--- title: "After the AI Infrastructure Buildout" author: David Hernandez (@davernandez) date: 2026-07-17 url: https://x.com/davernandez/status/2078124835746513127 source: X article --- # After the AI Infrastructure Buildout Originally published on X on July 17, 2026: https://x.com/davernandez/status/2078124835746513127 This piece was written July 3rd, 2026. Every transformative, commoditized infrastructure buildout in history has overshot near-term demand and destroyed the capital of the people who got there first. The technology itself succeeded in the end, but its early assets were acquired at a fraction of build cost by capitalized second movers. Think of the examples of the past: the UK railway mania of the 1840s, the US rail busts of the 1890s, and the 1990s-2000s fiber buildout all follow this familiar pattern. The current AI capex cycle is the largest and fastest infrastructure buildout in history (roughly $725 billion combined 2026 capex for the four largest hyperscalers, up 77% from ~$410 billion last year). But currently the financing structure is showing some question marks the BIS recently warned about in its annual report: circular vendor financing, SPV debt taken off balance sheets, GPU-collateralized loans against a rapidly depreciating asset, and widening credit spreads at the most levered names. The durable fortune from the buildout likely comes from what's left after the dust settles - after the bubble pops - and those well-positioned for distressed situations would be targeting depreciation-resistant assets (grid interconnection rights, power purchase agreements (PPAs), energized shells, fiber, land) bought from forced sellers (overlevered neoclouds, single-tenant developers, ex-Bitcoin miners, private credit-backed shells), while avoiding depreciating GPUs. The historical precedent would be Cogent buying roughly $14 billion of telecom capital for about $60 million, and Chevron buying Noble Energy's reserves for under $5 per barrel of oil in 2020. Auditing the demand quality This year’s capex is justified by at least four distinct kinds of demand that differ in durability, credit quality, and elasticity. This is important to distinguish what kind of AI demand is able to service long-duration liabilities during a downturn. Paid demand with unit economics, like enterprise API consumption, consumer subscriptions, and usage-based inference that customers renew because it consistently produces value above its price. This is the foundation of revenue, and it's small relative to the capital deployed against it. The gap between revenue and capex is not proof of a bubble, as early infrastructure always frontruns revenue, but the size of the gap defines how much of the buildout depends on the other layers of demand holding. Venture-subsidized demand. A meaningful share of GPU consumption comes from startups whose compute bills are paid from raised capital rather than customer revenue, making this demand a derivative of capital market conditions. This is the demand class that vanished in 2001, when the dot-coms that were the carrier customers stopped raising and therefore stopped spending, converting growth projections into stranded capacity within a few quarters. The 2026 version is larger and burns faster, and sustained by the assumption of continued access to new funding. Circular/double-counted demand. This category of demand has three layers; Vendor financing: NVIDIA committing capital to OpenAI, xAI, and CoreWeave, its own largest customers, which means chip demand is partly funded by the chip seller. Customer-supplier circularity: Oracle borrows tens of billions to build capacity for OpenAI, whose ability to pay rests substantially on future fundraising, which rests on growth expectations, which rests on access to compute. Each step looks rational in isolation but the loop means one misstep propagates as both a demand shock and credit shock. The leverage here is magnified. Double counting: the same handful of frontier labs anchor the forecasts of Microsoft, Oracle, CoreWeave, Google, and Stargate all at once. Every provider books the same future demand, and every provider's lenders underwrite it. To again compare to the telecom days, ten carriers each planned to carry the same traffic. The traffic was real, but the sum of the plans was not. Strategic and sovereign demand. Hyperscalers train frontier models for competitive position rather than direct revenue, and nations increasingly buy sovereign compute as akin to national defense. This demand is strong credit-wise and insensitive to price, but arrives infrequently and can pause abruptly. It supports the hyperscalers but cannot be modeled as recurring revenue for neoclouds and other smaller players. All of this happens while token prices deflate. The price of an average token has collapsed, driven by model efficiency, hardware optimization, and pricing floors set by open-weight models that make last year's frontier capability literally free. This is a technology priming for mass adoption, but the math is rough for compute asset owners: revenue per unit of installed compute trends down even as physical demand trends up, so owners must run ever harder to service fixed costs. Similar to a few decades ago, bandwidth demand compounded for twenty years without interruption, but bandwidth prices fell faster, which is why traffic growth and carrier bankruptcy happened at the same time during the tech burst. The hope that enterprise token adoption converts into durable revenue at scale is running slower than the capex assumes. The evidence is pretty clear: PwC's January 2026 survey of 4,454 chief executives across 95 countries found just 30% reporting any revenue increase from AI in the prior twelve months, with 56% saying AI delivered zero cost or revenue improvement. McKinsey's November 2025 survey found over 80% of organizations reporting no meaningful enterprise-wide EBIT impact despite adoption, with its 2026 follow-up showing fewer than 20% of pilots ever crossing into enterprise-scale production. BCG's 2026 survey of 1,800 executives put the share of companies generating meaningful financial value at 26%. S&P Global found 42% of companies abandoned most of their AI projects in 2025, more than double the prior year. Enterprises have been experimenting broadly and only a quarter are seeing value added: adoption which usually resolves on the order of multiple years. That's an issue. The liabilities created today amortize on schedules that don’t assume extensions into the days after tomorrow. See Alex Karp’s latest appearances. Real paying demand with genuine unit economics is compounding, and sovereign compute purchases are authentic. The problem is what the financing assumes. The debt and capex are sized against gross demand, which is every dollar of usage regardless of where it comes from, but only a percentage of that demand survives a venture funding downturn and keeps paying as unit prices fall. The gap between that percentage and the full demand picture is where the distressed opportunity lives: liabilities were built against the big number, but only the smaller number will be there to service them in a downturn. Supply overshoots even the honest demand If demand quality were the whole story, the infrastructure buildout would slow as weakness develops. This won’t happen for a few reasons: No individual player can underbuild. For each hyperscaler, building too much is a margin problem while building too little is a strategic catastrophe: losing the platform position of the next computing era. They each behave rationally, and the sum of their individual reasoning is collectively guaranteed to produce excess. This is the same logic that gave every 1990s carrier a plan to win a fifth of a market shared by ten carriers. It’s game theory. Builders must spend money years before they can know whether the demand will exist. Energy buildout takes time: interconnection queues in major US markets run years, and large transformers and switchgear have multi-year order books. Forecast error is a near-certainty in this structure, and the game theory biases every participant's error in the same upward direction. The forward numbers keep ratcheting up: JPMorgan's June 2026 midyear outlook lifted projected global AI capex through 2030 to $5.5 trillion and sees hyperscaler spending passing $1.1 trillion in 2027; BloombergNEF counts more than 23 GW of data center capacity under construction and expectations for 2027 developer spending rose 56% in just six months. Commitments of this size made this far ahead of demand visibility are the raw material of every historical overbuild. The capital structure gun has switched off its safety. Typical overshoot is checked by balance sheet pain. Today, the strongest spenders have been funded from operating cash flow so vast that discipline was discretionary, but that buffer is rapidly eroding: Amazon's FCF is projected to turn negative. Oracle's nine-month FCF has run to -$24.7B. Hyperscalers raised roughly $108B of debt in 2025 and around $100B more in early 2026 while investors demanded record CDS protection. JPMorgan now projects AI-related debt financing reaching $4.1 trillion as loan-to-cost ratios rise. Meanwhile projects have started migrating off balance sheet: SPV structures like the Meta/Blue Owl Hyperion deal, private credit funds hunting long-duration assets, GPU-collateralized lending to neoclouds, and vendor financing from the chip supplier itself. CoreWeave, carrying roughly $18.6 billion of debt against about $1.9 billion of trailing revenue and a near $1.5 billion net loss, trades with five-year CDS around 675 basis points, which is a market-implied default probability of roughly 42% over five years. Regulators have noticed this leveraging too: FSOC has flagged the sector's debt accumulation as a potential financial stability concern. Contract backlogs are fake demand. Take-or-pay agreements make supply look pre-sold, and lenders underwrite backlog as revenue. Oracle's backlog hit $638B in May, up almost 4x in a year, and BofA estimates more than half of it is OpenAI alone. FCF ran negative $23.7B, debt is climbing, and Oracle plans to raise $40B more in fiscal 2027. Granted, $75B of that backlog is prepaid or customer-supplied hardware, which softens the funding burden, but doesn’t change the concentration. The credit market has been pricing this risk in real time: Oracle's five-year CDS ran from 43 basis points to over 110 in two months of late 2025, then widened to a record near 198 basis points by April, wider than during the global financial crisis. Financing announcements buy time in quarterly intervals, but the fundamentals reassert positioning. This is how a demand pause becomes a credit shock. Supply arrives irregularly while demand grows in smoother curves. A buildout with 2-3 year construction cycles can’t track the demand curve smoothly as it tends to alternate between shortage and glut. Shortage is likely where we are now: GPU capacity coming online through late 2026 is fully booked, H100 one-year contract rates rose roughly 40% between October 2025 and March 2026, NVIDIA has pushed through rental price increases on a three-year-old chip, and data center vacancy in primary markets sits near record lows. Today's pricing is an artifact of scarcity, not evidence of steady economics, and it is exactly what pulls forward the next wave of commitments. The gigawatts announced over the past two years land over the next two, against announced capacity that exceeds both power supply and recurring demand. What does and doesn’t last A distressed buyer's discipline is knowing what not to buy, because the shakeout will be full of assets that look cheap but are actively losing value. GPUs are the largest depreciating line item. H100 rental rates collapsed roughly 60-75%, from a 2023 peak above $8 per hour to the $2 to 4 range by late 2025, as supply caught up. They then rose into the 2026 shortage. Secondary-market H100s currently trade around $25,000-$35,000, down from over $40,000 in early 2025. Real value remains, but the direction is still one way only. Goldman treats the 4-6 year useful life assumption as a variable that moves cumulative capex by hundreds of billions. The aftermath buyer underwrites silicon to salvage value and never pays for a chip what its lender did, because by the time a levered holder is forced to sell, the collateral is 1 or 2 generations old. Scarcity and durability lives on the power side. Grid interconnection rights have years of queues that no capital can avoid. Secured power contracts and below-market PPAs are essentially annuities, and their value rises in exactly the scenario where their holder fails for other reasons. Substations, transformers, and switches carry lead times so long they are worth more installed than invoiced. A powered shell awaiting its GPUs replaces at roughly $10-15M per megawatt, the same order as Goldman's $15M baseline. Its next buyer skips the utility timeline entirely and deploys whatever compute generation is current when demand returns. Fiber routes, water rights, land adjacent to transmission, and contracted customer relationships are also on the list of what’s durable. This is the lesson from the fiber-era: the fortunes went to buyers of the routes and conduit, not to buyers of the routers that ran on them. Owning the arteries around the replaceable beating hearts. As the strongest hands, hyperscalers and sovereigns will compete as buyers. The sellers will be the smaller levered players mentioned above and their lenders, not their management teams, will control how the spoils move. Three playbooks ideas Prior cycles give us ideas for plays to execute: Buying the debt, not the assets. The consolidators of the railroad era took control through reorganization rather than purchase, and the modern equivalent is buying a distressed developer's loans at a discount and converting to ownership through restructuring, taking control of the facility, the power supplies, and the contracts. Debt is shown to known counterparties before it is auctioned, which means it’s important to map which funds, banks, and financing arms hold the debt, and become the counterparty they already know when re-marks force a sale. Buying routes, not routers. Telecom-era bankruptcies sold at 1 to low-2 digit percentages of invested capital, and the buyers who compounded bought conduit and routes and rented them to the streaming era, while equipment bought at the same auctions went to zero anyway. After the AI buildout, when forced sales begin, bid only on interconnection positions, energized shells, power contracts, and tenant relationships, with any included GPUs marked to salvage. The near-term action is to identify which specific neoclouds, converts, and single-tenant developers sit on genuinely scarce power in constrained markets, so that the bidding decision is already made and specifically targeted. Acquiring capacity for the next era's product. Google's early-2000s dark fiber purchases were a product thesis, not a distressed trade: it knew what it would run on cheap capacity. The modern equivalent is that structurally cheaper inference enables application classes that cannot exist at current compute prices, and that owning low-cost capacity into the deployment phase is a growth position wearing a value hat. The action would be to fund deployments where unit economics work at today's prices, cheap contracted power and conservative underwriting, both because low-cost operators survive shakeouts and because operating capacity provides live pricing, utilization, and demand data that makes the eventual distressed underwriting happen faster. Where this can go wrong If fundamental demand compounds fast enough to absorb new capacity, GPU rates will stay consistently stay firm, spreads on the levered names tighten through 2027, sublease inventory staying scarce, and model-developer revenue closing toward commitments. The bull case is still pretty strong: Microsoft's $80 billion Azure backlog says demand currently exceeds deliverable supply, JPMorgan's midyear view is that the economics remain attractive and increasingly profitable for now, with hyperscaler operating cash flow projected above $900 billion by 2027, and arguing Jevons paradox, that collapsing inference costs expand usage faster than they compress unit revenue, is still genuinely supported in 2026. The timing is insecure. Today’s mania has deep-pocketed sponsors, genuine underlying adoption, and a financing machine still accelerating. But the timing is why identifying the first steps are valuable, since it takes time for these playbooks to build before the crises appear. The AI infrastructure overshoot is not just a tragedy to predict but an asset transfer to prepare for. The demand is real and deflationary, and the supply is structural and pro-cyclical. The gap between them will be closed by someone well-prepared. Disclaimer: This article is not financial advice. The views expressed are solely my own and do not reflect those of any other person, firm, or entity.