Are the major chip buyers running out of money soon?

On October 28, Alphabet, Google's parent company, released its Q3 financial report: operating activities generated approximately 42.3 billion USD in cash, while the expenditure on equipment and property acquisition reached around 51.4 billion USD, resulting in a negative free cash flow of roughly 9.1 billion USD. Search, advertising and cloud businesses are still growing, but the cash generated from operations in the quarter is no longer sufficient to cover capital expenditures.

For new cloud service providers that specialize in purchasing GPUs and renting out computing power, such as CoreWeave and Nebius, capital directly determines their expansion speed. On October 22, Jensen Huang mentioned this group and stated that the demand for computing power is extremely strong, their service capabilities are also robust, and what they lack is working capital; multiple segments of the AI industry are ramping up investment simultaneously, and the capital demand has exceeded the available capital.

Three days later, Elon Musk put forward a more pessimistic judgment. He said that in the past few weeks, the debt market has hardly provided funding for companies that lack large-scale revenue other than from AI data centers, and predicted that more than 200 new cloud service providers will experience large-scale closures and consolidation, with only a few able to survive. Jensen Huang believes the capital issue can be resolved, while Elon Musk argues that many enterprises may not live to see that day.

With continuously growing computing power demand on one side, and the mounting pressure of investment and debt repayment drawing increasing attention on the other, are the major bulk buyers of chips really running out of money?

 01 

Orders are surging, but cash flow fails to keep up 

Over the past few years, new cloud players including CoreWeave and Lambda have expanded their clusters through GPU mortgage loans, with capital providers extending from private credit institutions to large banks. Operators use equipment and customer contracts as the basis for financing, obtain procurement funds in advance, and then repay debts with computing power rental income. This arrangement has supported rapid expansion, but also made the next round of procurement increasingly dependent on how lenders assess equipment value and customer credit.

In early October, in the computing power financing projects promoted by NVIDIA, some banks and asset management institutions require stronger guarantees than the original plan. Lenders are not only concerned about whether GPUs can be rented out today, but also about how much profit can be made in a few years, whether customer payments can be sustained, and how much the equipment can be sold for after default.

The two sides already have divergences in judging the value of GPUs. Tony Trzcinka, senior portfolio manager at Impax Asset Management, pointed out that banks usually evaluate GPUs based on a 3 to 4-year depreciation schedule, while NVIDIA believes high-end GPUs can generate revenue for up to ten years. How long the equipment can still be used is not the same as whether it can continuously generate sufficient revenue. For new cloud players that lack other business revenue, more conservative collateral valuation and higher guarantee requirements may mean that they need to contribute more of their own funds, or accept more expensive loans.

CoreWeave has also encountered pressure in the bond market. A batch of bonds issued by the company in June had a yield of about 9.7% at issuance, and by mid-October, as the trading price fell, the market yield had risen to about 15%. Investors demand higher returns to take on its credit risk, and the company may need to accept higher financing costs for subsequent bond issuance or refinancing.

Its second-quarter accounts show that the net cash inflow from operating activities was about 679 million USD, while the cash expenditure on equipment and property acquisition reached 6.422 billion USD. In the same quarter, net interest expenses were about 640 million USD, equivalent to a quarter of revenue. The company not only needs to purchase equipment for the computing power reserved by customers, but also bear the interest of existing borrowings, and continuous expansion cannot only rely on the income from operations in the current quarter.

On October 20, when CoreWeave responded to debt concerns, it emphasized that most of its debts are supported by long-term customer contracts. The company first obtains procurement commitments from customers with strong credit, and then conducts financing and builds infrastructure accordingly; the agreed revenue of these contracts is sufficient to repay the relevant principal and interest, without relying on customer renewal. The revenue corresponding to the unfulfilled contracts at the end of the second quarter reached 104 billion USD, and 25 billion USD of new commitments were obtained in early Q3.

Existing contracts provide repayment sources for relevant debts, but new procurement still requires continuous fundraising. If lenders demand higher returns and more guarantees, even if operators get new orders, they have to recalculate whether it is cost-effective to expand the cluster. The pressure faced by new cloud players has thus extended from the large investment scale to the cost and conditions of each round of financing.

The sustainability of these procurement plans has also been included in TSMC's capacity assessment. C.C. Wei said at the earnings call on October 15 that the company will take into account factors such as cloud vendors' capital expenditure and financial status, and communicate with end customers. The direct customers of wafer foundries are chip designers, but advanced process and packaging capacity need to be arranged in advance, and the demands put forward by designers ultimately need to be supported by the investment of end buyers. Verifying the financial status of these buyers helps TSMC judge how long the demand can last.

 02 

Chip vendors have also begun to participate in fundraising 

NVIDIA has provided support in some projects. Transaction information disclosed on October 28 shows that Indonesian computing power platform Zankore is introducing more participating banks for its 3.1 billion USD GPU loan. The six-year agreement it signed with NVIDIA includes revenue sharing and credit support, covering the unsold computing power capacity in the first batch of 100MW facilities. The relevant financing was announced in September, and the recent disclosure has made the specific way for suppliers to participate more concrete.

According to the general model previously introduced by NVIDIA, the company will provide a minimum revenue commitment for part of the computing power capacity, so that financial institutions have a basis to support project construction. New cloud players can still sell computing power to third parties, and NVIDIA will participate in revenue sharing after meeting the conditions. Operators get support beyond their own credit, while NVIDIA gets the opportunity to share subsequent rental income in addition to hardware sales, and assume corresponding support responsibilities at the same time.

Model companies also need to raise funds for equipment. At the custom chip end, Broadcom has participated in Anthropic's capital arrangement. On October 26, Broadcom and its financial partners announced a 63 billion USD financing transaction, in which a newly established project company will purchase its custom AI accelerators and network equipment and then lease them to Anthropic, with delivery expected to start in 2027. Banks, Blackstone and other institutions provide upfront procurement funds, model companies obtain equipment through leasing, and Broadcom organizes supply according to the plan.

To make this arrangement work, Broadcom has provided equipment residual value support. Regulatory documents submitted on October 27 show that after all racks are deployed, its maximum potential liability is about 43 billion USD, calculated on an undiscounted basis. This liability increases with the delivery of racks and decreases as Anthropic pays rent; if the customer defaults, Broadcom may bear the difference between the agreed outstanding amount and the equipment disposal value.

43 billion USD is the upper limit of potential liability, not the cash that Broadcom has already paid. The corresponding changes are very specific: after the supplier completes the equipment sales, it still needs to pay attention to whether the customer can pay the subsequent rent. Financial institutions provide funds, customers bear the equipment cost in installments, and Broadcom exchanges partial risk bearing for procurement advancement, thus making chip supply and financing arrangements more closely linked.

Broadcom supports equipment value, while NVIDIA supports part of the computing power revenue in the above model. Both practices provide an extra layer of guarantee for lenders, and also give some customers who cannot independently afford huge investments the opportunity to expand procurement. For suppliers, to win such orders, in addition to product performance, delivery and price, it also involves how much support they are willing to provide for customers.

But lenders still have their own judgments. In early October, some banks and asset management institutions participating in NVIDIA-related financing projects required stronger guarantees than the initial plan. Supplier participation does not make risks disappear, but changes who bears the risks. Whether transactions can continue to expand depends on whether financial institutions accept these conditions, and how much subsequent liability chip manufacturers are willing to reserve for new sales.

 03 

Capital is still entering, but the conditions vary 

The day after Musk expressed pessimism about new cloud financing, SpaceX announced that it had obtained approximately 40 billion USD in GPU financial leasing commitments. An entity managed by Apollo will purchase GPUs and related equipment and then lease them to SpaceX, with leases expected to start in batches in the first half of 2027 following equipment delivery. The financing still needs to meet the agreed closing conditions, but the committed scale indicates that large-scale GPU procurement can still attract financial institutions to participate.

When Musk explained SpaceX's financing advantages, he emphasized investment-grade credit, as well as the revenue from Starlink and launch businesses. It has a different source of funds that can be demonstrated to lenders from new cloud players that mainly rely on GPU rental income to repay debts. Being pessimistic about the financing prospects of some operators while raising funds for its own expanded procurement is not contradictory; the difference lies in who has the conditions to sustain continuous investment.

The cash generated by mature businesses also gives large cloud vendors more room for maneuver. Microsoft disclosed on October 28 that its net cash inflow from operations in the quarter ending at the end of September was 58 billion USD, the cash expenditure on equipment and property acquisition was 39.4 billion USD, and the free cash flow still reached 18.6 billion USD. The company expects capital expenditure in the 2026 calendar year to be around 170 billion USD, and expects free cash flow to remain positive in fiscal year 2027, while stating that it will use additional financing when conditions are appropriate.

Alphabet, mentioned earlier, raised a net of approximately 41.1 billion USD through common stock and senior unsecured notes in the same quarter. Operating cash flow did not cover the current capital expenditure, but that did not stop it from continuing to raise funds from the capital market. For this type of buyer, fundraising is part of arranging huge investments, and the procurement capacity cannot be judged to be declining solely based on financing activities.

Not all new cloud players can call on such business revenue, but some of them have obtained support based on large customer contracts. On October 22, Nebius signed a 3.5 billion USD financing line with Citi for data center construction, GPU procurement and other purposes. The company said that its long-term contracts worth more than 40 billion USD from investment-grade customers, including Microsoft and Meta, can support subsequent asset financing.

This also presents another layer of relationship between large cloud vendors and new cloud players. Microsoft and Meta not only invest in infrastructure on their own, but also purchase computing power from external operators. New cloud players take on hardware procurement and operations for their customers, and then use the multi-year payment commitments of these customers to seek financing. What lenders value is not only the size of the new cloud player itself, but also who is on the other end of the lease and whether the contract revenue is reliable.

The term of Nebius's financing this time is 364 days, which will be used in combination with customer advance payments and operating cash flow to support recent expansion. Long-term funds for 2027 and beyond still need to be arranged continuously. Large customer contracts do not solve all capital needs at once, but provide conditions for phased procurement, and also make its fundraising capacity different from operators that lack stable tenants.

On October 27, BUZZ Alpha Cloud, under HIVE, also announced a three-year loan of approximately 129 million USD with a fixed annual interest rate of 9% for purchasing GB200 clusters, and plans to repay it with the revenue from a three-year contract with Cohere. The loan covers about 80% of the acquisition cost, and the remaining contribution comes from the proceeds of previous convertible bond financing, with the equipment expected to be delivered before the end of the year.

With customers and contracts matching the term, and the operator contributing part of the funds at the same time, this procurement has a specific capital arrangement. The 9% fixed interest rate also reminds the market that being able to raise money and having cheap capital are two different things. New cloud players need to compare not only GPU rental rates and equipment prices, but also borrowing costs, and how these costs affect the affordable scale of the next batch of procurement.

 04 

Conclusion 

The major bulk buyers of chips have not fallen into capital depletion together, but continuous expansion of procurement is testing the fundraising capacity of each party. Mature business revenue, long-term customer contracts, and support provided by suppliers determine under what conditions different buyers can continue to invest. For the chip industry, in addition to how large the demand is, it also depends on who can pay as planned and complete procurement, and how long this kind of investment can last.

This article is from the WeChat official account "Semiconductor Industry Review" (ID: ICViews), written by Jun Xi, and published with authorization from 36Kr.