AI in Healthcare: The Anatomy of a $250 Billion Race

Imagine a hospital where doctors spend one in every five hours typing notes instead of seeing patients. A quarter of job postings go unfilled because staff are burning out.

Imagine a hospital where doctors spend one in every five hours typing notes instead of seeing patients. A quarter of job postings go unfilled because staff are burning out. This single image explains why healthcare is sprinting toward AI faster than almost any other sector.

2025 marked an inflection point for healthcare AI. For the first time, more than half of all digital health investment — 54% — went to AI companies. The year before, that figure was 37%. The year before that, under 20%. The trajectory is clear: this is not a trend, it is a structural shift.

Three forces are converging at once. Large language models have begun to genuinely understand medical language. The FDA opened its approval process to iterative development in late 2024. And hospitals, crushed by pandemic debt and a workforce crisis, have become unusually open to cost-saving technology. The simultaneous alignment of all three is what makes this moment different.

1. Ambient AI Scribe: Hottest, Fastest, Most Precarious

The concept sounds sophisticated but the mechanics are simple: while a doctor talks with a patient, AI listens in the background, writes the notes, and enters them into the system. The doctor looks at the patient, not the screen.

In H1 2025, clinical and non-clinical workflows, including ambient scribe, captured 55% of all digital health funding. The number looks outsized but the rationale is sound: pilots are generating real ROI. A Microsoft study involving 879 clinicians found that doctors saved five minutes per encounter, and 70% reported reduced burnout and fatigue.

The defining company in the category is Abridge, which raised $300 million led by a16z in June 2025, lifting its valuation to $5.3 billion — exactly double what it was four months earlier. Competitor Ambience Healthcare signed a five-year exclusive partnership with Cleveland Clinic and reached unicorn status with a $243 million Series C in July 2025. Clinical AI assistant OpenEvidence represents the most extreme case: its valuation reached $12 billion in January 2026, a 12x increase in roughly eleven months.

"Valuations are running well ahead of revenue — but this is also the first product category to meaningfully address clinician burnout."

The risk is equally real. Epic, the world's most widely deployed hospital software with 42% market share, announced its own native ambient scribe in February 2026. The question for investors is straightforward: what can a standalone company offer that Epic cannot build in? Companies without a clear answer will face a difficult 2026 and 2027.

2. Diagnostics and Imaging AI: Quiet but Profitable

Radiology AI generates fewer headlines than ambient scribe, but it is the most mature corner of the sector. By the end of 2025, the FDA had cumulatively cleared 1,451 devices in this space, 76% of them in radiology. Clearance volumes hit a new record every year.

Tempus AI became the category's reference story with its June 2024 IPO. The platform combines genomic data with clinical data for oncology and reported $1.27 billion in revenue for 2025, up 83% year over year, alongside a $200 million multimodal oncology deal with AstraZeneca. Israeli company Aidoc raised $150 million with NVIDIA backing and announced its own healthcare foundation model.

The category's unresolved inflection point is reimbursement: how will insurers pay for AI-based diagnostics? When Cleerly received a CPT code from Medicare for coronary plaque analysis, the industry celebrated — because it was an exception. A systemic solution remains far off.

3. Drug Discovery: Biggest Bets, Least Evidence

Drug discovery AI writes the most spectacular headlines. Xaira Therapeutics raised $1 billion on its founding day. Isomorphic Labs, spun out of DeepMind as an Alphabet company, secured deals worth approximately $3 billion combined from Novartis and Eli Lilly. Numbers at this scale shape perception, making clinical approval feel like a matter of time.

The real picture is more complicated. The category's most significant development came in June 2025, when Insilico Medicine's rentosertib — a compound fully discovered and designed by AI — passed a Phase 2a clinical trial, showing meaningful improvement in pulmonary fibrosis. Published in Nature Medicine, this result entered the record as the field's first clinical proof of concept.

Outside that single data point, clinical validation remains scarce. In the same period, BenevolentAI cut 30% of its workforce and cancelled its lead program. Recursion's share price trades more than 90% below its 2021 peak. The practical lesson for investors: round size is not a substitute for clinical evidence.

Underlying the entire category is Google DeepMind's AlphaFold 3, which predicts the interactions of DNA, RNA, and drug candidate molecules beyond protein structure alone. Published in 2024 with the stated ambition of illuminating the "dark matter" of biology, it contributed to the Nobel Prize in Chemistry awarded to Hassabis and Baker for the broader body of protein structure prediction work. The long-term impact will likely exceed the sum of all VC deployed in the space so far.

4. Mental Health and Genomics: Two Separate Stories

This final category houses two entirely distinct narratives under one roof.

On the digital mental health side, a structural engine has switched on. CMS began reimbursing FDA-cleared digital therapeutics in early 2025, transforming the "download an app and get better" promise into a real business model. Spring Health and Lyra Health hold strong positions in the enterprise market.

Within the same category, however, a dangerous tension is building. In 2025, multiple courts ruled that AI chatbot outputs constitute a "product," not "protected speech." These were the precedent-setting cases filed against Character.AI in connection with adolescent suicide. The industry must build its safety framework quickly.

On the genomics side, the technology has leapt forward while the business model has fractured. DeepMind's AlphaGenome has begun reading the 98% of the genome previously considered uninterpretable. NVIDIA and Arc Institute's Evo 2 model was trained on 9.3 trillion base pairs. In the same period, 23andMe — which made the most ambitious mainstream bet on consumer genetics — filed for bankruptcy. Value is migrating from consumers to those who can platformize the underlying data.

Summary: Four Sub-Sectors, State of Play 2025

So, What's the Takeaway?

The concerns are real. Ambient scribe companies are doubling their valuations in 90 days, running well ahead of ROI data. Rock Health's 2025 report flags an important detail: most of the year's growth came from just nine mega-companies. Strip those out and total investment falls below the prior year. The "haves vs. have-nots" divide has never been sharper.

The optimist case is equally strong. Insilico's clinical proof of concept, the pace of FDA clearances, ambient scribe systems now operating across more than 400 hospitals, and reimbursement inflection points are collectively moving these from "promising" to real product categories.

For investors, one conclusion stands out: in this sector, category selection always comes before company selection. In ambient scribe, defensibility rests on integrations Epic cannot replicate. In drug discovery, clinical scale is non-negotiable. In imaging, the market will not expand until reimbursement is resolved.

Türkiye presents a striking paradox. A centralized health data infrastructure like e-Nabız, with more than 78 million active users, offers a starting advantage that is genuinely rare globally. Yet open, standardized access mechanisms that would allow startups to use this data directly do not yet exist. Türkiye's real competitive edge will come not from generating data, but from opening it to innovation in a secure and anonymized way. Early examples like Virasoft and Albert Health show this is achievable, but competing at global scale requires data access, capital, and clinical validation capacity to grow in parallel.

Applying AI in healthcare, where the margin for error approaches zero, is among the most demanding challenges across any sector, given the regulatory intensity, the conservatism of buyers, and the requirement for clinical evidence. Scaling the sector requires three structural steps: national data infrastructures that enable safe and standardized access to quality clinical data; reimbursement frameworks that accelerate the integration of AI-based products into insurance systems; and the expansion of joint approval processes between authorities such as the FDA and EMA. For startups, the real bottleneck is not building a good product. It is feeding it with data, proving it works, and getting it in front of doctors and patients.

Perhaps the strongest signal of all: hospitals, among the most conservative institutions on earth, are now signing five-year exclusive agreements with AI companies. Healthcare is, in the simplest terms, the world's largest sector and the one that has preserved its inefficiency the longest. As long as those two facts coexist, there is no obvious reason for the opportunity to run dry. At Boğaziçi Ventures, we believe this transformation will only accelerate, and that AI in healthcare will produce one of the most enduring breakthroughs of the next decade.