GPU Rental Prices Have Gone Vertical The real risk to AI infrastructure is not demand.

The real risk to AI infrastructure is not demand. It is supply. And the market has it completely backwards.

Since October 2025, GPU rental prices have been climbing. In January, the increases were gradual. By mid to late February, they went vertical. Year to date, rental rates for AI compute infrastructure are up 15 to 20 percent. The trajectory suggests a cumulative increase of roughly 40 percent through 2026. AI labs are accepting renewal price hikes of 40 percent to hold onto existing clusters. On demand capacity is sold out. Lead times for new deployments extend into the middle of the year.

This was not supposed to happen. The consensus view was that new supply would push prices down. The opposite has occurred.

The scarce resource in a gold rush is never the ambition. Ambition is abundant. The scarce resource is the shovel.

Right now, Microsoft has billions of dollars of AI chips sitting in warehouses. Not because nobody wants them. Because the electrical grid cannot absorb them. Satya Nadella said it himself: the constraint is not compute supply. It is the ability to get the builds done fast enough, close to power. The company has an $80 billion backlog of Azure orders it cannot fulfil. Not because demand softened. Because physics intervened.

Meanwhile, Google was forced to cut its own chip production targets. Not because customers cancelled. Because the foundry that makes every advanced AI chip on earth does not have enough capacity. Nvidia has locked up more than half of that foundry’s advanced packaging lines through 2027. Everyone else is fighting for the remainder.

The market is debating whether demand for AI infrastructure is real. The rental rates have already answered the question. The market just has not looked up yet.

I understand the instinct to distrust this. When prices move this fast, the natural reaction is suspicion. I share that instinct in most contexts. But I have a reference point that most financial analysts do not.

Before I managed capital, I built a pilot training business. Not a software company. A physical infrastructure business. Hangars, simulators, aircraft, fuel systems, runway access. I learned early that the person who owns the airfield is in a different position than the person who owns the plane. Planes come and go. Models change. Operators rotate. But the airfield, the physical capacity to receive them, endures. And when demand for flight hours exceeds available airfield capacity, the airfield owner does not need to negotiate. The scarcity does the negotiating for him.

That is the position AI infrastructure owners occupy right now. The chips will evolve. The models will change. The customers will rotate between architectures. But the rack space, the power, the cooling, the interconnection, these are the airfield. And the airfield is full, with a waiting list that stretches into next year.

$650 Billion and It Is Not Enough

In 2026, the four largest hyperscalers are on track to spend roughly $650 billion in capital expenditure. Amazon alone has committed approximately $200 billion. Alphabet has guided to between $175 billion and $185 billion. Meta sits at $115 billion to $135 billion. Microsoft is tracking toward $145 billion annualised. Their combined spend will exceed four times what the entire US energy sector deploys in a year. Amazon’s capex alone is larger than the entire US energy sector’s annual spend.

And it is not enough. Microsoft disclosed a remaining performance obligation of $625 billion. Azure grew 39 percent year over year while being explicitly capacity constrained. Google’s cloud backlog surged 55 percent in a single quarter to over $240 billion. Every hyperscaler has said the same thing on earnings calls: we are supply constrained, not demand constrained.

These are the most profitable enterprises in history. Microsoft generated $25.7 billion in free cash flow in a single quarter while deploying $37.5 billion in capex. Their liabilities to assets ratio actually declined to 48 percent in late 2025, near decade lows. Roughly 75 percent of aggregate capex in 2026 is directed at AI infrastructure. Even if AI revenues grew more slowly than expected, the traditional cloud business alone would sustain enormous demand for the capacity being built. The downside case is not zero demand. The downside case is somewhat slower growth in a market that is already larger than most countries’ GDP.

Goldman Sachs noted that AI capex relative to GDP remains well below the peaks seen in prior technology investment cycles. During the late 1990s telecom buildout, equivalent spending would have required roughly $700 billion in 2026 terms. And that cycle, for all the excess it produced, laid the physical foundation of the modern internet. The current cycle is arguably more durable because the demand signals are more concentrated, more contractual, and more visible than anything the telecom boom produced.

I will state my view plainly: hyperscaler capex in 2027 will approach one trillion dollars. Not because these companies want to make headlines, but because the backlog demands it, the competitive dynamics require it, and the balance sheets can support it. Microsoft’s remaining performance obligations grew 51 percent year over year. Google’s backlog surged 55 percent in a single quarter. If each of the five major hyperscalers increases spending by 40 to 50 percent in 2027, which is below the growth rate they achieved from 2024 to 2025, the aggregate figure approaches $900 billion to one trillion. And this excludes sovereign wealth funds, Chinese technology companies, and the emerging class of AI focused capital allocators.

The $650 billion is not evidence of exuberance. It is evidence of rationing.

Building Ahead of Demand vs Building Behind It

This is the distinction that matters most, and the one the market is getting wrong.

There are two kinds of infrastructure booms. In the first, companies build ahead of demand. They lay cable across oceans, pour concrete into speculative towers, and bet that the customers will eventually arrive. The fibre optic boom of the late 1990s was this kind. Companies built massive capacity on the assumption that internet traffic would fill it. For years, it did not. Utilisation rates sat in the single digits. The revenue was speculative. The bust was brutal.

The AI infrastructure cycle is the second kind. This is a build behind demand. The customers are already here. The contracts are already signed. Microsoft has $625 billion in remaining performance obligations. Google cut its own TPU production targets not because demand fell, but because it could not secure enough packaging capacity from TSMC. The hyperscalers are not building and hoping. They are building and still falling behind.

The difference is not a nuance. It is the entire difference between a speculative bubble and a supply constrained expansion. A bubble collapses when demand fails to appear. A supply constrained expansion resolves when capacity catches up. In one case, the asset becomes stranded. In the other, it becomes more valuable during the constraint period and normalises when supply equilibrates. The failure mode is entirely different.

Every time someone draws a comparison to the dot com bust, ask them one question: did Cisco have $625 billion in signed backlog when the music stopped? Did the fibre optic companies have customers willing to accept 40 percent price hikes to retain capacity? The answer to both is no. The structural conditions are not comparable. The comparison is convenient, but it is wrong.

This does not mean the spending is riskless. I want to be clear about that. GPUs depreciate. Technology evolves rapidly. Useful lives may prove shorter than assumed. Regulatory environments will shift. And the market will, at times, price these assets as though the demand were imaginary, even while the contracts sit in plain sight on the balance sheet. These are real considerations, and they deserve rigorous analysis. But they are engineering risks, not demand risks. And engineering risks are precisely the kind that favour companies with strong balance sheets, long duration contracts, and physical assets that retain value beyond any single generation of silicon. A farmer does not abandon his land because one season is dry. He assesses the aquifer, checks the soil, and plants again.

One Foundry, One Node, Every Chip

The reason the hyperscalers cannot spend more is specific and technical, and it has a name.

Every major AI accelerator launching in 2026 requires fabrication on a single process node at a single foundry. TSMC in Taiwan. The N3 family. Nvidia’s next generation, Google’s TPUs, Amazon’s Trainium, AMD’s Instinct. All of them. Competing for the same wafers from the same production lines. This convergence is not a temporary scheduling conflict. It is a structural collision. The entire AI industry is attempting to pass through a single gate at the same time.

There is one bakery in town that makes the bread everyone needs. It runs at full capacity every day. There are enough customers to buy three times the output. More ovens are being built, but they will not be ready for eighteen months. The bread is allocated, not sold. The customers who secured their orders earliest get served. Everyone else waits.

Nvidia has secured more than half of TSMC’s advanced packaging capacity through 2027. Google was forced to cut its 2026 TPU production target from four million to three million units because of limited packaging access. The advanced packaging process known as CoWoS, which integrates processors with high bandwidth memory on a silicon interposer, remains a separate and equally binding bottleneck. TSMC has been scaling this from roughly 35,000 wafers per month in late 2024 toward 130,000 by the end of 2026. That sounds like rapid growth. It is. But demand has grown faster. Every additional wafer is consumed before it comes online. High bandwidth memory from SK Hynix, Samsung, and Micron is fully allocated through 2026 with double digit price increases already underway.

Here is a number that puts the fragility of this supply chain in perspective. AI chips represented less than 0.2 percent of global wafer starts in 2024, yet already generated roughly 20 percent of semiconductor revenue. That extreme concentration on a single space explains why these shortages feel different from anything that came before. Today’s AI accelerators require leading edge logic, exotic memory stacks, and advanced packaging that cannot be expanded quickly. The bottleneck is not a temporary disruption. It is an architectural fact.

Behind the silicon wall sits the power wall. Data centres require enormous electricity. AI workloads require materially more than traditional cloud computing. The US grid was not built for this. Connection timelines stretch beyond four years in major markets. Northern Virginia and Texas are turning away new projects because they have exhausted available power capacity. Microsoft’s processors sit in warehouses waiting for the grid to catch up. Building a data centre now takes over three years. Securing utility interconnections takes longer. These are physical facts that do not compress in response to capital expenditure. For infrastructure providers that already have power, already have energised sites, and already have tenants, this is not a problem. It is a moat that widens with every month of delay for everyone else.

Why Rental Prices Are the Signal That Matters Most

I opened this letter with the rental rate data because I believe it is the single most important signal in AI infrastructure investing right now. More important than earnings beats, more important than capex guidance, more important than analyst sentiment. Let me explain why.

Rental prices are a real time clearing price for physical compute. They cannot be manufactured by press releases or inflated by accounting. When a GPU hour trades at a higher price than it did six months ago, that is the market telling you, in cash, that demand exceeds supply at the current price. There is no way to fake that signal.

What makes the current pricing environment especially striking is that the increases are no longer explained by rising input costs alone. Server component prices have risen, reflecting memory shortages and tariff pressures. But rental prices are now increasing well beyond what those input costs would justify with a constant margin. The implication is clear: margins for infrastructure owners are expanding, not compressing. And they are expanding at a pace that the market has not priced in.

From what I can observe, it has become a genuine challenge to procure not only next generation Blackwell capacity but even the prior generation Hopper capacity. At the higher performance tiers, availability is spoken for through August or September. At lower tiers, through May or June. The constraint is data centre space and GPU shipments. The market is tight across the entire stack. Anyone who needs compute and does not already have it under contract is facing a difficult few quarters.

The reason is structural. AI consumption has changed. The first wave was conversational: a question, an answer, a predictable amount of compute. The current wave is agentic. AI systems running multi step workflows autonomously, planning, executing, verifying, iterating at high concurrency. A single agentic session can consume more tokens than dozens of traditional conversations. The adoption curve for agentic coding tools echoes the original ChatGPT moment, with some tools already accounting for a meaningful share of public code commits. Media generation tools producing billions of images per month add another demand layer that did not exist eighteen months ago. Video generation, where each output requires processing enormous numbers of frames and pixels, is even more compute intensive. Users iterate rapidly to refine results, multiplying consumption with each attempt. The demand curve is not flattening. It is steepening.

For infrastructure owners, the margin story is transformative. Project level operating margins, already strong at roughly 40 to 45 percent, are expanding toward 50 percent and above on renewal. When an existing compute fleet reprices into the current rental environment, the EBIT increase on that capacity can approach 40 percent. Every tranche of capacity that comes off an older contract and re-enters the market generates materially higher cash flow. This is not a story of survival. It is a story of strengthening economics in a supply constrained environment. And the providers that have capacity coming off contract in the current pricing environment are the direct beneficiaries, even as the market punishes their stock prices on sentiment alone.

The Market Has It Backwards

Here is what I find remarkable. Investor sentiment toward AI infrastructure providers has turned increasingly negative at the exact moment the underlying economics have strengthened the most. Rental rates are surging. Margins are expanding. Capacity is sold out. Backlogs are growing. And the stocks are falling.

The market is pricing these companies as though $625 billion in signed backlog does not exist. That is not scepticism. That is innumeracy.

I have seen this pattern before. The physical economy and the financial economy diverge for a season, sometimes for several seasons. Sentiment leads price in the short term. Fundamentals lead price in the long term. And the reconciliation, when it comes, tends to be swift and uncomfortable for those on the wrong side.

I learned something crossing the Bass Strait at night in a single engine aircraft that has never left me. When you are suspended over black ocean with no visible horizon, your inner ear will lie to you. It will tell you the world is tilting when the instruments say you are level. The pilots who survive are the ones who trust the instruments over their own senses. The ones who follow the feeling into the darkness do not come back.

The market’s inner ear is telling it that AI infrastructure is tilting toward a bust. The instruments say otherwise. The contracts are binding. The rental rates are surging. The capacity is sold out. The backlog is growing. I trust the instruments.

The risks are real, and I take them seriously. GPUs depreciate. Technology evolves. Useful lives may prove shorter than assumed. Regulatory environments will shift. These are engineering risks, and they deserve careful analysis. But they are not demand risks. And engineering risks favour companies with binding contracts, strong counterparties, and physical assets that retain value beyond any single chip generation. A farmer does not abandon his land because one season is dry. He assesses the aquifer, checks the soil, and plants again.

The tenants are the most creditworthy companies in the history of capitalism. The contracts are binding, multi year, and prepaid. The utilisation is at or near one hundred percent. The rental rates are surging. And the supply chain cannot deliver fast enough to satisfy what the hyperscalers have already committed to purchase.

In most market environments, the primary risk for infrastructure assets is demand. Will the tenants show up? Will the contracts be renewed? Will the utilisation hold? In the current cycle, those questions have been answered. The risk has inverted. The risk is not that demand evaporates. The risk is that the supply chain cannot deliver fast enough. For those who own the physical assets, this is the position of the landlord in a city where every building is full, where the waiting list is growing, where new construction will take years, and where the tenants are Fortune 5 companies who have signed long term leases at rates the market did not expect.

The world wants to build the infrastructure for the age of intelligence. The shovels are in short supply. The gold is real. And we own the ground where both converge.

I do not know what the market will do next quarter. I never have. But I know what the contracts say, I know what the rental rates are doing, and I know the difference between a thesis built on hope and a thesis built on arithmetic. Ours is built on arithmetic. That is enough.

Neel Khokhani

Founder and CEO, Epochal Corporation

@neel_epochal