Who Pays for the Buildout?
AI demand is visible. The financing channel underneath it is not.
Who’s ready for a long weekend? I know I am, but while you are out doing what you do for the Memorial Day weekend how about a little light reading? So without delay, let’s dive into that layer sitting below all of this CAPEX buildout!
So everyone is still trying to figure out the next layer of the AI trade, and I get why. The first layer was easy enough to see. GPUs were scarce, hyperscaler capex exploded, the largest companies in the world started spending like small governments, and the market rewarded the companies sitting closest to the shortage. Then the bottleneck moved into memory, packaging, power, networking, and data-center infrastructure, which is why we keep ending up back in the same conversation even when we are trying to move on from it.
But I think there is another layer underneath all of this that is getting less attention.
The question may not be simply what needs to be built, because we already know a lot of that answer. AI needs compute, memory, data centers, power, cooling, transmission, and chips. Defense needs production capacity, electronics, drones, shipyards, sensors, secure communications, and supply-chain redundancy. Energy security needs grids, turbines, transformers, substations, natural gas backup, nuclear optionality, storage, and equipment.
The better question is who finances it, because eventually all of these themes stop being stories and start becoming balance-sheet problems. So this is what I mean:
Someone has to absorb Treasury issuance.
Someone has to absorb the hyperscaler bond issuance.
Someone has to finance infrastructure.
Someone has to fund utilities, data centers, manufacturers, defense suppliers, and project developers.
Someone has to hold duration, extend credit, and take the other side of a capital cycle that is no longer asset-light.
So this is where I think the market may still be missing something. The capex cycle is being discussed mostly as an earnings story, but it may become a credit story. If that is right, then the next phase of AI, power, defense, and overseas industrial capacity may not be decided only by demand, but by whether the financial system is willing and able to fund the physical buildout already embedded in expectations.
Our current liquidity posture does not say we are there yet. The model is still defensive, and I do not want to pretend otherwise. This is not an all-clear signal, and it is not a green light to assume that every infrastructure, power, or AI-adjacent company suddenly deserves a higher multiple because the theme sounds right. If anything, the current setup is more uncomfortable than that, because the old regime is still visible, the new one has not fully arrived, and the transition between the two is where i’m starting to feel a little uneasy.
Now the reason the plumbing matters is that the composition of liquidity has changed, even if the market has not yet been forced to care about it. Reverse repo balances have been drained down to nearly nothing, which means the easy money-market liquidity cushion that helped absorb prior Treasury issuance is mostly gone. Bank reserves remain large, but Treasury funding needs are still very much with us. Treasury’s latest quarterly financing estimate, released May 4, showed expected April through June borrowing of $189 billion in privately held net marketable debt, assuming a $900 billion end-of-June cash balance, and that borrowing estimate was $79 billion higher than previously announced. But what’s a few billion amongst friends?
That doesn’t mean we have a funding crisis sitting on the doorstep. It does mean the buyer base matters more than it did when the Fed was the dominant balance sheet in the room and money-market liquidity had more room to absorb the pressure.
The April Senior Loan Officer Opinion Survey is important because it does not yet confirm the bullish version of this thesis. Banks reported tighter standards for commercial and industrial loans and weaker demand from firms of all sizes, which means we are not yet seeing broad credit deployment. This is not a piece saying banks are already lending aggressively into a new boom. It is a piece saying the conditions may be forming where liquidity eventually moves from hoarding to deployment, and the confirmation won’t be coming from another AI press release, it‘ll be coming from lending standards, project-finance appetite, bank earnings commentary, infrastructure credit demand, Treasury auction absorption, and whether reserves remain ample as the balance-sheet runoff phase matures.
The term premium is part of the same story. When the long end of the Treasury curve starts demanding more compensation, the market is telling us something about fiscal supply, inflation uncertainty, oil risk, foreign demand, and the question of who absorbs duration when the Fed is no longer the automatic marginal buyer. The exact decimal point matters less than the direction. The long end is not just trading the next Fed cut anymore; it is trading the cost of financing a much larger physical economy.
So, that’s the tension running underneath the market right now. We have a market that wants lower rates because AI, power, infrastructure, and defense all become easier stories when financing costs fall, but we also have a Treasury market reminding everyone that capital is not free. We have enormous demand for physical investment, but we don’t yet have clear evidence that private credit creation is accelerating enough to fund it smoothly. We have banks with large reserve balances, but we also have tighter lending standards and a regulatory backdrop that does not exactly invite reckless balance-sheet expansion.
So the thesis has to stay probabilistic. The opportunity is not that a credit boom is guaranteed but rather the opportunity is that the market may be underpricing the possibility that the end of the liquidity drain eventually creates the conditions for a new private-sector balance-sheet expansion. If that happens, the beneficiaries are not just banks. The beneficiaries are the parts of the economy that need financing to turn themes into assets.
The way I would frame the AI and infrastructure cycle now is in three layers.
The first layer was GPU scarcity. That was the easiest phase to see because the shortage was visible, the winner set was narrow, and the earnings revisions were immediate.
The second layer has been infrastructure bottlenecks. That’s where the market moved into memory, packaging, power, cooling, transmission, substations, grid equipment, and the overseas industrial stack that supports the physical buildout.
The third layer may be the financing channel, and that’s where the detective work still sits. Most AI leaders are no longer early, memory is no longer ignored, and power infrastructure has already been discovered in many places. But the financing layer is less obvious because it is not about whether the demand exists, but whether the credit system is willing and able to fund the demand already embedded in prices.
That is why the current model signals matter. Our gates are not passing yet, term premium remains too high, and the liquidity posture is still defensive, which means this is not the point where we declare the turn. It is the point where we start looking for evidence that the next regime is forming before the broader market sees it.
Power infrastructure is probably the cleanest example because data centers need electricity, and that part of the story is no longer obscure, but building the power stack requires enormous financing. Generation, transmission, substations, transformers, grid hardening, backup power, cooling, and interconnection are capital-intensive, and utilities can’t just wish this into existence and for that matter, neither can hyperscalers. If private balance sheets begin deploying capital more aggressively, the power theme gets more than a demand story, it gets a financing channel!
That distinction matters because the market has already started to understand AI power demand, but it may not fully understand the credit impulse required to fund the solution.
The same logic applies to defense and strategic autonomy, which are no longer just geopolitical talking points. Europe is increasing defense commitments, Japan is expanding its strategic industrial base, and the Middle East has reminded everyone that energy routes, shipping lanes, production capacity, and military readiness aren’t theoretical concerns. But defense modernization is not only about budgets, it’s about procurement cycles, production capacity, supplier financing, working capital, and the ability of private industry to expand without suffocating under its own capital requirements.
A theme can be right and still fail as an investment if the balance sheet can’t carry it.
That’s very true for AI infrastructure as the market has spent two years rewarding demand, but i’m seeing signs that demand isn’t enough anymore. The next phase requires the ability to finance data centers, power contracts, custom silicon, memory capacity, advanced packaging, cooling systems, and the industrial base behind all of it. This is where the overseas angle still matters, but in a more disciplined way than simply saying AI is moving east.
Korea matters because of memory. Taiwan matters because of foundry and advanced packaging. Japan matters because of materials, tools, components, industrial policy, and energy-import sensitivity. Europe matters because of power equipment, automation, defense electronics, and grid infrastructure. But all of those stories depend on financing conditions. If credit stays tight, the theme is still real, but the timing stretches. If credit begins to open, the physical buildout can accelerate.
This is also where we have to be careful with the sector labels. This isn’t about owning every bank, every utility, or every company with AI infrastructure in the description. The better question is which institutions can deploy capital into infrastructure lending before the cycle becomes obvious, which power and grid companies have the balance sheet and financing access to convert backlog into earnings, and which industrial companies can fund growth without having to dilute shareholders or depend on heroic assumptions around capital costs.
For power infrastructure, demand alone is not enough anymore. We need to know who has secured credit facilities, who can finance backlog, who can pass through costs, and who can keep projects moving if rates stay higher for longer. For banks and financial infrastructure, the question is not traditional net interest margin alone; it is whether liquidity moves from reserves into productive credit, especially in areas tied to infrastructure, industrial capex, data centers, power, and defense production.
This is where our process has to separate itself from the crowd. We aren’t simply trying to find a new slogan, but rather trying to identify where the next source of earnings power forms before it becomes obvious in the estimates. The first phase of AI was a demand shock, the second phase was a bottleneck shock, and the next phase may be a financing test.
Can the system fund what the market is already pricing?
Right now, the answer isn’t totally clear to me. The lending survey doesn’t yet say yes, treasury supply is still heavy, term premium is elevated, oil and Middle East risk are keeping inflation psychology uncomfortable and the long end isn’t giving the market an easy pass. So that’s why the liquidity model remains defensive even though the forward setup is becoming more interesting.
Our confirmation gates are fairly clear. The next Senior Loan Officer Opinion Survey matters, but so do bank earnings calls, especially any commentary around commercial and industrial lending, infrastructure project pipelines, construction loan demand, and project-finance appetite. Treasury auction tails and bid-to-cover ratios matter because they tell us whether the market can absorb supply without forcing yields materially higher. Utility and infrastructure earnings matter because capex guidance is only useful if it comes with believable financing access. Credit spreads, term premium, reserve volatility, construction and development loan growth, infrastructure bond issuance, and backlog-to-financing ratios all become part of the same question.
Now can the system fund what the market is already pricing?
If those indicators start improving together, the market may have to move from “higher rates are choking the capex cycle” to “credit is beginning to fund the capex cycle.” If they don’t, then the theme may still be right, but the timing stretches and valuation discipline becomes even more important.
That would be a meaningful regime shift, and it would change how we think about sector leadership. The winners would not simply be the companies with the most exciting theme exposure. They’d be the companies with projects that can actually get financed, backlogs that can convert, balance sheets that can survive the transition, and pricing power that can turn capital scarcity into earnings rather than dilution. That’s quite a different story from chasing every AI-adjacent, power-adjacent, or defense-adjacent name that has already re-rated.
Do valuations still matter? I wonder sometimes, but yes, valuation still matters. In fact, valuation matters more in this part of the cycle because the obvious themes are no longer ignored. Memory is definitely not unknown, grid equipment isn’t unknown, European defense is less known but still known, and power is no longer a secret either. So the edge isn’t in discovering that these things matter. The edge is in figuring out where the financing environment can still improve the forward earnings path enough to justify the price.
For memory and semiconductor supply chains, the question is whether earnings revisions can keep outrunning the re-rating. For power and grid equipment, the question is whether backlog converts into margins or gets eaten by costs, rates, politics and wars. For defense, the question is whether announcements turn into orders and orders turn into production. For banks and financial infrastructure, the question is whether liquidity moves from reserves into productive credit or simply stays trapped in defensive balance-sheet management.
So this is where our three-phase investor model helps us determine a crowded story. Some parts of the AI trade are already late Phase 2 or Phase 3. The obvious GPU and memory leaders no longer require detective work. They may still be very good companies, but they are definitely not undiscovered. Power infrastructure is somewhere in Phase 2, with some names already behaving more like Phase 3 momentum trades. Robotics and physical AI are earlier, but the hard deployment evidence is not yet strong enough to make them the center of this thesis.
The financing layer may still be closer to Phase 1, not because nobody understands banks or Treasury supply, but because fewer investors are connecting the plumbing setup to the next phase of industrial capex. The market can see AI demand, but who can’t? It can see power demand and it can see defense urgency. What it may not be fully pricing is that the next leg depends on the balance-sheet channel that turns those demands into funded projects.
The Fed may think it’s managing the final phase of balance-sheet normalization and Treasury may think it is simply funding the government. Banks may be waiting for cleaner lending conditions while Investors may think they are still choosing between AI winners and losers. However, underneath all of that, a more important question is forming: if liquidity stops being drained and private balance sheets begin to expand, what gets funded first?
My guess isn’t the speculative edges of the economy but rather the strategic middle which includes power, data centers, defense, grid equipment, energy security, industrial automation, supply-chain redundancy, and the infrastructure required to make the AI cycle physically possible.
Now I’m not saying to abandon defensive discipline, but I do think we should start watching for the transition.
The next Senior Loan Officer Opinion Survey
Bank earnings commentary
Project-finance language
Treasury auction tails and bid-to-cover ratios
Term premium, reserve volatility, and credit spreads
Utility and infrastructure companies talking about financing access
None of these signals alone tell the whole story, but if you see them moving together the market may be forced to update its view.
The AI trade may still be about chips, memory, power, and overseas industrial capacity, but the next regime may be decided by credit as much as demand. If the physical buildout is going to validate the earnings expectations already embedded in the market, someone has to finance it, and that is why the obscure question under the capex cycle may be the most important one: not just what gets built, but whether the credit cycle is ready to underwrite it.
Luke Perry
Whalen Financial, Portfolio Manager