Wiki/Economics
Yield-on-cost vs. payback: reconciling the Sacks math with REIT math
This page exists to settle the most common public argument about AI capex economics: whether 1 GW of AI datacenter pays back in ~2 years (the David Sacks "boom is real" framing) or ~10 years (the colocation-REIT framing). Both math walks-through are arithmetically defensible. They differ on what they count, who's counting it, and over what time horizon.
The underlying numbers are now disclosed clearly enough — by Digital Realty, Equinix, Meta, CoreWeave, the Blue Owl Hyperion SPV — to do the reconciliation honestly.
The two claims, side by side
Claim A — Sacks (October 2025, paraphrased from tweet)
| Line | $ |
|---|---|
| All-in capex for 1 GW | ~$50B |
| Enterprise revenue/year | ~$25–30B |
| Electricity cost/year | $1–2B |
| Implied "payback" | ~2 years |
Claim B — "Stripped to NOI" critique
| Line | $ |
|---|---|
| All-in capex for 1 GW | ~$50B |
| Gross revenue/year | ~$25–30B |
| Opex (power, staffing, maintenance, taxes, insurance) | ~$3–5B |
| Stabilized NOI | ~$5B |
| Yield on cost | ~10% |
| Implied payback | ~10 years |
Both claims start from the same revenue assumption. They diverge on (1) what gets stripped out as cost before calling it "return," and (2) whether the $25–30B revenue figure is even realistic outside of a fully-leased anchor scenario.
What the disclosures actually say
Datacenter REIT yield-on-cost is ~11.9%, not 50%
The single most load-bearing primary datapoint we've found:
Digital Realty Q4 2025: $10B / 769 MW under construction at 11.9% expected stabilized yield-on-cost (DLR Q4 2025 earnings transcript). That figure is essentially flat against the 11–12% range DLR has guided to for years. AI demand has not moved the project-level yield-on-cost number upward — pricing and cost have inflated roughly in tandem.
Equinix discloses a different metric: 26% cash-on-cash return on gross PP&E for its 180 stabilized assets (Q3 2025) (Equinix Q3 2025 earnings presentation). That number is on legacy gross book — i.e. capex spent years ago at lower per-MW prices — not a forward yield on new builds. Different metric, different denominator.
The reconciliation: DLR's wholesale/hyperscale model targets ~12% on fresh capital today; Equinix's retail interconnection model converts to ~25%+ once an asset has fully ramped over 5–7 years. Neither is reporting a number remotely close to Sacks's 50%+ implied yield.
For full detail: operators/colocation-reits.md.
Build cost per MW is now ~$20–27M, not $50M
JLL's 2026 Global Data Center Outlook puts AI-optimized turnkey build at $20M+ per MW, vs. $10.7M global average for traditional turnkey colo. The Meta Blue Owl Hyperion JV implies ~$27M/MW for a fully-financed 1 GW AI-ready campus in Louisiana (Meta press release Oct 2025).
So Sacks's $50B for 1 GW pencils only at $50M/MW — roughly 2x the disclosed JLL figure and ~85% above the Hyperion datapoint. Possible read: the $50B figure includes GPU stack (typically separate from real-estate capex) plus generous land/substation provisioning plus phasing for future upgrades. With GPUs in, $50M/MW becomes plausible.
That recasts the comparison. Sacks's "$50B" is a server-included number; the REIT $20M/MW is real-estate-and-power-shell only. The right comparable for the REIT model is the M&E + buildings line — call it $20–27B of the $50B. The remaining $20–30B is GPU stack capex that the operator (CoreWeave, hyperscaler, or co-located customer) is putting in on top.
This is why the two camps talk past each other:
- REIT math measures yield on the building. NOI is rent from a tenant.
- Sacks math measures yield on the building + the chips inside it, assuming the chips are revenue-generating compute capacity.
If you're the hyperscaler with the GPUs and the customer in one entity (Microsoft running OpenAI workloads on its own metal), Sacks's framing is closer to reality. If you're a colo REIT renting powered shells to others, the REIT framing is closer to reality.
Stabilized opex per MW
Pulling from colocation operator disclosures and industry rules-of-thumb, the run-rate opex for a 1 GW AI-ready facility looks roughly like:
| Line | $/MW/yr | $/GW/yr |
|---|---|---|
| Power (8,760 hrs × ~1.3 PUE × $50–80/MWh wholesale) | ~$0.6–0.9M | ~$0.6–0.9B |
| Staffing (1 engineer / ~5 MW; ~$200k loaded) | ~$0.04M | ~$0.04B |
| Maintenance + service contracts | ~$0.3–0.5M | ~$0.3–0.5B |
| Insurance + property taxes | ~$0.2–0.4M | ~$0.2–0.4B |
| Total operating opex | ~$1.2–1.8M | ~$1.2–1.8B |
| Power-only (matches Sacks) | ~$1–2B |
Sacks's $1–2B power-only line is roughly right. But total stabilized opex including non-power lines is closer to ~$2–3.5B/yr at the building level, before GPU refresh and before lease/cost-of-capital. Sacks's revenue – power = NOI proxy is therefore overstating NOI by $1–2B/yr.
GPU depreciation is the missing line
The biggest hole in Sacks's framing: no depreciation charge.
If the GPU stack inside that 1 GW facility costs $25–30B (the differential between $50B all-in and $20–27M/MW real estate), and useful life is 3 years (the case the bears make — Hopper→Blackwell→Rubin cycle; rental-rate collapse evidence; see $8–10B/yr**.economics/gpu-depreciation.md), depreciation runs **
Even at hyperscaler-booked 6-year lives, depreciation on $25–30B of GPUs is ~$4–5B/yr.
Sacks's revenue–power = "2-year payback" math ignores this entirely. Including it:
| Scenario | Revenue | Power | Other opex | GPU depreciation | "Cash NOI" (revenue – cash opex) | "Economic NOI" (after dep'n) |
|---|---|---|---|---|---|---|
| Sacks math (cash NOI on capex, no dep'n) | $27.5B | $1.5B | — | — | $26B | (n/a) |
| Cash NOI with full opex | $27.5B | $1.5B | $1.5B | — | $24.5B | (n/a) |
| Hyperscaler booked (6-yr GPU life) | $27.5B | $1.5B | $1.5B | $4.5B | $24.5B | $20B |
| Bear case (3-yr GPU life) | $27.5B | $1.5B | $1.5B | $9B | $24.5B | $15.5B |
On the bear-case economic-NOI line: $15.5B / $50B = 31% yield-on-cost. Still far above the REIT 11.9% figure, but for a different asset class (GPU-laden compute factory, not powered shell). And that 31% assumes Sacks's $27.5B revenue figure holds — which is itself contested (see demand/demand-and-revenue-gap.md).
What's actually going on: three regimes, not one
The Sacks-vs-REIT argument confuses three different unit economics by treating them as the same business:
Regime 1 — Colocation REIT (DLR, Equinix, IRM)
- Builds shells and powered shells.
- Rents to hyperscalers and enterprises on 10–15-year leases with 3–4% escalators.
- Stabilized yield: ~11.9% on fresh AI development.
- Payback in cash terms: ~8–9 years; in NPV terms longer because of escalator structure.
- This regime has not been transformed by AI. Scale up, cost up, rent up — yield flat.
Regime 2 — Neocloud (CoreWeave, Lambda, Crusoe, Nebius)
- Leases the powered shell from regime 1 or builds its own.
- Buys GPUs with debt (~5–15% cost of capital — CoreWeave's DDTL ladder).
- Rents GPU hours to hyperscalers and AI labs on 2–7 year contracts.
- Revenue density ~$8M/MW/yr (CoreWeave-disclosed approximation).
- Tail risk dominates: customer concentration (CRWV ~67% MSFT in FY25), GPU obsolescence, contract renewal cliff.
- Booked yield is high (~25–40% gross) but expected return is much lower after risk-adjustment.
Regime 3 — Hyperscaler self-build (MSFT, META, GOOG, AMZN, ORCL)
- Builds shells and provisions GPUs in one stack.
- Uses internally (training, inference for own models, customer compute).
- Revenue side is vertically integrated: MSFT compute consumed by OpenAI → MSFT books revenue → MSFT depreciation runs against same revenue.
- Effective "payback" depends on whether internal usage drives upstream revenue (Copilot seats, Azure OpenAI consumption, ad-targeting improvement).
- This is the regime Sacks is implicitly describing — where GPU rental rates and own-use shadow prices are interchangeable.
The 2-year payback math is closest to true in Regime 3, under the assumption that internal consumption can be valued at external rental rates. That's a strong assumption. Microsoft's internal cost-of-capital is roughly the WACC (~8%), not the 25–40% gross margins CoreWeave charges third parties — so MSFT's effective shadow price is lower, and "payback" is correspondingly longer.
The 10-year payback math is closest to true in Regime 1, and approximately true in Regime 2 once GPU depreciation is honestly modeled and customer-concentration risk is priced.
Both critics are right about a different business.
The Hyperion datapoint — a cross-check
The Meta Blue Owl Hyperion JV is the cleanest disclosed deal-economics on a 1+ GW AI datacenter we have (DCD coverage; Meta press release).
- $27B all-in for a gigawatt-scale campus in Louisiana → ~$27M/MW real estate + power
- A+-rated debt anchored by PIMCO ($18B) and BlackRock ($3B) at investment-grade rates
- Blue Owl 80% / Meta 20% equity split
- Initial lease term: 4 years (with extensions)
Two things to note:
The deal is priced like REIT economics on the real-estate side, not like Sacks economics. A+ debt prices around 5.5–6% these days. If lenders thought this was a 50%-yield asset they would not be lending against it at 6% in a sleeve where they front 80% of the equity. The debt market is pricing Hyperion like an ~11–13% asset, consistent with the REIT comp.
The 4-year initial lease is anomalous and tells a separate story: Meta wants optionality if AI economics deteriorate. If Sacks's 2-year payback were correct, Meta would happily sign 15 years. The structure suggests Meta — the company building Hyperion — has more uncertainty about the long-run revenue trajectory than the public bull case implies.
Bottom line
The argument isn't really 2 years vs. 10 years. It's whether you're underwriting a building or a compute factory, and whether you're using cash-NOI or economic-NOI as your numerator.
- Building, cash NOI → 11–12% yield, ~9-year payback. (Disclosed at DLR/EQIX.)
- Building, after depreciation → 6–8% return on assets. (Hyperion-style structures.)
- Building + GPUs at external rental rates, cash NOI → ~50%+ gross, "2-year payback." (Sacks math, but excludes opex and depreciation.)
- Building + GPUs at external rental rates, after honest depreciation → ~25–35% economic NOI, ~3–4 year payback, contingent on Sacks's $27.5B revenue assumption holding for the full useful life of the GPUs.
That last contingency is the heart of the debate. If the $25–30B/yr revenue per gigawatt is sustainable as several hundred gigawatts come online simultaneously, the bull case stands and most of the buildout pencils. If aggregate revenue is capped well below the implied $250–500B/yr total industry revenue at full hyperscaler build-out, the math breaks not at the project level but at the industry level — see demand/demand-and-revenue-gap.md.
The REIT numbers tell you the buildings will get built and the rent will get paid. They don't tell you whether the rent itself is being financed by sustainable end-customer revenue or by the next round of foundation-model financing.