The AI Healthcare Trade Has a Hidden Risk: What a Dog's Cancer Vaccine and an FDA Rejection Reveal

By Leo Jaeyoul Lim | CEO, SEJONG DnC | Global Macro Analyst · 20 Years Experience
July 9, 2026 · Healthcare & Investment Strategy · 12 min read

In March, a Sydney data analyst with no biology background used ChatGPT, Gemini, AlphaFold, and Grok to help design a personalized mRNA cancer vaccine for his dog. The tumor shrank by roughly 75%, and the story went viral. Two months later, in May, a Korean AI device that had already cleared Korea's own fast-track regulatory system hit a wall at the US FDA. Read together, these two stories are not a curiosity — they're a preview of the gap investors need to price into the AI healthcare trade: the distance between what AI can compress in a lab and what regulators, insurers, and clinicians will actually accept at scale. I've spent two decades tracking how capital moves around regulatory friction, and this sector is currently mispricing that friction in both directions.

What this piece covers
1) Why the "AI drug discovery" narrative is real but overstated at the margin
2) The market-sizing numbers four independent research firms actually agree on
3) A regulatory case study — Korea's fast-track system vs. the US FDA — that functions as a natural experiment for where AI healthcare capital should flow
4) A newly documented risk (clinical "deskilling") that almost no valuation model currently accounts for

Four-stage AI pipeline diagram showing ChatGPT, Gemini, AlphaFold and Grok used to design a personalized mRNA cancer vaccine for a dog

Figure 1. The multi-model AI pipeline behind Rosie's personalized mRNA vaccine

Why Rosie's Story Is a Better Data Point Than a Headline

Strip away the viral framing and what's left is a genuinely interesting supply-chain insight. No single AI model carried the project. ChatGPT functioned as a research assistant, helping interpret immunology literature and plan next steps. Gemini cross-checked and cleaned raw genomic sequencing data. AlphaFold modeled the 3D structure of the mutated c-KIT protein driving the tumor. Grok then designed the final mRNA vaccine construct that UNSW's RNA Institute turned into an injectable vaccine in under two months, using the same lipid nanoparticle delivery technology behind the COVID-19 vaccines.

That's the bull case in miniature: drug discovery timelines compressing from years to months when multiple specialized AI systems are chained together. The bear case sits right next to it. The vaccine was administered alongside a checkpoint inhibitor, so there is no control arm isolating the vaccine's independent effect. AlphaFold's confidence score on the c-KIT structure came in at 54.55 out of 100 — well below the threshold researchers would want for clinical decision-making. And UNSW itself has been explicit: Rosie's cancer remains incurable, and at least one tumor didn't respond at all. Investors chasing headline-driven AI-biotech names should treat this as a template for the gap between demo and product.

The Market-Sizing Numbers That Actually Hold Up

Set the anecdote aside and look at four independent forecasts for the global AI-in-healthcare market. They don't agree on the exact dollar figure, but they agree on the shape of the curve.

Research Firm	2026 Market Size	Long-Term Forecast	CAGR
Fortune Business Insights	$56.01B	$1,033.27B by 2034	43.96%
MarketsandMarkets	$36.67B	$194.79B by 2031	39.70%
Grand View Research	$50.7B	$505.6B by 2033	38.90%
Research and Markets	$31.97B	$91.85B by 2030	30.20%

Bar chart comparing four research firms forecasts for global AI healthcare market size in 2026 versus long-term target years

Figure 2. Absolute figures diverge; the growth trajectory does not — a high-30s to mid-40s percent CAGR across every estimate

For a global macro reader, the practical takeaway isn't the specific terminal value — forecasts this far out always carry wide error bars. It's that four firms using different methodologies converge on the same order of magnitude for growth, which is a stronger signal than any single number. North America still captures roughly 44% of the market today, but the growth rate tells a different story.

Donut chart showing 2025 regional market share of global AI healthcare market: North America, Europe, Asia-Pacific and rest of world

Figure 3. North America leads on share; Asia-Pacific leads on growth rate

Korea's Regulatory Arbitrage — and Its Limits

This is where Korea becomes a genuinely useful case study for how regulatory design shapes capital formation in healthcare AI. Korea's Ministry of Health and Welfare and Ministry of Food and Drug Safety merged what used to be three sequential approval steps — device authorization, non-reimbursed fee-code eligibility, and health technology assessment — into a single integrated review track. The result: time-to-market for innovative medical devices dropped from 390-plus days to roughly 80 days. Devices meeting clinical-data thresholds can also skip formal cost-effectiveness review entirely for up to two years, generating real-world evidence while already earning revenue from patients.

The payoff shows up in the market data: Korea's domestic AI healthcare market is projected to grow from $377 million in 2023 to $6.67 billion by 2030 — a 50.8% CAGR that outpaces every global estimate above. Companies like Lunit (oncology imaging), VUNO (multi-modal vital-sign prediction), and JLK (cerebrovascular imaging) have used this domestic runway to build commercial traction before attempting international expansion.

Bar chart comparing time to market in days for Korea's integrated device review system versus the US FDA standard 510k review pathway

Figure 4. A fast domestic track doesn't guarantee a fast international one

But domestic regulatory speed is not a passport abroad, and VUNO's experience makes that point concretely. Its cardiac-arrest early-warning tool, DeepCARS, became the first Korean medical AI software to secure a non-reimbursed clinical fee code domestically, and it earned FDA Breakthrough Device Designation. In May 2026, though, its 510(k) submission came back with a "Not Substantially Equivalent" determination from the FDA — the agency's way of saying the device couldn't establish sufficient comparability to an existing predicate device. DeepCARS accounted for an estimated 25.7 billion won of VUNO's roughly 34.8 billion won in 2025 revenue, so the delay isn't cosmetic; it also pushed back potential inclusion in CMS's new technology add-on payment program.

Three Korean AI Companies, Three Very Different Q1 Outcomes

Company	International Strategy	Q1 2026 Trajectory
Coreline Soft	Entered Germany's national lung cancer screening program	Revenue +48.7% YoY (62.4% from overseas)
Lunit	Partnerships with major North American radiology networks	Record quarterly revenue
VUNO	FDA 510(k) NSE determination on flagship product	Revenue declined

The dispersion in this table is the real investment signal. "Korean medical AI" is not a monolithic trade — a company's overseas regulatory pipeline, not its domestic fast-track status, is what should be driving position sizing.

The Risk Almost No Model Is Pricing: Clinical Deskilling

The most underpriced risk in this sector isn't regulatory — it's cognitive. A study published in The Lancet Gastroenterology & Hepatology in August 2025 tracked 19 endoscopists across four centers before and after months of exposure to AI-assisted polyp detection. When those same physicians performed colonoscopies without AI assistance, their adenoma detection rate fell from 28.4% to 22.4% — a six-point drop the authors describe as the first real-world evidence linking AI exposure to measurable clinical skill erosion.

Card-style infographic summarizing Lancet study data on AI-related clinical deskilling including adenoma detection rate decline

Figure 5. The first documented case of AI-linked deskilling with a direct clinical outcome

This matters for valuation in a way that isn't yet reflected in most AI-diagnostics multiples. If reliance on AI systems measurably degrades unaided clinician performance, liability exposure, malpractice insurance costs, and the regulatory push toward "explainable AI" (XAI) design standards all become real cost lines for device makers over a five-to-ten-year horizon — not just for the AI vendor, but for hospital systems and insurers underwriting the technology. Companies that build governance and explainability into their product roadmap now are building a moat; companies that don't are building a future liability.

Where This Leaves Global Investors

None of this argues against the AI healthcare growth thesis — the market-sizing data across four independent firms is directionally too consistent to dismiss. It argues against buying the theme indiscriminately. The more durable framework is to weight exposure toward companies with (1) demonstrated international regulatory traction, not just domestic fast-track approval, and (2) visible investment in explainability and governance infrastructure ahead of the deskilling and liability risk becoming a mainstream headline. Coreline Soft and Lunit currently sit closer to that profile than VUNO does — though regulatory setbacks are rarely permanent, and re-filings under the De Novo pathway remain an option after an NSE determination.

Search This Blog

Senior Wisdom: Navigating Capital and Power