Have you noticed how every board meeting now has a slide about artificial intelligence, even when the company does not train a model or own a single graphics card? That is not just fashion. Something quieter is happening underneath the slogans. Money is being pulled toward servers, power plants, chips and cooling systems at a scale that would have looked theatrical a few years ago. The productivity story may still be incomplete. The bill for the buildout is already on the table.
Why The Buildout Hits Capital Before It Hits Prices
Most conversations about AI capital repricing start at the wrong end. People talk about jobs that might vanish, software that writes itself, or valuations that look stretched. Fair topics. They are not the first-order effect I keep coming back to. Before an economy can produce more with less, someone has to pour concrete, string transmission lines and lock in multiyear power contracts. Investment arrives first. Efficiency, if it arrives at all, comes later.
That sequencing is not a footnote. It changes the price of money for firms that will never touch a data hall. An apartment building, a regional manufacturer, a hospital system expanding a wing. None of them need an AI strategy for the opportunity set around them to shift. If capital can earn a more compelling risk-adjusted return financing compute and power, every other project has to compete with that alternative.
I have found that this point gets lost because it is less cinematic than robot labor. It is also more immediate. You can already see it in the mix of business spending. Categories tied to digital infrastructure are running hot. Offices and some traditional structures look tired by comparison. That is not a morality tale. It is a relative price story.
The Spending Wave Is Physical, Not Abstract
It is easy to treat AI as a cloud floating above the real economy. Look closer and it is steel, copper, transformers, land and water. Industry estimates put global data center capital needs in the multi-trillion range through the end of the decade, with the bulk tied to AI workloads. That is not a software license. That is a construction cycle.
The investment comes first. The productivity comes later. Confuse the two and you will misread both inflation and asset prices.
Recent official commentary has already flagged how much of the jump in business fixed investment traces back to infrastructure that supports AI services. Growth in that bucket can look robust even while other forms of structures lag. If you only watch the headline investment number, you miss the rotation inside it.
Power is the binding constraint that keeps surprising people. A model does not run on a press release. It runs on electricity that has to be generated, moved and cooled. That pulls capital into generation, transmission and related equipment. Governments are still heavy borrowers. Traditional infrastructure still needs financing. Companies still want to expand. Stack those claims and you do not need a shortage of savings for the opportunity cost of capital to rise.
Capital Does Not Need To Be Scarce To Get More Expensive
After the financial crisis, money felt abundant and cheap for a long stretch. Investors hunted for yield. Firms borrowed because the coupon barely stung. Real estate enjoyed low required returns. Cash sitting in an operating account did not feel like a decision. The penalty for sloppy liquidity was tiny.
That world is fading, and AI is one reason among several. Data centers need capital. So do semiconductor plants, grid upgrades and the vendors that feed them. If those uses look productive, they bid against everything else. The economy does not have to run out of dollars. Alternatives only have to look better.
Think about a mid-rise rental property. Occupancy might be fine. Rents might be steady. Operating costs might not have jumped in a dramatic way. Still, the building now competes with a different menu of projects. An investor who can finance power and compute with a cleaner spread will demand more from the apartment deal. The asset did not change. The hurdle around it did.
- Investors compare risk-adjusted returns across sectors, not inside one silo.
- A higher outside option lifts required yields even when local cash flows are stable.
- Real estate, private credit and ordinary corporate projects all feel that comparison.
- Cash on the balance sheet stops being inert and starts looking like a choice.
In my experience, people underestimate how quickly that comparison travels. You do not need every pension fund to become an AI specialist. You only need enough capital at the margin to prefer the new story. Prices adjust at the margin. Always have.
Hurdle Rates Move From The Market Into The Firm
Bond yields grab headlines. Inside companies, the quieter shift is the internal hurdle. A project that cleared the bar when money cost five percent may fail when the bar sits nearer eight. A plant expansion slips a year. An acquisition no longer pencils. Paying down debt starts to look like the grown-up move. Inventory and working capital get a harder look.
That is not austerity theater. It is arithmetic. Higher required returns change which ideas get funded. Management becomes pickier. That pickiness shows up as delayed capex in sectors far from silicon. The AI boom can crowd other projects without anyone writing a memo titled crowding out.
Cash changes character too. When policy rates hovered near zero, idle balances earned almost nothing. Waste was cheap. When safe assets pay a real coupon and borrowing still costs something, every dollar on the balance sheet has a measurable opportunity cost. Treasury work stops being a back-office chore and becomes part of capital allocation.
A firm can put that dollar into the business, cut debt, send it to owners, hold it for liquidity, or park it in a market instrument until it is needed. Those are strategy choices, not clerical ones. I have watched finance teams rediscover this the hard way after years of treating cash as wallpaper.
Three Prices Of Money That Do Not Move In Lockstep
It helps to separate the clocks. The central bank sets an overnight policy rate. Markets set longer yields. Boards set hurdle rates from those benchmarks, plus risk, plus whatever else they can earn. Those three do not have to travel together.
Policy rates can ease if inflation cools while investors still demand a fat premium to lock money up for ten or thirty years. A soft patch can pull both policy rates and required returns down. The interesting question is not only whether AI pushes the overnight rate up or down. It is whether AI lifts the marginal cost of capital across the economy before the full productivity dividend shows up in unit costs.
| Price of money | Who sets it | What it governs |
| Overnight policy rate | Central bank | Short funding and sentiment |
| Long-term market yields | Investors and supply | Discount rates for long assets |
| Internal hurdle rate | Management | Which projects actually get built |
That table looks simple. It is where a lot of confusion lives. Commentators treat one rate as a stand-in for all three. Then they act surprised when corporate behavior does not match the latest policy statement. Different clocks. Different audiences.
Do Not Mix The Construction Phase With The Mature Economy
None of this tells you where rates settle once AI is ordinary. Rapid labor displacement could weaken demand and eventually pull rates lower. The infrastructure boom could overshoot, leaving empty halls and stranded generation. Or the technology could work unusually well, lifting productive capacity, helping some manufacturing niches, and pressing down the cost of goods.
Several of those paths can run at once. That is the messy part. Those are questions about a mature AI economy. They matter. They are also harder to forecast than the capital cycle sitting in plain view.
Today the demand for investment is observable. Trillions are being committed to physical and digital plant. Labor, energy, equipment and finance are being deployed now. Much of the promised productivity still sits ahead of us. That gap is the story. The economics of the buildout can look nothing like the economics of the finished system.
Perhaps the most interesting aspect is how little this depends on whether you personally love the technology. You can be a skeptic on model quality and still accept that the capex is real. You can be a true believer and still admit that the payoff is uneven and delayed. The financing happens either way.
Who Feels It First Outside Tech
Commercial real estate is the cleanest classroom. Cap rates are a story about required returns. If alternative uses of money look richer, cap rates have a reason to stay firmer even when local rents are not collapsing. That does not doom every building. It does change which buildings clear and at what price.
Manufacturers that need long-lived equipment face a similar squeeze. A tool that looked fine at a lower cost of capital can wait. Private credit that used to fund ordinary mid-market deals now competes with infrastructure stories that sound more strategic. Even municipal borrowers feel the comparison when large private projects bid for the same pools of long money.
- Map your projects against the new outside option, not last year’s hurdle.
- Treat cash as an allocated asset, not leftover change.
- Stress power, land and grid timelines if your sector sits near the buildout.
- Assume long yields can stay sticky even if overnight rates drift.
- Ask which of your assets are substitutes for AI-linked claims and which are not.
That list is practical on purpose. Theory is cheap. Boards fund things. If the comparison set has changed, the model has to change with it.
Inflation, Energy And The Awkward Overlap
AI did not single-handedly force the latest policy move, and it would be sloppy to pretend otherwise. Inflation, energy prices and broader demand still matter. The coincidence is still worth sitting with. Capital spending is firm. A large share of the extra growth in that spending looks tied to the AI complex. Inflation has not simply vanished. The mix is unusual.
Energy sits in the middle of the knot. Training and inference eat power. Power markets were already tight in some regions. Add a multiyear wave of demand that is relatively price-insensitive in the short run and you get pressure on both kilowatt hours and the capital that builds new supply. That can feed into measured inflation even as software types talk about deflation later.
I do not think that means AI is inherently inflationary forever. Timing again. Construction is a demand shock. Diffusion is a supply shock, if it works. We are living in the first act and arguing about the third.
What Companies Get Wrong About Opportunity Cost
Plenty of firms still budget as if the last decade’s cheap money was a law of nature. They approve projects with thin spreads because “this is how we have always grown.” They let working capital sprawl. They treat the treasury stack as plumbing. That habit gets expensive when the outside option improves.
Another mistake is copying the hyperscale playbook without the hyperscale balance sheet. Not every company should own compute. Some should rent it. Some should wait. Some should partner. The capital cycle rewards discipline more than slogans. I have seen more value destroyed by fashionable capex than by cautious delay.
An apartment building does not need an AI strategy for AI to change its valuation. The same is true of a factory, a warehouse and a mid-market lender.
A third mistake is collapsing all rates into one number from the morning news. Overnight cuts do not automatically rescue a thirty-year project. Credit spreads can widen while the policy rate falls. Hurdles inside the firm can stay high because management does not trust the cycle. Watch the whole stack.
Productivity Is The Prize, Not The Present
Give the optimistic case its due. If capabilities diffuse, firms should eventually produce more with the same or fewer resources. That can restrain unit costs and ease price pressure. Research accelerates. Software gets cheaper to write. Decisions improve at the margin. Those are real possibilities, not science fiction.
They are also uneven. Diffusion takes management time, process redesign, legal review and worker adaptation. Some sectors will capture gains quickly. Others will spend years wrestling with messy data and half-finished tools. Average productivity can rise while many firms feel nothing but higher power bills and pickier lenders.
That is why I keep separating the buildout from the equilibrium. We are financing construction. We are not yet living in the finished city. The first broad economic imprint of AI may not be that everything gets cheaper. It may be that capital itself gets more valuable.
A Practical Filter For Allocators
If you allocate capital for a living, the filter is blunt. Ask whether a claim is a substitute for the AI infrastructure complex, a complement, or something orthogonal. Substitutes have to pay more to keep you. Complements can ride the spend. Orthogonal assets live or die on their own cash flows, but they still face a higher outside option.
Duration matters. Long assets feel required-return shifts more than floating short paper. Leverage matters. Thin equity cushions leave less room when discount rates tick up. Power exposure matters, even for businesses that think they sell something else. In tight grids, energy is no longer a background cost. It is strategy.
Simple allocator checklist: 1. What is the competing use of this dollar? 2. How long is the cash locked? 3. Who supplies the power and on what terms? 4. Is the productivity story in the price already? 5. What happens if the buildout overshoots?
That last item is the one people skip. Booms create capacity. Capacity can arrive late and all at once. If halls, chips and turbines overshoot, the capital cycle can flip from scarce returns to stranded steel. That does not cancel the present squeeze. It is a reminder that cycles have two sides.
Policy Can Ease And Still Leave Capital Dear
There is a temptation to wait for the next policy pivot and assume the old world returns. Maybe parts of it will. Overnight rates can come down if inflation behaves. That would help floating-rate borrowers and short funding. It would not automatically restore the era when almost any project cleared because money was nearly free.
Heavy public borrowing, large private infrastructure claims and a richer set of productive uses can keep term premia from collapsing. I might be wrong about the size of that effect. I do not think I am wrong about the direction of the question. The question is relative returns, not a single administered rate.
Households feel a version of this too, even if they never say “hurdle rate.” Mortgage math, car loans, and the return on cash in a savings vehicle all sit inside the same family of prices. When the opportunity cost of money rises, patience gets more expensive and so does stretching for yield in the wrong place.
The Human Texture Of A Capex Boom
Strip away the macro language and you get ordinary frictions. Towns arguing about water and land use. Utilities scrambling for transformers that suddenly have a waiting list. Contractors bidding against each other for electricians. University labs competing with private labs for the same scarce researchers. This is what a buildout looks like on the ground. It is not a slide deck.
Those frictions are part of why productivity lags spending. You cannot cool a hall that is still waiting on equipment. You cannot train at scale without power that is contracted and delivered. You cannot diffuse tools into a hospital or a factory overnight. The lag is structural, not just statistical.
I keep coming back to that lag because it is where markets get impatient. Equity stories want the end state now. Credit stories have to fund the middle. The middle is messy, capital intensive and politically noisy. That is also where the repricing of capital actually happens.
What This Means If You Run A Normal Business
You do not need a moonshot. You need a clearer cost of money. Revisit project IRRs with a higher outside option. Tighten the process for approving long-lived assets. Make treasury report opportunity cost the way operations reports scrap rates. If a deal only works because last decade’s discount rate is still hiding in the spreadsheet, it does not work.
Be honest about whether AI tools change your unit economics this year or merely decorate a pitch. Plenty of useful software will never justify a data campus. Rent intelligence where it pays. Build it only where control and scale truly matter. That sounds obvious. It is not how a boom psychology usually operates.
Watch your lenders. Covenants, tenors and spreads can shift even if your own story is unchanged. A bank that can place money into infrastructure-linked credits may get choosier with vanilla borrowers. Relationship still counts. Arithmetic counts more when the alternative is glittering.
Holding Two Ideas At Once
AI may eventually lower the price of many goods and services. It is already changing the price of capital. Both statements can be true. The error is collapsing them into a single timeline.
If the technology delivers, the later acts can look deflationary in pockets and wildly abundant in others. If it disappoints, we will still have paid for a lot of concrete and copper. If it overshoots, the hangover will rhyme with other capex busts, even if the gadgets are new. None of those endings erase the present fact that claims on money are being reordered.
We are not yet living in the mature AI economy. We are financing its construction. That is a less glamorous sentence than the ones that trend. It is also the one that should sit at the top of an investment memo this year.
So here is the practical close. Look through the slogans. Follow the capital. Ask what else that capital used to fund. Then decide whether your projects, your properties and your balance sheet can clear a higher bar. The models can wait. The cost of money is not waiting with them.
