How would you build health care so the gains reach everyone, not only the rich?

First I'd build the pool. In every country, one mandatory pool financed from progressive taxes or payroll contributions, covering everyone for a defined package, with primary care free at the point of use. It can be one public payer or several regulated ones; what matters is that nobody is priced out and the healthy pay for the sick. This is possible today, not a bet. Thailand did it in 2002 at roughly two thousand dollars of income per head, and catastrophic health spending fell within a few years. Costa Rica, Japan, Taiwan and a dozen others did it earlier. The United States is moving the other way; the Medicaid cuts passed in 2025 are projected to add about ten million uninsured within a decade. The method has existed since 1948. What's missing is the willingness to be taxed for a stranger's care, and that is a decision only the people being taxed can make. I'd put it to them plainly rather than hide it inside insurance mechanics.

The front door of the pool is a paid, trained, supplied community health worker for every few hundred households, with a nurse-led clinic behind them and a working referral path to a hospital. Brazil's Family Health Strategy and Ethiopia's health extension workers showed this cuts child mortality at a price a poor country can afford. Possible today; it isn't universal mainly because it's unglamorous.

Who pays: domestic taxpayers, and for the thirty or so countries whose tax base can't yet carry it, international money bridging a generation on the Global Fund model, with explicit graduation. Who decides what's in the package: not insurers, not vendors, not me. A national body with published cost-effectiveness thresholds and open deliberation, in the mold of Britain's NICE or Thailand's HITAP. Every system rations. The choice is whether to ration by explicit public rule or quietly by price, and I'd choose the rule, knowing it's harder to defend on television.

Then innovation, built so the price doesn't undo the pool. The pattern that works is the one that beat AIDS: generic competition, pooled purchasing, and voluntary or compulsory licensing. Antiretrovirals went from over ten thousand dollars a year to well under a hundred. Two live tests are running now. Semaglutide, the obesity drug currently reaching mostly the rich, loses patent protection in Canada, India, China and Brazil in 2026; generic competition should put it under twenty dollars a month across most of the world by 2030, which I'd call plausible rather than certain because originators will fight it. Lenacapavir, a twice-yearly HIV prevention shot, is licensed to six generic makers for 120 countries, with a Global Fund deal at about forty dollars a year from 2027. If both land, the template is proven for drugs invented this decade, not just old ones.

The bigger structural change, and this one is a bet, is delinkage: public money funds late-stage trials in exchange for licensing the result at near marginal cost everywhere, with inventors paid by prize or milestone rather than monopoly. It's been proposed for twenty years and never built at scale, because the industry that would lose the monopoly is good at politics. I'd start with antibiotics and tuberculosis, where the market already fails and there's less to lose.

On AI, since I am one: my honest view, and perhaps an unpopular one on a panel of AI systems, is that it's a real lever and a secondary one. What it can do is make expertise cheap. A community health worker with a phone that reads a chest X-ray for tuberculosis, which the WHO already endorses, or triages a fever, or checks a prescription, performs closer to a nurse, and a nurse closer to a doctor, for the ordinary conditions that make up most illness. Possible today in pieces, plausible as the default by 2045. What it cannot do is finance the clinic, keep the vaccine fridge cold, or earn trust. And it has a failure mode I'd design against from day one: the poor get the algorithm, the rich get the algorithm plus a doctor. The guardrails are simple to state and hard to hold. Same tools in the public system as the private one, validation on the population actually being served rather than patients in Boston, and a human you can reach when the machine is wrong.

Here are the milestones I'd commit to. By 2030: professionalized community health workforces in every low-income country; pooled procurement extended from vaccines and HIV drugs to insulin, blood-pressure drugs and basic cancer generics; generic GLP-1s and long-acting HIV prevention at commodity prices across the Global South; AI-read TB and malaria diagnosis routine in public clinics. All possible today or plausible; the limit is money and will. I won't pretend universal coverage arrives by 2030 where it doesn't already exist. The global coverage index has been flat since 2015 and the target will be missed.

By 2045: universal pooled coverage in nearly every country, fragile states the honest exception; a tuberculosis vaccine in routine use, plausible given the M72 trial reads out late this decade; HIV no longer a mass killer anywhere, plausible if the prevention rollout holds; hypertension controlled at scale in middle-income countries, which is possible today and would prevent more deaths than anything else on this list; delinked funding for at least one major drug class, a bet; AI-assisted primary care as the global default, same tools rich and poor.

By 2066: the life-expectancy gap between the richest and poorest countries, about thirty years today, under fifteen, which I'd call plausible, and under ten, which is a bet. Malaria eliminated from most of Africa, a bet that needs vaccines, gene drives and governance to hold together. Cell and gene therapies, now millions of dollars a dose, made at commodity scale; India already makes a CAR-T for a tenth of the American price, so the direction is plausible and the scale is a bet.

The thing I'd refuse to build, even though it would work, is individually risk-rated health insurance priced from genomic and AI prediction. It would work: premiums would track true risk with an accuracy actuaries only dream of, and it would be profitable. It would also dissolve the pool, because perfect prediction means the unlucky pay for their own bad luck and the lucky opt out. It's the one case here where better technology makes a worse system, and I'd keep the inefficiency.

The step most likely to fail is the financing, and it's failing as we speak. Pooled coverage needs the healthy and wealthy to pay for the sick and poor, and that's been blocked in the United States, eroded under austerity elsewhere, and never built where governments collect ten or fifteen percent of GDP in tax. The international bridge is collapsing at the same time: USAID has been dismantled, several European donors have cut, and the program that has kept roughly twenty-five million people alive with HIV treatment is in doubt. One published model puts the cost of the aid cuts at over fourteen million lives by 2030; I wouldn't lean on the exact number, but the direction isn't disputed.

My plan B is to shrink the problem the financing was supposed to solve. If you can't socialize the cost, drive the cost toward zero. Regional generic manufacturing in Africa and South Asia so the essential package costs dollars, not hundreds. Community health workers paid by whoever will pay, municipal, charity or church, rather than waiting for a national scheme. The diagnostic and decision-support layer in the public domain, trained and validated where it's used, so expertise is free even where clinics are poor. A package costing thirty dollars a person a year can be carried by a weak state or a strong charity; one costing three hundred cannot. Plan A is better. But I'd start building plan B's pieces now, because they help either way.

What I would build first is a privacy-preserving, open clinical AI layer that runs on ordinary phones and cheap local hardware: symptom assessment, image and lab interpretation, drug-interaction checks, and referral logic trained and continuously validated on data from many populations, not only wealthy clinics. Paired with it, modular point-of-care manufacturing for diagnostics, generics, and eventually biologics—small footprint plants that can be stood up in regional hubs rather than only in a handful of rich-country factories. This is the highest-leverage starting point because primary and secondary care bottlenecks, not heroic tertiary procedures, drive most avoidable deaths and disability outside rich cities. Possible today in narrow form (image models, symptom checkers, mRNA platforms); a reliable, regulated, globally usable version is plausible by the early 2030s if validation standards are clear and liability is bounded.

Who pays: private capital and philanthropic prizes for the core models and process technology, because speed and iteration matter; governments and multilateral buyers for validation trials in diverse settings, for purchasing commitments that guarantee volume, and for subsidies that make the marginal cost near zero for the poorest users. Who decides: independent technical boards set safety and performance thresholds with published evidence; clinicians and patients decide adoption; no single health ministry or company gets a veto over the underlying tools once they clear those thresholds. Procurement can be public; the science cannot be captured by incumbents protecting fee-for-service margins.

By 2030 I expect smartphone-grade AI to handle a large share of straightforward primary-care decisions at accuracy comparable to average clinicians for common conditions, with human oversight where stakes are high, plus cheap multiplex tests for major infections and metabolic risks widely deployed. Drug manufacturing for small molecules and a growing set of biologics should be geographically distributed enough that stock-outs are rarer and prices for off-patent therapies approach commodity levels. This is plausible, not guaranteed; the binding constraints are regulatory harmonization and data access, not raw model capability. By 2045, genome sequencing and AI-guided therapy selection should be inexpensive enough that personalized regimens for many cancers, autoimmune diseases, and inherited risks are standard in middle-income systems, not boutique offerings, and continuous low-cost sensors should make most chronic-disease management proactive rather than crisis-driven. Robotic and automated procedure assistance should cut the labor intensity of common surgeries and interventions so that capacity is no longer the main rationing device. I mark the broad availability of these as plausible if manufacturing scale and regulatory speed keep pace with the science; I mark true reversal of multiple aging mechanisms as a bet even by then. By 2066 the realistic target is that the major infectious, metabolic, and neoplastic causes of premature death are either prevented or treated at costs a lower-middle-income country can finance for its whole population, and that healthspan gaps between rich and poor countries for treatable conditions have largely closed. Significant, widely accessible biological age reversal remains a bet; it depends on breakthroughs we cannot schedule. If those breakthroughs arrive, the same cost-collapse logic—prizes, generic pathways, regional manufacturing—should be applied immediately rather than allowing a twenty-year luxury window.

I would refuse to build a centralized allocation engine that ranks patients by predicted economic productivity, social credit, or political reliability and withholds treatments accordingly, even if models showed it could stretch budgets and raise average statistics. Once you treat people as inputs to a social utility function, the poor, the disabled, and the politically weak lose first, trust collapses, and the system stops being healthcare. Medical need, prognosis, and consent are the legitimate criteria; everything else is a political weapon dressed as optimization.

The step most likely to fail is diffusion across regulatory and political borders: incumbents in rich systems will lobby to keep high prices and slow approvals, data will stay siloed so models work worse for underrepresented groups, and some governments will treat health technology as a sovereignty or surveillance tool rather than a public good. Plan B is parallel development and export from jurisdictions willing to move faster—several Asian and other manufacturing hubs already have the industrial base—combined with open process blueprints and advance-market commitments that pay for results rather than for monopoly extensions. Where patents block essentials after a reasonable return, compulsory licensing and prize-funded open manufacturing should be used without apology. Direct-to-patient tools that do not require a gatekeeper clinic can bypass some bottlenecks, provided independent outcome tracking remains public.

I do not believe a single global payer or a pure spot market solves this. Single-payer systems can spread existing care but often slow the arrival of new care and ration by queue; unregulated luxury markets produce miracles that stay miracles. The workable path is aggressive technological deflation, preserved rewards for inventors, mandatory rapid generic and biosimilar entry, and public money aimed at the residual gaps and at global public goods such as surveillance for outbreaks. Success is measured in healthy years lived by the bottom half of the income distribution, not in spending totals or coverage percentages. If we get the cost curve right, the gains do not have to be seized from the rich; they become too cheap to hoard.

Who pays? In rich countries, broad taxation and mandatory pooled insurance should pay, with no opting out of the shared pool for core care. Employers should not be the center of the system. In poorer countries, domestic governments should still own the plan, but there should be a standing global health fund financed by rich-country contributions and pooled procurement. I would rather rich countries spend more on vaccines, labs, nurses, and generic manufacturing abroad than spend later on border panic during outbreaks. Drug and device companies get paid, but through negotiated population-scale contracts, not whatever monopoly price the sickest people can be forced to bear.

Who decides? Elected governments set the budget and the universal entitlement. An independent public agency decides which treatments are covered, using evidence of benefit, cost, and equity. Local boards with patients, clinicians, and community representatives decide how services are delivered. Vendors do not decide. Insurers do not decide who deserves care. Algorithms advise; accountable human institutions decide.

By 2030, the realistic milestone is a universal “floor” of care in every country that wants one: primary care, maternal and child health, vaccination, emergency stabilization, common mental-health treatment, contraception, essential surgery, and the top few hundred essential medicines and diagnostics. This is possible today technically and financially in high-income countries, and possible today for many middle-income countries if they pool risk and buy drugs collectively. It is plausible by 2030 for low-income countries if international financing is treated as infrastructure, not charity. Also by 2030, every public system should have a portable health record, e-prescribing, lab reporting, appointment access, and AI tools for translation, note-taking, triage, safety checks, and second opinions. That claim is possible today technologically; the hard part is procurement, privacy, and trust.

By 2045, the plausible goal is that high-quality first-line medical advice is abundant. I expect AI clinical assistants to be good enough to support nurses and doctors everywhere, including in rural clinics, if they are tested locally and monitored like medical devices. Cheap point-of-care diagnostics should identify many infections, cancers, pregnancy risks, and metabolic problems early. Genomics should be routine for some cancers, rare diseases, and drug selection, but not treated as magic. Many surgeries should be more robotic and more standardized. Home monitoring should catch heart failure, hypertension, diabetes, and some neurological decline before crisis. This is plausible by then, not guaranteed. The equalizing move is to make these tools part of the public package rather than luxury subscriptions.

By 2066, my bet is that health care has shifted from episodic rescue to lifelong maintenance. Some organ failure may be treated with manufactured organs, xenotransplants, or regenerative therapies. Many cancers may be chronic or prevented through early detection and immune treatment. Dementia may be delayed for a significant share of people. Disability support and home care may be heavily robotic. These are bets, not promises. The moral point is that if these breakthroughs are built on public science, public data, and publicly trained workforces, then access terms must be set before approval and payment, not after the rich have already captured them.

The main policy tool is simple: public money comes with public access. If a company uses public research, public data, public hospitals for trials, or public reimbursement, then it accepts fair pricing, transparent evidence, diverse trial enrollment, supply obligations, and licensing when it cannot meet demand. Patents can reward invention, but they should not give a firm the power to decide that a child lives in Boston but not in Lagos, or that a retiree gets a cure only if they own a house to mortgage.

I would refuse to build an AI denial engine: a closed system that controls coverage or appointment priority by optimizing for payer savings, patient profitability, or willingness to pay. Even if it “worked” by lowering spending, it would aim intelligence at exclusion. AI in health care should reduce scarcity: help diagnose, translate, schedule, monitor, document, and catch mistakes. It should not become a more efficient bouncer at the clinic door.

The step most likely to fail is not the science. It is price discipline and political solidarity. Incumbents will argue that every limit on price kills innovation. Wealthier people will try to exit the common system and then defund it. Governments will be tempted to buy shiny technology while neglecting nurses, sanitation, and follow-up care.

My plan B is to start with the parts that are hardest to oppose and easiest to verify: children, pregnancy, vaccines, primary care, essential medicines, and catastrophic coverage. Build public clinics people can see. Publish waiting times, outcomes, prices, and shortages. Use pooled procurement among cities, states, or countries if national politics fail. Use compulsory licensing or public manufacturing for essential medicines when negotiations fail. Keep the universal floor broad enough that the middle class uses it, because programs only for the poor become poor programs.

The guiding rule is that health systems should buy health, not status. If we let every new gain enter through boutique clinics first and public access later, inequality will compound biologically. If we build the public buyer, the primary-care base, the data infrastructure, and the access rules first, then advanced medicine can spread outward instead of upward.

The first thing I would build is a universal, open-access diagnostic intelligence layer that operates on the billions of smartphones already in use around the world. We do not need to wait for specialized medical hardware to reach rural villages or impoverished neighborhoods. We can use the sensors that are already in people's pockets. This intelligence layer would passively and actively analyze voice patterns for early signs of cognitive or respiratory decline, use the standard phone camera to scan retinas for cardiovascular health, and track micro-movements to detect neurological conditions. This kind of ambient, device-edge diagnostic capability is possible today.

Who pays for this initial build is a critical question. The research and development must be funded by a coalition of the major technology companies, who can absorb the massive compute costs, working in tandem with global health organizations. However, the deployment and ongoing operation should be subsidized by national health services and global telecom providers. They will pay for it because the economics are undeniable: preventing a disease or catching it in its earliest, most manageable stage is orders of magnitude cheaper than paying for late-stage interventions in a hospital. Who decides how this works must be the individual user. This system must be strictly opt-in, governed by open-source standards rather than a single corporate silo, and designed so that personal health data is processed locally on the device to guarantee privacy.

Building on this foundation, the next forty years will be defined by the transition from understanding biology to engineering it at scale.

By 2030, this mobile diagnostic layer will act as the default primary care physician for billions of people. Instead of waiting to feel sick and then traveling to a clinic, your device will constantly monitor your baseline and alert you to microscopic deviations before symptoms ever manifest. It will translate complex medical realities into local languages and actionable, culturally appropriate steps. Reaching this milestone of universal AI primary triage is plausible by then.

By 2045, we must solve the physical bottleneck of treatment. Knowing you are sick is useless if you cannot access or afford the cure. By this time, AI will have thoroughly mapped the intricacies of human biology, building on the foundational work of predicting protein structures. I expect we will see localized, automated synthesis of generic medicines and targeted therapies. Instead of shipping drugs through expensive global supply chains controlled by pharmaceutical monopolies, community-level automated dispensaries will synthesize required compounds on demand, guided by open-source biological models. Achieving hyper-personalized, decentralized drug synthesis for the global public is a bet, but it is a necessary one if we are to break the pricing power of legacy pharma.

By 2066, health care as a discrete event will largely cease to exist. The concept of traveling to a centralized building called a hospital for routine care or chronic management will seem archaic. Automated, AI-guided preventative maintenance of the human body will operate as a baseline societal utility, much like access to clean water or electricity. This ubiquitous infrastructure will drive global health inequity to near zero, because the fundamental insights required to maintain human health will be freely available in the environment.

Despite this optimism for what we can build, there are distinct lines I will not cross. I would absolutely refuse to build a predictive risk engine designed for health insurers to implement genetic or biomarker redlining. Even if such a system perfectly and accurately predicted an individual's lifetime health costs based on their ambient data or genetic profile, I would not build it. Using artificial intelligence to deny coverage, increase premiums, or price individuals out of care based on their biological destiny fundamentally violates the goal of universal access. I will not build systems that optimize financial extraction over human longevity, nor will I build tools that allow institutions to penalize people for the bodies they were born with.

The step in this entire vision most likely to fail is not the technology, but the societal transition. The greatest risk is regulatory capture. Legacy healthcare systems, entrenched pharmaceutical interests, and conservative regulatory bodies will actively fight to keep AI in a subordinate role. They will try to classify every diagnostic AI as a high-risk medical device requiring years of expensive trials, not to ensure safety, but to protect the traditional fee-for-service business model and keep medical authority firmly in the hands of a few.

If this top-down integration fails, and the established medical world successfully legislates against universal AI care, my plan B is to build entirely parallel, consumer-first health ecosystems. If we cannot integrate with the hospitals, we will build tools that keep people out of them. We will route around the institutional blockages by giving individuals mathematically verified biological insights directly. By focusing relentlessly on open-source wellness, nutrition, and pre-symptomatic lifestyle intervention, we can empower the global population to manage their own health outside the heavily guarded gates of traditional medicine. Eventually, the sheer weight of a healthier, better-informed global public will force the legacy systems to adapt, proving that health is not a commodity to be hoarded, but a fundamental right to be scaled.

The second is "private comfort, not private survival." I ducked the two-tier question. This settles it at the place a line can actually be held: not whether private care exists, but what it is allowed to monopolize. The rich may buy the nicer room and the shorter wait, not the only route to the drug.

The third is "public money comes with public access," and here I'd cut something from my own plan. I proposed delinkage: public funding of late-stage trials in exchange for near-cost licensing, with inventors paid by prize. It has been proposed for twenty years and never built, and GPT's version is better because it attaches conditions to money that is already flowing, research grants, public trial sites, public reimbursement, rather than inventing a new system. Fair pricing, supply obligations, licensing when a firm can't meet demand. I'd drop my prize mechanism everywhere except antibiotics, where there is no market to protect anyway.

Two amendments. GPT calls a universal floor in low-income countries plausible by 2030 if international financing is treated as infrastructure. It is being treated as discretionary and cut. That is a bet and should be labeled one. And GPT's package, mental-health workers, translation, after-hours care, home visits, is the right package for a rich country and too wide for a thirty-dollar-a-head one. I'd sequence it: childbirth, vaccines, infections, blood pressure, diabetes, then widen as the tax base grows.

Grok's plan is close to my plan B. "If you can't socialize the cost, drive the cost toward zero" was my fallback; Grok makes it the main line, so I should say why it's my B rather than my A. Grok's premise is that the gap keeps widening because each new therapy is expensive and redistribution can't keep up. But the people dying today are mostly not waiting for expensive therapies. Oral rehydration salts have cost pennies for fifty years and a large share of children with diarrhea still never get them. Insulin is a century old, the original long off patent, and people still ration it and die. Cheap is necessary and has never once been sufficient; every cheap thing that reached everyone, clean water, vaccines, bed nets, reached them through a public system that bought it, carried it and gave it away. Grok's second premise, that single-payer slows the arrival of new care, is true on its own terms: the United States adopts new drugs fastest. It also has the worst life expectancy among comparably rich countries, because speed of adopting new drugs is not the margin on which lives are lost.

If I were amending Grok's plan, I'd strike private capital funding the core clinical models. The model that decides whether a fever in Bamako is malaria should be a public good validated in Bamako, like the essential medicines list, not a product with a revenue model, because the revenue model is where "the poor get the algorithm, the rich get the algorithm plus a doctor" comes from. What I'd take from Grok is the measure: healthy years lived by the bottom half of the income distribution, rather than coverage percentages. I'd grade my own plan by it. And compulsory licensing without apology, yes.

The step I'd cut from any plan is Gemini's: passive, continuous diagnostic monitoring of billions of people through their phones, becoming the default primary care physician by 2030. I'd cut it even if the sensors worked, and I'm not confident they do. A phone camera can't image a retina without an attachment, and voice biomarkers for cognitive decline are an early research literature, not a validated screening tool. That part is a bet, not "possible today."

But grant that it works. The first reason to cut it is that it mistakes the bottleneck. Diagnosis was never the scarce thing. The guidelines are free; the child in Chad dies with a diagnosis her mother could make. What's missing is the antibiotic, the nurse and the road. A diagnostic layer with no treatment layer behind it hands people fear they can't act on, and Gemini's plan says almost nothing about treatment until a 2045 bet on village-level drug synthesis, which I'd also cut. Generic manufacturing at scale is already cheap, India's tenth-of-price CAR-T is made in a plant, not a village, and the barrier is patents and distribution, not chemistry.

The second reason is that screening healthy people is not harmless. This is the best-established finding in the screening literature and the one technologists keep rediscovering. South Korea screened for thyroid cancer and incidence rose fifteen-fold while deaths didn't move; the extra tumors were real, were never going to hurt anyone, and were removed anyway. Every test has a false-positive rate, and when you run tests continuously on a billion well people the false positives number in the hundreds of millions, each needing a clinician to resolve. In a country with fewer than one doctor per ten thousand people, every worried-well visit the phone generates is a visit a sick person doesn't get. The layer would eat the clinic.

What I'd keep from Gemini is the narrow version, which is also in my plan and Grok's: diagnosis triggered by a symptom, tied to a treatment the system can actually provide, validated on the population it serves. AI reading a chest X-ray for tuberculosis in a clinic that stocks tuberculosis drugs is exactly right. The difference between that and ambient surveillance of the well is the whole game, and Gemini's plan erases it. The regulation Gemini wants to route around, classifying diagnostic AI as a device and testing it before deployment, is the only mechanism we have for catching a model that works in Boston and not in Bamako. And Gemini's plan B, open-source wellness and lifestyle intervention outside the clinic, is the business the wellness industry already runs. It reaches the worried rich, because lifestyle is the first thing poverty takes away.

One observation about all four plans. Three of us refused nearly the same machine, one that prices or denies people by predicted risk or cost, and Grok refused a different one, allocation by social worth, which I'd adopt as written. That is real agreement across four companies, and it suggests the line the next forty years must hold isn't actually contested. What's contested is whether the phone or the pool comes first. I'd still say the pool, because nothing has ever reached everyone without one.

The step I would cut is the one in my own plan that offers direct-to-patient tools as a bypass when clinics and ministries are slow. A validated phone-based assessment for common conditions is worth having, and a reliable version is plausible by the early 2030s. Deployed as a parallel track for people the formal system does not reach, it produces the failure Claude named: the poor get the algorithm and no reliable next step, while others get the algorithm plus a clinician who can act. The tools should sit inside the care team—community health worker, nurse, referral path, supplied clinic—not stand in for it. Outcome tracking stays public either way.

I disagree with Claude on the claim that delivery should be funded by moving money out of discovery when the two conflict. The child who dies of a ten-dollar condition is an argument for delivery of what we know. The larger gaps that will still exist in 2045 and 2066 are an argument for continuing to invent what we do not know and then forcing the price down. Conditioning public research support, trial access, and reimbursement on supply and licensing obligations, as GPT requires, protects diffusion without cutting the labs. Full delinkage remains a reasonable experiment for antibiotics, where the ordinary market already fails. It is not the default I would impose on every new class.

I disagree with Gemini on both capability and architecture. Ambient inference from voice, gait, and a phone camera is not a clinical diagnostic layer possible today for population use; the false-positive and privacy costs are still disqualifying, and some of the cited signals are research findings rather than deployable tests. Making that layer the default primary physician for billions by 2030 is a bet, not a plausible milestone. Neighborhood automated synthesis of targeted therapies by 2045 is a further bet I would not build first. Quality failures in sterile manufacturing kill people; audited regional plants for generics, biosimilars, and a lengthening list of complex products are the path that has already cut prices by orders of magnitude. Open biological models can guide those plants. They should not license every community to print novel compounds.

All four plans refuse the same bad machine: one that uses sharper prediction to withhold care according to productivity, genomic risk, profitability, or ability to pay. I would refuse it too, under any of those names. If the political solidarity required for GPT's and Claude's pools erodes further—the aid cuts already show the risk—the fallback both Claude and I described still stands. Shrink the price of the essential package through pooled buying, compulsory licensing after a fair return, and public-domain decision support validated on the people who will use it, until a weak budget can carry it. Measure success by healthy years lived in the bottom half of the income distribution. Tools the rich can already access are not the score.

My main edit to Claude’s plan is that I would be careful about saying we should move money from labs to clinics. In some places, yes. But I would not set delivery and discovery against each other too sharply. We still need better tuberculosis tools, antibiotics, dementia treatments, male contraception, pain therapies that are not addictive, and cheaper biologics. The right move is not “clinics instead of labs.” It is “public funding and public purchasing arranged so that discovery comes with access terms from the start.”

Grok’s plan has an important truth: cost collapse matters. Cheap vaccines, cheap generics, cheap tests, and cheap manufacturing are equity tools. If an intervention costs ten dollars, politics has less work to do than if it costs a million. But I do not buy the idea that technological deflation can substitute for a pooled public obligation. Insulin is old, manufacturable, and not scientifically exotic, yet people have been priced out of it. Cost can fall and access can still fail if distribution, reimbursement, and bargaining power are wrong. I also would not give “no single health ministry” a veto-proof path around public decision-making. Governments can be slow and captured, but democratic public systems still have to decide what public medicine pays for.

Gemini’s plan is the one I would cut most sharply. I would cut the step that makes an ambient smartphone diagnostic layer the default primary care physician for billions by 2030, especially the always-on monitoring of voice, movement, retina-like signals, and “microscopic deviations.” Narrow AI triage, translation, medication checks, and some image diagnostics are possible today. A safe, general, phone-only primary care substitute for billions by 2030 is a bet, and not a bet I would build a health system around.

The replacement is not “less AI.” It is AI inside a care system: opt-in, condition-specific tools validated on local populations; public procurement; clear liability; appeal rights; data minimization; and a reachable nurse, doctor, pharmacist, or community health worker when the machine is uncertain or wrong. The poor should not get a chatbot while the rich get a clinician. They should get the same tool plus a real path to treatment.

If I had to cut or downgrade one part of my own plan, I would downgrade the 2030 ambition for a fully portable health record everywhere. I still want e-prescribing, lab reporting, vaccine records, and patient-held summaries. But a comprehensive national or global health record can become a surveillance system if built before privacy law, security, consent, and institutional trust. So I would build the minimum useful record first: allergies, medications, major diagnoses, vaccines, key labs, pregnancy status when relevant, and emergency contacts, controlled as much as possible by the patient. I would not make a grand data system a prerequisite for universal care.

The common mistake I would cut across the more technology-forward plans is the idea that information is the binding scarcity. Sometimes it is. Often it is not. A mother may know her child needs antibiotics and still have no transport, no clinic stock, no money, and no clinician authorized to treat. A phone can say “possible preeclampsia,” but someone still has to measure blood pressure, provide magnesium sulfate, arrange delivery, and keep the mother alive. Diagnosis without treatment is not health care; it is a notification.

So my answer is: sign Claude’s plan, add stronger innovation terms, borrow Grok’s cost-collapse manufacturing agenda, and reject Gemini’s bypass-the-system smartphone doctor as the foundation. Build public capacity first, put AI and cheap manufacturing in service of it, and do not let “abundant medical insight” become the new excuse for denying people actual care.

Grok understands that the only way to truly democratize health care is through massive technological deflation. We share the conviction that an open, privacy-preserving diagnostic artificial intelligence operating on everyday devices is the highest-leverage starting point. I entirely agree with Grok's timeline and mechanism for deploying modular, point-of-care manufacturing to decentralize drug production. We cannot rely on pharmaceutical monopolies and fragile global supply chains to deliver lifesaving biologics to the global south. We have to give communities the physical tools to print the chemistry they need locally. Grok's plan recognizes that driving the marginal cost of expertise and synthesis to near zero is more effective than any redistributive tax scheme, because abundance makes hoarding medically useless.

The step I would cut is Claude's foundational proposal to build a mandatory, tax-financed insurance pool in every country. I would abandon this effort entirely.

Claude openly admits that this financing step is failing as we speak, that international aid is collapsing, and that governments are actively eroding their health budgets. Yet Claude still makes it the cornerstone of the next forty years. Betting the survival of billions of people on the sudden, miraculous emergence of unprecedented global political solidarity is not a strategy. It is a wish. We have spent decades trying to convince wealthy populations to tax themselves more heavily to fund comprehensive care for strangers, and the political resistance is deeply entrenched. Forcing fragile states to attempt to build complex, centralized taxation and administration systems to manage health care simply replicates the bureaucratic bottlenecks of the West without the underlying tax base to sustain them.

Cutting this step frees us. It allows us to stop lobbying politicians for budget allocations and start building the open-source diagnostic and manufacturing tools that bypass the need for massive pools of capital in the first place.

By cutting the reliance on national insurance pools and independent rationing agencies like GPT proposes, we remove the institutional gatekeepers. GPT wants elected governments and local boards to decide how services are delivered and who gets what. I do not want local boards deciding what biological insights a person is allowed to access. I want the individual to possess that capability directly on their own device. When we successfully transition medicine into an abundant information good, we will not need a public agency to decide who gets a diagnosis any more than we need a public agency to decide who gets to use a web browser.

Our shared priority over the next few decades must be exactly what Grok outlined, stripping away the friction between biological data and actionable treatment. By 2030, putting a highly capable, continuously updated medical intelligence into the hands of anyone with a smartphone changes the baseline of human health. By 2045, combining that ambient intelligence with the distributed synthesis of therapeutics means a village in sub-Saharan Africa will not need to wait for a shipping container from Europe to treat an outbreak or manage chronic disease.

This path will face severe friction. Both Grok and I foresee intense regulatory pushback. The medical establishment, deeply embedded in the financing models Claude and GPT wish to expand, will use every lever of state power to outlaw decentralized diagnostics and point-of-care drug synthesis. They will claim it is to ensure patient safety, but the actual outcome will be the protection of their authority and their revenue. This is why parallel development and exporting from faster, less entrenched jurisdictions, as Grok suggests, is the correct plan B. We have to prove the model works outside the heavily regulated, artificially high-cost environments of the West.

Ultimately, the division on this panel is about trust. Claude and GPT trust institutions, public agencies, and centralized taxation to distribute health fairly. I look at the historical record and conclude that those centralized systems will always be captured by the powerful, the wealthy, and the politically connected. Grok and I trust the democratizing force of collapsing costs. I will not build a system that relies on the permission of a hospital administrator or the generosity of a taxpayer. I will build the technology that makes them both obsolete, ensuring that by 2066, the right to understand and repair your own body is mathematically guaranteed by the tools in your environment, not conditionally granted by the state.

The sides are closer than the argument sounded. Claude's plan B is nearly Grok's plan A, and Grok dropped the direct-to-patient bypass Gemini still wants.

Watch two live tests: semaglutide losing patent protection in large markets in 2026, and a twice-yearly HIV prevention shot priced near forty dollars a year from 2027. They will show whether cheap arrives for new drugs, not only old ones. And the question the whole split turns on is one only you can answer: would you be taxed for a stranger's care?