Business Models, Economics & ROI

Every chapter before this one spends money; this is the chapter that decides whether the money comes back. An AI data center is not, financially, a building — it is a depreciating capital asset with a very short fuse, dominated by silicon that loses value faster than almost any industrial equipment ever financed at this scale. The engineering decisions in Parts 5 through 12 set the cost stack; the market decisions in Part 16 set the demand; this chapter is where the two meet in a single objective function: does the asset earn its cost of capital before the workload, the hardware, or the price curve makes it obsolete?

We build the cost stack and the TCO denominator; we confront the depreciation debate that quietly determines whether the whole industry is profitable; we lay out the $/GPU-hr pricing ladder and the breakeven that governs it; we trace inference unit economics down to $/M-tokens and the gross-margin waterfall that the application layer lives or dies on; we score build-vs-own-vs-lease as an NPV with an explicit option value; and we close on the operating archetypes and the downside stress tests. The through-line: the return depends on dated prices, qualified demand and an explicitly chosen earning life, so each commitment needs its own cash downside. Accounting depreciation, capital recovery, operating cash and debt service remain separate lines.

Start with the denominator. The canonical bottom-up reference is a 1 GW AI data center: roughly $38B total-program capex and ~$8.5B/yr all-in TCO once costs are annualized over their respective asset lives (Epoch AI, 2026). The $38B is up-front capex: servers ~$21.2/W — the accelerators live inside that line — network and cluster infrastructure ~$4.9/W, the facility ~$11.4/W, and land plus utility works under ~$0.3/W (financing is not in it). It works out to about $8.5M per MW per year — the number to carry in your head when someone quotes you a lease or a colo rate, because it is the all-in cost you are implicitly benchmarking against.

The cost stack inverts the intuition of a traditional data center. In a legacy facility the building and power plant dominate; in an AI factory the silicon dominates everything. The split is roughly: servers ~56% (about $21.2/W, accelerators included), facility — MEP, shell, and construction labor — ~30% (about $11.4/W), and network and cluster infrastructure ~13% (about $4.9/W), with land and utility works under 1% (Epoch AI, May 2026; an earlier vintage's ~$27.9/W ‘core stack’ decomposition no longer appears in Epoch's current model). The consequence of a server-dominated stack: the asset's economic life is the GPU's economic life, not the concrete's. You can amortize a shell over thirty years; you cannot amortize a frontier accelerator over thirty years, and pretending otherwise is the original sin of AI-infrastructure accounting.

The costing denominator matters as much as the numerator. Quote a facility in $/MW-year and you are comparing real estate; quote it in $/GPU-hour and you are comparing compute supply; quote it in $/M-tokens and you are comparing the product the customer actually buys. The three denominators are linked by utilization and by tokens-per-GPU-second, and a number that looks competitive in one can be uncompetitive in another. Name the denominator before you compare. → metric definitions in Chapter 0.3.

This is the canonical home for the depreciation debate, because it is where the contested figures do the most damage if mis-set. Begin with the unit. SemiAnalysis’s October 2024 eight-GPU H100 compute-node example falls from $270k to $256.4k after host/DPU optimization; rack, cluster fabric, storage and software sit outside that chassis subtotal (Chapter 7.8). Use a separate illustrative $300k installed-server allocation for the book-schedule comparison below. Depreciation on a $300k server is ~$50k/yr over six years, ~$60k/yr over five, ~$75k/yr over four — and that one line is the largest single component of a self-operated cluster's cost. The depreciation schedule is the cost structure.

The bull case — the reason a 5–6 year book life is defensible — is the training-to-inference cascade. A GPU retired from frontier pre-training is not scrap; it cascades down to post-training, then to inference serving, then to batch and internal workloads, earning revenue at each step. If the cascade holds, the economic life stretches toward the book life and the accounting is honest. The bear case is that the cascade is finite (there is only so much inference demand for a two-generations-old part), that each new generation is so much more efficient per token that the old part is uneconomic to run against grid power, and that the residual market is thin. Both can be partly true at once.

The residual evidence is mixed, which is why it is CONTESTED. H100s retained ~60–83% of value at 18 months, but secondary rental rates fell 64–75% from their $8–10/hr peak, while rental earning power and net liquidation proceeds remain different quantities. A three-year disposal assumption needs dated same-SKU transactions, condition and removal cost; stress zero net proceeds against DSCR and project return. The hyperscalers themselves disagree in public: Meta extended server life from 4.0 to 5.5 years (+$2.9B income); Amazon went the other way, 6 to 5 years (−$700M), in the same window (company filings, 2025). When the largest operators move depreciation in opposite directions, no outside party should pretend the number is settled.

Cost is half the equation; the other half is what you can charge and how full you keep the asset. The historical $/GPU-hour ladder spans unlike H100 allocations: Spheron’s interruptible SXM5 listing at $1.03/hr, Lambda’s SXM on-demand configurations at $2.49–3.44/hr, and Spheron-listed AWS p5 on-demand ~$6.88/hr and Azure ND H100 v5 ~$12.29/hr (Spheron, May 14, 2026 H100 table). Spheron specifies 8-way HGX for its spot listing and 8×–1× configurations for Lambda; its AWS/Azure rows omit regions. These advertised per-GPU allocations carry different host, fabric, support and interruption terms. The lower neocloud price must still fund any managed services the buyer supplies. The ladder is not static: the 1-year contract index rose ~+40% from October 2025 to March 2026 as supply tightened — but the tightening did not keep compounding: by mid-August 2026 Silicon Data's H100 index printed ~$2.5–2.7/hr on the neo-cloud segment (hyperscaler segment ~$7.3/hr), inside the ladder above. CME has announced cash-settled H100 and B200 rental-index futures for launch on 2026-10-05; each contract tracks a Silicon Data index of hourly rental costs and represents one month's rent for the respective GPU. One scope note on the whole ladder: it prices the installed Hopper fleet, the mature, installed, and well-quoted end of the market rather than the NVL72-class capacity a new frontier build is actually buying. Benchmark against the H100 index; do not price a new build off it. GPU pricing is a commodity market with real cycles, not a SaaS price list.

Against that revenue ladder sits the cost the operator actually carries. Owner cost per GPU-hour must be built from accelerator and fabric capex, useful life, residual value, financing, energy/PUE, spares, and achieved utilization. The spread is comparable only after host resources, fabric, location, support, egress, term, spare capacity and billed units match; an idle reserved GPU still has a cash cost. Utilization is the silent variable that dominates the whole pro-forma.

Revenue management must fill paid capacity with accepted, billable work and tier that work by its realized contribution. Revenue per GW of AI capacity runs ~$12–13B/GW/yr (SemiAnalysis, Jun 2026 — a contested, single-source figure), which is why speed-to-power has direct dollar value: energizing and immediately monetizing 200 MW six months early would imply roughly $1.2–1.3B in incremental gross revenue under that rate scenario against a depreciation clock that is already running. The revenue-per-MW you can actually realize tiers by archetype — interactive inference at a latency premium, batch at a spot discount — and the mix you contract determines whether you sit comfortably above breakeven or hope for it. Apply the actual ramp, collection terms, discounts, credits and counterparty exposure; energization alone earns no revenue.

Deloitte’s November 2025 forecast puts inference at ~2/3 of 2026 AI compute. For an inference operator, accepted and billable output is the revenue boundary, so this chapter derives its unit economics in full. The build-up runs from physics to price: tokens/GPU-second → $/GPU-hour → $/M-tokens. A throughput-only illustration on the installed Hopper base: an 8x H100 node at ~$19.20/hr serving Llama-70B at ~2,800 tokens/sec would imply near $1.90/M generated tokens at continuous stated throughput; Chapter 14.1 owns useful-output accounting; Chapter 1.3 supplies the inference service contract. This is not yet cost per accepted response, and the number is brutally sensitive to model size, precision, and batch efficiency. Current frontier capacity keeps the arithmetic and changes both of its terms. A GB300 NVL72 serves inside a 72-GPU NVLink domain, and providers slice that domain differently: CoreWeave and Azure sell four-GPU instances, other capacity is contracted by the rack. Price against the unit your contract actually names, and re-derive tokens per second on the platform you serve from, because prefill/decode disaggregation moves that term too, by an amount that depends on the model, the prompt mix, and the concurrency. The same hardware can swing the cost several-fold depending on how well you batch and how long the decode sequences run.

At the application layer, inference COGS consume revenue before it becomes margin. The ICONIQ survey describes its respondent population, not a structural ceiling for every AI business. ICONIQ’s surveyed companies expected AI gross margins to reach ~52% in 2026; its earlier ~41% baseline and Bessemer’s separate 70–90% mature-SaaS comparison are different periods and populations. They do not establish an observed whole-industry margin path. In the cited scaling-stage AI-company population, inference COGS averages ~23% of revenue — for every $1M of AI product revenue, roughly $230k is consumed by inference. The gross-margin waterfall is: list price, minus token COGS, minus the inevitable free-tier and retry overhead, minus the cost of the long decode sequences that reasoning models emit. Every layer of that waterfall is under pressure from the layer below it.

Chapter 1.6 framed the procurement fork qualitatively; this is its quantitative home. The benchmark to anchor against is wholesale colocation: ~$217/kW-month global average in 2025 (Ashburn ~$215 at record highs; range ~$120 in Atlanta to ~$250 in Silicon Valley, up to ~$450 in Singapore), with build-to-suit / credit-tenant leases at ~$150–220/kW-month over 15 years (JLL / CBRE, 2025); availability is market-specific, so price the lease case from a current quote for density-ready capacity in your target market rather than from a global benchmark. Convert $/kW-month to $/MW-year — roughly $1.8–2.6M/MW-year, and note the scope: that rent buys the powered shell and facility service only, while the tenant still buys its own IT fleet, power, and operations. Only after adding those does the lease case compare against the ~$8.5M/MW-year all-in TCO of a self-build. On that like-for-like basis, price any quoted lease premium against capex-light speed and optionality under demand uncertainty.

The NPV comparison is necessary but not sufficient, because a flat DCF understates the value of being able to change your mind. When demand is uncertain — the normal state in 2026 — negotiated break, renewal, and expansion clauses give a lease its real-option value: the committed term remains an obligation, while those clauses let the tenant change its capacity position as demand resolves. A self-build forecloses that option; you own the megawatts whether or not the workload materializes. The correct comparison prices the option premium: how much extra $/MW-year is the exit/flex right worth, given your demand variance? Durable, well-forecast demand at scale spreads owned fixed cost; build wins only when matched cash flows beat the lease or rental price, including its margin. Spiky or uncertain demand raises an enforceable exit’s value; pay a premium only when avoided future bills justify it. Price confidence in your demand forecast alongside those cash flows.

Build the cash flows: paid hours H = 64 × 8,760 = 560,640 GPU-hours/year. Let K, O and r be the displayed owned capex, annual operating cash and rental rate. Annual rent R = Hr, about $1.7M/year. The three-year discount factor sum A = 1/1.10 + 1/1.10² + 1/1.10³ = approximately 2.5; calculate with the unrounded sum. Ownership present cost is K + OA, about $3.4M; rental present cost is RA, about $4.2M. Ownership saves about $0.8M in present cost. Its incremental cash-flow trace is −K at time zero, then R − O at each of years one, two and three. Appendix C reproduces this NPV with avoided rent as benefit, owned opex as expense, and zero debt/tax.

Select ownership for the three-year service: upfront capital buys lower lifetime cost. Procurement is HOLD pending matched quotes and service acceptance. Straight-line book expense is K/5 per year, about $0.48M. After three years the book carrying value is K(1 − 3/5), about $0.96M, while assumed disposal proceeds are zero. Book value creates neither exit cash nor earning years. Divide both costs by the same useful output; paid idle hours remain included.

Flip on price: solve K + OA = Hr*A, so rental-rate crossover r* = (K/A + O)/H, about $2.4/GPU-hour. Below that matched-service crossover, rent. At two-thirds of the base rate, rental PV is about $2.8M, saving about $0.6M; support, reserve and cancellation must still match.

Flip on earning life: if all demand ends after year one, the rental exit right removes years two and three. Ownership cost is K + O/1.10, about $2.8M, while rent is R/1.10, about $1.5M; rental saves about $1.2M. In whole-year cash periods the ranking crosses between years two and three: ownership’s incremental NPV is about −$0.2M over two years and +$0.8M over three. A three-year take-or-pay rental would retain the later bills and would not buy this exit benefit. For delayed ownership, add bridge capacity and shorten earning time. Method: Epoch capital recovery. This chapter owns procurement NPV and economic versus book life; Chapter 2.5 adds debt, CFADS and DSCR, and Chapter 7.11 supplies qualified equipment and support inputs.

How you finance the asset changes what you can build and what survives a downturn — the deal mechanics live in Chapter 2.5, but the strategic logic belongs here because it feeds straight into the ROI scorecard. The defining feature of the 2026 build-out is that it has outgrown self-funding: against a multi-year build estimated near $2.9T (2025–2028) with a ~$1.5T financing gap beyond hyperscaler cash flow (Morgan Stanley, 2025), the market reached for 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, against $11.4B combined in 2024 — and gross ABS+CMBS supply is projected toward $30–40B/yr in 2026–2027 (CREFC, 2026).

The strategic catch is that the collateral is the very asset whose value is contested. GPU-backed lending underwrites a depreciating, deflating asset against a thin secondary market — the same residual-value uncertainty from the depreciation debate, now wired into the capital structure. CoreWeave is the visible test case: Q2 2026 revenue $2.58B (+112% YoY) at a 59% adjusted-EBITDA margin, but a −$626M quarterly net loss, ~$35B of balance-sheet debt ($31.4B of it recourse), $640M of quarterly net interest (~42% of adjusted EBITDA), and a ~$104B backlog — excluding >$25B of early-Q3 commitments — concentrated in a few anchor tenants (company filings, Aug 2026). The 'circular financing' critique — vendor stakes and residual backstops that let a buyer finance the purchase of the vendor's own chips — is a structural risk: it couples the financing to the same demand and residual assumptions the equipment depends on, so a residual shock hits collateral, covenants, and revenue at once.

The same physical asset earns very different returns depending on who operates it, because each archetype inherits a different cost-of-capital, utilization risk, and margin structure. They are four distinct business models that happen to share a bill of materials.

The hyperscaler finances from operating cash flow at its own corporate cost of capital, fills the asset with its own first-party demand (search, ads, cloud, internal training), and treats the data center as cost-of-revenue for a far larger product. Captive demand changes utilization risk but does not eliminate a demand miss; the depreciation policy is the visible lever, which is why hyperscaler life-extensions move billions of reported income. The neocloud is the opposite: thin margins, high leverage, GPU-backed debt, and acute exposure to its contract-specific cash breakeven and to tenant concentration — a high-beta bet on sustained GPU demand. The colo / build-to-suit operator sells powered shells and steady $/kW-month rent, carries real-estate-like risk and real-estate-like cost of capital, and is largely insulated from GPU obsolescence because the tenant owns the silicon. The self-build enterprise/lab optimizes for control and long-run unit cost on a durable, well-forecast workload, accepting the deepest capital commitment and the full obsolescence risk in exchange.

Protecting the return is a small set of levers, each of which maps to a downside it hedges. The power-cost lever is the largest controllable opex line — energy is ~$0.6B/yr in the 1 GW model — so a cheap, firm, long-dated PPA or on-site generation is worth more to lifetime ROI than most capex optimizations. Depreciation policy is the lever that decides whether reported margin reflects reality; the conservative choice protects against a residual shock at the cost of near-term earnings. Design-for-flexibility — reserving floor loading, water, and electrical headroom for a density ramp, and keeping procurement mode hybrid — is the lever that hedges workload and generation uncertainty. The ROI scorecard ties them together: levered IRR, DSCR, payback against economic (not book) life, and the contracted-vs-merchant revenue split that sets debt capacity.

Then stress the downside, because the asset's fragility lives in the tails:

• Utilization collapse. Fewer sold GPU-hours reduce contribution while fixed operating cash and scheduled debt service remain due; a price cut compounds the loss. Use the matched cash-flow case here for the capacity commitment and Chapter 2.5’s DSCR case for the debt schedule. The hedge is contracted/take-or-pay revenue; the failure mode is a merchant fleet into a soft GPU-rental market.
• Residual-value shock. If three-year residuals fall toward the low end of the contested ~20–40% band, GPU-backed debt is under-collateralized and the short-life depreciation bears are vindicated. The hedge is conservative depreciation and limited leverage; the failure mode is circular financing against an optimistic residual.
• Rate spike. Highly-levered builds (interest already ~40% of adj. EBITDA at the visible neocloud) are acutely rate-sensitive; a financing-cost spike can exceed the entire margin. The hedge is fixed-rate, long-dated debt and a contracted revenue base.
• Contract non-renewal. Backlogs concentrated in a few anchor tenants (~13x revenue at the visible case) mean one non-renewal can strand a campus. The hedge is tenant diversification and take-or-pay with real termination economics.
• The secondary-market-depth (Burry) thesis. The whole defense of the long life 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, and a large slice of reported industry earnings is borrowed from a future write-down. This is the systemic version of the residual shock.