AI in Healthcare: Can Tech Scale Biology?

AI may not be a magic bullet, but it is undoubtedly impacting healthcare

• AI in healthcare is one of the longest-standing areas of development. It took ~60 years to build to single-digit billions in annual revenue; that figure is now estimated at ~$37–39bn (2025) and compounding at ~35–40% a year as the technology matures.
• According to estimates, 97% of hospital data remains unused. Inefficiencies are pervasive - spanning systems, providers, data management, and legal frameworks. Short-term hurdles hide longer-term potential.
• The differing mindsets, incentives, and technological infrastructures of the medical and tech sectors often clash, making seamless integration hard at first sight. But it is a matter of when, not if. 

AI's primary contribution to human lives will be in healthcare 

• Diagnostics alone - the sine qua non for personalized medicine - represents a trillion-dollar opportunity. Misdiagnoses cost the US economy $750bn annually, nearly twice the entire CMS yearly spend on prescription drugs.
• Patient management is highly inefficient due to manual data entry. Primary-care physicians spend more than half of an ~11- hour workday on electronic-health-record entry instead of patient care - an administrative burden that adds an estimated extra $1tn to healthcare costs.
• Drug discovery and therapy development are multi-trillion-dollar opportunities. Bringing a drug to market takes $2–4bn and around 10 years. AI can compress parts of that, but the gains are concentrated at the front end, and the step that determines success (clinical efficacy) is largely untouched. We treat the "10x, half the cost and time" claims with caution.

The road ahead is complex and lengthy, but the premises are encouraging

• The intuition that tech makes healthcare more expensive may mislead: keeping a patient healthy for an extra year costs less today than a decade ago.
• In the US, the first AI-related healthcare lawsuits are already underway, while federal policy has swung from oversight to deregulation: Biden's 2023 AI executive order was rescinded in January 2025, and the 2025 "AI Action Plan" now directs the FDA to build regulatory sandboxes to speed adoption rather than constrain it. Binding, generative-AI-specific rules remain absent.
• And beneath it all, a red-hot, increasingly global biotech market and a US-China race for innovation leadership fought on two fronts, frontier AI and drug innovation, where the US lead is real but no longer assured. 

AI may not be a magic bullet, but it is undoubtedly impacting healthcare

AI in healthcare is a very long story

The use of AI in healthcare goes back to 1965, when Edward Feigenbaum created the expert system DENDRAL to help organic chemists identify unknown molecules. Numerous systems followed in the 1970s and '80s - QMR for diagnostic support, PAPNET for medical imaging. These did not diagnose disease; they produced scores to help physicians gauge the likelihood of a condition.

None became usable in everyday clinical operations. This was not for lack of use cases - though scarce data and limited compute did not help - but mostly a lack of trust, regulatory oversight, and clinicians' resistance to change. As a striking example, PAPNET took over 15 years to reach FDA approval in 1995, and it took another 15 years to reach a further ~10 FDA-approved AI/ML devices.

The rise since 2010 reflects new algorithms and access to more data

After a lengthy 15-year process to approve the first 11 devices, the FDA has since authorized roughly 1,400 AI/ML-enabled medical devices cumulatively through end-2025 - with a record ~331 cleared in 2025 alone - a rapid acceleration reflecting real advances in the underlying technology. The concentration is telling, and sets up a theme we return to below: radiology accounts for roughly three-quarters of them.

The early 2010s marked a turning point, with the rise of Natural Language Processing (NLP), voice synthesis/recognition, computer vision, and large language models. NLP in particular expanded machine learning by enabling systems to understand human language, unlocking the vast potential of Electronic Health Records (EHR) as training data.

The parallel digitization of EHRs drove an unprecedented surge in healthcare data. By 2020, healthcare had produced over 2,000 exabytes, up dramatically from 153 exabytes in 2013 - roughly 30% of global data volume. That volume demanded advanced analytics to turn data into actionable insight. Companies like IQVIA and Veeva Systems have become leaders here.

Yet challenges remain. While 99% of hospitals have adopted EHRs, ~97% of hospital data goes unused due to a silos-based approach, and only ~76% of US non-federal acute-care hospitals can send, receive, find, and integrate data. This lack of interoperability complicates the training of LLMs for generative AI (GenAI).

Regulatory hurdles compound the problem. The FDA still treats any AI/ ML software intended to treat, diagnose, cure, mitigate, or prevent disease as a medical device. But the earlier vacuum is closing. In January 2025, the agency issued comprehensive draft guidance on AI-enabled device software functions across the total product lifecycle (finalization still pending), finalized its Predetermined Change Control Plan guidance in August 2025, and convened its Digital Health Advisory Committee twice, specifically on generative AI (November 2024 and November 2025). The first foundation-model device, Aidocs's CARE1, was cleared in early 2025, and an LLM-based patient chatbot won Breakthrough Device Designation later that year. What is still missing is finalized, generative-AI- and LLM-specific binding guidance; the gap for the most powerful multimodal models is narrowing rather than closing, and that residual uncertainty still complicates adoption. 

The slow rate of change is the product of clashing mindsets

It has proven very hard to reconcile all the stakeholders' motto: Big Tech's "move fast and break things," MedTech's "low risk beats big profits," the FDA's "slow first, safe after," doctors' "do no harm," and the government's "least costly possible." The pace has improved sharply since 2010: market-sizing estimates now put AI-in-healthcare revenue at roughly $37–39bn in 2025 (methodologies vary widely), rising toward ~$50bn in 2026 and growing at a ~35–40% CAGR up from the ~$6–10bn attributed to AI as recently as 2023. Even so, that is only a few tenths of a percent of the ~$11tn spent on healthcare worldwide: the revolution is real and accelerating, but still early. 

Big tech's healthcare forays have mostly disappointed, from IBM Watson in oncology to Alphabet's DeepMind in diabetic retinopathy. The exception is Apple: by focusing on consumer health and wellness rather than complex clinical applications, the Apple Watch (heartrate tracking, ECG, fall detection) achieved wide adoption and eventually FDA clearance for atrial fibrillation. That path is hard to reproduce because the healthcare consumer experience remains among the least user-friendly in modern life. 

Doctors have, in fact, embraced AI faster than the institutions around them: physician use has roughly doubled yearly since ChatGPT's launch, reaching around 81% by 2026. But much of it is unsanctioned "shadow AI": consumer-chatbot use that surged through 2025 because health systems lacked the tooling and governance to run generative AI safely on patient data. Formal, sanctioned adoption is far narrower and still concentrated in radiology, where imaging tools are largely AI-based and, crucially, where comparative studies exist to enable benchmarking. That is why AI imaging software, alongside data brokerage by Veeva, IQVIA and Tempus AI, remains the bulk of today's AI-in-healthcare vendor revenue: what AI companies capture lags what clinicians already extract off the books.

The FDA has been reticent to fully exploit multimodal AI (image, sound, text). Two concerns dominate: hallucinations and data security. Recent hacks of healthcare providers justify caution. Ironically, AI can quickly strengthen cybersecurity via providers like CrowdStrike, but addressing hallucinations will take years, notably building proper benchmarks for LLM accuracy and integrating them into clinical processes. 

AI's primary contribution to human lives will be in healthcare

As managers investing in tech-heavy companies, we firmly believe that one of AI's primary contributions to human life will be in healthcare. AI can improve diagnostics, patient management, and drug discovery by analyzing complex data with accuracy and efficiency - reducing misdiagnoses, streamlining care, and accelerating therapies. Despite data-access and regulatory challenges, the transformative potential is real.

Diagnostics is the key bottleneck for the healthcare system

Today's system is still built around acute care: a sudden dysfunction is felt, diagnosed, and treated as fast as possible. That is why hospital care takes the biggest slice of expenditure - ~31%, or ~$1.6tn of the US's $5.3tn total in 2024 (CMS). But with an aging, more obese population, the real picture is chronic care, where initial conditions go unnoticed for years. Rising blood pressure, a leading symptom of kidney and cardiovascular disease, goes undiagnosed in 50% of patients until it is classified as hypertension. In a different field, only 8% of mild cognitive impairment leading to dementia is diagnosed before the final diagnosis.

Today's ~$200bn diagnostics market (in-vitro diagnostics, pathology, medical imaging, genomic testing) remains undervalued relative to its potential. Monitoring devices - continuous glucose monitors, heart rate monitors, oximeters - paired with AI capable of crunching large data streams could detect abnormal patterns and lower hospital bills through preventive treatment, a market dominated by entrenched players like Dexcom, iRhythm, and Masimo. The same holds for imaging, where even top clinicians' tumor detection rarely exceeds 75%, and where AI suites like Pro Medicus's can lift it by 10–20%. 

For the full picture: medical errors are responsible for ~250,000 deaths per year in the US, the third leading cause of death. Nearly 30% of those are due to inaccurate, delayed or incomplete diagnoses; misdiagnosis drives 17% of preventable deaths in hospitalized patients. Beyond accuracy and safety, cost matters: misdiagnoses alone cost the US economy $750bn annually - almost twice total CMS prescription-drug spending.

This is why diagnostics is one of our strongest convictions in Atonra's Bionics strategy.

Patient management is done inefficiently due to time-consuming, error-prone manual data entry

One reason misdiagnosis is so high is that physicians spend ~50% of their time on administrative tasks such as documentation rather than care, a major source of burnout. The AMA found primary-care physicians spend more than half of an ~11-hour workday on EHR entry (~4.5 hours during clinical hours, ~1.5 off-hours).

At scale, roughly 15–25% of healthcare spending is administrative. With so much data unstructured, workers manually enter and transfer it, cutting productivity and inviting errors. Researchers estimate ~25% of healthcare spending is wasteful - roughly $760–935bn per year - of which ~$266bn is administrative complexity and $231–241bn pricing failure. Per CMS, improper payments were ~6.5% and 21% of Medicare and Medicaid payments respectively in 2020.

For administrative staff, AI can accelerate prior authorization and eligibility checks, reduce claims errors, improve billing, and automate data transfers between systems. A famously complicated payment system is fertile ground to streamline processes and tie outcomes to payments.

Drug-treatment discovery can be improved - but not as easily as the hype suggest

Big biopharma, despite a ~90% failure rate, still generates ~$800bn in combined yearly revenue on a model of 10–12 years of R&D and only 8–10 years on-market. Any cut in time-to-market or lift in probability of success dramatically enhances ROI. 

AI can, in principle, touch many stages: speeding trial recruitment (often the biggest bottleneck), managing large data volumes, predicting efficacy, improving safety through better patient targeting, aiding molecule synthesis via simulation, and assisting regulatory compliance. One-third of Phase III trials - the most expensive phase - fail due to enrollment challenges, and matching patient records to enrolling trials (a specialty of Tempus AI) plus remote monitoring can cut dropout and cost. In design, ~4,000 genes have experimental links to human disease but lack the structural data for traditional drug design - a potential fourfold increase in FDA-approved drugs. Meanwhile, AI investment for drug design is still meager: only ~$18bn on AI-first biotech over ten years, less than 10% of what big biopharma spent on R&D last year alone. The momentum is nonetheless visible: the AI-specific drug-discovery software segment roughly doubled from ~$0.9bn (2023) to ~$1.86bn (2024), and by 2025 around two-thirds of life-sciences executives reported investing in generative AI for research.

A necessary caveat: where AI moves the needle in drug discovery - and where it does not. None of this dims our conviction on AI in healthcare; it simply argues for discipline about where, within 7 discovery, the technology changes outcomes. The bull case implicitly assumes the industry's bottleneck is information processing - finding targets, designing molecules, reviewing literature, and running experiments faster. That is precisely what AI, and LLMs in particular, do well. But it is not where the value, or the failures, actually sit. 

Target and molecule generation were never the binding constraint. Schrodinger Inc, with a world-class physics-based discovery engine, and Recursion Pharmaceuticals, with an industrialized wet lab generating biological data at scale, are the proof: both have produced abundant candidates and remarkably few approved drugs. Nor has lab automation scaled as the market underwrote it to - Ginkgo Bioworks is the cautionary case that more throughput does not translate into more medicines.

The same logic holds downstream. Better patient selection requires better biomarkers - new, validated, predictive biology - not a better literature review, which is the task AI actually accelerates. Better trials require the FDA to accept new endpoints and evidentiary standards, a regulatory process no model can shortcut. And while AI will genuinely lighten the load in QA, regulatory affairs, medical writing, and marketing, that is a gain in execution quality and efficiency, not in efficacy, and administrative overhead was never the dominant cash burn. That burn is clinical, concentrated in Phase II/III failures driven by drugs that simply do not work in humans. AI compresses the cheap, fast front end that was never the problem, while leaving the expensive, slow, value-determining step largely intact.

The one real exception is structural biology: AI-driven protein structure prediction has genuinely expanded the druggable target space, including many of the ~4,000 disease-linked genes that lacked structural data. That is a real contribution - but it enlarges the menu upstream; it does not relocate the bottleneck downstream. This is why claims of halving costs and timelines deserve scrutiny: they apply mostly to a pre-clinical stage that accounts for only a small share of total spend and time. The prize is real, but it will be won in the clinic, not the compute cluster.

The clearest test of the thesis is Isomorphic Labs. If any team can prove that technology scales biology, it is this one: spun out of DeepMind, built on the Nobel-winning AlphaFold, led by Demis Hassabis, and capitalized with roughly $2.6bn after a $2.1bn Series B in May 2026 - the second-largest biotech raise on record - backed by 8 Alphabet and a roster of sovereign-wealth funds. It has the field's best pedigree, deepest pockets and loudest narrative. 

And yet, nearly five years in, not a single molecule has entered the clinic and no human has been dosed. First-in-human trials have slipped to end-2026, a year past a target originally set for 2025 (since recast as pre-clinical). Its marquee collaborations with Lilly (up to $1.7bn in milestones), Novartis and Johnson & Johnson are genuine, but they are discovery deals in which partners apply the engine to their own programs; to date there is no disclosed co-developed candidate in the clinic and no announced milestone confirming a molecule advancing. The internal pipeline itself remains undisclosed. The validation so far, in other words, is contractual and computational - not clinical. And this is the very company that owns the one genuine AI breakthrough in discovery, structure prediction, which only reinforces the point: the edge is real, and it still has not moved the bottleneck to where molecules actually succeed or fail - in humans.

For investors, this reaches beyond one private company. The listed AI-discovery complex is priced on a scaling narrative, not on current cash flows, and Isomorphic is its bellwether. A stall there would not stay contained: it would re-rate the whole cohort, because a failure at the best-resourced, best-credentialed attempt reads as evidence that technology cannot yet scale biology - precisely the premise the market is underwriting. With risk appetite elevated and equity indices near record highs, tolerance for that kind of thesis-break is thin; euphoric markets punish broken narratives hardest. Our concrete watch-item: if Isomorphic is neither in the clinic nor advancing a partnered candidate within roughly three years, we would expect a material, sector-wide de-rating of AI-first drug discovery - a drag on the theme, not a footnote. In fairness, drug timelines are inherently long and a clean 2026 clinical entry would still be fast; but the burden of proof now sits squarely on delivery, not on models.

After the 2020–21 boom-and-bust in the few listed names, we await the upcoming clinical results to see whether these R&D efforts truly reshape the biotech landscape.

AI-assisted robotic surgery can improve patient care and surgeons' lives

Among medical errors, surgical errors have the most direct and serious consequences. Worldwide, an estimated 4.2 million people die 9 within 30 days of surgery each year, roughly 7.7% of all deaths and the third-leading contributor after heart disease and stroke, more than HIV, malaria and tuberculosis combined. In the US, surgeons leave a foreign object inside a patient, operate on the wrong site, or perform the wrong procedure an estimated 4,000 times a year, resulting in death in 6.6% of the cases, hence why ~75% of US lawsuits against surgeons stem from intraoperative errors.

Robotic surgery has proven able to reduce errors and costs. On average, robot-assisted patients stay eight days in hospital versus 10 for open surgery - a 20% reduction - with a lower complication rate (13.2% vs 23.7%), roughly half the readmission risk, and a striking fourfold reduction in blood clots (deep-vein thrombosis and pulmonary embolism).

The first barrier is cost, but the real challenge is upskilling surgeons - where AI plays a crucial role, training on best procedures and accelerating learning via live feedback. Demonstrations of Intuitive Surgical's latest robots are remarkable. With ~300mn surgeries performed worldwide last year and Intuitive, the leader, performing only ~3.1mn, the potential is immense.

Today, in our Bionics strategy, robotic surgery is our second most significant investment.

The road ahead is complex and lengthy, but the premises are encouraging

US health spending has bent well below its earlier trajectory

In 1997, CMS projected healthcare spending per elderly would reach $20k by 2010. Around 2005, spending growth slowed markedly; gradually, but visibly. In dollar terms, spending increases fell by more than half; in percentage terms, by two-thirds. When recomputed in 2010, the $20k figure was pushed to 2020; it stood at just $13k in 2022. Across 2011–2022, the spread between projected and actual spending reached ~$3.9tn, all with an aging, more obese population in a forprofit system.

$3.9tn is a staggering sum. As the director of healthcare policy at the Committee for a Responsible Federal Budget put it, there is essentially no recent precedent for budgetary savings of the kind seen in the 10 Medicare slowdown. Reframing healthcare as pure cost misses the point: with better health, over the last 20 years the world has worked ~7% fewer hours while output per hour rose ~25%. Healthcare is not a cost, it is an investment.

The role of Biotech and Medtech is crucial

One leading factor is the biopharma business model, where innovation reaches the most people after patent expiry. Roughly half of the reduction in major cardiovascular events, for instance, is directly attributable to medications controlling cardiovascular risk, statins and antihypertensives among them. We expect the next drop in spending-vs-projection around 2028–2030, at the peak loss of exclusivity for the blockbusters that made biopharma a mint over the past decade.

Second, biotech has historically leaned toward rare diseases which are not rare as a group. By definition, each affects fewer than 1 in 2,000 people in any WHO region; yet with more than 7,000 types, roughly 300 million people live with a rare disease.

Third, MedTech improves iteratively: from mini-invasive treatment to better diagnostics. Over the last ten years, half of yearly FDA requests concern improvements to existing products. MedTech compounds year over year, a feat impossible in biotech, where drug patents last 20 years and only the production method can be slightly amended. 

Data and legal challenges await

All these factors have limitations that AI can help address, but AI's primary needs are data and accountability.

Data is hard to come by: only ~70% of US non-federal acute-care hospitals have fully interoperable systems. Even setting privacy aside (which is far from resolved), hospital IT infrastructure still lags. 

More importantly, even with accessible data and well-trained models, what about accountability? In November 2023, families of former UnitedHealth members filed a class action alleging the company used an AI model with a known 90% error rate to wrongfully deny extended-care coverage to Medicare Advantage patients, forcing families to pay out of pocket or go without care. Decisions in this and similar cases will be crucial in defining algorithmic accountability, and 11 for now, accountability is being shaped more by litigation and state law than by comprehensive federal rules. The federal posture has, in fact, reversed: Biden's October 2023 executive order, which set broad safety standards including provisions for healthcare AI, was rescinded on the first day of the Trump administration in January 2025 and replaced by a deregulatory agenda. The July 2025 "America's AI Action Plan" explicitly casts slow adoption in sectors like healthcare as the bottleneck and directs the FDA to stand up regulatory sandboxes and AI Centers of Excellence to accelerate deployment. One Bidenera measure that survives is the ONC's HTI-1 rule, the first federal transparency requirement for AI and machine-learning predictive software in healthcare. The direction of travel, in short, is toward speed over constraint, which leaves accountability comparatively under-defined.

The AI outsiders make their move: OpenAI, and now Anthropic

Managing data and navigating legal exposure is OpenAI's forte, and it holds the lion's share of AI revenue, with reported figures north of $25bn, up from $1.6bn in late 2023. More importantly, while most AI providers in healthcare have built narrow, single-data-type models, the latest ChatGPT can handle the multimodality of healthcare data (unstructured text, sound, images, lab results, bills) and OpenAI has since released a purpose-built biological-reasoning model (GPTRosalind), drawing partners such as Amgen, Moderna, Thermo Fisher Scientific and Novo Nordisk. Sanofi SA and Eli Lilly were the first biopharma to move beyond simply enabling ChatGPT internally and sign extensive deals directly with OpenAI; tellingly, none of the sellside desks would put a number on the R&D or process-ROI impact of those deals. 

As of mid-2026, OpenAI is no longer the only outsider in the fray. Anthropic, the maker of Claude, has entered along a distinct axis. Rather than sell only a horizontal capability into pharma, it has launched Claude Science, a research workbench that consolidates databases, computing, and domain tools (genomics, proteomics, 3D protein structures) into a single environment, and, more strikingly, has stood up its own internal drug-discovery program targeting "neglected" diseases the commercial market overlooks. Backstopped by its acquisition of drug-discovery startup Coefficient Bio (reported at ~$400m) and building on Claude for Life Sciences (2025) and Claude for Healthcare (2026), Anthropic is positioning to own the 12 discovery "operating layer" and, unusually, to run its own programs rather than merely equip others.

We read the two strategies as complementary evidence of the same thesis: that large, multimodal models are the only tractable way to attack healthcare's unstructured, multi-format data at once. Which approach compounds faster, OpenAI's enterprise-distribution play or Anthropic's build-and-operate stance, is precisely the question the next few years of clinical and commercial data will answer. As the Isomorphic case above shows, the harder test is not shipping a capable model; it is turning that capability into an approved medicine.

A red-hot market and a global race the US must lead

Biotech has come roaring back. After a brutal multi-year drawdown, the S&P Biotech index (XBI) closed June 2026 around $158, up 90.5% over the trailing twelve months and 30.3% year-to-date; the large-cap IBB has posted comparable gains, and the sector's late-June run was its strongest monthly stretch since 2023. Capital has followed: US biopharma deal-making reached ~$106bn across ~200 transactions in the first half of 2026 alone, late-2025 follow-on issuance was the strongest in years, and the IPO window has reopened. The sector trades far from euphoria relative to the S&P500 by its own history, still allowing the rally to have legs as we wrote in our outlook (Healthcare: From Underdog to Front Runner). The drivers are familiar: a looming patent cliff forcing M&A, the obesity/GLP-1 wave, oncology innovation, and a policy/FDA backdrop that the market has learned to look past. 

Crucially, this is not a US-only phenomenon, and the more consequential shift is where innovation now originates. By early 2026, roughly 30% of all novel drug candidates in global clinical development came from Chinese-headquartered companies, up from low single digits a decade ago, with over 1,200 novel candidates in trials and China taking the global lead in new study initiations. The capital markets have noticed: greater-China outbound licensing reached an unprecedented ~$137.7bn in 2025 (nearly 10x the 2021 level), average deal size has jumped to ~$1.3bn, and Chinese-originated assets already accounted for ~28% of large-pharma innovator deals in 2024 (~$41.5bn). China now sits behind nearly 90% of global ADC licensing activity. Facing ~$230bn of US patent-cliff erosion through 2030, Western pharma is licensing Chinese molecules "at a fraction of the cost of full M&A" lile Astrazeneca/CSPC Pharmaceutical (up to $18.5bn), Bristol-Myers Squibb/Hengrui Medicine (up to $15.2bn) and the Pfizer and Lilly/Innovent Biologics collaborations are the new template.

This redraws the competitive map. China's ecosystem has "gained the lead" in generating early, promising candidates, and expects that edge to persist as US early-stage assets stay underfunded. The US retains formidable advantages (the deepest clinical capital, the largest and highest-value market, the FDA gold standard, and the world's leading AI ecosystem) but leadership is no longer a given. Washington's response has so far been defensive: the BIOSECURE Act became law in December 2025, the US International Trade Commission opened an investigation into Chinese biotech state support and pricing in February 2026, and the proposed Biotech 14 Investment National Security Act of 2026 (H.R. 9102) (June 2026) would extend outbound-investment screening to biotech licensing, joint ventures and equity stakes. Chinese executives largely dismiss this as "noise" so long as the data keeps reading out.

Which is exactly why the AI race matters. If there is a lever that lets the US out-innovate on speed and cost rather than merely fence off competitors, it is artificial intelligence and it is telling that the frontier labs (Anthropic, OpenAI) and the best-funded AI-native drug shop (Isomorphic, via DeepMind) are overwhelmingly American. But as argued above, that lever is real at the front end and still unproven at the step that decides winners: the clinic. The US's structural edge in AI is its best card for keeping the biotech lead this decade, provided it can convert compute into approved medicines faster than China converts chemistry, cost and speed into the same. For now, both remain to be proven, and that contest, not any single product launch, is the real story beneath the rally.

Conclusion

The transformative potential of AI in healthcare is undeniable, yet the journey is complex and multifaceted. Despite substantial challenges - data accessibility, regulatory hurdles, and the need for interdisciplinary collaboration - advances in AI offer promising solutions to longstanding inefficiencies. AI's ability to sharpen diagnostics, streamline patient management, and improve surgical outcomes can meaningfully raise the quality of care while reducing cost. In drug discovery, we are more measured: the technology already compresses the front end, but the value-determining step remains clinical efficacy, and the flagship AI-native efforts have yet to put a molecule in a human. As investors in science and technology, we are confident AI will play a pivotal role in healthcare, while remaining disciplined about which parts of the value chain it actually changes.

For our part, this is why Atonra treats AI in healthcare as a thematic conviction rather than a single trade. In our Bionics strategy, we are positioned where the value is already measurable, across diagnostics, patient monitoring, and robotic surgery, and in the data and workflow infrastructure that monetizes AI regardless of which model wins. In Biotech 360° we back therapeutic innovation while staying deliberately disciplined on the drug-discovery hype, favoring assets whose value is proven in the clinic over platforms whose value is asserted on a slide. And because innovation is globalizing, we seek 15 exposure to it wherever it originates, mindful that the US's edge in AI and capital is the lever most likely to keep it in front. 

The path to fully integrated AI will be gradual. Robust regulatory frameworks and hard-won trust among providers and patients are essential to unlocking its potential. Against a backdrop of a red-hot, globalizing biotech market, the deeper question is not whether AI helps, but whether the US can convert its AI and capital advantages into a durable innovation lead before lower-cost, faster-moving competitors close the gap. Addressing these challenges and leveraging AI's genuine strengths points toward a future where healthcare is more efficient, effective, and accessible, ultimately improving human lives at scale.

Catalysts

• FDA guidelines for LLMs. GenAI, well managed, can speed nearly all processes but needs an approval path.
• Clinical successes of AI-based therapies. Whether AIdesigned drugs or software-as-a-medical-device, positive patient outcomes will boost the sector. The first AI-native molecules (Isomorphic and peers) reaching the clinic, and any partnered milestone confirming a candidate advancing are the key proof points to watch.
• Measurable hospital savings from AI. Documented savings in hospital financials will trigger wider adoption.

Risks

• LLM hallucinations lead to harm. Misdiagnosis or a misfired AI-powered device could reduce trust in AI.
• Lower-than-expected productivity gains. If AI deployment at hospitals or pharma fails to improve efficiency, the hype fades.
• Flagship failure in AI drug discovery. Isomorphic Labs or Anthropic, reaching neither the clinic nor a partnered clinical candidate, would likely trigger a sector-wide de-rating of listed AI-discovery names, since the investment case rests on scaling rather than delivered assets.