Remember how calm everything felt in the housing market during the summer of 2006? Prices kept climbing, delinquencies sat near record lows, and spreads stayed tight. Most people on trading floors would have laughed if you suggested the whole system was about to seize up. Yet the instruments themselves had already written the ending. Millions of borrowers sat inside low introductory rates on adjustable mortgages, paying comfortably until the scheduled step-up arrived. Nothing looked wrong because the system lived inside its teaser period.
I keep coming back to that moment when I look at today’s artificial intelligence build-out. The parallels feel almost too clean. Trillions of dollars in compute capacity commitments have been signed, celebrated, and capitalized into equity valuations. Yet the actual bills have not started arriving in full. The payments begin only when the data centers finish construction, get powered, and the capacity is formally accepted. That gap between signing and commencement creates a multi-year window that looks remarkably like the old teaser phase. And just as before, a clear wall of scheduled obligations sits on the calendar for 2027 and 2028.
The Structural Parallel No One Wants to Price
The core instrument this time is the take-or-pay compute contract. A frontier lab or large buyer signs a multi-year commitment for capacity that does not yet exist. Construction, power procurement, and campus build-out typically stretch twenty-four to thirty-six months. During that interval the seller books the full amount into remaining performance obligations or contracted backlog. The buyer books almost nothing as expense because no service has been received. Equity markets treat the booked figure as if it were already earned revenue. Everyone wins on paper until the day the capacity goes live and the invoice clock starts running regardless of actual utilization.
That is the reset. The obligation steps up on a construction schedule, not on a demand schedule. The parallel to the old adjustable-rate mortgage is precise. The low introductory phase gets the deal done. The higher fully indexed phase is deferred far enough that markets largely ignore it. In residential credit the teaser lasted two or three years because that was the product design. In compute the interval is dictated by the physical reality of building gigawatt-scale facilities. Signings concentrated in 2025 and 2026 therefore lock in a commencement surge two to three years later.
More than two trillion dollars of these contracts now sit across the major American cloud providers as remaining performance obligations and backlog. A large share traces to a handful of private frontier labs that currently burn cash at extraordinary rates. The market capitalizes the booked numbers while the billed numbers remain a fraction of the total. During this window the reported economics and the committed economics diverge by the full value of everything signed but not yet delivered.
How a Single Contract Behaves Across Its Life
Walk through a simplified example. A lab signs a twelve-billion-dollar, ten-year capacity deal. Delivery requires two years of construction. On day one the seller adds the entire amount to backlog. The market treats it as contractual future revenue. Cash flow at the seller looks terrible because capital expenditure runs far ahead of receipts, often with financing structured off balance sheet. Reported earnings feel little of the build-out pain.
The buyer announces secured access to the compute needed for its scaling roadmap. Valuation rises on the news. The commitment remains a future obligation, disclosed lightly if at all. No expense hits the profit-and-loss statement yet. A lab that has signed tens of billions across multiple providers still shows a cost structure that only reflects capacity already commenced. The market sees explosive backlog growth on one side and secured supply on the other, and prices both as pure growth stories.
Then commencement arrives. Capacity comes online in tranches. The buyer’s cash obligation steps from near zero to the full contractual rate, often as a series of sharp increases rather than a smooth ramp. The seller starts recognizing revenue and the backlog begins to drain. The asymmetry is the entire point: the payment schedule is fixed by construction progress and remains indifferent to whether the lab’s revenue has grown enough to cover it. I’ve found that this feature is the one most consistently under-appreciated by equity investors who focus mainly on technology roadmaps.
Why These Contracts Function as Debt
The common pushback is that this cycle lacks the leverage of 2008. Look only at reported gross debt and the numbers appear manageable, roughly in the mid-hundreds of billions across the broader complex. Look instead at the present value of disclosed non-cancellable compute and capacity commitments and the picture changes. The economic claim on future cash rises into the low trillions. Rating agencies have treated take-or-pay style agreements as imputed debt for decades in pipelines, shipping, and power purchase arrangements. The same economic substance applies here even if the accounting treatment differs.
Three mechanisms keep most of these obligations off traditional balance-sheet debt lines. First, remaining performance obligations are primarily a seller-side disclosure. Buyers face no symmetric public requirement to list every commitment. Second, the largest buyers remain private and file no periodic reports, so their obligations surface mainly when a counterparty announces a deal. Third, the contracts are usually structured as service agreements rather than formal leases, the same maneuver that once kept operating leases off balance sheets.
The result is a system in which reported leverage looks light while the contractual claims on future cash are large. The single most important credit fact is straightforward: a substantial portion of the largest technology capital cycle in history is underwritten by a small number of private companies that do not yet generate positive free cash flow. Their ability to meet the payments rests on continued access to capital markets and on revenue growth arriving on a specific schedule.
The Concentrated Signing Wave
One major lab alone executed an extraordinary volume of commitments in a compressed window. In less than a year the notional value of deals announced reached levels that, on certain days, moved the combined market capitalization of several large technology suppliers by hundreds of billions. Almost every major commitment was timed so that the bulk of the billing would begin in 2027 and 2028. Signing was cheap relative to the valuation uplift it produced. The market treated each announcement as evidence of insatiable demand rather than as an obligation event.
That framing matters. When capacity is still under construction the announcement signals strength. Once the same capacity starts billing, the same dollars become a fixed cost that must be met whether or not utilization matches the original underwriting assumptions. The early vintages of these contracts have largely cleared because demand still outran available supply. Success in the early rounds encouraged larger later rounds, exactly as strong early housing performance encouraged more aggressive mortgage originations.
Building the Commencement Schedule
The reset wall can be reconstructed from public announcements and typical construction timelines even though the full contracts remain private. Two views are useful. The first tracks annual cash obligations as successive tranches go live. The second tracks the notional amount of contracts that transition from deferred to full pay in each calendar year. Both show a clear peak across 2027 and 2028.
Synchronized signing produces synchronized commencement. Mortgage originations peaked in 2005 and 2006; the teaser period lasted roughly two years; resets clustered in 2007 and 2008. Compute signings concentrated in 2025 and 2026; construction intervals run two to three years; the wall therefore lands in 2027 and 2028. The arithmetic is the same even if the collateral differs.
Compare the projected annual compute cost against plausible revenue paths for the largest private counterparties. Even under optimistic management plans the coverage ratio at the peak leaves little room for other operating expenses. In some scenarios committed compute alone exceeds projected revenue in the critical year. The gap does not require demand to collapse. It only requires growth to decelerate below the rate assumed when the contracts were signed. A fixed obligation meeting a lagging revenue stream creates a funding need that must be met through new capital raises or renegotiation.
What the Numbers Imply for Equity Valuation
Take the largest private lab as an illustration. Publicly discussed non-cancellable commitments run into the hundreds of billions. Carry even a discounted present value of those obligations as a senior claim and the residual equity value begins to look thinner than the headline private valuation suggests. The market still prices the equity as if the obligations were ordinary service agreements rather than fixed claims on future cash. That gap can close in either direction: either revenue grows into the obligations and the equity remains valuable, or funding conditions tighten and the senior claims assert themselves.
Perhaps the most interesting aspect is how little of this discussion appears in mainstream valuation models. Analysts focus on model capability roadmaps and total addressable market forecasts. The contractual payment schedule receives far less attention even though it is the more certain of the two curves. Revenue is a projection. The commencement dates are written into the contracts.
The Hyperscaler Side of the Same Coin
Frontier labs face a contractual teaser. Hyperscalers face an accounting one. Capital under construction sits in construction-in-progress accounts. Depreciation does not begin until the asset is placed in service. Interest incurred during construction is often capitalized into the asset basis. The moment a tranche goes live, depreciation starts on the full capitalized cost. The income statement absorbs a fixed charge that was invisible while the asset was still under construction.
A live data center carries substantial fixed monthly costs whether utilization runs at ninety percent or thirty percent. Depreciation, power, interest, staff, cooling, and maintenance form the bulk of the stack. Operating leverage works beautifully when utilization stays high. The same fixed base produces sharp margin compression if utilization falls short of the underwriting case. If a major counterparty struggles to meet its take-or-pay obligations, the hyperscaler does not simply miss revenue. It absorbs a fixed-cost shortfall that its own operating leverage magnifies.
In my experience watching previous capacity cycles, this is the point where narrative and numbers start to diverge most clearly. Markets celebrate the placement of new capacity into service as progress. The income statement begins to feel the fully indexed cost at the same moment.
Available Options When Bills Arrive
Once billing begins, unused capacity changes character overnight. Strategic optionality becomes cash burn. A finance team facing that line item has limited levers. Cancellation is generally not available under take-or-pay terms. Growing into the capacity depends on demand that sits largely outside management control. That leaves capital raises, renegotiation, or attempts to sublease excess capacity.
Renegotiation is the path of least resistance for systemically important counterparties. Amendments can be framed as partnership adjustments, capacity rephasing, or efficiency-linked pricing. Yet the moment one major contract is quietly amended, every remaining performance obligation in the complex carries a demonstrated probability of revision. “Contracted” ceases to mean the same thing across the board. The entire backlog can be re-rated without a single formal default.
Subleasing offers another route. A tenant that cannot use capacity will often accept almost any rate above zero because each recovered dollar reduces cash outflow. Short-term sublease markets currently look firm because overall supply still lags demand. That balance shifts once the multi-gigawatt deliveries of 2027 and 2028 land. Bridge demand disappears and shadow vacancy can appear just as the larger wave of commitments converts from booked to billed.
Index Concentration and Passive Exposure
The equity of the entire complex is heavily concentrated in a small group of names that dominate market-capitalization-weighted indices. Artificial intelligence related companies now account for a large share of major equity benchmark market value. Semiconductor suppliers and hyperscalers together drive a disproportionate portion of recent earnings growth. That growth is closely tied to new order books. When labs shift from signing fresh contracts to servicing existing ones, the second-derivative demand for new chips can slow first.
A passive retirement account that holds a broad equity index therefore owns, whether intentionally or not, a concentrated exposure to the clean conversion of contracted backlog into billed revenue. It also owns exposure to the ability of a small number of cash-burning counterparties to fund those payments. Almost no individual investor chose that specific risk profile. The index construction did it for them.
Living Inside a Teaser Period
The hardest part of 2006 to explain to anyone who did not live through it is how good the contemporary data looked. Origination volumes were strong, homeownership rates high, delinquencies low, and builder earnings robust. Every statistic confirmed strength. What few priced was that the data were generated by a system whose real payment test had not yet begun. The instruments themselves guaranteed that credit metrics would look healthy until the calendar forced the reset.
Read today’s compute cycle through the same lens. Record backlog growth, capacity that appears fully sold, and vendor revenue that continues to beat estimates. All of these observations are consistent with a construction-phase environment in which delivery still lags contracted demand and billed payments remain well below booked commitments. The strength of current data is real. It simply does not yet answer the question that will matter when the commencement wall arrives.
A teaser period manufactures the very evidence used to dismiss concerns about the future schedule. The bear who points to the wall is answered with today’s utilization rates and revenue growth. Both can be true at the same time. The early vintages worked. That success itself encouraged the larger later vintages that create the concentrated wall.
Why the Specifics Differ but the Structure Does Not
This cycle is not identical to the mortgage episode. Graphics processors are not houses. Take-or-pay compute contracts are not mortgage-backed securities. Artificial intelligence may prove one of the most consequential technologies of the century. All of that is true and none of it changes the financing structure. A scarcity narrative justifies large fixed obligations. A teaser interval makes those obligations feel costless at the moment of signing. Fixed commencement dates convert the obligations into cash due. The bet is that income will have grown enough by then to meet the bill.
The bull case remains open. Demand could scale into the committed supply on schedule. Capital markets could remain open and willing to fund any interim gaps at non-punitive terms. The technology could deliver productivity gains large enough to justify the entire build. The narrower point is that the financing architecture can experience stress before the ultimate demand picture is fully known, simply because the obligations are fixed and front-loaded in commencement while the revenue is variable and back-loaded in adoption.
Construction delays can shift the peak from one year to the next without reducing the total obligation. Deferral changes the timing, not the amount. That is exactly what happened when mortgage servicers pushed resets later in 2007. The problem was postponed, not solved.
The next eighteen months will feature endless debate about whether artificial intelligence is overvalued. That question sits at the wrong layer. The technology itself can remain transformative while the specific financing instruments underwritten during the teaser period meet their scheduled reset. The wall is already visible in the contract announcements and construction timelines. What remains is whether the market chooses to look at it before the bills arrive, or only afterward.
In the end the structure is simple. Booked compute becomes billed compute on a date fixed years earlier. The payment is owed whether or not utilization matches the original forecast. The counterparties that signed the largest shares of those contracts currently fund themselves through continued capital raises rather than operating cash flow. That combination does not guarantee failure. It does guarantee that the true cash cost of the build-out will become visible for the first time in a concentrated window. How the system navigates that window will determine a great deal about the next phase of this cycle.
