The most consequential transformer in a housing market may not be the one that predicts the next word. It may be the one behind a fence, waiting to connect a building to the grid.
One belongs to a technology that appears to move at the speed of thought. The other belongs to a world of factory slots, crews, approvals and delivery dates. A homebuyer lives in that second world. So does the electricity bill. So does the mortgage.
The connection is easy to miss. Artificial intelligence usually enters the housing conversation through a screen: a listing description written in seconds, a property valuation, a chatbot that can explain an amortization schedule. Yet the larger effect may arrive before the buyer uses any of those tools. It can come through the price of capital, competition for physical resources, the geography of new wealth, or uncertainty about the salary that will support the loan.
On September 28, Federal Reserve Governor Lisa Cook put a difficult timing problem at the center of a speech in Oakland. AI investment, she argued, was adding inflation pressure now; productivity improvements might modestly ease that pressure over the next few years, but she did not expect them to offset broadening pressure later this year. She distinguished expensive AI-specific inputs from demand spilling into shared resources such as construction labor and energy. That distinction matters: she warned that blunt monetary policy aimed at a narrow sector's relative price increases could be a mistake. These are her assessments, not a committee promise about its next decision. Cook's speech
This is not a story about a chatbot setting your mortgage rate. No source in this article provides a defensible estimate of how many basis points of a particular mortgage quote are caused by AI. The useful question is subtler. Can the investment required to make a future economy cheaper make parts of today's economy more expensive first? And if it can, who pays during the interval?
I. The cloud needs an address
Software seems weightless until someone builds the place where it runs. The demand becomes an order for equipment, a claim on land, a construction schedule and an application for power. Whatever the business eventually earns, those obligations have to be dealt with in the present.
The International Energy Agency's April 2026 assessment puts global data-center electricity consumption at 485 terawatt-hours in 2025. Its central projection reaches about 950 in 2030. The second number is a projection, not electricity already consumed. Neither is an AI-only total: data centers also run ordinary cloud services and other computing workloads. IEA, executive summary
Those qualifications do not make the physical demand trivial. They make it legible. A terawatt-hour has no neighborhood. A substation does. National totals can obscure the places in which a load arrives faster than a connection, a supplier or a local planning system can accommodate it.
Consider a hypothetical town approving a large computing campus. There is no named project hidden inside this example. During construction, contractors need workers and places for them to stay. The finished buildings need an electricity connection. The town may anticipate tax revenue. Nearby owners may anticipate higher land values. A household considering a purchase has a different question: will its income change along with the costs around it?
That household can lose ground without losing a dollar of salary. If a higher-paid buyer enters a market with few homes for sale, the existing household's wage may be unchanged while its next feasible purchase becomes dearer. If a local resource becomes more costly, the household's disposable income can shrink even when the sticker price of a home stands still. These are possible mechanisms, not outcomes established for every data-center town.
We need to keep at least three maps apart. The places that build AI infrastructure are not necessarily the places where highly paid AI workers live. Those are not necessarily the places where firms capture the eventual productivity gains. A power-intensive campus, an equity-rich neighborhood and a business using cheaper software can occupy different states, with different housing supplies and very different residents.
A national boom is not a local instruction to raise the asking price.
II. The bill arrives before the dividend
There is a deceptively comforting sentence in almost every discussion of a new technology: it will make us more productive.
Perhaps it will. But a future benefit cannot be spent twice, once on the equipment needed to obtain it and again on the household bills due before it arrives. There is an interval between buying a capability and reorganizing an economy around it. During that interval, the inputs are real, the invoices are real, and the eventual distribution of the benefit remains unsettled.
Imagine a small contractor who buys a tool that will eventually cut estimating time. The subscription can begin today. Staff may need training. Existing job records may be too inconsistent to use. A faster estimate may expose a delay elsewhere: permits, an inspection, a scarce subcontractor. The software saves time on one task, but the completion date of the house barely moves. This is an illustrative adoption problem, not a reported contractor's experience.
Economists Erik Brynjolfsson, Daniel Rock and Chad Syverson describe a related problem in their research on the productivity J-curve. General-purpose technologies require complementary investments, including changed business practices and new skills. Some of that work is intangible and difficult to measure. The purchase of technology and the full measured payoff need not coincide. The paper is a framework for understanding adjustment, not a countdown to an AI dividend. The productivity J-curve
The household equivalent is painfully ordinary. A salary is paid on one calendar. A rate lock expires on another. An expected industry transformation has no obligation to arrive before either.
THE ADJUSTMENT / TWO DIFFERENT CLOCKS
Investment now. A possible dividend later.
THE PRESENT
Orders. Connections. Construction. Capital.
Physical inputs and obligations can arrive before the economy reorganizes around a new capability.
THE POSSIBLE PAYOFF
New skills. New practices. More useful output.
The size, timing and distribution of future benefits remain uncertain. They need not match the construction calendar.
Symbolic clocks, not a measured time series or an estimate of the lag. Based on the productivity J-curve framework and Lisa Cook's September 28 assessment.
III. A mortgage is not an overnight loan
Here the story needs a brake. It is tempting to draw a clean arrow from AI investment to inflation, from inflation to the Fed, and from the Fed to a mortgage quote. Each connection has economic content. The finished chain is still not a one-for-one machine.
A thirty-year fixed mortgage is priced in a market looking far beyond the next policy meeting. Investors assess future interest rates and the compensation they want for holding a long-lived asset. Mortgage securities add their own complications; a household loan also involves origination, servicing and borrower-specific terms. The rate offered to one applicant is not the policy rate with a universal markup pasted on top.
Paul Willen's May 2026 Boston Fed research explains why comparing mortgages with ten-year Treasury yields requires care. The cash flows differ, and the mortgage spread includes intermediation and credit-related costs. A particularly important difference is the borrower's ability to prepay. When rates fall, borrowers may refinance, returning investors' principal just as reinvestment becomes less attractive; when rates rise, borrowers can keep their older loans. That asymmetry has a price. Interest-rate expectations and volatility can therefore affect the mortgage spread as well as the underlying bond yield. These are research findings, not lending advice or an official policy forecast. Why mortgage rates exceed Treasury yields
This leaves room for apparently contradictory days. A central bank can make a decision the bond market already expected. Longer-term yields may then move on a change in the outlook rather than on the announcement alone. A mortgage spread can widen while a benchmark yield declines. A borrower can receive a worse quote while a headline says policy is easing. To understand the quote, one needs the date, the loan terms and the market behind it.
AI belongs in that wider outlook alongside other forces, not above them. Treating it as the sole explanation would erase the very evidence that makes the story worth reading. It also creates a practical trap: a buyer begins to wait for a single technological or policy event to repair a budget that depends on several moving parts.
For someone who already owns a fixed-rate loan, a change in market rates does not mechanically reset the contracted principal-and-interest payment. The next buyer faces a different rate. The current owner may still face changing taxes, insurance, utilities or income. Those positions are not interchangeable. A town can contain protected incumbents, vulnerable would-be movers and first-time buyers absorbing today's cost of money at the same time.
The technology is shared. The exposure is not.
IV. Who gets the electricity bill?
A computing campus may need new infrastructure. It does not follow that every nearby family must finance it. That is a question about contracts, tariff design, capacity and regulation, not a natural law.
The IEA explicitly cautions against assuming that new data-center demand necessarily raises electricity prices. Spare capacity and better utilization can help lower average costs; constrained systems or poorly matched investment and demand can create pressure in the other direction. How upgrade costs are allocated is a policy choice. Geography and the agreement matter. IEA on electricity affordability
That qualification is important enough to stop the story for. It changes the villain from a building to an allocation rule.
Suppose a utility expands for a customer that promises a large load, then uses much less than expected. In an illustrative version of that problem, somebody still has to service the investment. The household's exposure depends partly on whether the industrial customer has enforceable commitments, whether adequate security was posted, and which costs other customers are permitted to bear. A press release announcing a project will not answer those questions.
AEP Ohio's current data-center tariff provides a concrete example of attempting to deal with this mismatch. It sets a load ramp and a contract term, imposes minimum billing-demand rules and specifies cancellation and exit obligations. Some minimum-demand calculations use 85%, with capacity bands and other conditions. This is not the simple claim that every data center must pay for 85% of electricity it never consumes: demand capacity charges and energy usage are different things. The utility also describes reimbursement of buildout costs when a customer cancels or substantially delays before energization. Those are terms for its service territory, not a nationwide guarantee against residential bill increases. AEP Ohio's tariff
The distinction between power and energy is a useful one to carry away. Capacity concerns the system's ability to meet demand; consumption concerns how much electricity is used over time. A reserve built for a peak can be costly even if the peak rarely occurs. Conversely, counting an announced connection as though it were already a full-time operating load can exaggerate what has happened.
For a household, the practical inquiry is not simply whether AI is being built nearby. It is whether the project is operating, what the connection requires, how it is financed, and what the regulator has actually approved. Taxes and local employment need the same scrutiny. Promised revenue is not received revenue. Temporary construction work is not the same as permanent operating employment. A benefit to one public budget is not automatically a saving on every residential bill.
These are less thrilling questions than how intelligent the next model will be. They are also closer to the price of living somewhere.
OBSERVED / JULY 2026
Two tech metros. Not one housing story.
Year-over-year change in median sale price. The same measure, in the same month, moving in opposite directions.
−10%0+10%
San Francisco+6.0%
Active listings: -18.4% year over year
Seattle-3.6%
Active listings: +16.7% year over year
Redfin, September 2, 2026 report · Metro-area data for July 2026. Median sale prices can change with the mix of homes sold. These observations are not estimates of AI's causal effect.
V. Two tech cities, two housing markets
Even the phrase “tech city” is now too blunt to serve as a housing forecast.
In its September 2 report, Redfin compared July housing measures for the San Francisco and Seattle metro areas. Its table shows San Francisco's median sale price up 6.0% from a year earlier; Seattle's was down 3.6%. Active listings moved in the other direction: down 18.4% in San Francisco and up 16.7% in Seattle. These are metro-area observations, not city-boundary figures. The report links the divergence partly to concentrated AI wealth in San Francisco and employment uncertainty in Seattle. It does not identify a controlled causal effect of AI. Redfin's comparison
The chart is a reason to abandon an easy generalization, not to adopt a new one. A median sale price can change because the mix of homes sold changes. Inventory can change because sellers enter, buyers complete, owners withdraw, or the observation window moves. Different neighborhoods within either metro can behave differently again.
Yet the contrast exposes a meaningful possibility. The same technology can enlarge one buyer's down payment and make another buyer question the durability of their income. An employee with valuable equity and an employee waiting to hear whether their team will be reorganized may read the same AI headline and reach opposite decisions about a mortgage.
Neither has to be irrational.
Now put a third household between them. It does not work in technology. It earns a steady wage and wants to live near family. If affluent new buyers compete for a limited number of homes, the third household can face the consequences of wealth it never received. Housing affordability is partly a contest over scarce locations, not merely a reward for society's average productivity.
This is why more money and more homes are not substitutes. A gain in purchasing power expands choice only to the extent that there is something to buy. Where supply can respond, demand can support construction. Where it cannot, some of the gain can be reflected in land and prices. That is an analytical implication of constrained supply, not a claim that the Redfin figures measure the size of either channel.
It is also why a booming housing market and a worsening affordability problem can be the same event viewed from different kitchens. A seller sees a better price. A would-be buyer sees a larger required deposit. A renter sees the risk that new earnings nearby will become a rent increase before becoming an opportunity of their own.
VI. The number a household can actually use
One useful response to a large, uncertain macroeconomic story is to make a smaller calculation exact.
Take an illustrative $750,000 home, a 20% down payment and a thirty-year fixed loan. The principal is $600,000. At 6%, monthly principal and interest are about $3,597. At 7%, they are about $3,992. That roughly $395 difference does not measure AI's effect on rates; it measures the sensitivity of this loan to a one-percentage-point change. Taxes, insurance, fees, maintenance and utilities are outside these figures.
The comparison becomes more useful when the budget, rather than the advertised price, is held still. If a buyer could afford the first payment but faces 7% instead, the same principal-and-interest budget supports a smaller loan. Whether a seller adjusts, a buyer finds more cash, or no transaction happens is a different question. A calculator can solve the financing arithmetic. It cannot make the market accept the result.
The experiment below lets the reader change the price, rate and down payment, then compare neighboring rate scenarios. No scenario is assigned a probability. There is no button labelled “AI surcharge,” because we do not possess an estimate that would justify one.
That absence is part of the explanation.
INTERACTIVE / THE HOUSEHOLD LEDGER
Move the rate. Feel the difference.
A mortgage cannot borrow from a future productivity gain. Change the assumptions to see the payment due today.
MONTHLY PRINCIPAL & INTEREST
$3,597/ month
$600,000 loan · 30-year fixed · 6.0%
5.0%
$3,221
6.0%
$3,597
7.0%
$3,992
One percentage point higher: $395 more per month.
$100,000$2,500,000
0%80%
0%15%
Illustrative payment sensitivity, not an AI forecast or a lending quote. Principal and interest only. Excludes taxes, insurance, mortgage insurance, fees and closing costs. The comparison holds price, down payment and the 360-month term constant. Comparison rates are bounded between 0% and 15%. A zero rate uses principal ÷ 360.
VII. Faster software, slower houses
The hopeful version of this story deserves equal care. AI does not have to stop at competing for scarce inputs. It could help people use those inputs better.
For housing, the relevant promise is not that a model can draw an attractive building. It is that fewer hours might be lost between a viable idea and a home someone can occupy. An application assembled more accurately might avoid a preventable resubmission. A contractor might detect a conflict between specifications before work reaches the site. A lender might find a missing document earlier. Better information can matter without replacing the people responsible for each decision.
These are potential applications, not claims that they have already reduced national homebuilding costs or that Oftu operates them. Each would need tests against the full process. A task completed in half the time is not a house completed in half the time. If the next stage remains the bottleneck, some of the saving is absorbed by a queue. A faster form cannot authorize a use the planning rules forbid.
There is another question after efficiency: who receives the saving? Less work can become a lower customer price, a higher margin, a higher wage, or a mixture. The tool's capability does not select the distribution. Competition, contracts and bargaining do. A household needs the benefit to reach income or the cost of housing, not just a firm's internal productivity chart.
An August 2026 discussion paper by Tania Babina, Alex He and Renhao Jiang offers a useful counterweight to the idea that all the benefits are perpetually in the future. Using a firm-level AI-investment measure based on AI-skilled employment, the authors find associations with productivity growth in recent years and investigate the creation of durable organization-specific knowledge. It is evidence about firms and a particular measure, not a forecast for home prices or proof that every firm adopting a chatbot becomes more productive. Organization capital and AI
The optimistic path is therefore neither a miracle nor a postponement forever. It is a sequence of things that have to work: useful capability, adoption, complementary changes, measurable improvement, and a route by which the gains reach people. Housing adds another requirement. More effective demand needs room to become more supply rather than merely a more expensive claim on the same location.
A town that approves computing capacity but cannot accommodate homes for an expanding workforce may discover the mismatch in rents. A town that makes both kinds of investment possible may have a different outcome. The comparison is a policy and planning question, not an argument against the technology.
Abundance has to be built in the places where it is needed.
VIII. A better question for the next headline
The next AI headline will probably offer a very large number. An investment commitment. A company valuation. A power requirement. A productivity estimate. Before translating it into a housing conclusion, ask what kind of number it is.
Is the investment announced, contracted, under construction or operating? Is the power figure capacity or consumption? Is the housing geography a city, a metro or a county? Does the price measure follow the same homes over time or the median of the homes that happened to sell? Is a forecast describing a national economy while the purchase under consideration sits on one constrained street?
Then ask which household it describes. The one with equity to realize? The one exposed to a changing job? The one with an old fixed-rate loan? The one applying today? Without that second question, an accurate statistic can still tell the wrong story about affordability.
For an owner who may need to sell, add a date. How much usable cash might the property produce by then, and what could prevent the sale from completing? A rise in local wealth can coexist with a thin pool of buyers for a particular home. A rate decline can fail to repair a buyer's income uncertainty. An improved valuation does not move money into a settlement account.
This is where simulations can help, if their assumptions are exposed and their performance is tested. One can rehearse different rate paths, buyer incomes, costs and sale timelines, including paths in which no buyer completes by the deadline. The point is not to predict an AI future with theatrical precision. It is to find out which commitments depend on that future arriving conveniently.
The project would need observed listing histories, outcomes, financing conditions and defensible local assumptions. Common exposures matter: several buyers may depend on the same employer, stock price or lending environment. Repeated draws cannot turn missing evidence into knowledge. None of the illustrations here is a validated Oftu forecast for a specific property.
There is a temptation to end this story by choosing a side. AI will save affordability. AI will destroy it. Both endings are easier to write than the one households may actually live through.
A useful technology can impose costs before it distributes gains. It can generate wealth without generating enough nearby homes. It can make work more efficient while leaving a particular worker less secure. It can also help people remove costly frictions, build better systems and widen access. Timing, location and institutions decide how those possibilities meet.
Return to the transformer behind the fence. When it finally arrives, a project can move forward. That is a real achievement. It is not yet an answer to the household on the other side of the road, asking whether it can still afford to stay.
The promise of a more productive economy is worth pursuing. A home is paid for in the economy that exists while we pursue it.
Methodology and reading notes
Research checked September 30, 2026. The news peg is Cook's September 28 speech. The published judgments are hers; they are not a promise about future FOMC decisions. We make no estimate of an AI-attributable change in mortgage rates, electricity bills or home prices.
The Redfin graphic uses the report's table, not its rounded headline: San Francisco +6.0% and Seattle -3.6%, median sale price year over year, July 2026, metro areas. Median prices are affected by the mix of sales. This is not a repeat-sales index or a causal comparison.
The particle forms and two-clock graphic are schematic. A particle is a sampled point on an illustrative surface, not a home, dollar, watt or statistical observation. The clock hands do not represent measured times or estimate the duration of investment or productivity lags. IEA's 2030 electricity figure is a central projection, not an observed value or an AI-only total.
Methodology: The experiment computes monthly principal and interest for a fully amortizing fixed-rate loan over 360 payments. With principal L, monthly rate r and payment count n, payment equals L × r / (1 − (1 + r)^−n); at zero interest, payment is L / n. Principal equals price × (1 − down payment / 100). Comparisons use the selected mortgage rate minus one percentage point, the selected rate, and plus one percentage point, holding principal and term fixed. Comparison rates are bounded between 0% and 15%, so edge cases may repeat a rate. Outputs are rounded for display; calculations retain precision. CSV records inputs, assumptions and scenario outputs. No forecast probabilities or causal AI attribution are included.
The experiment excludes property taxes, insurance, PMI, closing costs, points, fees, maintenance, utilities, refinancing and income qualification. It is an educational sensitivity calculation, not a borrower quote, recommendation or approval. The hypothetical town, contractor and household examples are explanatory constructions, not reported transactions.
