post-pandemic, a JAMA study finding 30% of nurses didn’t want to be nurses within 12 months, Medicare cuts compounding financial pressure — “we were pre-selling the market… and in 2023 when the sky opened, we ran right in.” Now it’s table stakes: doctors won’t sign with hospitals that don’t have “a bridge,” a shift that happened “almost overnight, over the course of 3 to 4 years.”
• The hardest raise was the Series A1 — “there’s a digit at the end of it… obviously things are tough over there.” A handful of no’s, then Whittington Ventures alongside USV and Bessemer did it at a 2X, at 100, pre-inflection: “folks believed that it was coming soon.”

4. Vintage discipline and the foundation-model question

• Rao’s taxonomy of AI-native vintages: post-Transformer pre-LLM, post-LLM pre-agent, post-agent. “Whatever your vintage, you have to become the latest variant as fast as you possibly can” — and that means the product and the way the company is organized both change.
• His answer to “won’t OpenAI/Anthropic build healthcare apps?”: “If you are fighting against them, you’ve already lost.” The choice is coexist or collaborate — “if the tailwinds that they create are not yours to leverage, then you’re screwed.”
• The reframe: healthcare is $5.3 trillion, 18-19% of US GDP — “I don’t know if it’s a vertical AI company… we are an AI company” serving one of the biggest opportunities there is. The defense is going “millions of miles deep in a regulated industry with proprietary data sets” built into workflows “really really hard to replicate.”
• The 2023 misread, corrected: post-training — learning from every user’s daily edits, across all doctor types, care settings, and spoken languages — “people in 2023 thought that was last mile. That’s actually most of it. The model piece is much less.” The vertical AI companies with the most upside “reach farther down into the stack and own their destiny.”

5. 40% in-house models: the binary-task rule

• About 40% of model outputs inside the product are generated by in-house models, varying month to month: “next month it might be 60% because we’ve distilled a new open source model and fine-tuned it… and we’ve just replaced a frontier model.”
• The allocation rule: binary, ring-the-bell tasks (get the nurse’s data point into the right discrete field at the right moment) go in-house — faster, cheaper, and “set and forget.” Tasks “you’re never going to be perfect on,” where the market rewards being “less imperfect than you were before” every month, “ride the frontier wave.”
• Why build at all: milliseconds in workflow. “We want to be like good air conditioning, where when it’s set right, we’re in the background” — and for high-stakes moments like a doctor okaying an order or a visit diagnosis before the patient leaves the room, frontier-model latency simply couldn’t deliver. (He points to the new Thinking Machines in-the-moment agent as the kind of experience requiring “insane performance and latency.”)
• On cost discipline: “I don’t care now. But I also know that my future self might care more” — the recent weeks of angst about model costs rising make having “little levers” on the P&L useful. And on when optimization matters, the Henry Kravis story: asked in their first meeting whether he wanted to IPO, Rao fumbled until Kravis interrupted — “No, you don’t. You don’t need to. Why are you even thinking about it?”

6. The wedge is the conversation; the note is a bill

• The load-bearing insight: clinicians “are not compensated for the care that they deliver. They’re compensated for the care that they documented that they deliver.” So the note was always a bill, and the roadmap ran along jobs nobody wanted: notes → orders → billing. “No doctor went to billing school” — as a health-system corporate VC, Rao watched revenue-cycle lunch-and-learns with pizza and PowerPoints draw the “thousand yard stare.”
• The enterprise sale means threading three lenses at once — CMIO, CIO, and CFO — and architecting note generation from day one so it extends “into the CFO’s world.”
• On Epic: “we are not competing with Epic.” They have “a product or a feature” for notes, but the company never wanted to be an EMR — it builds the intelligence layer on top, wedged into “that sacrosanct moment in health care where the actual value is getting exchanged.”
• On competition: Microsoft, via its $20-some-billion Nuance acquisition, was the first big competitor; now “we don’t see either of those companies much anymore” (the other, called Nabler by Rao) — Rao thinks category creation played a role, though he isn’t sure bundling played none: doctors at Emory using “a bridge” as a verb meaning “it did all these different jobs for me and unburdened me.” His general advice invokes Hamilton Helmer’s counter-positioning: build “in a way where the competitor couldn’t build because it would impact their current business.”

7. Trust economics and the vertical AI signal

• What the company could monetize but won’t: data. “The industry moves at the speed of trust” — and they’ve built in four years “what a lot of companies take 15 to 20 years” to build. Harry’s moral pushback — wouldn’t a healthcare data market improve models for global benefit? — gets the “earn the right” doctrine: health-system partners bless the roadmap in advance, papered into contracts. A greater-good foundation model? “Absolutely. We would just want to do that with everybody’s eyes wide open.”
• The OpenAI/Anthropic forward-deployed-engineer announcements and PE-portfolio partnerships read to Rao as a clear signal: “If that wasn’t a sign that there is absolutely an incredible opportunity for the foreseeable future for vertical AI, I don’t know what is.” Why: enterprise reality is brutal — data access, cleaning, workflow integration, and compliance where “it’s not SOC 2, it’s HITRUST,” and psych-encounter data must be treated differently from a primary-care visit. No FDA process required in their case — but building a scalable machine for all this is the moat.

8. “Just replace the doctors” — the 30-hour day and Jevons

• Harry’s pushback on the company’s new clinical-cues feature: “if you’re giving them prompts for questions and diagnoses, for goodness’ sake, just replace them.” Rao’s answer runs through a journal study showing doctors need 30 hours a day to finish their work — the company parsed those tasks and is “picking them off” with teams of background agents. High-frequency, low-stakes care (he cites the Utah controversy over automating medication refills) will be progressively automated toward “the most boring variant of primary care that nobody wants to practice” — but the hard cases are where judgment gets real, and “invoke Jevons paradox or whatever, we’re in for a tsunami of healthcare needs and the system is not prepared to deliver it.”
• On Goldman’s estimate that agents will increase token consumption 24X in five years: “I don’t know about the specific number, but directionally, I think so… you just can’t get enough of this technology once you start to use it.” Jobs you’ll never be good enough on are “bottomless pits.”
• Forced to choose between six months of early frontier-model access and six months of the best researchers and engineers: “B. No question about it.” Great people build your own models and primitives; the frontier arbitrage window “is not the be-all and end-all for us.”

9. Founder mode is tours of duty; the org is flattening

• Jensen — an investor in the company — cold-called Rao at midnight, responding to a two-line email sent that day (“his SLAs are insane”). The lesson Rao took: “your job is to fall in love with whatever the job is” — you “CRISPR new DNA” that loves living on airplanes. Founder mode isn’t micromanaging: “it’s about tours of duty” — go crush whatever’s on fire, where you’re best. Not yet CRISPRed: HR reporting to him, where he’s not “operating top of license.”
• Wartime or peacetime CEO? “There’s no other way to be… who’s a peacetime CEO?” Maybe Costco’s — “Ron something,” the forklift-to-executive-office lifer he calls the most underappreciated CEO, alongside Ali Ghodsi “in the way he plays his chess.” Hiring filter: Josh Wolfe’s “chips on shoulders makes chips in pockets” — insane slope from somewhere deep, because “it’s always wartime and it’s a different kind of war now… not everybody’s a warrior.”
• What he’s changed his mind on in 12 months: how flat a company can be — fewer managers, “super ICs” taking on much more, coordinated “in a way that was impossible before.” At 450 people he still thinks all-A-players is possible; he stopped meeting every hire around 100. The hardest hire now: high-judgment executives who can pattern-match against priors yet go against them, as “the time between a decision and an action is getting compressed.”
• On the $300M round (about a year ago) and spending discipline: hire principled finance leaders, and against goals that matter “you don’t want to blink” — though the fancy speakers he ordered for the New York office get a concession: “sometimes that stuff creates culture, too.”

10. The end state: change the business model, not science fiction in a clinic

• His map of US healthcare is the old XKCD Conway’s Law cartoon — the Microsoft org chart of “silos pointing guns at each other”: providers, insurers, and life sciences misaligned with each other and with patients. “The opportunity that we’ve got right now with AI is not to deploy the latest model and make science fiction happen in a clinic, it’s to change the business model.” Both ends of the spectrum move: AI doctors and automated low-stakes workflows below, human experts when you’re really sick, and new models that “incent prevention” — “care… not sick care.” Timeline: “the beginnings of that in the next three to five” years, not ten.
• On inevitability — the placebo effect of great investors betting on you: “the more you can feel inevitable, the more you will be.” Now: “I feel our mission is absolutely inevitable… but we’re at war to make sure that we’re the ones to do it and we do it in the next two to three years.”
• The personal ledger, stated without varnish: “the folks who say you can have everything are lying.” He’s in San Francisco Monday-Wednesday, with customers after, family on weekends — anchored in Pittsburgh for aging parents (“my dad has heart failure”), and Sundays with his parents are the thing he refuses to miss. On Harry’s claim that SF is the worst place to start a company, he disagrees: they moved from Pittsburgh in 2022-23 “exactly at the right time” as they inflected — “the amount of ideas that you just osmotically absorb by being there” matters.

Verification Notes

• The raw captions say “Abacus” in the introduction but render “a bridge” elsewhere; the company name is unresolved.