Two numbers describe two markets that took off in 2025. Neither one looks like it has anything to do with your fridge. Both of them decide what ends up on the shelf, and both decide what you pay later to fix what broke because of what you ate.
Two molecules booked $71 billion. Semaglutide and tirzepatide, the GLP-1 drugs, brought in more than double the combined revenue of OpenAI and Anthropic. Tirzepatide is now the best-selling drug on earth.
In the same year, four hyperscalers — Amazon, Microsoft, Google, Meta — spent $410 billion on capex, money laid out to build the thing before anyone has paid for it, and have guided to roughly $725 billion for 2026. That is capital pointed at an AI market whose revenue is still less than half of what two molecules collected.
Those are not the same kind of number. One is revenue. The other is capital expenditure. One is money that arrived. The other is money that left.
We know AI is different. The comparison still tells you something.
The GLP-1 number is proof: demand for reversing metabolic disease runs so deep that one intervention became the biggest product in pharmaceutical history.
The AI capex number is conviction: money spent today against demand somebody is forecasting.
We are not making the case against AI. The build-out is mostly right. We are making the case that the largest mispricing on the board is in health, and not GLP-1 itself, but the infrastructure that makes GLP-1 unnecessary: the spending that would eliminate the $2 to $3 trillion the United States lays out on chronic disease every year.
Collected from customers
Spent against a forecast
The penetration argument
Every AI thesis carries adoption risk. Every one. The model gets better, and then a human being has to change what they do, a company has to change a process, a budget has to move. That is the whole ballgame for the $410 billion, and it is why the debate is about seat counts and token consumption and whether the enterprise pilots convert.
Food does not have adoption risk. Food has 100% market share.
Three times a day, 365 days a year, from birth to death, every human being on earth is a customer. No trial. No onboarding. No procurement cycle. No switching cost, because the switch already happened, every day, forever. Americans spent $2.58 trillion on food in 2024. They spent $5.3 trillion on healthcare. Put those two together and you are past a quarter of GDP, running through a system where food is the gateway drug for healthcare. The $2.1 trillion Americans now spend on wellness sits alongside that, and we do not add it in: healthy eating and preventive medicine are two of the eleven sectors inside it, so it overlaps both lines above rather than extending them.
AI is a TAM-creation story, total addressable market. It has to conjure demand that does not exist yet.
Food is Health is a mix-shift story inside a market that is already fully penetrated.
One hundred percent share means zero organic growth. Food's unit growth is population growth, half a percent a year. Every dollar has to be taken from an incumbent with enormous scale, entrenched shelf position, and a century of formulation science aimed at exactly this fight. Consumer staples share wars are historically brutal and slow, and anyone selling you a food thesis on TAM math is selling you something.
That is also why the capital does not come. Full penetration means there is no growth story to fund, so the money goes to chasing trends, or better yet creating them. And here is the part most models cannot hold: the return on changing what is on our plates never lands on the food receipt. It lands in the total cost of healthcare, in the claims that never get filed. Revenue in one industry, payoff in another, and there is no growth model built to price that.
The externalities of cheap calories are not priced upstream. They land downstream, in healthcare costs and premiums, which is how a household ends up financially locked into the tractor beam of chronic disease.
No trial, no onboarding, no switching cost.
each budget line, one at a time.
They were never sixty-one conditions
In January 2025, Washington University followed 215,970 patients on these drugs, matched against controls, across 175 conditions. Risk fell on 42. It rose on 19.
Sixty-one conditions moved on one molecule. Even the nineteen that moved the wrong way make the same point as the forty-two that improved: this is one system wearing sixty-one names.
A hormone is a chemical messenger that travels in the blood, so it reaches every tissue the blood reaches. Nothing about the molecule changes between tissues. Only the cell receiving it. GLP-1 receptors sit on the pancreas, in the hypothalamus, on the stomach lining, and in the brain's reward circuitry. The drug does not report to a department. It goes everywhere at once.
Healthcare is not designed that way. It is built by organ, by sub-system. Endocrinology, cardiology, nephrology, pulmonology, hepatology, addiction. Separate boards, separate journals, separate clinics, separate billing codes. That structure came from how you train a doctor, how you pay one, and how you bill a symptom. It never came from human health.
Now read the approval sequence with that in mind. Diabetes. Then obesity in 2021. Then a 20% cut in major cardiovascular events. Then 24% on kidney outcomes. Then sleep apnea. Then liver disease. Addiction is in the data and waiting on confirmatory trials.
That is not a drug pipeline. It is an inventory of the American sick-care economy, and each new approval crosses off another line item.
Insulin and continuous glucose monitoring. Bariatric surgery. Cardiac devices and stents. Dialysis. CPAP and hypoglossal nerve stimulators. Transplant. Rehab. Every one of those is a public company or a division of one, and every one of them has revenue that is a direct function of how many Americans stay sick.
Read the totals as exposure, not arithmetic. Categories overlap (a diabetic heart-failure patient is counted in both), vintages span 2000–2026, and the societal figures lean on value-of-life accounting. Every anchor is labeled with its scope and year, and the toggle separates them.
One by one, GLP-1 breaks entrenched markets
It took one afternoon.
On the day Novo said the kidney trial had been stopped early for benefit, DaVita and Fresenius each fell about 17%. Dialysis runs about 1% of the entire federal budget, two firms hold most of it, and three quarters of kidney failure traces back to diabetes and high blood pressure. A drug that slows kidney decline shrinks the population those companies exist to serve.
Nobody had to win that argument. The stock settled it before lunch.
Then the market went back to modeling next quarter's treatment volumes.
The repricing of healthcare is being ignored. No one can wrap their mind around the idea that chronic disease might actually go away. A 2040 thesis does not pay a 2026 bonus. It is why our favorite unicorns feel they can wait to innovate around human health: the fire drills of chronic disease will always keep running hot.
Health and food markets stay mispriced
GLP-1 is the acute version of an upstream intervention. Food is the chronic one, the input that decides whether the metabolic failure happens at all. If one molecule hitting one receptor moves sixty-one conditions, the substrate the body runs on for eighty years is not a smaller lever. It is a slower, larger one, and it is the one that matters most.
But notice what the DaVita afternoon did not do. The value that left dialysis did not flow to food. It vanished into savings nobody owns.
A mispricing needs a reason to persist, or it is not a mispricing, it is your error. Here are five, and they are all structural.
Duration.
The effect of food on health compounds over fifteen to twenty years. The average holding period for US equities is under a year. Nobody gets paid for this.
There is no ticker for the beneficiary.
You can buy stock in Lilly. You cannot buy “the household saves twelve thousand dollars.” The savings from food-is-health accrue to families and to Medicare. Neither one has a clean shareholder payout. When the value of a change lands on entities with no securities attached, the market has no instrument to express it, so it expresses it late and through proxies.
Sector coverage is aligned to System B.
Sell-side covers this as a pharma story. Pharma analysts model Lilly against Novo. Medtech covers Inspire and ResMed. Staples covers Nestlé and the basket data. Managed care covers the medical loss ratio. Nobody owns the cross-sector call, and the cross-sector call is the entire thesis.
The value of health does not show up. It is an absence.
There is no revenue line for the heart attack that did not happen, the clinic that was not built, or the transplant that was not needed. Markets price cash flows, not avoided ones. Avoided cost is real money and it is invisible to a screen.
AI has a capex cycle and this does not.
That $410 billion buys turbines, transformers, GPUs, concrete, and cooling, which creates a visible chain of suppliers with order books you can model. Momentum needs something to attach to. Food-is-health has no capex cycle, so there is no chain of visible winners, so there is no flow. It is not that the return is worse. It is that the return has nowhere to land.
The household ledger
A family of four on an employer PPO costs $35,119 a year in healthcare. Not the premium, the total: employer share, employee share, out of pocket. The employer share is foregone wages, so the household pays all of it. Food for that family runs about $20,000. Call it $55,000 a year for food and health together.
Now take out the diet-driven chronic disease. Nobody publishes a clean number for its share of a commercial family's claims, so hold it as a range. Call 35% of claims diet-driven, an estimate built from components rather than a published figure, and the health bill drops to about $22,800. Call the share half instead, and $17,600.
Everyone assumes healthy eating costs more. The figure behind that belief is $1.50 per person per day, from a 27-study meta-analysis. It is measured per 2,000 calories. So is nearly every study in the literature.
USDA priced 4,439 foods three separate ways and found the answer changes with the denominator. Per calorie, produce looks expensive. Per 100 edible grams and per average portion, grains, vegetables, fruits and dairy come in cheaper than most proteins and most less-healthy foods. On every metric except price of food energy, healthy food costs less.
Pricing food per calorie is the right method only if your problem is getting enough of them.
A Big Mac meal averages $8.13, so two of those plus two Happy Meals is about $29 with tax for a family of four. Salmon for four at $10 a pound, with rice, frozen broccoli, and olive oil, is $19.80. Chicken thighs make it $8.80. Lentil soup and bread, $5.50.
The most expensive healthy protein in the store beats McDonald's by a third.
So food as health is not a trade-off between food and health. On cost it is a straight win on both lines. Roughly $55,000 today, roughly $38,000 to $43,000 in a whole-food, low-chronic-disease household. Twelve to seventeen thousand dollars a year, per family, recurring.
AI can sell you more intelligence as a single service. Nobody sells food and health as a bundle. The grocer keeps the food margin. The insurer keeps the claims savings. Different companies, no contract between them, and the family changes jobs every three years while the payoff lands on Medicare in 2045.
What survives as the real barrier is not money. It is time. That salmon dinner is 35 minutes with cleanup, and McDonald's is six minutes and no dishes. At median wage a family cooking nightly is spending something like $7,000 a year in household labor, which is larger than any grocery gap anyone has ever measured. That is the actual price of eating well, it is why the affordability story survives grocery prices that say the opposite, and it is the one input in this entire model that automation attacks directly.
for the rest of their lives
What we are willing to fund
Big Tech put $410 billion into AI infrastructure last year.
NIH nutrition research got $2.2 billion. That is 4.07% of NIH, down from 4.90% in 2019.
One hundred eighty-four to one, and the small side is shrinking.
And yet the prize, on our assumptions, is $1.5 trillion a year of diet-driven US chronic disease spending, recurring, which capitalizes somewhere between $7.5 and $15 trillion depending on your discount rate. We hold that as the conservative read, and plenty of people put the annual number above $2 trillion. It is a constructed estimate rather than a published figure, which is why the next figure lets you set the share yourself, lower than ours included.
We are not short on knowledge about what causes chronic disease. We have known for forty years. We are short on anyone whose balance sheet improves when chronic disease goes away.
So who could that be?
The blue one is six pixels wide.
The trade
The short side is the easy half, and it is the half that will hurt you first.
Anything whose revenue is a direct function of the prevalence of diet-driven chronic disease. Dialysis. Sleep apnea devices. Insulin and glucose monitoring. Bariatric programs. The long tail of ambulatory surgery centers and specialty clinics built around metabolic failure. These are good businesses with real cash flows and defensible positions, and that is precisely the problem, because a business whose moat is a growing sick population has an asset that is now in run-off and does not know it yet.
Two warnings on that side. First, several of these get a near-term volume boost. Patients who were too heavy for a knee replacement become eligible after they lose weight, and orthopedics has said so out loud. Losing on the way to being right is still losing. Second, discontinuation is high, real-world persistence at one year is poor, and weight comes back. The drug is not the thesis. It is the proof of concept.
The long side is harder to name, which is exactly why it is cheap.
Who makes money when a hundred million people get ten more good years?
Costco is the archetype, and not because of the stock. Because of the structure. It is the rare entity that owns the food margin and the recurring membership relationship at the same time, with a vertically integrated private label it can reformulate at scale, and a pharmacy attached. It is the closest thing in public markets to the bundle that does not exist. We wrote about whether Kirkland could end the chronic disease epidemic, and the answer is that Costco is one of maybe five organizations on earth with the standing to try.
The rest of the long side sorts into three buckets.
The supply chain that has to exist if the mix shifts: whole-food producers, protein, produce, cold chain, and the measurement layer that proves a food did something, because you cannot patent broccoli but you can patent the sensor that scores it.
The consumption of years 65 to 85 lived in health rather than decline: travel, hospitality, leisure, education, second careers. That is a demand curve nobody is modeling because the years currently get spent in a waiting room.
And the asymmetry nobody talks about at all. Life insurers pay later when people live healthier. Annuity writers and defined-benefit pensions pay longer. Same demographic input, opposite signs, and one of those two groups is not marking it.
Then the macro version, which is the one that should actually get an allocator's attention. Every developed economy's fiscal projection rests on a dependency ratio. Healthspan changes the denominator. Ten more productive years per worker, in a world of shrinking workforces, is not a consumer trend. It is a labor supply event.
Revenue that depends on chronic disease staying
- Dialysis
- Sleep apnea devices
- Insulin, glucose monitoring
- Bariatric programs
- Metabolic specialty clinics
Revenue from thirty years handed back in health
- Whole-food retail, membership models
- Travel, leisure, education, ages 65–85
- Life insurance (pays later)
- Labor supply in shrinking workforces
- The layer that measures what food did
Long biotech is long more medicine, which puts it on the same side of the ledger as the dialysis operator.
The index that does not exist
Look at what is currently sold as longevity exposure. The funds are long biotech.
Long biotech is long more medicine. It is a bet that we get better at treating the sick, which puts it on the same side of the ledger as the dialysis operator.
The actual healthspan trade is the opposite. Long the absence of medicine and the consumption that follows it. Short the infrastructure of managed decline, long the thirty years of life that get handed back.
Nobody has built that index. The inputs are all public, the constituents are all liquid, and the reason it does not exist is the same reason the household bundle does not exist. The benefit has no owner, so no product manager can find the fee.
What kills this
State it plainly, because a thesis without a kill condition is a pitch.
Discontinuation rates stay bad and the population effect never materializes. Compounded and generic semaglutide collapses the economics before the cascade finishes running. The near-term surgical volume boost lasts longer than the position does. Food behavior at population scale has never been demonstrated to change, ever, in the direction of health, and pointing at a drug that works does not prove that a food system can.
And the ordinary one. This is right and fifteen years early, which for almost everyone is the same thing as wrong.
Bottom line
The AI number is $410 billion of spending against demand somebody is forecasting. The food and health number is $7.9 trillion of spending against demand that shows up three times a day and has never missed.
One of those has an adoption problem. The other has an ownership problem.
Adoption problems get solved by better products. Ownership problems get solved by somebody building the entity that can hold both sides of the household long enough to collect, and that is a much smaller list of people, most of whom are reading something else right now.
The food was never the expensive part.