Vector 1
AI Strategy & EconomicsHealthcare

The Price of an Hour: How AI Could Make Life Cheaper, Healthier and Less About Work

Brian J. Curry
Brian J. Curry
October 17, 2026 · 18 min read

A practical model for where AI will lower the cost of living and buy back time, where it won’t, and the numbers to watch.

Since 2000, the price of a television in the United States has fallen 98 percent. Over the same stretch, the price of hospital services rose 275 percent, college tuition 196 percent and child care 185 percent. Average hourly pay for production and nonsupervisory workers rose about 130 percent. So a TV now costs about 1 percent of the hours of work it did in 2000, while a day in the hospital costs 63 percent more hours.

That split is the most useful lens I know for thinking about AI and abundance. The things that got cheap were made by machines on learning curves. The things that got expensive are made mostly of human hours that did not get more productive. Health care, education, care for children and the elderly, legal help, much of government: these are where household budgets and household stress now live.

The optimistic case for AI is not that it makes gadgets cheaper still. It is that, for the first time, the human-hour-heavy services are exposed to the same kind of cost curve that flattened electronics. If that happens, it changes three things people actually care about: what life costs, how healthy it is, and how much of it has to be spent at work.

This piece builds that case with numbers rather than adjectives. It sets out a simple model you can run yourself, walks through the evidence in medicine, science and energy, works through the math of prices and leisure in Python, and then spends real time on where the optimistic story breaks.

It is also a companion to my Autonomy Economy series. Part I proposed a full economic architecture for a world where work is optional, built on a Universal Autonomy Dividend and a large drop in consumer prices. Later installments modeled the rent extraction that would capture those gains first, and a Sovereign AI Wealth Fund to route them to people instead. Those pieces argued for the destination. This one checks the road: which of the price declines that framework depends on are already visible in the data, which ones AI can plausibly deliver, and which ones it cannot touch without policy.

The frame: the price of a human hour

In the 1960s the economists William Baumol and William Bowen explained why a string quartet costs more every decade even though it plays the same music with the same four people. Wages rise across the whole economy because factories and offices get more productive. The quartet has to pay those wages to keep its players, but it cannot play faster. Its price climbs relative to everything that can.

The cleanest way to see this is to measure prices in hours of work rather than dollars. If p is the price of one unit of a good or service and w is the hourly wage, the time price is:

In a sector where labor is the main cost, the price tracks unit labor cost: p ≈ w / a, where a is output per hour of work in that sector. Substitute and the wage cancels:

That one line is the whole argument of this piece. The time price of anything depends on the productivity of the people who make it, not on how rich the economy is. A TV’s time price collapsed because each hour of factory and design labor produced vastly more television. A hospital stay’s time price rose because an hour of nursing still delivers roughly an hour of nursing.

Run the 2000 to 2025 data through it and the pattern is stark. Wages grew 3.29 percent a year. Hospital services grew 5.28 percent a year, so their time price rose 63 percent. That gap, about 1.9 percent a year, is the rate at which the effective hours behind a unit of hospital care grew, through more staff per patient, more administration and more complexity. The overall CPI basket did better: its time price fell 16 percent, carried by goods.

For the rest of the article, a in the formula above is the variable that matters. AI is interesting for living standards to exactly the degree that it raises a in the sectors where a has been stuck. The chart below shows where those sectors are.

Everything above the zero line is a sector where human hours did not get cheaper to produce. Those are the sectors this article is about.

Medicine: raising a where it has been stuck longest

Health care is the largest test of the frame, and it is also where the early evidence is strongest. Three results show AI raising output per hour of skilled human work in different parts of the pipeline.

Screening. The MASAI trial in Sweden randomized more than 105,000 women to standard double reading of mammograms or to AI-supported reading, where software triaged exams and flagged suspicious areas. Results published in The Lancet Digital Health in February 2025 found the AI arm detected 29 percent more cancers (6.4 versus 5.0 per 1,000 women screened) with no rise in false positives, while cutting radiologists’ screen-reading workload by 44 percent. In the frame’s terms, that is a better product with roughly 0.56 of the hours. Few productivity results in any sector look like that.

Discovery. AlphaFold, whose creators shared the 2024 Nobel Prize in Chemistry, has put predicted structures for more than 200 million proteins into a free database used by over 3 million researchers in more than 190 countries. A structure that once took a doctoral student years of lab work is now a lookup. Downstream, Insilico Medicine reports that its AI platform took drug programs from project start to preclinical candidate in 12 to 18 months on average, against 2.5 to 4 years for traditional discovery (a company figure). Its drug rentosertib, aimed at a target that its AI identified for idiopathic pulmonary fibrosis, completed a 71-patient, 12-week Phase IIa trial published in Nature Medicine in June 2025. Patients on the 60 mg daily dose gained a mean 98.4 mL of lung capacity (FVC) while the placebo group lost 20.3 mL, a gap of about 119 mL.

Frontline work. In a study of 5,179 customer support agents, a generative AI assistant raised issues resolved per hour by 14 percent on average and by 34 percent for the newest and least skilled workers. Support desks are not hospitals, but the shape of the result matters for care work: the biggest gains went to people at the bottom of the experience curve, which is exactly where staffing shortages in nursing, home care and primary care bite hardest.

The caveats are real and belong right here. A 71-patient, 12-week trial is a signal, not a cure; most drugs that look good in Phase II fail in Phase III. Discovery is the cheap part of drug development; clinical trials are the expensive part, and AI has not yet shortened them much. And the MASAI result comes from one well-run national screening program. None of that undercuts the direction. It sets the pace.

Science, weather and energy: cheaper answers to physical problems

The second channel is less direct but probably larger over decades: AI lowering the cost of finding answers in the physical sciences, which then lowers the cost of things built on those answers.

Weather is the cleanest example because the output is measurable every day. Google DeepMind’s GenCast, published in Nature in December 2024, beat the European Centre’s ENS ensemble, long the best operational system, on 97.2 percent of 1,320 forecast targets, and on 99.8 percent at lead times beyond 36 hours. It produces a 15-day ensemble member in about 8 minutes on a single TPU, where traditional ensembles run for hours on supercomputers. Better extreme-weather forecasts translate directly into fewer crop losses, earlier evacuations and cheaper grid planning. A forecast is a public good whose marginal cost just fell by orders of magnitude.

Materials science shows both the promise and the need for skepticism. DeepMind’s GNoME predicted about 2.2 million new crystal structures in 2023, 380,000 of them flagged as stable, and outside labs went on to synthesize 736 of them. But a 2024 critique in Chemistry of Materials by Anthony Cheetham and Ram Seshadri argued that many entries were trivial variants of known compounds and that none had demonstrated useful properties. The fair reading: AI has made the search for candidates cheap, while the bottleneck has moved to synthesis, testing and judgment. Cheap candidates are still valuable. They are not the same as cheap batteries.

Energy is where the cost-curve logic is most familiar. Solar module prices have followed Wright’s law for half a century: each doubling of cumulative production cuts cost by a fixed fraction, the learning rate. The industry roadmap ITRPV puts that rate at 26 percent since 1976, and modules sat around 9 cents per watt-peak at the end of 2025. In formula form, with x as cumulative production:

A 26 percent learning rate gives b ≈ 0.434: three doublings cut cost to about 41 percent of where it started, ten doublings to about 5 percent. AI’s role here is less about the panels than the system around them. The IEA estimates that AI-based fault detection could shorten grid outages by 30 to 50 percent and that AI grid tools could free up to 175 GW of transmission capacity without new lines. Transmission, not generation, is now the binding constraint on cheap power in many markets.

Food sits downstream of all three: weather, materials (fertilizer, packaging, refrigeration) and energy. There is no single headline result for AI in agriculture comparable to MASAI or GenCast, and I would distrust anyone who offers one. The honest claim is that food prices are mostly energy, logistics and labor, and AI is now working on each input.

Prices: the fastest cost curve ever measured, and what it touches

The input AI itself provides, machine cognition, is getting cheaper faster than anything in the CPI. Epoch AI measured the price of reaching a fixed level of model performance across six benchmarks and found declines of 9x to 900x per year, depending on the milestone, with the fastest drops in the most recent year. Matching GPT-4 on PhD-level science questions (GPQA Diamond) got 40x cheaper per year. On the MMLU benchmark, the cheapest GPT-4-level model went from $37.50 per million tokens in March 2023 to $0.18 in February 2025, a factor of about 208 in under two years.

TVs fell 14 percent a year for 25 years and that was enough to take them from a major purchase to an afterthought. Cognition is falling at least an order of magnitude faster. The question for living standards is how much of the price of a hospital visit, a semester of college or a legal filing is cognition that can ride this curve, and how much is something else: a licensed human’s time, a building, land, liability, regulation.

The model from the frame section makes that question concrete. If a service’s price tracks unit labor cost, its time price after n years is the inverse of its cumulative productivity gain. The script below computes time prices from the BLS data and projects them under different productivity scenarios. It is self-contained and runs with the standard library.

Time prices: hours of work needed to buy a good or service, and how
sector productivity changes them.

Data: BLS CPI category changes, Jan 2000 to Sep 2025 (via Visual
Capitalist); average hourly earnings, production and nonsupervisory
workers (FRED AHETPI): $13.74 in Jan 2000, $31.56 in Sep 2025.

import math

WAGE_2000, WAGE_2025 = 13.74, 31.56
YEARS = 25 + 8 / 12  # Jan 2000 to Sep 2025

PRICE_CHANGE = {  # percent, 2000 to Sep 2025
    "Hospital services": 275,
    "College tuition and fees": 196,
    "Child care": 185,
    "Housing": 111,
    "Food and beverages": 104,
    "All items (CPI)": 92,
    "Clothing": 2,
    "Toys": -74,
    "Computer software": -75,
    "TVs": -98,
}

def time_price_change(pct_price: float) -> float:
    """Change in hours of work needed to buy one unit, in percent."""
    wage_ratio = WAGE_2025 / WAGE_2000
    return ((1 + pct_price / 100) / wage_ratio - 1) * 100

def annual_rate(total_pct: float, years: float = YEARS) -> float:
    return ((1 + total_pct / 100) ** (1 / years) - 1) * 100

def project_time_price(wage_growth: float, sector_productivity: float,
                       years: int) -> float:
    """Baumol model: p = w / a, so the time price p / w = 1 / a.
    wage_growth cancels out, which is the point."""
    w = (1 + wage_growth) ** years
    a = (1 + sector_productivity) ** years
    return (w / a) / w

def years_to_hours(h_now: float, h_target: float, g: float) -> float:
    """Years until annual hours fall from h_now to h_target if all
    productivity growth g is taken as leisure instead of income."""
    return math.log(h_now / h_target) / math.log(1 + g)

if __name__ == "__main__":
    print(f"Wages: {annual_rate((WAGE_2025 / WAGE_2000 - 1) * 100):.2f}%/yr")
    for name, pct in PRICE_CHANGE.items():
        print(f"{name:26s} price {pct:+5d}%  ({annual_rate(pct):+.2f}%/yr)"
              f"  hours of work {time_price_change(pct):+6.1f}%")

    print("\nTime price of hospital care after 20 years:")
    for a in (0.0, 0.02, 0.04):
        m = project_time_price(0.03, a, 20)
        print(f"  sector productivity {a:.0%}/yr -> {m:.2f}x today")

    h_now, keynes = 1789, 15 * 52
    for g in (0.02, 0.03, 0.04):
        print(f"g={g:.0%}: {years_to_hours(h_now, keynes, g):.0f} years "
              f"to a 15-hour week; after 20 years: {h_now / (1 + g) ** 20:.0f} h/yr")

The hospital projection is the line to look at. With no productivity growth in the sector, hospital care costs the same hours of work in 20 years as today, which would already beat the last 25 years, when its time price rose 63 percent. At 2 percent a year of sector productivity growth, the time price falls to 0.67 of today’s. At 4 percent, it falls to 0.46: care would cost less than half the work it costs now. For scale, a 44 percent cut in hours like MASAI’s, if it applied to a whole hospital rather than one task, would equal about 15 years on that 4 percent path.

This is also the right test for the price side of the Autonomy Economy. Part I projected sector price declines from 2025 to 2040. Converting them to annual rates and setting them beside the last 25 years shows where the framework rides an existing curve and where it asks for a reversal:

Electronics is the easy row: the target is slower than what TVs already did. Energy has a working cost curve behind it. Food and housing are the hard rows. Both need a swing of roughly 6 to 8 percentage points a year from their historical path, and housing is mostly land and permitting, which no model makes more productive. The targets can stand, but they come with conditions: food needs cheap energy and automated logistics to compound, and housing needs the land-value and zoning reforms that Part IV of the series paired with the AI Dividend, or the gains will be captured as rent.

Time: what productivity bought, and what it could buy

In 1930 John Maynard Keynes predicted that by 2030 living standards in rich countries would be four to eight times higher and people would work about 15 hours a week. A 2024 reappraisal summarized by the BLS found he got the first half right: industrialized economies did grow 4 to 8 times. The second half missed badly. Average annual hours per worker in the United States were 2,022 in 1950 and 1,789 in 2023, a decline of only about 12 percent across seven decades of rising output.

The gains went to consumption instead of time. Some of that was choice: bigger houses, more travel, more stuff. Some of it was the Baumol trap from the frame section. When health care, education and child care absorb a rising share of hours, a household has to keep working just to stand still on the things it cannot do without.

The arithmetic of buying time back is simple. If productivity grows at g a year and all of it is taken as leisure rather than income, annual hours fall as:

where H₀ is today’s hours, H* the target, and t* the years to reach it. Keynes’s 15-hour week is about 780 hours a year. Starting from 1,789:

At g = 2 percent, close to the 2.1 percent the BLS measured for 2025, the 15-hour week arrives in 42 years, and after 20 years hours would be about 1,204 a year, roughly 23 a week.

At g = 3 percent, it takes 28 years, with about 991 hours after 20 years.

At g = 4 percent, a pace AI optimists consider plausible and skeptics do not, it takes 21 years, with about 816 hours after 20 years.

Part IV of the Autonomy Economy series went further with a Leisure Transition Curve, a logistic model in which hours fall steeply once a universal dividend crosses a sufficiency threshold. In that model’s projection, hours reach about 580 a year (roughly 11 a week) when the dividend reaches $84,000 in real terms. The simple formula above shows why the sufficiency threshold matters so much. Productivity growth alone moves hours slowly, a couple of percent a year, because most people trade it for income. A dividend changes the trade: once a baseline is covered, an extra hour of work buys extras rather than security, and people choose differently.

The frame adds one more lever that gets less attention. Leisure depends on two things: how productive your own hour is, and the time price of what you need. If AI cuts the time price of care, schooling and health by half, a household could cover its essentials in far fewer hours even if wages never rose. Cheaper essentials are a leisure policy.

Where it breaks

I’d push back on anyone, including me, who presents the abundance case without the following five problems attached.

The gains may not reach workers. Productivity in the nonfarm business sector rose 2.2 percent in the year to Q2 2026, while real hourly compensation fell 0.1 percent. In the same quarter, labor’s share of nonfarm business income fell to 52.8 percent, the lowest since the series began in 1947, against a postwar norm of roughly 60 to 66 percent. My Follow the Money piece traced where the difference goes: rising markups concentrated in a small set of firms, health care administration, land and finance. Cheaper production does not become cheaper living if the savings are captured as margin or rent along the way. This is the strongest argument for treating distribution as a design problem, which is what the Sovereign AI Wealth Fund and Universal Autonomy Dividend in Part IV of the series are for.

The macro estimates are modest. Daron Acemoglu’s widely cited model puts AI’s total factor productivity gain at no more than 0.66 percent over ten years, under 0.07 percent a year. Goldman Sachs estimated a 1.5 percentage point boost to annual productivity growth and about 7 percent higher global GDP over a decade. The gap between those two numbers is the whole debate, and neither has been settled by data yet. The firm-level studies in this piece (14 percent, 29 percent, 44 percent) are large, but they cover specific tasks, and adding up tasks into an economy is where estimates diverge.

Some costs are not made of hours. Housing rose 111 percent since 2000, and much of that is land. Land is fixed, and its price absorbs any income gain in a desirable place. Licensing, liability and regulation also set floors under prices that technology does not lower on its own. The time-price model applies to labor-heavy costs; it says nothing about rent.

AI has its own physical bill. The IEA estimates data centers used about 415 TWh of electricity in 2024, around 1.5 percent of the global total, and projects about 945 TWh by 2030, more than double. If that demand outruns new generation, AI raises electricity prices in the short run, which feeds into food and goods, the opposite of the story told here.

The transition is the dangerous part. Even if the destination works, displacement arrives before any dividend does. An earlier piece in this series called it the transition gap: the years when automation is removing jobs faster than any distribution system is paying out. Price declines in TVs never threatened anyone’s income. Price declines in services will, because the price of a service is someone’s wage.

A smaller caveat on the data itself: the BLS adjusts CPI for quality, so part of the TV and software declines reflects better products at similar sticker prices. The direction is right; the magnitudes overstate what a shopper sees at the register.

The path forward

The optimistic case survives the caveats, but it changes shape. AI will not make life cheaper by default. It will make cognition nearly free, and whether that turns into cheaper care, cheaper schooling and shorter weeks depends on where it is pointed and who keeps the savings.

For builders, the frame gives a simple filter: point AI at the sectors whose time price has risen fastest. Hospital services, tuition and child care are where the unmet need, the household stress and the economic upside all sit. A tool that shaves 40 percent off the clinician hours in a workflow is worth more to living standards than one that makes an already cheap product cheaper. Measure success in hours per unit of service, not in seats sold.

For anyone measuring progress, track time prices instead of dollar prices. The script above takes ten minutes to extend with the latest CPI release. If hospital services stop outrunning wages over the next few years, AI is working where it matters. If they keep outrunning wages while productivity rises, the gains are being captured somewhere, and the labor share will say where.

For policy, pair the technology with the distribution. The Autonomy Economy series proposes a specific architecture: a Sovereign AI Wealth Fund capitalized from automation gains, a Universal Autonomy Dividend paid from it, land value taxation so the dividend is not absorbed by rent, and removal of the extraction layers documented in Follow the Money. You do not have to accept every parameter of that model to accept its premise, which this piece’s data supports: 75 years of productivity growth bought back only about 12 percent of working hours, and there is no reason to expect AI-driven growth to behave differently without a mechanism.

Keynes was right about the wealth and wrong about the weeks. AI gives us a second chance at the half he missed. The model, the code and the full framework are in the Autonomy Economy repository for anyone who wants to run the numbers.

About the Author

Brian James Curry is a Kansas City AI researcher and data scientist, the founder of Vector1 Research, an independent applied AI lab, and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and grew to more than 1,500 members. He has more than twenty years of experience leading analytics, data science and machine learning at enterprises including Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source work, including Papilon, PyCausalSim, MeaningFlow, Daedalus and tincast, is on GitHub. He grew up in a small Kansas farming town and attended Kansas State University. He writes about agentic AI, causal inference and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.

Sources

Related research