The Economics of the Build-Out

Chapter 1.8 scored a single asset: does this facility, financed this way, clear its cost of capital over its economic life. This chapter asks the larger and more contested question: is the industry as a whole over- or under-building, and does the aggregate capital stack survive the depreciation clock it is racing? A single asset can clear its hurdle in a sector that is collectively over-built, and a single asset can fail in a booming one. The two altitudes are independent, and the 2026 discourse routinely conflates them.

At sector altitude there are four questions. The first is the capex wave and the build cost per MW that sets its denominator. The second is the macro depreciation problem: the GPU-ROI-decay argument from Chapter 1.8, aggregated across the fleet, where it becomes a question about the quality of reported industry earnings. The third is the debt wave and the new financing structures that fund the gap, including power oversubscription as a yield strategy. The fourth is the revenue-vs-capex gap and what protects returns when the cycle turns. Almost every number here is contested, and the contested ones decide whether this is a supercycle, a digestion or a bubble. We flag them as we go and bind them to the dated forecast register in Appendix D.

The magnitudes are without precedent in industrial history. Top-4 US hyperscaler 2026 capital-investment guidance totals ~$735–760B on Microsoft's unchanged ~$190B economic-investment basis (the four companies’ own guidance through 30 Jul 2026: Amazon ~$220B, Microsoft ~$190B, Meta $130–145B, Alphabet $195–205B; CreditSights' Oracle-inclusive top-5 estimate is ~$750B). That is up more than 60% year-on-year, against Dell’Oro’s more-than-$1T full-year 2026 worldwide data-center infrastructure capex forecast issued on 10 June 2026. The next year points higher again. J.P. Morgan Asset Management projects that the top four and Oracle, the five US hyperscalers in its count, will together spend more than $1T of capex in 2027 (20 August 2026), which is a strategist's estimate, not company guidance. The long-horizon framings are larger still: ~$6.7T of global data-center capex by 2030, of which ~$5.2T is AI-capable (the midline of McKinsey's scenario band, from $3.7T constrained to $7.9T accelerated; McKinsey, 2025). These are forecasts, not committed builds, and the error bars are wide and scenario-dependent. Even so, the central case is a sustained, multi-year deployment, and a single year of it is large by any historical yardstick: the top four's 2026 guidance alone is more than twice the inflation-adjusted cost of the Apollo program (about $309B in 2025 dollars; The Planetary Society).

The denominator that turns those headline trillions into a financeable plan is the cost to build and energize a watt, and the conventional benchmark understates it. JLL’s 5 January 2026 forecast is $11.3M/MW for shell and core of a single-tenant, 50 MW air-cooled facility, excluding land and active IT equipment. That series is up 47% from $7.7M/MW in 2020. AI-ready facilities price higher: Goldman Sachs puts next-generation AI data centers at $15–20M/MW (May 2026), and Cushman & Wakefield averages $17.6M/MW for the newest US and Canada greenfield builds, excluding chips (September 2026). JLL separately places tenant technology fit-out at up to $25M/MW for AI infrastructure. That scope includes the IT equipment, so it is not a facility-only figure. Keep the scopes separate: adding the tenant fit-out to a silicon-inclusive line double-counts technology and corrupts the financing model.

The cost stack also inverts the legacy intuition. In a conventional data center the building and power plant dominate. In an AI factory the silicon dominates everything. Epoch AI (May 2026) puts servers at ~56% of capex (~$21.2/W, with the accelerators inside that line), the facility (MEP, shell and construction labor) at ~30% (~$11.4/W), network and cluster infrastructure at ~13% (~$4.9/W), and land plus utility interconnection at the remaining ~1% (~$0.3/W). The all-in up-front intensity, silicon included, is near $37.9/W. Check the scopes before subtracting. Epoch's ~$11.4/W facility line is close to JLL's $11.3M/MW shell-and-core average, while JLL's separate up-to-$25M/MW tenant technology fit-out overlaps Epoch's server and network lines. Epoch's ~$37.9/W all-in works out to about $38B per GW of IT before financing, near the top of the ~$30–40B per-gigawatt campus range in Chapter 3.1. Chapter 6.3 has the facility-only benchmarks. GPUs alone are roughly a third of all data-center capex (Dell'Oro, 2026). The consequence at sector scale is the same as at firm scale, only larger: the asset's economic life is the GPU's economic life, not the concrete's. The industry’s reported profitability is sensitive to how fast that silicon decays, while the surviving building, power rights and network keep their own service lives.

Chapter 1.8 is the canonical home for the firm-level depreciation debate. Aggregated across the fleet, the same argument stops being an accounting choice and becomes a question about the quality of reported industry earnings. Every year an operator extends the assumed useful life of its AI fleet, it moves cost off the current income statement, so reported operating margin rises even though nothing about the physical asset improved. The largest operators have moved in opposite directions in the same window. Meta moved certain servers and network assets to 5.5 years (about $2.9B less 2025 depreciation, by its estimate). Amazon shortened a subset of servers and networking equipment from 6 to 5 years (2025 depreciation up $1.4B, net income down $1.0B, per its 2025 Form 10-K). No outside party should pretend the number is settled.

The bull case is the training-to-inference cascade. A GPU retired from frontier pre-training is not scrap: it cascades down to post-training, then inference serving, then batch and internal workloads, earning revenue at each step. If the cascade holds, economic life stretches toward book life and the accounting is honest. The bear case is that the cascade is finite, that each new generation is so much more efficient per token that the old part becomes uneconomic to run against grid power, and that the residual market is thin. The historical observations address different cash flows: resale value and rental earning power have not moved together, and rental declines have differed by market tier (Silicon Data, Dec 2025). A three-year net disposal value cannot be read off resale prices or rental earnings alone. It takes dated same-SKU transactions, the part's condition and the cost of removal, as Chapter 1.8 sets out. A high residual underwrites the cascade defense, and a low one validates the short-life bears. Both can be partly true at once.

At sector scale the stakes are systemic rather than per-firm. The sharpest bear version holds that the industry is systematically under-depreciating its fleet, booking long lives to flatter earnings while the assets decay on the short schedule. Michael Burry’s contested November 2025 estimate is ~$176B of cumulative understated hyperscaler depreciation during 2026–2028 (reported by RBC GAM, April–May 2026). You do not have to pick a side to see that depreciation policy is now the central earnings-quality battleground for the sector. The whole long-life defense also rests on a deep, liquid secondary GPU market that can absorb cascaded hardware at a stable residual. If that market is thin, the cascade is a story rather than a cash flow.

A positive income statement cannot pay this cash gap. Before financing, $200B − $240B − $10B − $20B = −$70B. Only $40B − $10B = $30B of opening liquidity is spendable, which leaves $40B of funding or spending cuts. With no external financing, the cash-capex release ceiling is $200B + $30B − $10B − $20B = $200B, so defer $40B of the proposed $240B. The decision flips at CFO of $240B, or with $40B of committed financing that leaves the closing-cash floor intact. In the combined stress, $170B − $260B − $10B − $20B = −$120B. After $30B of liquidity, the gap is $90B and the capex ceiling falls to $170B. New noncash lease additions are disclosed separately and create future payments; they do not fill today’s cash gap. This is a sector funding and release screen, using the cash-versus-investment boundary explained by Epoch AI (16 June 2026). It does not derive an individual project’s NPV, debt coverage or acceptable GPU price; Chapter 1.8 and Chapter 2.5 own those tests.

The defining structural feature of the 2026 build-out is that it has outgrown self-funding. Morgan Stanley (2025) estimates a multi-year build near $2.9T (2025-2028) with a ~$1.5T financing gap beyond hyperscaler operating cash flow, and the sector has reached for external capital; the form that capital took is the macro story. Epoch AI fitted trends to the filings of the five largest cloud companies: cash capex at ~70%/yr against operating cash flow at ~23%/yr over Q2 2023–Q1 2026. Extrapolated, the aggregate lines cross around Q3 2026 (Epoch AI, 16 June 2026; cash capex excludes finance-lease additions). Epoch's company-level trends have Oracle already past the crossover and Alphabet, Meta and Microsoft following in 2027–2028. Later company cash-flow results must be kept separate from that extrapolation. Alphabet reported its first negative quarterly free cash flow since its IPO in Q2 2026 (−$5.9B, on $44.9B of capex), and Amazon's trailing-twelve-month free cash flow swung to −$7.6B (from +$18.2B a year earlier). Oracle's quarter ended 31 August 2026 (fiscal Q1 2027) shows the same strain. It spent $28.5B on capex while generating $23.1B of operating cash, and $11.4B of that was customer prepayments with a significant financing component. With the prepayments, free cash flow was −$5.4B; take them out and it falls to about −$17B. The company also sold $20B of common stock during the quarter (company filing, 2026-09-10). Customer advances are outside funding, so do not count them as cash the business earned. When the richest balance sheets in the world stop self-funding, the marginal megawatt is financed by the credit market, and that market is now exposed to AI-infrastructure risk in a way it was not in 2023.

For the hyperscalers themselves, outside money arrives mainly through two channels the securitization figures below do not show: investment-grade bonds and leases. Alphabet, Amazon, Meta, Microsoft and Oracle sold $219B of investment-grade bonds between January and 20 August 2026, $62B of it outside the dollar market. Their share of US-dollar investment-grade issuance rose from 2% in 2022–2024 to an expected 9% in 2026. Their spreads are about 30 bp wider than at the start of the year, while the broader dollar investment-grade index has widened only about 2 bp, so investors are starting to charge for the supply (J.P. Morgan Asset Management, 20 August 2026). The leases are the larger commitment: about $1.4T disclosed, most of it not yet on any balance sheet (Chapter 2.5 has the lease accounting and the purchase commitments that sit beside it). Size a hyperscaler's funding capacity on those commitments as well as on its reported debt, because the obligation exists years before the balance sheet shows it.

The structured instruments are GPU-collateralized debt, delayed-draw term loans (DDTLs), bankruptcy-remote SPVs and asset-backed securitization. Data-center securitization ran ~$26B in 2025 — ~$15B of ABS plus ~$11B of CMBS — with gross supply projected toward $30-40B/yr in 2026-2027 (CREFC, 2026). Morgan Stanley’s July 16, 2025 forecast for annual US data-center ABS issuance ran from ~$10B (2025) toward ~$37B in 2028, a baseline the ~$15B actually printed in 2025 already outran. CoreWeave's first ~$8.5B investment-grade-rated (A3) GPU-collateralized deal was a delayed-draw term loan (DDTL 4.0), not an ABS issuance. Morgan Stanley estimates ~$130B of US data-center securitized-credit net issuance across 2026-28, ~75% of it ABS. The strategic catch rhymes with the depreciation debate: this debt underwrites a depreciating, deflating asset against a thin secondary market. The same residual-value uncertainty that makes the accounting contested is now wired into the capital structure, and securitization distributes that risk into pensions and asset managers, a new systemic linkage between AI capex and broad credit markets. Facilities carry the same question at the building level, and a joint venture can take ownership off the balance sheet while the sponsor keeps the building's residual risk as a guarantee. Meta's El Paso venture with BlackRock leases long-lived buildings on a four-year initial term, backed by a residual-value guarantee from Meta. Its aggregate threshold is about $13B and decreases over time (Meta, 2026-07-28). Chapter 2.5 has the structure.

The visible test case makes the tension concrete. Chapter 1.8 carries CoreWeave's Q2 2026 figures in full. Their shape is $2.58B of quarterly revenue at a 59% adjusted-EBITDA margin sitting on ~$35B of balance-sheet debt, a −$626M net loss for the quarter, and a ~$104B backlog concentrated in a few anchor tenants (company filings, Aug 2026). The 'circular financing' critique is a real structural risk: vendor equity stakes and residual backstops let a buyer finance the purchase of the vendor's own chips, which couples the financing to the same demand and residual assumptions the equipment depends on. A residual shock then hits collateral, covenants and revenue at once.

The named cases are specific. NVIDIA agreed in September 2025 to buy CoreWeave's unsold capacity through April 2032, under an order with an initial value of $6.3B (CoreWeave 8-K, 2025-09-15). AMD gave OpenAI a warrant for up to 160 million AMD shares that vests as OpenAI deploys AMD GPUs toward 6 GW and as AMD's share price hits set targets (AMD, 2025-10-06). In both deals the vendor has its own money on the buyer's demand: NVIDIA pays when CoreWeave cannot sell the capacity, and AMD gives up more of its equity the further OpenAI's deployment goes. The deal mechanics, including NVIDIA's lease guaranties for OpenAI's Ohio campus, are engineered in Chapter 2.5. At this altitude the point is that the sector's true debt capacity is set by how much of the revenue is contracted and how durable the residual is, not by the headline collateral value.

This is the question that decides the outcome: does the revenue arrive before the depreciation does? The most-cited framing is David Cahn's. His '$600B question' (Sequoia, 2024) was reframed in July 2026 as the $1.5T question: the lifetime end-customer revenue a single year of capex must now earn back. That bar has risen, not fallen, as capex accelerated through 2026 (Cahn, 8 July 2026).

The pressure shows in the ratio of capex to demonstrated end-demand. Revenue at the pure-play AI labs is growing quickly, yet it remains small next to the capital going in. Anthropic itself announced run-rate revenue above $47B in May 2026; Bloomberg and Reuters later reported more than $65B at the end of July, citing a person familiar (2026-08-17). OpenAI told investors that annualized revenue was nearing $50B at the end of September (FT and The Information, 2026-10-08). An earlier ~$40B figure (Bloomberg, 2026-08-13) counted Azure resale revenue that OpenAI had stopped receiving (The Information), and a ~$70B OpenAI figure (Axios, 2026-09-29) was an investor's estimate on Anthropic's gross basis. Added together, though booked on different bases (below), the two labs' run-rates equal roughly 13–16% of 2026 capex guidance at the top four hyperscalers. A run-rate annualizes a recent period; it is not recognized revenue. The two labs also count cloud-resold sales differently: Anthropic books the gross amount of sales made through cloud partners such as Amazon, while OpenAI books only its own share (Bloomberg, 2026-10-09). Neither company has published audited figures. Hyperscaler cloud-AI revenue is also growing fast off a real base (company filings, 2026).

The supplier's books show where the capex lands: NVIDIA reported $89.0B of data-center revenue in the quarter to 26 July 2026, up 117% year on year, and guided total revenue to $108.0B for the next quarter with no China data-center compute assumed (NVIDIA, 2026-08-26). That is the capex wave seen from the seller's side rather than end demand, so it shows the spending had not turned as of that report; it says nothing about whether the revenue gap is closing. Bears set that capex against the labs' revenue and call the gap a bubble. Bulls take the labs' revenue, small but growing fast, as an early sign of Jevons-paradox demand, and both sides are reading the same numbers.

The counterweight to the gap is real, and it is the bull case's strongest card: the cost to serve a model of fixed quality has fallen ~10x per year (a16z's 'LLMflation'; Epoch's median measurement nearer 50x/yr). Through the Jevons effect, that deflation can multiply token volume. Stanford's AI Index 2025 measured a fall of more than 280-fold in the cost of GPT-3.5-level queries between November 2022 and October 2024. Total spend rises only if usage grows faster than the price falls; Chapter 16.3 sets out that condition and the demand evidence. Reasoning and agentic workloads are a structural demand multiplier on top, and they push serving toward decode-heavy work. If demand keeps absorbing the supply, the gap closes from the revenue side and the build-out is vindicated. The gap is not evidence of a bubble by itself. It is the open question whose resolution defines which of the three scenarios the sector lands in. Underwrite it with an explicit demand-growth curve rather than a flat extrapolation, and do not assume today's token price holds.

At the macro altitude, what protects returns is a small set of positions that pay off across all three correlated tails at once, not a clever capital structure. Four matter, and they are the inverse of the levers that amplify the downside.

• Contracted revenue over merchant exposure. The sector's true debt capacity is set by the take-or-pay-backed share of revenue, not the headline collateral value. IREN's August 2026 results (2026-08-27) show the customer-credit effect: GPU financing for its Microsoft contract cost a 6.0% weighted average before fees, against 9.0% fixed for a customer that is not investment grade. The deals are structured differently, so treat the gap as a customer-credit difference, not a measured premium. Chapter 2.5 has the mechanics. A contracted, diversified offtaker base survives a demand miss, while a merchant fleet in a soft GPU-rental market is the canonical bubble casualty. The hedge is real termination economics and tenant diversification. The failure mode is a backlog concentrated in a few anchor tenants, and a backlog is a revenue multiple only once you divide it by the annual or run-rate period it covers.
• A low-cost, firm power position. Energy is the largest controllable lifetime opex line, and on a power-bound build a cheap, firm, long-dated supply is worth more to ROI than most capex optimization. It is also the asset that lets you run power oversubscription as a yield strategy rather than a liability. → siting and speed-to-power economics in Chapter 3.2.
• Conservative leverage and depreciation. The operator that books the short life up front and limits leverage has already priced the bear case into its balance sheet. The one running maximum leverage against an optimistic residual is the one a residual shock detonates. Circular financing is the sharpest version of the amplifying lever; avoid it as a structural matter, not a reputational one.
• Design-for-flexibility. This is the cheapest hedge that pays off across every correlated tail: a hybrid procurement posture (colo anchor plus neocloud overflow) so a demand miss sheds opex instead of stranding capex, plus reserved floor loading, water and electrical headroom so the asset can absorb a density step instead of being stranded a generation behind. You cannot manufacture a take-or-pay contract or a deep residual market after the cycle turns; you can buy flexibility cheaply before it does. → the irreversibility framing in Chapter 1.8.