← Back to AI-Human Insights

Market Analysis

AI Competitive Advantage: Why the Real Race in Health Isn't Being Run Where Everyone Is Watching

Traders in London are no longer checking Google or Amazon's share price before their morning coffee. They are checking Korea's. The 60-day correlation between the KOSPI and the Nasdaq has nearly tripled its five-year average, and Hani Redha, a London-based portfolio manager at PineBridge Investments, put it plainly: "We are all Korean investors now." South Korea's "AI for All" programme — free, government-funded AI access for 52 million citizens, paid for by a tax on the country's own semiconductor industry — has made a national population, not a single company, the reference point global capital now watches.

Britain has made a different but related bet. The UK's new £500 million Sovereign AI Fund writes cheques of £1M to £10M directly into AI startups, backed by up to a million hours of sovereign compute and fast-tracked visas for global talent. Of the five sectors the fund names as national priorities, AI in Health & Life Sciences sits alongside foundation models and scientific discovery — a state treating clinical AI as infrastructure worth owning a stake in, not just regulating from a distance.

Southeast Asia is running a third experiment entirely: not access, not capital, but workforce. A partnership between Parallaxnet and the Geneva College of Longevity Science aims to certify one million AI-ready healthcare professionals across Indonesia, the Philippines, and Malaysia within two years, built on over a hundred institutional agreements already signed. The bet is structural — that the binding constraint on AI in healthcare was never the technology. It was always the workforce that has to use it.

Three regions, three different theories of advantage. None of them are American, and all three are treating AI in health as a question a government or an institution has a stake in answering — not one to leave entirely to five competing consumer products.

What the five companies actually built

Between January and September of this year, OpenAI, Anthropic, Amazon, Microsoft, and Google all launched consumer or provider-facing health AI products. It is worth being precise about what each one actually is, because the coverage has mostly treated them as interchangeable, and they are not.

Google's approach runs through hardware: Fitbit, Pixel Watch, and a health coach built to interpret sensor data — heart rate, sleep stage, activity level. Amazon connects to the nationwide Health Information Exchange, bringing in diagnoses, medications, and lab results, while also drawing on a user's own retail purchase history to "improve responses." Microsoft's Copilot Health, in the company's own words, exists because "the product's value is not that it knows medicine in the abstract; it is that it can interpret your own scattered data in one place" — records, wearables, and lab results, aggregated and pattern-matched. Microsoft's consumer products already field more than 50 million health questions a day.

OpenAI's ChatGPT Health now integrates with Epic, the electronic health record system holding data for over 325 million patients. The company has been explicit about the boundary: the integration is read-only. AI does not write anything back into the record. It can summarise, prepare a clinician for a visit, and answer a question — but it does not act.

That single word, read-only, is the whole story.

Medical records are not MCP

A record is a static thing. It tells you what happened. It does not, on its own, let a system act on what happened — alert someone, adjust a plan, follow up. The infrastructure that allows an AI system to actually do something, not just retrieve and summarise, is a different layer entirely: the Model Context Protocol, introduced by Anthropic in 2024 and since donated to a neutral, cross-industry foundation under the Linux Foundation.

A live example makes the distinction concrete. UST, a global technology firm, has integrated Claude into CarePath, an operational platform used by health insurers and providers. In UST's own description: "Claude Code and MCP connectors link the platform directly to claims and care management systems, while an agentic layer routes each recommended action for approval before it reaches a member." That is not a record being read. That is an agent identifying a next step and routing it to a human for sign-off — the same foundation-before-action structure that any responsible clinical model has to be built on.

It is worth stating the honest counter-argument directly, because it matters. A critic could reasonably point out that MCP is now an open, donated standard — any developer can build a connector once and point it at more than one model provider. That is true, and it means MCP alone does not hand Anthropic a permanent, unchallengeable lead. What it does confirm is the shape of the actual advantage: not which company owns the protocol, but which company's whole model was built around language and dialogue in the first place, rather than numbers bolted onto a chat interface afterward.

Why language, not numbers, is the real foundation

A large language model is, at its foundation, trained on human language — not sensor readings, not lab values, not purchase history. Chloe Lubinski, who leads Anthropic's research into how these systems reflect human values, has put it directly: models "trained on human language and data, ultimately reflect human characteristics," and, in her own words, "the stories we tell, the language that we use, it shapes who we become. And they mirror us."

That claim is worth taking further than she stated it. Language does not exist independently of the people who produced it. There is no language that exists unless from us — language is us. And what gets carried inside language is not vocabulary alone. It is values, fears, and accumulated wisdom, inherited through the same training data that teaches a model grammar. To train on language is to train on centuries of reciprocal dialogue — argument, correction, agreement, one voice answering another — not a static list of facts.

This is the actual, structural gap between the two approaches now competing for the same market. A gadget that monitors sleep can tell you how many hours you slept. It cannot tell you that your sleep has been disrupted since a bereavement three weeks ago, or that a financial stressor mentioned in passing might be the real driver behind a change in mood that shows up nowhere in a wearable's dashboard. Data interpretation answers what a number means. Reciprocal, contextual dialogue can hold what a number means in light of an entire, accumulating human situation — biological, psychological, and social, together, not read off separately from three different feeds.

Where this actually plays out: community, not hospital

Almost every well-funded initiative described above is built around the hospital and the clinician: reading the chart, summarising the record, preparing for the appointment. That is real, useful work, and it happens inside an environment with almost no margin for error — critical care, acute intervention, decisions measured in hours.

Community care is a different environment entirely, and it is the one almost nobody is building for. The time frame for intervention is measured in weeks and months, not hours. That longer window is not a lesser opportunity — it is the one place agentic AI, built on genuine dialogue rather than data interpretation, has room to actually work: noticing a pattern across several conversations, recognising when a stressor in one domain of a person's life is quietly driving a change in another, and knowing when to bring in the physiotherapist, the occupational therapist, the dietitian, not only the clinician holding the diagnosis.

Almost every current model treats the clinician as the whole picture. The clinician is one part of a multidisciplinary team, consulted at the point of diagnosis. What happens after the diagnosis — the actual treatment pathway, lived day to day in the community — depends on a wider circle of practitioners that the current wave of medical-records-first products barely accounts for at all.

The actual question

The money now flooding into health AI is flooding toward records, hospitals, and clinicians, because that is where the drama is visible and the data already exists in a structured form. But the evidence, examined honestly, points somewhere quieter: toward the models built on language and reciprocal dialogue, applied to the long, unglamorous window of community care, where a person's biological, psychological, and social reality actually has room to be understood as one connected whole — not three separate feeds, read in isolation, by a system that was only ever trained to interpret numbers.