| Scenario | IT load (GW) | B200-equivalent GPUs | Notes |
|---|---|---|---|
| Bear | ~8 GW | ~5–6M | Hyperion phases slip; one major campus delayed 2+ years; permitting/grid issues |
| Base | ~11 GW | ~7–8M | Hyperion hits ~2 GW by 2027, ~5 GW by 2030; Prometheus full; fleet of mid-size campuses online |
| Bull | ~14 GW | ~9–10M | Hyperion full 5 GW by 2030; second mega-campus announced and partially online; MTIA displaces a large share of Nvidia spend |
Zuckerberg's Jan 2025 Facebook post explicitly framed the build as Hyperion reaching ~5 GW eventually, with footprint "covering a significant part of the footprint of Manhattan"[4]. The Blue Owl JV press release confirms ~5 GW at full build for Hyperion[1]. Combined with the existing 25+ campus fleet, base-case 2030 lands at 10–12 GW.
Reference points: Meta's installed base going into 2025 was ~5 GW across the global fleet (estimates from SemiAnalysis and DCD)[5]. Google's 2030 path is comparable (10–15 GW), Microsoft is in a similar range largely via leases, Amazon is the biggest but most fragmented. Meta is in the top tier — not Stargate scale (5+ GW from a single project), but in the small club of operators that will own/operate multi-GW AI campuses.
The Oct 27, 2025 announcement is structurally important and was the load-bearing error in v1[1][2][6]:
Economic interpretation: This is the same financing pattern Microsoft uses for some of its CoreWeave-adjacent capacity and that Oracle uses with Crusoe — get the campus built using third-party capital, pay rent rather than capex, keep more cash for GPUs and OpEx. The difference is Meta has a 20% equity strip and an RVG, so it retains meaningful upside if the campus is worth a lot in 2042. But the headline "Meta owns Hyperion" is wrong — Meta is the 20% LP and the tenant.
Why Meta did this: the ~$3B distribution + the ability to expense lease payments rather than depreciate ~$27B of capex over 20+ years materially improves near-term reported earnings and frees cash for the bigger ticket item — GPUs. Wall Street's reaction was positive precisely because it offloaded capex intensity at a time when 2026 capex was about to be guided to $115–135B[3].
| Campus | Status | Capacity | Ownership |
|---|---|---|---|
| Prometheus — New Albany, OH | Phasing online 2025–26 | ~1+ GW | Meta-owned |
| Hyperion — Richland Parish, LA | Under construction | ~5 GW at full build | 20% Meta / 80% Blue Owl JV — Meta leases[1] |
| Existing fleet (~25 owned campuses) | Operational | ~5 GW today, growing | Meta-owned |
| Next mega-campus(es) | Site-hunting / partially announced | 2–5 GW each | Likely structured similarly to Hyperion JV given the financing precedent |
Existing fleet (operational, owned): Forest City NC, Prineville OR, Altoona IA, Luleå Sweden, Clonee Ireland, Henrico VA, Newton GA, Los Lunas NM, Eagle Mountain UT, Papillion NE, Sarpy NE, Fort Worth TX, Temple TX, DeKalb IL, Huntsville AL, Kansas City MO, Cheyenne WY, Gallatin TN, Rosemount MN, Jeffersonville IN, Montgomery AL, Stanton TX, Idaho Falls ID, Beaver Dam WI[7].
Historically small. Meta has used some build-to-suit deals (Digital Realty, CyrusOne, QTS in earlier years) but third-party colo is a small minority of its footprint. With the Hyperion JV, the operating-lease share of the balance sheet just jumped materially.
Meta is still an outlier here vs. peers. Microsoft has multi-tens-of-billions committed to CoreWeave; Oracle is leaning heavily on Crusoe; Google has selectively contracted. Meta has done little of this publicly at scale. The Blue Owl JV is the inverse of a neocloud contract — Meta uses third-party capital but operates the asset itself.
| Year | Capex | Source |
|---|---|---|
| 2023 | $28.1B | 10-K |
| 2024 | $39.2B | 10-K |
| 2025 (actual) | $72.22B (incl. principal on finance leases) | Q4 2025 release[3] |
| 2026 (guided) | $115–135B (incl. principal on finance leases) | Q4 2025 release[3] |
| 2027 | Plausibly $140–170B | Trajectory — Meta has flagged "notably larger" growth and Superintelligence Labs build-out[8] |
| 2028–2030 | Likely plateau in $130–170B/yr band | Hyperion 5GW build alone is ~$27B over its lifecycle; new sites likely structured JV-style |
| Cumulative 2025–2030 | ~$750B–$1.0T | Implied by trajectory above — meaningfully higher than v1's $550–750B estimate |
Important: this is the largest single-company capex program in history — comparable in real-dollar terms to the Apollo program, the US interstate system, or peak-era AT&T. It is funded by a combination of (a) Family of Apps operating cash flow, which still runs $90–110B/yr in EBITDA, (b) the Hyperion-style JV structures that move shell capex off-balance-sheet, and (c) a modest increase in debt issuance (Meta floated $30B of bonds in late 2025).
| Layer | Owned? | Detail |
|---|---|---|
| Land (legacy fleet) | Yes (freehold) | Tens of thousands of acres aggregated across ~25 campuses |
| Land (Hyperion) | JV-owned, Meta 20% | ~2,000+ acres in Richland Parish, LA — owned by Blue Owl-led JV |
| Buildings & mechanical (legacy) | Yes (self-designed) | Open Compute reference design, modular |
| Buildings & mechanical (Hyperion) | JV-owned, Meta tenant | Meta provides construction management services to the JV |
| Power infrastructure (legacy) | Mostly yes | Owns substations and backup gen on the legacy campuses |
| Power infrastructure (Hyperion) | JV-owned | Including the long-lived power/cooling/connectivity infra — explicitly enumerated as JV assets[1] |
| Energy contracts (PPAs) | Meta-contracted | 30+ GW of clean-energy PPAs cumulatively. Constellation Clinton nuclear restart ~1.1 GW, 20-year[9]. Multiple SMR developer agreements signed 2024–2025 |
| GPUs (Nvidia) | Owned by Meta | ~600K H100-equivalents end of 2024; targeting 1.3M GPU-equivalents end of 2025 per Zuckerberg[4]; trajectory to 5–8M B200-equivalents by 2030. These sit on Meta's balance sheet even when housed in the Hyperion JV's shell |
| Custom silicon (MTIA) | Owned IP, fab via TSMC | MTIA v2 ramped 2024–25 for ranking/recommendation inference. v3 in development for training. Plausibly handles 30–50% of inference workload by 2028, compressing the Nvidia bill |
| Networking | Owned | Custom in-house DC switch fabric, optical interconnect |
| Cooling | Owned (legacy) / JV (Hyperion) | Shifting to direct-to-chip liquid for AI racks |
| Models & data | Owned | Llama family, training data corpus, RLHF infra |
Key reframing: Meta still owns the value-dense piece (the GPUs) and the optionality piece (the models, data, MTIA IP). What it has partially sold off is the lowest-IRR, highest-capital-intensity piece (shell + power infra at Hyperion). That's actually rational capital allocation if you believe GPUs and IP carry the strategic premium and shells are a real-estate cost.
Sanity-check the asset on a rebuild-from-scratch basis. Industry rule of thumb for greenfield AI DC construction in 2025–2026[5][6]:
| 2030 capacity | Gross replacement (shell) | Gross replacement (all-in) | Meta equity-weighted share (all-in) |
|---|---|---|---|
| 8 GW (bear) | $320–480B | $600–880B | $430–630B |
| 11 GW (base) | $440–660B | $825B–$1.21T | $575–845B |
| 14 GW (bull) | $560–840B | $1.05–1.54T | $735B–$1.08T |
The "Meta equity-weighted" column applies a haircut: for the ~6 GW Meta fully owns, count 100%; for the ~5 GW in Hyperion-style JVs, count GPUs (~100%, since GPUs aren't in the JV) plus 20% of the shell. That gives roughly 70% of gross all-in value as Meta's economic share, which is the column shown.
~$575–845B in 2030, base case. Conservative but a defensible lower bound.
CoreWeave at peak traded 8–15x forward revenue on contracted AI capacity. At ~$25M/MW-yr fully utilized rental revenue for B200-class capacity, 11 GW = $275B/yr theoretical revenue. Apply a 10x multiple: $2.75T gross. Even with a steep haircut for Meta consuming most of its own capacity (so no actual revenue) and the Hyperion JV split, the opportunity-cost framing says Meta's economic interest in the compute fleet is worth well over a trillion if compute itself stays scarce.
If recursive self-improvement starts in 2026–2027 and the binding constraint on intelligence becomes compute, then 7+ GW of Meta-controlled fleet by 2028–2030 is not a depreciating asset — it's a seat at the table. Likely 5 entities globally with this scale: Google, Microsoft, Amazon, Meta, and the OpenAI/Stargate complex. Anyone training a frontier model must negotiate with one of them or wait years.
Under that lens the right comparable isn't EQIX/DLR (20x EBITDA) or even CoreWeave. It's "essential strategic infrastructure during a phase change" — Standard Oil's pipeline network, AT&T's long-distance lines pre-divestiture, TSMC's leading-edge fabs today. Such assets at peak strategic importance don't carry a normal multiple — they're priced on optionality and bargaining power.
| Component | Implied value | Comment |
|---|---|---|
| Family of Apps (ads business) | ~$1.2–1.4T at 18–22x earnings | The market essentially prices Meta as ads + a low residual for AI |
| Cash & investments (net of debt) | ~$30–50B | Down from prior years after $30B bond issuance + Hyperion-related $3B inflow |
| Reality Labs | Net negative ($15B+/yr drag) | Market gives it zero or negative value |
| Implied value of Meta's compute infrastructure equity | ~$250–500B | The residual after ads + cash – RL drag. Below Lens A replacement cost |
None of these change the qualitative thesis: Meta is one of 4–5 globally significant compute operators of the AGI era and the market is paying for it as if compute is a cost line. They change the magnitude — the right framing is "compute infrastructure adds $300–500B to fair value above the standalone ads business" rather than v1's "$1T+."