Will AI Make Everything Cheaper? Who Wins, Who Loses

I belong to the camp that believes AI will trigger one of the biggest price collapses in a century, doing to many services what globalization did to TVs. My caveat is that the collapse will be selective. AI is already collapsing the price of thinking work, and companies that sell billable hours or headcount are exposed. But the sectors people complain about most are expensive for reasons AI doesn't touch: regulation, land and physical labor.

The investing question is not "will things get cheaper" but "who keeps the savings." Below we sort sectors by that question, using price data rather than predictions.

What the price data actually shows

Prices already split in two over the past 20 years. Between 1998 and 2018, overall US inflation was 56%, but traded, lightly regulated goods fell while regulated local services tripled (Kottke, citing Mark Perry's BLS chart):

The pattern is clear: what got cheap was exposed to competition and technology; what got expensive was shielded by licensing, subsidies and restricted supply. AI is a new technology force, but it does not remove those shields.

The second data point is AI's own price. The cost of running a language model at a fixed level of ability has fallen roughly 10x per year (a16z), and Epoch AI measured declines from 9x to 900x per year depending on the benchmark. That is faster than PCs or internet bandwidth ever fell. Anything that is mostly reading, writing, summarizing or coding is being repriced against that curve.

One macro caveat: the Federal Reserve targets 2% inflation and will fight broad deflation with easier money. So expect falling prices in specific sectors, not a falling overall price level. Relative prices move; the dollar still loses value.

AI builders: more revenue, thinner margins over time

The builders sell the very thing that is deflating fastest. Their total revenue will likely keep growing, but each unit they sell gets cheaper every year. Whether they get more profitable depends on which layer they sit in.

Chipmakers (NVDA, AMD, AVGO). Today they are the most profitable layer because supply is scarce and every hyperscaler is racing. The four biggest buyers plan to spend about $725 billion on capex in 2026, up 77% from 2025. That money is chip revenue. The risk is that their customers are building their own chips: Amazon's in-house chip business already runs at a $20 billion annual rate. Scarcity margins fade once supply catches up.

Hyperscalers (MSFT, GOOGL, AMZN, META). They are spending far ahead of revenue. Microsoft plans $190 billion of capex this year against an AI business running at about $37 billion a year. Free cash flow is already falling: down 22% at Microsoft and 38% at Alphabet year over year in the latest quarter, and two-thirds of Microsoft's spending goes to GPUs that wear out in a few years. They win only if AI demand grows faster than AI prices fall.

Model labs. They sell tokens whose price drops roughly 10x a year. Their profit depends on staying a step ahead of open models.

The historical rhyme is the internet build-out. The internet changed everything, but the companies that built it overspent. Cisco lost 90% after its March 2000 peak and needed 25 years to reclaim that share price; its market value is still about $250 billion below the 2000 level. Cisco stayed a good business. The stock was simply priced as if boom-time margins were permanent.

Our answer: AI makes the builders bigger, not necessarily more profitable per dollar invested. The risk is in the price you pay, not in whether AI succeeds.

The most disrupted industries

The common thread: revenue is tied to human hours, headcount or information that AI now produces for pennies. The market has already started repricing them.

A warning for short sellers: many of these already fell hard. Being disrupted and being a good short today are different questions.

Margin capturers: how Visa and Moody's actually gain

A company keeps AI savings only if its price is not set by its labor costs. Both of these companies charge for something customers can't easily get elsewhere, so cheaper internal work becomes profit instead of a price cut.

Moody's (MCO): the clearest case

• Cheaper ratings, same price. Bond issuers must buy ratings from a recognized agency, and investors trust Moody's track record. Moody's now automates roughly 25% of pre-committee checks, and its ratings arm reached a 68.3% adjusted operating margin in Q2 2026, up from 64.2% a year earlier, while rating more than $2 trillion of debt in the quarter.
• It shows up in results. Company-wide operating margin rose to 47.9% in Q2 2026 from 43.1% a year earlier, per Moody's own earnings release.
• AI is also a customer. Data centers are funded with debt, and every bond needs a rating. Moody's CEO said hyperscalers have issued more debt this year than in the prior three years combined.
• AI as a sales channel. Moody's data now plugs into Microsoft Copilot, Amazon and Claude, so AI agents become distributors of its data rather than replacements for it.
• The risk: the analytics arm sells research that AI can partly imitate. FactSet's 30% margin shows what data firms without a regulatory moat look like.

Visa (V): gains volume more than margin

• Its price isn't tied to labor. Visa takes a percentage of each payment, so AI cost savings don't push its fee down.
• Fraud prevention. Visa scores every one of its roughly 300 billion yearly transactions with AI and blocked about $40 billion of fraud in 2023, nearly double the prior year. Fewer fraud losses and fewer false declines mean more approved payments, and each approval earns a fee.
• AI shoppers. When AI agents start buying on people's behalf, they need a trusted way to pay. Visa's agent-payment platform, Intelligent Commerce, aims to make its network the default, and Visa has since partnered with OpenAI to bring it into ChatGPT-style shopping. That adds volume.
• The limit: Visa already runs at very high margins with a small workforce, so cost savings add little. The upside is volume. The risk is AI agents routing payments to cheaper rails such as bank transfers or stablecoins.

Cheaper services, more usage: real, but who profits?

When a service gets much cheaper, people use far more of it, but only where demand was held back by price. Economists call this the Jevons paradox. The catch for investors: the extra usage rarely enriches the old providers.

Legal help is the best example, and it splits into three markets. This does not mean lawyers disappear.

• Complex work stays with lawyers. Litigation, mergers, regulatory fights and anything that ends in court still need a licensed lawyer who carries liability. AI makes these lawyers faster, so a firm may need fewer junior associates per case. It may also shift from hourly billing to flat fees, since clients won't pay 20 hours for work that now takes 5.
• Routine, standardized work gets cheaper. Document review, basic research, first drafts of contracts and template forms are what AI does well. The losers are businesses that sell this routine work: template-document sites, outsourced document review, and legal research databases priced per user (see the disrupted list above).
• A huge new market opens. Low-income Americans get no or inadequate legal help for 92% of their serious civil legal problems, and nearly half of those who don't seek help cite cost. Think tenant disputes, debt collection letters, benefit denials. Law firms never had this revenue, so they lose nothing here. A $400 lawyer review that becomes a $20 AI review creates millions of jobs nobody was buying before, and the money goes to whoever owns the customer and the tool.

Software follows the same pattern. Developers are still needed, but their job changes from typing code to deciding what to build, reviewing AI output and owning the result. Entry-level roles feel it first. Total demand for software grows, because when code costs a fraction of what it did, every small business can afford custom tools and companies fund projects they would have skipped. The losers are those selling programmer hours for routine work, such as offshore IT outsourcers, and software priced per user seat. The winners are the infrastructure layers the extra software runs on: cloud hosting, data storage and payments.

What history suggests for lawyers and developers. When spreadsheets made bookkeeping nearly free in the 1980s, clerk jobs shrank, but demand for financial analysis grew and analysts multiplied. The same split is likely here. We don't expect people to book more lawyer appointments. We expect far more legal questions to get answered, mostly by AI tools, with lawyers taking only the cases that need judgment or a court. Likewise, businesses won't necessarily hire more software firms. They will get more software built, much of it by small in-house teams using AI. The total amount of legal help and software rises; the share delivered by billable human hours falls.

Where it does not work: healthcare. AI can read scans and triage symptoms cheaply, but licensing rules and liability still require a doctor's sign-off, so prices fall slowly and volume barely moves.

Our answer: cheaper services leading to more usage is real, but "volume winner" is not a sector you can buy. Look instead for companies that own the customer relationship or the infrastructure the extra usage runs on.

Pass-through retailers: savings go to shoppers, scale goes to the leaders

Big retailers will use AI heavily, but competition forces them to hand most of the savings to customers. Their reward is market share, not fatter margins.

Where AI helps a retailer: better demand forecasting (less spoiled food and fewer markdowns), smarter store staffing, faster warehouse picking, delivery routing and automated customer service. Each cuts a fraction of a percent from costs.

Why the savings don't stick: grocery and general retail are among the most price-competitive businesses there are. Walmart earned 4.2 cents of operating profit per dollar of sales in fiscal 2026, and Costco about 3.8 cents per dollar of revenue in fiscal 2025. Costco's whole model is to cap its markup and pass every saving through, and Walmart's is "everyday low prices." If one retailer cuts prices with AI savings, the others must follow.

The precedent: in the 1990s Walmart led the industry in barcodes, satellite inventory tracking and supply-chain software. Its margin barely changed, but it took share from Kmart, Sears and regional chains that couldn't keep up. The efficient giant grew; the laggards disappeared.

What actually raises retailer profit: side businesses built on their customer data, such as advertising sold to brands and membership fees. These carry software-like margins and benefit from AI targeting.

Our answer: expect the scale leaders (WMT, COST, AMZN) to keep growing and the mid-size chains to struggle. Don't expect margin expansion in the core retail business. The consumer is the real winner here.

Utilities and power: they win because AI needs them

Both reasons matter, but in different ways. AI's demand for electricity drives the profits. AI's inability to produce electricity is why prices don't fall.

The demand. US data centers used 176 TWh in 2023, about 4.4% of the country's electricity, up from 58 TWh in 2014. The Department of Energy expects 325 to 580 TWh by 2028, or 6.7% to 12% of US power. After decades of flat demand, that is a structural shift.

The scarcity. A power plant takes years to permit and build, and transmission lines take longer. AI can design a better turbine but can't conjure megawatts. So unlike text or code, electricity can't be deflated by software, and new demand runs into fixed supply.

How each type of company profits is different:

• Regulated utilities (DUK, SO, AEP, D). Regulators let them earn a set return on what they invest. Data centers force them to build plants and lines, which grows the base they earn on. Steady, slow gains.
• Independent power producers (CEG, VST, TLN). They sell power at market prices or through long contracts with tech companies. Tight supply raises prices, and hyperscalers pay a premium for guaranteed, carbon-free power, as in Microsoft's deal with Constellation to restart a Three Mile Island reactor. Bigger upside, bigger swings.
• Grid equipment makers (ETN, GEV, PWR). They sell the transformers, turbines and construction work everyone needs.

The risks: a leap in AI efficiency would cut forecast demand. That fear has already hit once: on January 27, 2025, after China's DeepSeek released a far more efficient model, Vistra, Constellation and Talen each fell more than 20% in a single day. Regulators may also push data-center costs back onto the tech companies to protect household bills.

Dividend payers: MCD, XOM, and which ones actually benefit

"Dividend payers" is too broad a category. A dividend is just a payout. What matters is whether the business behind it gets cheaper to run without having to cut its prices. Three different cases:

McDonald's (MCD): a modest, slow winner

• How it earns: most of McDonald's profit comes from rent and royalties that franchisees pay as a percentage of their sales. About 95% of its restaurants are franchised, and McDonald's owns roughly 56% of the land and 80% of the buildings under restaurants in its consolidated markets.
• Where AI helps: drive-thru ordering, kitchen scheduling and inventory. The savings land mostly with franchisees, since they pay the crew. Healthier franchisees open more stores and can hold prices down, which brings in more customers and more royalties.
• Reality check: it's slow. McDonald's ended its AI drive-thru test with IBM in 2024 after accuracy problems; franchisees wanted at least 95% accuracy and real savings over a human worker. It has since partnered with Google Cloud.
• Verdict: mild positive. AI cuts franchisee labor costs, but McDonald's royalties are a cut of sales, not of profit, so it gains only through volume.

Exxon (XOM): neutral, and AI isn't the driver

• How it earns: selling oil, gas and chemicals at market prices. AI cannot make oil, and the oil price is set by global supply, OPEC and geopolitics.
• Where AI helps: seismic imaging, well placement and plant maintenance cut costs. But every producer gets the same tools, which raises supply and pushes prices down. The savings flow to drivers at the pump.
• One real plus: many data centers run on natural gas plants, and Exxon is a major gas producer.
• Verdict: neutral. Own it for an oil or geopolitical view, not for an AI view. In a genuinely deflationary economy, commodity prices usually fall, which hurts.

Who actually benefits: contracted cash flow

If AI does push prices down in parts of the economy, the dividends that hold up best are those backed by contracts or regulation, not market prices:

• Gas pipelines (WMB, KMI): paid a fee per unit shipped, and data centers raise gas volumes.
• Regulated utilities: an allowed return on a growing asset base (see the power section).
• Net-lease REITs (O): long leases with fixed rent increases.

Our answer: to be AI-resilient, a dividend payer needs a price nobody can undercut with cheaper labor. MCD partly has one; XOM doesn't need one, but AI doesn't help it either.

Scorecard

Where I land: AI is a real deflationary force, and the price data and token costs prove it. But it deflates thinking work, not land, licenses or physical labor, so don't expect healthcare, housing or childcare to follow TVs. And the Fed will keep the overall price level rising. The opportunity is not in betting on falling prices. It is in owning businesses whose costs fall while their prices hold, and avoiding those whose product AI now gives away.

This post reflects personal analysis, not investment advice. Do your own research before trading.