The build is no longer a spending decision. As of Sept. 30, 2026, Meta carried $454.52 billion of non-cancellable contractual commitments and roughly $355.73 billion of operating and finance leases not yet commenced for data centers, energy infrastructure, colocations and network gear, running through 2036. On Nov. 11, the company said infrastructure and co-head of engineering Santosh Janardhan will leave in early 2027 and hand the program to Dave Brown, the former AWS compute chief hired four months ago. That combination moves the risk from spending to execution: the question is no longer whether Meta spends the money, but who converts a decade of signed commitments into earning capacity.
2026 capital expenditure guidance
$135B–$145B
Including finance-lease principal, narrowed with Q3 2026 results on Oct. 21, 2026
Q3 2026 free cash flow
-$2.12B
Quarter ended Sept. 30, 2026, reported Oct. 21, 2026
Commitments falling due in 2027
$121.41B
Equal to 84–90% of the full-year 2026 capital expenditure guide
S&P forecast 2027 cash-flow deficit
$71B
Free operating cash flow, per the Oct. 30, 2026 outlook revision
Contracted nuclear capacity
1,121 MW
Clinton Clean Energy Center under Meta's 20-year agreement, starting June 2027
The commitment book
The Build Stopped Being Optional Somewhere Around $454 Billion
Meta's capital program stopped being a plan and became a contract. At Sept. 30, 2026 the company carried $454.52 billion of non-cancellable contractual commitments — mostly third-party cloud capacity arrangements, investments in servers and network infrastructure, energy service agreements for data centers, and some consumer hardware in Reality Labs — plus roughly $355.73 billion of operating and finance leases not yet commenced. That makes the 2027 build a contractual obligation rather than a choice, and it sets the frame for the succession: the question is no longer whether the company spends, but whether the fleet gets run.
| Commitment or obligation | Amount | Terms and timing |
|---|---|---|
| Non-cancellable contractual commitments | $454.52B | Cloud capacity, servers, networking, energy services, consumer hardware |
| Operating and finance leases not yet commenced | $355.73B | Data centers, energy infrastructure, colocations, network gear through 2036 |
| Contractual commitments due in 2027 | $121.41B | 84–90% of the $135–145B 2026 capital expenditure guide |
| Contractual commitments due in 2026 | $45.85B | Remainder of this year |
| Contingent cloud-capacity purchase obligations | $14.72B | Over five years; reducible if the provider resells capacity |
| Maximum exposure to unconsolidated data center ventures | $68.40B | Includes $41.49B of residual value guarantees; ~$27B Louisiana and ~$14B Texas ventures |
| Long-term debt, non-current | $80.92B | After $25.0B of senior unsecured notes issued in May 2026 |
| Unused letters of credit supporting infrastructure purchases | $1.53B | As of Sept. 30, 2026 |
Two figures matter more than the guidance itself. $121.41 billion of commitments fall due in 2027, equal to 84–90% of the $135–145 billion Meta expects to spend this year, and $45.85 billion is due in the remaining months of 2026. That rules out spending restraint as a defense of the balance sheet: if free cash flow stays negative, the answer has to come from revenue and utilization rather than from a smaller build, and the option to shrink the program has already been surrendered.
- Cutting 2027 capital spending to protect cash — $121.41 billion of obligations come due no matter what the earnings call says
- Buying optionality on power, shells and land — $355.73 billion of leases not yet commenced already run through 2036
- Arguing the risk is cyclical demand — the commitments are signed for a decade, so the only variable left is execution
- Partnering to reduce exposure — the company has already taken $68.40 billion of maximum exposure to unconsolidated data center ventures, including $41.49 billion of residual value guarantees
The cash engine
The Cash Engine Flipped Before the Handover Did
Meta's quarterly capital spending has more than doubled in a year
Purchases of property and equipment per Meta's quarterly cash flow statements. Including finance-lease principal, capital expenditures were $34.91 billion in Q3 2026 alone.
Unit: $ billions
Q1 2025
12.9
Q2 2025
16.5
Q3 2025
18.8
Q4 2025
21.4
Q1 2026
19
Q2 2026
30.1
Q3 2026
33.9
Cash generation stopped covering capital spending before the departure was announced. Operating cash flow was $32.79 billion in Q3 2026 against $33.94 billion of property-and-equipment purchases, and Meta's own free cash flow definition, which also deducts finance-lease principal, turned the quarter negative at -$2.12 billion. On a trailing twelve-month basis operating cash flow of $133.09 billion still covers capital spending of $104.44 billion, but only by $28.65 billion — and that gap shrinks once lease principal and dividends are paid.
| Meta (USD billions) | Q1 2026 | Q2 2026 | Q3 2026 | Trailing 12 months |
|---|---|---|---|---|
| Operating cash flow | $32.23 | $31.86 | $32.79 | $133.09 |
| Purchases of property and equipment | $19.00 | $30.12 | $33.94 | $104.44 |
| Free cash flow after capital spending | $13.23 | $1.75 | -$1.16 | $28.65 |
| Depreciation and amortization | $5.99 | $6.36 | $7.15 | $24.92 |
| Cash and equivalents, period end | $31.10 | $29.27 | $24.35 | $24.35 |
| Net debt issued in period | -$0.84 | $23.95 | -$0.96 | $51.29 |
Meta's cash cushion has thinned while debt funded the build
Cash and equivalents at each quarter end. Over the trailing twelve months the company funded the gap with $51.29 billion of net debt issuance, leaving $80.92 billion of long-term debt outstanding.
Unit: $ billions
Mar 2025
45.4
Jun 2025
30.1
Sep 2025
11.9
Dec 2025
39.1
Mar 2026
31.1
Jun 2026
29.3
Sep 2026
24.4
Underwriting the build now happens in the credit market as much as the equity market. Meta carried $80.92 billion of long-term debt at Sept. 30, 2026 after issuing $25.0 billion of senior unsecured notes in May 2026, and it repurchased no stock in 2026 after $16.8 billion of buybacks in 2025. Depreciation is the quieter drag: Q3 2026 capital spending ran 4.7 times depreciation of $7.15 billion, so as leases commence across the next decade, the income statement absorbs the cost of capacity that is not yet earning.
The succession
A Cloud Operator Takes the Wheel of a Fixed Program
Santosh Janardhan joined Meta in 2009 and ran production engineering before becoming head of infrastructure and co-head of engineering. On Nov. 11, 2026 he told colleagues in an internal memo that it was time to hand over, and that he would stay through the first quarter of 2027 to support the transition. The memo gave no reason beyond the handover itself, and no disagreement over strategy is disclosed. His successor, Dave Brown, joined Meta in July after nearly 19 years at Amazon Web Services, where he worked on EC2 in its earliest days and later ran the compute and machine-learning services group; he will also serve as chief technology officer of Meta Compute, the initiative unveiled in January 2026 to run gigawatt-scale AI capacity. The appointment hands the build to an operator of a commercial cloud rather than the architect of an internal fleet.
…time to pass the baton.
- The objective changes: a fleet built to lower Meta's cost per training token must now also earn revenue per rack from paying external customers
- Meta gains an insider's view of its landlord's cost curve, worth more on a $14.72 billion cloud bill than most efficiency programs can deliver
- Meta inherits a channel conflict, buying capacity from the market while selling capacity under its own brand
- Meta Compute is spearheaded by Dina Powell McCormick and Daniel Gross, with Brown as chief technology officer, so infrastructure leadership now sits inside a commercial structure
- Amazon responded to Brown's exit by handing AWS compute to Dave Treadwell, an executive moved over from e-commerce foundation, effective Aug. 1, 2026
Here is what investors should sit with. The same import that makes Meta better at operating its fleet makes it likelier that the fleet is optimized for utilization and revenue rather than for the internal economics that justified 17 years of Janardhan's work. With $454.52 billion of commitments that cannot be unwound, a leader accountable for filling external contracts can add capacity to fill contracts rather than cut the cost of a token. The risk the handover introduces is therefore not a slower build but a build aimed at the wrong target, and it lands in the one year the cash flow statement has no slack.
Demand
Demand Is the One Variable Nobody Disputes
The order book is not where the problem lives, which is exactly why the execution question is the one that matters. In the same quarter Meta reported negative free cash flow, NVIDIA posted $89.0 billion of data center revenue, up 117% year over year, after $75.2 billion in the prior quarter. Amazon grew AWS net sales 39% to $46.0 billion and held full-year cash capital spending at roughly $220 billion, while Microsoft grew Azure 48% and spent $45.2 billion including finance leases in a single quarter. Nothing in the order book argues for a demand shortfall.
| Company and metric | Latest reading | Period and source date |
|---|---|---|
| Meta, capital expenditures guidance | $135B–$145B | FY2026 guide, Oct. 21, 2026 |
| Amazon, cash capital expenditures | ~$220B | FY2026 guide, Oct. 29, 2026 |
| Amazon, AWS net sales | $46.0B, up 39% | Q3 2026 |
| Microsoft, Azure revenue growth | 48% | Quarter ended Sept. 30, 2026 |
| Microsoft, capital expenditures incl. finance leases | $45.2B | Quarter ended Sept. 30, 2026 |
| NVIDIA, data center revenue | $89.0B, up 117% | Q2 FY2027 |
| CoreWeave, expanded Meta capacity agreement | ~$21B through Dec. 2032 | Announced Apr. 9, 2026 |
| SK hynix and NVIDIA, AI factory initiative | $500B+ | Announced Jul. 25, 2026 |
| Meta, contracted cloud-capacity obligations | $14.72B | As of Sept. 30, 2026 |
The constraint has moved to the physical layer, and to the people who operate it. Meta's power arrangements are already contracted: 1,121 megawatts of Clinton Clean Energy Center output under a 20-year purchase agreement with Constellation Energy beginning in June 2027, plus energy service agreements and colocations inside the $355.73 billion of uncommenced leases. For a company whose committed 2027 obligations are $121.41 billion, power, not silicon, sets the schedule that matters.
Talent
Fleet-Operating Talent Is the Scarce Input, and It Just Moved
Amazon's own succession shows how thin the bench has become. When Brown left after nearly 19 years, Amazon handed AWS compute and machine-learning services to Dave Treadwell, an executive moved over from e-commerce foundation, effective Aug. 1, 2026. For the company that built EC2 from scratch, promoting an executive from outside the compute organization shows there was no ready internal successor for the job. That is what scarce compute-operating talent looks like from the inside, and it is why a hyperscaler losing one to a customer-turned-competitor is a bigger signal than a personnel change.
The irony is structural rather than personal. Meta still buys external AI capacity it does not build, carrying $14.72 billion of contingent cloud-capacity purchase obligations over five years, while simultaneously standing up Meta Compute to sell AI compute to enterprises. A chief technology officer for Meta Compute who ran the same business at Amazon Web Services makes Meta both AWS's customer and its competitor — and hands the company a cost curve it can arbitrage on its own cloud bill. The procurement benefit is real; so is the incentive to keep capacity in-house that would be cheaper rented.
Who wins
Suppliers Get Paid on Delivery; Meta Gets Paid on Utilization
| Entity | Role in Meta's build | What changes for its economics |
|---|---|---|
| NVIDIA | Multiyear deal for millions of Blackwell and Rubin GPUs, first large-scale Grace-only deployment, Spectrum-X networking | Paid on delivery, not on Meta's returns; risk shifts to deployment timing |
| SK hynix | HBM4 supply inside the SK Group–NVIDIA initiative; first SK AI factory due 2027 | Memory pricing holds firm through the 2027 build peak |
| Constellation Energy | 20-year purchase agreement for 1,121 MW of Clinton nuclear output from June 2027 | Contracted output converts into revenue as Meta's leases commence |
| CoreWeave | ~$21B expansion of Meta's AI cloud capacity agreement through December 2032 | Book is contracted, but Meta Compute competes for the same enterprises |
| Amazon | Meta's cloud supplier and the incumbent Meta Compute targets | Keeps $14.72B of Meta capacity demand; loses its compute chief |
| Microsoft | Azure sets the price umbrella Meta Compute must undercut | 48% Azure growth leaves room to price rather than chase share |
| Meta shareholders | Owners of $454.52B of non-cancellable commitments | Carry the utilization risk with no ability to shrink the program |
Follow the money and the risk splits cleanly. NVIDIA, SK hynix, Constellation Energy and CoreWeave are paid for delivering racks, memory, megawatts and contracted capacity, and none of their revenue depends on Meta earning a return on those racks. Meta shareholders own the utilization risk on $454.52 billion of commitments that cannot be cancelled, while Amazon and Microsoft own the pricing risk if Meta Compute sells capacity at thin margins. Suppliers get paid on delivery, not on utilization, which is why this handover matters far more to Meta's equity than to the AI supply chain.
Horizons
What to Watch, and What Would Break This Thesis
- Next two quarters: the 2027 capital budget. Street estimates cluster near $193 billion, and the guide lands in the same quarter the infrastructure transition ends, with Q4 2026 results due in January 2027
- Next two quarters: credit. S&P's negative outlook resolves against 2026–2027 cash burn, and Meta's $80.92 billion of long-term debt will be added to or refinanced
- 1–3 years: utilization. Meta Compute either converts installed capacity into external revenue or the $454.52 billion earns depreciation instead of profit
- 1–3 years: power and depreciation. Clinton's 1,121 megawatts starts in June 2027 while leases commence through 2036, and Q3 2026 capital spending ran 4.7 times depreciation
- What would break the thesis: a 2027 guide below the roughly $193 billion street baseline, which tells you the build is being trimmed rather than executed
Two risks run the other way, and both deserve naming. If Brown's arrival is followed by a firmer 2027 guide at or above consensus, the market will read the handover as a strengthening rather than a stumble, and the passing-of-the-baton framing becomes the whole story. And if Meta Compute's external contracts scale faster than expected, capacity that looks like a cash drain becomes a revenue line. Either outcome would break the execution-risk thesis built here, which is why the January guide carries more information than the memo did.
Meta has already been marked down for spending, not for management turnover. The stock closed Nov. 11 at about $649, roughly 17% below its 52-week high of $779.82, after a 5.7% intraday slide on Oct. 22 when the 2026 capital range was raised to $135–145 billion. The market has already absorbed the spending shock, which is precisely why the succession risk sitting underneath it remains, for now, unpriced.
The investable takeaways
- $121.41 billion of 2027 contractual obligations equal 84–90% of the full-year 2026 capital expenditure guide, leaving utilization, not spending restraint, as the swing factor for the equity
- Free cash flow turned negative at -$2.12 billion in Q3 2026, and S&P now forecasts a $71 billion free-operating-cash-flow deficit for 2027 — the year the infrastructure handover completes
- Near term the January 2027 capital guide is the binary catalyst; over 1–3 years Meta Compute must turn $454.52 billion of commitments into external revenue
- AWS net sales rose 39% to $46.0 billion in Q3 2026 with $17.8 billion of operating income, so Amazon's problem is talent, not demand, after losing its compute chief to a customer-turned-competitor
- Meta still carries $14.72 billion of contingent cloud-capacity purchase obligations, a relationship now managed by the man who ran Amazon's compute business
- Over 1–3 years Meta Compute selling excess AI compute presses on inference pricing, even as roughly $220 billion of 2026 capital spending keeps capacity scarce
- Meta's multiyear deal calls for millions of Blackwell and Rubin GPUs and the first large-scale Grace-only deployment, while NVIDIA data center revenue of $89.0 billion in Q2 FY2027 lifted the line 117% year over year
- Near term Meta's $135–145 billion 2026 guide and $454.52 billion of commitments keep accelerator orders contracted rather than optional
- Over 1–3 years the binding constraint on Meta's GPU deployments shifts to power and shells, with $355.73 billion of leases commencing through 2036
- The long-term HBM4 supply agreement inside SK Group's $500 billion-plus NVIDIA initiative ties Meta's accelerator volumes to memory supply, keeping HBM4 pricing firm through the 2027 build peak
- The first SK AI factory only comes online in 2027, the same year $121.41 billion of Meta commitments come due
- Near term Meta's 2026 capital spending keeps memory orders contracted; memory is paid on delivery, not on utilization
- Meta's 20-year purchase agreement takes 1,121 megawatts of Clinton Clean Energy Center output from June 2027, converting Meta's power demand into two decades of contracted output
- Energy service agreements sit inside Meta's $454.52 billion of commitments, so power contracts now set the schedule for data center capacity
- Over 1–3 years Meta's next sourcing decisions, with leases commencing through 2036, are the driver of incremental contracted megawatts
- Azure grew 48% in the September 2026 quarter while Microsoft spent $45.2 billion including finance leases, so the incumbent Meta Compute must beat has both scale and demand
- Meta remains a potential cloud buyer with $14.72 billion of purchase obligations, so Meta Compute competes with Azure only where Meta's own capacity wins
- Over 1–3 years Meta's $454.52 billion of commitments could turn into Azure demand or into captive capacity that never reaches the market
- Meta's roughly $21 billion expansion, on top of a $14.2 billion agreement signed in September 2025, puts about $35 billion of Meta capacity with CoreWeave through December 2032, a contracted revenue base that survives Meta's own build
- The same customer is now a competitor, with Meta Compute selling AI compute to the enterprises CoreWeave targets
- Near term the contract is intact; over 1–3 years Meta's captive capacity is the cannibalization risk to CoreWeave's book
