Meta의 8,250억 달러 규모 "청구서"는 AI 컴퓨팅 비용이 얼마나 높아졌는지를 보여주는 사례로 거론되고 있으며, 한편으로는 자금 조달 격차와 글로벌 시장 간 컴퓨팅 비용 차이도 나타나고 있다.

In the third-quarter 10-Q filed with the U.S. Securities and Exchange Commission on October 22, Meta disclosed commitments across three categories (off-balance-sheet leases, non-cancelable purchase contracts, and cloud capacity purchases) totaling up to approximately $824.9 billion, with terms extending as far as 2036.

The same filing contains a second set of figures. In the third quarter, Meta generated $32.79 billion in operating cash flow against $34.91 billion in capital expenditures, leaving free cash flow at negative $2.12 billion. In the same period a year earlier, free cash flow was positive $10.63 billion. Operating cash flow no longer covers the year's capital spending.

The bill is public, and so is the shortfall. The commitments are Meta's, but the financing challenge belongs to the entire AI compute industry.

**How the $825 billion figure is built**

The total comprises three components.

The first is off-balance-sheet operating and finance leases, totaling $355.7 billion. These leases run from their commencement dates through 2036, with individual terms ranging from one to 30 years.

The second is non-cancelable short- and long-term contract commitments as of September 30, totaling $454.5 billion. Of this, $45.85 billion falls due in 2026 and $121.41 billion in 2027.

The third is cloud capacity purchases over the next five years, up to $14.7 billion. Meta notes that if cloud providers resell this capacity to third parties, its own spending could be reduced accordingly.

Added together, the three components reach a maximum of approximately $824.9 billion. The market has rounded this to an $825 billion ten-year bill, a label Meta does not use in the filing.

Three months earlier, Meta had disclosed similar commitments of roughly $700 billion, an increase of about $125 billion in a single quarter. By comparison, Meta's long-term debt on the balance sheet is about $80.9 billion. Off-balance-sheet obligations are growing much faster. According to Larry McDonald, founder of the market research firm The Bear Traps Report, Meta's off-balance-sheet financing has risen from about $35 billion to about $600 billion.

The filing includes two other arrangements worth noting. First, $10.8 billion in money market funds has been reclassified as restricted cash equivalents, to be released in tranches between 2028 and 2030. Second, Meta has created a new class of non-voting Class C shares.

Meta did not explain how either arrangement relates to its capital spending. In the same period, the company raised its full-year capital expenditure guidance to $135 billion to $145 billion. Once the bill became public, the debate quickly shifted from the size of the commitments to their cost.

**The bond market has already priced the bill**

That AI capital spending is enormous, and that it burns cash quickly, is now widely accepted across the technology sector. On October 21, Nvidia CEO Jensen Huang told CNBC on the sidelines of an event in Berlin that the missing piece in AI's development is the working capital needed to drive these companies' growth. He said capital demand far exceeds capital supply and that the problem will be solved.

Broadcom CEO Hock Tan offered a similar assessment, arguing that the biggest constraint right now is a shortage of funding rather than a shortage of chips.

In practice, the bond market had already set its price before Huang spoke. Last week, CoreWeave's stock and bonds sold off together. Its bonds maturing in 2031 at one point yielded 15% to 16%, compared with a 9.675% coupon when the notes were issued in June 2025. The implied default probability in its five-year credit default swaps exceeded 60%. The company's shares fell about 25% that week.

The pressure on the numbers comes from both sides. CoreWeave has roughly $35.5 billion in debt. Its second-quarter revenue was $2.575 billion, with net interest expense of $640 million, capital expenditures above $9 billion, and free cash flow of negative $5.7 billion. Its full-year capital expenditure guidance is $35 billion to $39 billion.

The same week, another compute-leasing company secured new financing. On October 22, Citi led a $3.5 billion senior secured loan to Nebius with a 364-day term, priced at Term SOFR plus 2.75%, with a ceiling of 3.50%, to fund data center construction and GPU purchases. That day, Nebius shares fell 7.3% to close at $158.89.

The cost of that financing is higher than what CoreWeave paid when it issued bonds two years earlier.

The private market presents a sharp contrast. In March, OpenAI completed a $122 billion funding round at a post-money valuation of $852 billion. In May, Anthropic completed a $65 billion Series H at a post-money valuation of $965 billion, surpassing OpenAI. Its annualized revenue passed $47 billion in early May.

Looking more closely, capital reaches these companies through three channels, each more expensive than the last. The first is equity, such as the private rounds raised by OpenAI and Anthropic. The second is public bonds, such as CoreWeave's 9.675% notes issued in June 2025. The third is debt secured by assets or contract cash flows, such as Nebius's $3.5 billion senior secured loan. Nebius also has room for asset-backed financing against more than $40 billion in long-term contracts.

The third channel is the most expensive, and it has been the fastest-growing this year.

This cost structure appears in a different form among Chinese companies.

**China's balance sheets tell their own stories**

Chinese technology giants are also spending heavily on AI.

Alibaba's capital expenditures for the June 2026 quarter were RMB 67.678 billion, up 75% from RMB 38.676 billion a year earlier. Operating cash flow in the same period was RMB 22.945 billion, and free cash flow was a net outflow of RMB 44.67 billion. Financing has run in parallel: on August 23, Alibaba completed an HK$80 billion share placement.

Tencent's capital expenditures for the second quarter of 2026 were RMB 52.784 billion, up 176% year over year. For the first half, the total was RMB 84.72 billion, up 82%. Free cash flow for the quarter was negative RMB 13.8 billion. However, the company's announcement also presented an alternative measure: excluding prepayments for computing power procurement, free cash flow was positive RMB 37.6 billion, a gap of RMB 51.4 billion between the two figures. For the quarter, operating cash flow was RMB 52.7 billion, and capital expenditure payments were RMB 59.3 billion.

Unlisted companies rely more heavily on debt. In early September, ByteDance finalized a $29.6 billion syndicated loan. The facility was initially sized at $20 billion but was expanded to $29.6 billion due to strong demand. Nearly 30 banks participated, and Chinese lenders accounted for more than 60% of the total. The Industrial and Commercial Bank of China contributed $3 billion, Bank of China $2.5 billion, China Construction Bank $1.5 billion, and HSBC and others $1.5 billion each. The loan has a three-year term, extendable by two years, and is unsecured. At the same time, ByteDance raised its capital expenditure plan to as much as $70 billion.

The same spending looks different depending on how it is measured. Tencent's decision to report two free cash flow figures in a single announcement illustrates the point.

**What remains unknown**

AI commitments are now recorded through 2036, but the cash flow statements that can verify them extend only to the third quarter of 2026.

It is not yet clear who will ultimately absorb these commitments. It could be cloud providers reselling capacity, new debt, or operating cash flow. Huang has said capital demand far exceeds supply and that the problem will be resolved, but the mechanism for resolution has not appeared in any disclosure document.

There are two observable markers. One is Meta's free cash flow in its fourth-quarter results. The other is how the $121.41 billion in contract commitments due in 2027 will be renewed or replaced. An earlier signal may come from the bond market, which, ahead of CoreWeave's earnings, had already priced the risk in its 2031 bonds.