Summary

  • Meta reported approximately $278.99 billion in operating and finance leases that have not yet commenced as of June 30, 2026, up 53% from $182.88 billion a quarter earlier, and added roughly $68 billion more in July 2026, with most leases starting in 2027–2028 on 18–20 year terms.
  • Q2 2026 capital expenditures were $31.08 billion and full-year 2026 capex guidance was narrowed to $130–145 billion, while free cash flow was $784 million — a financing gap bridged by leasing and joint-venture structures rather than current cash generation.
  • Hyperion in Richland Parish, Louisiana, is announced at 5 GW compute capacity and more than $50 billion, but was structured in October 2025 as an 80/20 joint venture with Blue Owl Capital funds covering about $27 billion of development costs, with a Meta residual value guarantee for the first 16 years.
  • Operating evidence is narrower than announced capacity: at Prometheus in New Albany, Ohio, drone imagery confirms only the South portion operating while Prometheus North remains under construction, and Meta's Alberta data centre may start before its dedicated 932 MW gas plant.

The numbers behind Meta's AI infrastructure expansion have split into two registers that increasingly describe different realities. The first register holds commitments: figures reported in filings and press releases about what Meta has contracted or announced. The second holds operating evidence: what is energized and serving load today. This article keeps the two registers separate, because conflating them is how a $784 million quarterly free cash flow can appear to finance a $279 billion lease book.

Start with the commitment register. Meta's Q2 2026 disclosures (SEC Exhibit 99.1; company press release) show capital expenditures including finance-lease principal of $31.08 billion for the quarter and 2026 guidance narrowed to $130–145 billion from a prior $125–145 billion. The same quarter ended with $90.26 billion in cash and marketable securities against $83.66 billion of long-term debt — and just $784 million in free cash flow. The commitments, however, extend far beyond what capex captures. As of June 30, 2026, Meta had approximately $278.99 billion in operating and finance leases that had not yet begun, covering data centers, colocations and certain network infrastructure, with commencement dates running from the remainder of 2026 through 2036 and terms from over one year to 30 years (Seattle Times; Business Insider). That figure rose 53% from $182.88 billion in the Q1 filing. In July alone, Meta entered additional data-center leases of roughly $68 billion, expected to begin in 2027 and 2028 with 18- to 20-year terms. Bloomberg's synthesis of the filings put aggregate future spending commitments near $700 billion, split between $349.3 billion of non-cancelable contractual commitments — mostly third-party cloud, servers and network infrastructure — and roughly $347 billion in unstarted leases (Bloomberg). The two lease figures use different scopes and dates and should not be summed into a single headline total.

The physical register tells a more constrained story. At Hyperion in Richland Parish, Louisiana, Meta announced in July 2026 an expansion to 5 GW of compute capacity, its largest site, at more than $50 billion, supporting over 7,500 peak construction jobs and 1,000 operational roles (Meta). But that 5 GW is announced capacity, not operating load. The site's financing structure dates to October 2025, when Meta agreed with funds managed by Blue Owl Capital to develop the initial campus: the JV committed about $27 billion in total development costs, Blue Owl funds took an 80% interest against Meta's 20%, Meta received a roughly $3 billion one-time distribution, and Meta entered operating leases with a four-year initial term plus a residual value guarantee covering the first 16 years (Meta press release). In other words, the construction cost moved off Meta's reported capex and into future lease obligations, with downside on asset value returning to Meta through the guarantee.

Power sequencing shows where announced timelines depend on parties Meta does not control. In Louisiana, Meta's agreement with Entergy funds seven new natural gas plants, three grid-scale batteries, nuclear uprates and purchased power, with claimed customer savings of about $2 billion plus a prior $650 million, and Meta has contracted more than $1.6 billion with local Louisiana businesses. In Alberta, Meta's roughly 1 GW data centre north of Edmonton may start up before the adjacent 932 MW Greenlight gas plant — a Pembina-led, approximately C$4.6 billion project approved by its partners — making interim power arrangements part of the delivery plan rather than an afterthought (CBC; BNN Bloomberg; BTW). The Pembina Institute has said the new load could raise wholesale power prices even if Meta pays its direct connection costs. In Utah, Meta raised planned investment at Eagle Mountain to $3 billion in September 2026, adding a battery plant, while capacity, cooling design and power arrangements remain only partially disclosed (Hoodline).

What is actually operating? Third-party verification is thinner than the announcement volume. Synmax's drone imagery analysis confirmed Prometheus South in New Albany, Ohio, is operating while Prometheus North is under construction (Synmax), and local reporting describes the full site as slated to open in 2026, with Meta lining up nuclear power supply for the supercluster (NBC4; ABC6). That is the scale of verified energization against a commitment register in the hundreds of billions.

The binding constraint has visibly shifted from financing to delivery inputs. Reporting on the electrician shortage affecting Meta and Google buildouts describes skilled-trades scarcity as a direct constraint on construction timelines (eWeek); supply-chain analyses of the data-center industry describe procurement bottlenecks in transformers, switchgear and cooling equipment (Global Data Center Hub; SAVRN). Lease obligations and JV commitments accrue on schedule assumptions that a constrained labor and equipment market may not deliver.

The honest reading is that Meta has built a mechanism for converting announcements into contractual delivery obligations — leases, JV equity, guarantees — rather than a mechanism that guarantees the obligations become operating compute on schedule. That mechanism has real advantages: it locks in sites and power ahead of competitors, and it spreads construction cost across future periods.

It also concentrates risk in exactly the places the headline figures do not show: a residual value guarantee running 16 years, lease payments beginning in 2027–2028 against demand that must still materialize, and interconnection timelines owned by utilities and regulators. Neither 'execution proof' nor 'off-balance-sheet evasion' is a fair summary; both are claims that belong to their advocates, and the filings support a narrower conclusion — Meta has contracted its future, and the future now has to arrive on time.