Healthcare spending is a lingering political issue both because it has become a significant fraction of total GDP and also because of perceived inequities. Here, I consider various theories as to why costs have increased, the data that support or refute them, and the prospects of future cost control due to improved technologies.
Cost Trends since 1960
Many factors are blamed for high US healthcare costs. Some of the favorites are: high drug prices; insurance companies; regulations; private equity hospital ownership; over testing; gold plated care; chronic disease; aging population; cost of medical school and student debt. I look at these possibilities using government data, but first consider some prior work as a useful benchmark.
In https://www.commonwealthfund.org/publications/issue-briefs/2023/oct/high-us-health-care-spending-where-is-it-all-going
Using public data, they identify where 60% of the excess US healthcare spending vs the OECD-average
comes from. They summarize with the following chart:
- U.S. pays more in administrative costs of insurance ~15%
- U.S. providers spend more on administrative activities ~15%
- U.S. pays more for prescription drugs ~10%
- U.S. physicians earn more ~10%
- U.S. registered nurses earn more ~5%
- U.S. invests more in medical machinery and equipment <5%
- Sum of components estimated ~60%
They surmise that the remaining 40% of the gap could be at least partially explained by greater “service intensity”.
I think these numbers represent a decent benchmark. However, I do take issue with the definition of the gap as per-capita spending. A beater measure is spending to GDP. At one extreme, if one considers that US per-capita GDP is about $95K (in 2024) compared to the aggregate OECD average of about $48K, and there is almost no gap. But this is too extreme an adjustment. Instead, if we take the average of country-level ratio’s we find an OECD average of about 12.75% (in 2023) vs 17.7% in the US (in 2023, 17.25% in 2024).
First, an overview in graphs
It is immediate from the second graph that something interesting happened from1989-1993. The dip in cost fraction is attributed to the rise of HMO’s: Managed-care enrollment rose from just over 20% of insured Americans in 1985 to more than 70% by 1993. The dip was temporary. However, the first graph does indicated that the slow-down in per-capita spending has been sustainable. I suspect the reason is related to the following data:
Up until 1993 or so, insurance was paying an ever increasing share of total costs which lead to demand increasing faster than population. After 1993, the trend is much flatter, and so the boost to demand has petered out.
The main source is
Big pharma
I suspect suspicion and dislike of drug companies is a holdover from a time when they were one of the most significant out of pocket expenses for consumers. Another belief that may contribute to loathing is that they spend more on advertising than R&D. It is true that on-patent drugs cost more in the US than elsewhere. It is less appreciated that US generics are actually cheaper, and constitute 90% of prescriptions (albeit just 12% of total drug spending). On the other hand the R&D claim is false. I don't know if the wrong claim is intentional, or due to lack of understanding of basic accounting --- the difference between advertising and SGA.
According to the CBO, the % of sales spent by pharma has increased from 12% of sales in 2000 to 25% in 2021. This exceeds the levels of any other industry (for comparison, the next two R&D intensive industries, software and semiconductors, were both at about 15% R&D to sales in 2021). Even pre-covid (2019), pharma R&D was about 23% of sales.
In any event, the following graphs shows pharma sales as a % of total healthcare consumption
Big pharma has benefited from health care inflation, but at less than 10% of total spending, it is hard to see it as a driver. Looked at another way, big drug companies have been lousy investments for decades with a few exceptions such as Eli Lilly, benefiting from GLP’s. For example, Pfizer, at $28/sh, is the same price as it was in 1998. In short, with drug spending in line with healthcare growth, but R&D growing much faster, it seems unlikely that the future can see lower drug prices so long as US firms
remain dominant. If Chinese firms continue their rapid ascent, this may change, but may not be a good thing for the U.S..
Insurance companies
There is no question that, in contrast to drug companies, insurance companies have been excellent investments since the ACA (with some give-back in the last two years). For example, UnitedHealth Group increased about twenty-fold from roughly the time of the passage of the ACA until mid-2024. By law, ACA policies must pay 80 or 85% of revenues out to cover patient costs (or “improvements”). As an example, in their 2025 10K, UnitedHealth Group reports the medical loos ratio’s of 89.1 %, 85.5 % , and 83.2 % for the years 2025,2024, and 2023, respectively. They also reported operating costs (excluding cost of products sold --- they do more than just health insurance which supplies about 3⁄4 of revenues) of just under $60B. This works out to about $1100 per patient, which jibes with the figures in the commonwealth article cited above. This article also states US insurance administrative costs are a factor of five larger, per capita, than the OECD average. If even close to true, this suggests considerable
waste (albeit the employees of insurance firms may not consider themselves to be wastes).
If so, then the following chart suggests considerable savings can be obtained:
The other half of the 15% gap attributed to administrative costs in the commonwealth article is composed of regulatory costs, but also partially of costs related to insurance (e.g., wading through the occasional quagmire pre-approval process).
Regulation
True regulatory costs involve too many counterfactuals to allow accurate estimation --- there are not just direct compliance costs, but also costs associated with changing behavior, which should actually be beneficial if the regulation is a good one. But one can try and obtain order of magnitude estimates and the American hospital Association did just that in 2017 with a figure of $39B for compliance costs. In that year, total hospital spending was $1,08T, so just under 4%. Even 4% of total healthcare spending (so
$200B) does not seem like a game-changer.
Over Testing
In 2024 Medicare part B spent $8.6B on testing, or about 1.5% of total spending. While “defensive medicine” (or offensive medicine, given that physician `bans’ on owning medical testing facilities are easily skirted) is often cited as a driver of health care costs, I do not see it in the numbers.
Hospital
Hospital’s share of total spending has been mostly declining.
Of course, the average length of a hospital stay has fallen from 7.7 days in 1960 to about 4.5 today as well. And hospital rates increased faster than CPI (more on this soon).
Aging
If any story makes sense, it’s surely the aging population, no? Well, consider these three graphs:
The share spent on nursing homes is falling! This despite the fact that lifespans have greatly increased and (I suspect) children are less likely to take care of aging parents. And Medicaid is growing faster than Medicare.
Gold plated care
This is related to over-testing, but van also include over-treatment. The commonwelath article does document US patients get more procedures, but it is more likely they are getting better care than overtreatment.
Medical schools costs.
Many countries pay for doctor’s educations. This actually biases the healthcare spending gap because these costs will not appear as healthcare costs in other nations accounts, but will in the US (because Dr’s charge more to repay their loans). However, this represents at most 2% of annual healthcare consumption (assuming the annual 30K med school graduates all pay full annual tuition of about $90K/yr for 4 years).
Future Costs in an AI world
AI has the potential to impact healthcare spending in many ways. The most exciting, of course, is the discovery of new treatments. This may reduce costs (if, e.g., curing chronic disease) but could also increase costs by increasing demand. However, as discussed further below, this requires there actually being many useful things to discover about human disease that we are currently missing.
The more mundane cost reduction route is via improved back-office efficiency. Of this, it is interesting that healthcare AI spending represented less than 7% of total end-user AI spending in 2025, well below the near 20% of GDP. Clearly, the privacy issues and attendant regulatory maze will slow AI adoption here. An intermediate cost reduction route is via prevention or detection of medical errors.
However, the data clearly show that insurance could be the killer app for AI in healthcare. Already, providers are using AI to help get pre-approvals, and insurance companies are using AI to reject them. But, in particular, AI will be much better able to efficiently wade through the thousands of different plans and providers, each separately negotiated, and figure out what is covered and what is not. I look forward to this day, but perhaps I dream.
Difficulties in bringing the AI world to healthcare
Large language models have gone past mere syntax and grammar to discern meaning. This is quite impressive. However, it is important to note that these models were trained on enormous quantities of text that had meaning. If they had been trained on baby talk, I think it's obvious success would have ben lacking.
Unfortunately, our understanding of the human bio-chemical system is probably closer to baby talk than Shakespeare. It is only recently that "junk DNA" has been recognized as not junk, but a crucial element of controlling gene expression. Since introns (nee, junk DNA) are about 100 times more numerous than exons, even the DNA component of the control system of gene expression is very incomplete. The fact that there is substantial epigenetic control as well (also hard to understand) make the problem even harder.
Can AI make sense of this tangled web? Time will tell, but I suspect there will be many bodies on the journey. While there are currently some promising AI assisted drugs under investigation, there have also been a number of recent failures. These include: BenevolentAI’s BEN-2293, for eczema (failed on its primary Phase IIa measures); Exscientia/Sumitomo’s DSP-1181, for obsessive-compulsive disorder (discontinued after Phase I); Exscientia’s cancer candidate EXS-21546.
A final issue is the matter of data access. Healthcare data is lucrative. AI companies will not be allowed to scrape hospital and clinic data at will as they (to some extent) did when training LLM's. But, so long as the VC funding holds up, this should not be a serious roadblock.
Read Comments