文 | 半导体产业纵横
10月28日，谷歌母公司Alphabet公布三季度财报：经营活动带来约423亿美元现金，购置设备和物业却支出约514亿美元，自由现金流为负约91亿美元。搜索、广告和云业务仍在增长，但当季经营产生的现金，已经不足以覆盖资本投入。
对CoreWeave、Nebius这类专门采购GPU、向外出租算力的新云服务商，资金更直接地决定扩张速度。10月22日，黄仁勋谈及这一群体时表示，算力需求极其旺盛，服务能力也很强，缺少的是营运资金；AI多个领域同时扩大投入，资金需求已超过可获得的资本。
三天后，马斯克给出了更悲观的判断。他表示，过去几周，债务市场几乎不再向缺乏AI数据中心以外大规模收入的公司提供资金，并预言两百多家新云将经历大规模倒闭与整合，只有少数能够生存下来。黄仁勋相信资金问题能够得到解决，马斯克却认为，不少企业可能等不到那一天。
一边是持续增长的算力需求，另一边是越来越受关注的投入与偿债压力，芯片的大买主们，真的快没钱了吗？
订单涨得快，现金却没有跟上
过去几年，CoreWeave、Lambda等新云通过GPU抵押贷款扩大集群，资金提供方也从私募信贷机构扩展到大型银行。运营商以设备和客户合同作为融资依据，提前取得采购资金，再用算力租金偿债。这种安排支持了快速扩张，也使下一轮采购越来越取决于贷款人如何评估设备价值和客户信用。
10月初，英伟达推动的算力融资项目中，部分银行和资管机构要求比原先方案更强的保障。贷款人关心的不只是GPU今天能否租出去，还包括几年后能赚多少钱、客户付款能否持续，以及违约后设备还能卖多少钱。
双方对GPU价值的判断已有分歧。Impax资产管理公司的高级投资组合经理Tony Trzcinka指出，银行通常按3至4年的折旧安排评估GPU，而英伟达认为高端GPU可以创造长达十年的收入。设备还能使用多久，与它能否持续产生足够收入，并不是同一件事。对缺少其他业务收入的新云，抵押物估值更保守、担保要求更高，都可能意味着需要拿出更多自有资金，或接受更贵的贷款。
CoreWeave在债券市场也遇到了压力。公司6月发行的一批债券，发行时收益率约为9.7％，到10月中旬，随着交易价格下跌，市场收益率已升至约15％。投资者要求更高回报来承担其信用风险，公司后续发债或再融资，也可能需要接受更高的融资成本。
其二季度账目显示，经营活动带来的现金净流入约6.79亿美元，购置设备和物业的现金支出却达到64.22亿美元。同一季度，净利息费用约6．4亿美元，相当于营收的四分之一。公司既要为客户预订的算力购买设备，也要承担已有借款的利息，持续扩张不能只靠当季经营所得。
10月20日，CoreWeave回应债务疑虑时，强调多数债务有长期客户合同支撑。公司先取得信用较强客户的采购承诺，再据此融资、建设基础设施；这些合同的约定收入足以偿还相关本息，不必依靠客户续约。其二季度末待履行合同对应收入达到1040亿美元，三季度初又取得250亿美元新增承诺。
已有合同为相关债务提供了还款来源，但新增采购仍需继续筹资。如果贷款人要求更高回报、更多担保，运营商即使拿到新订单，也要重新计算扩大集群是否划算。新云面临的压力，由此从投入规模大，进一步延伸到每一轮融资的成本和条件。
这些采购计划能否持续，也进入了台积电的产能评估。魏哲家在10月15日法说会上表示，公司会考虑云厂商资本开支、财务状况等因素，并与终端客户交流。晶圆代工厂的直接客户是芯片设计商，但先进制程和封装产能需要提前安排，设计商提出的需求，最终仍要由终端买家的投入支撑。核查这些买家的财务状况，有助于台积电判断需求能够持续多久。
芯片厂商也开始参与筹钱
英伟达已经在一些项目中提供了支持。10月28日披露的交易信息显示，印尼算力平台Zankore正在为31亿美元GPU贷款引入更多参贷银行。它与英伟达签订的六年协议，包含收入分成和信用支持，涉及首批100MW设施中尚未售出的算力容量。相关融资在9月已经公布，近期披露使供应商参与其中的方式更加具体。
按英伟达此前介绍的一般模式，公司会为部分算力容量提供最低收入承诺，让金融机构有依据支持项目建设。新云仍可将算力出售给第三方，满足条件后，英伟达参与收入分成。运营商获得了自身信用之外的支持，英伟达则在硬件销售之外，有机会分享后续租赁收益，同时承担相应的支持责任。
需要筹措设备资金的还有模型公司。在定制芯片一端，博通已经参与到Anthropic的资金安排中。10月26日，博通与金融伙伴宣布一项630亿美元融资交易，由新设的项目公司购买其定制AI加速器及网络设备，再出租给Anthropic，预计2027年开始交付。银行和黑石等机构提供前期采购资金，模型公司通过租用取得设备，博通按计划组织供货。
为了使这项安排成立，博通提供了设备残值支持。10月27日提交的监管文件显示，全部机架部署后，其最高潜在责任约430亿美元，按未折现口径计算。这项责任随机架交付而增加，随Anthropic支付租金而减少；如果客户违约，博通可能承担约定未偿款项与设备处置价值之间的差额。
430亿美元是潜在责任上限，而非博通已经支付的现金。它所对应的变化却很具体：供应商完成设备销售之后，仍需关注客户能否支付后续租金。金融机构提供资金，客户分期承担设备成本，博通则以部分风险承担换取采购推进，芯片供应与融资安排由此联系得更紧。
博通支持的是设备价值，英伟达在上述模式中支持的是部分算力收入。两种做法都让贷款人多了一层保障，也让一些难以独立承担巨额投入的客户有机会扩大采购。对供应商而言，争取这类订单，除了产品性能、交付和价格，还涉及愿意为客户提供多少支持。
但贷款人仍有自己的判断。10月初，部分参与英伟达相关融资项目的银行和资产管理机构，要求比最初方案更强的保障。供应商参与并没有让风险消失，而是改变了风险由谁承担。交易能否继续扩大，要看金融机构是否接受这些条件，以及芯片厂商愿意为新增销售保留多少后续责任。
资金仍在进入，只是条件各不相同
马斯克看衰新云融资的第二天，SpaceX便宣布获得约400亿美元GPU融资租赁承诺。由Apollo管理的实体购买GPU及相关设备，再出租给SpaceX，预计2027年上半年随设备交付分批起租。融资仍需满足约定的交割条件，但已承诺的规模表明，大型GPU采购仍能吸引金融机构参与。
马斯克解释SpaceX的融资优势时，强调了投资级信用，以及星链和发射业务收入。它与主要依靠GPU租金偿债的新云，能够向贷款人展示的资金来源并不相同。看衰部分运营商的融资前景，同时为自身扩大采购筹资，并不矛盾；差别在于谁有条件承担持续投入。
成熟业务提供的现金，也让大云厂商保有更大余地。微软10月28日披露，截至9月底的季度经营现金净流入为580亿美元，购置设备和物业的现金支出为394亿美元，自由现金流仍有186亿美元。公司预计2026自然年资本开支约1700亿美元，并预期2027财年自由现金流保持为正，同时表示会在条件合适时使用额外融资。
此前提到的Alphabet，则在同一季度通过普通股和高级无担保票据净筹资约411亿美元。经营现金流没有覆盖当期资本投入，并没有阻止它继续从资本市场取得资金。对这类买家，筹钱是安排巨额投入的一部分，不能仅凭融资动作判断其采购能力正在衰退。
新云并非都能调用这样的业务收入，但其中一些企业已经凭大客户合同取得支持。10月22日，Nebius与花旗签下35亿美元融资额度，用于数据中心建设、GPU采购等。公司表示，手中超过400亿美元、来自投资级客户的长期合同，可以支持后续资产融资，相关客户包括微软和Meta。
这也呈现了大云厂商与新云之间的另一层关系。微软、Meta既自行投入基础设施，也向外部运营商购买算力。新云替客户承担硬件采购和运营，再用这些客户的多年付款承诺争取融资。贷款人看中的，不仅是新云自身的规模，还包括租约另一端是谁、合同收入是否可靠。
Nebius此次融资期限为364天，将与客户预付款、运营现金流搭配使用，支持近期扩张。2027年及以后的长期资金仍需继续安排。大客户合同没有一次解决所有资金需求，却为分阶段采购提供了条件，也使其筹资能力与缺少稳定租户的运营商有所区别。
10月27日，HIVE旗下BUZZ Alpha Cloud也公布了约1.29亿美元、固定年利率9％的三年贷款，用于采购GB200集群，计划以Cohere三年合同的收入偿还。贷款约覆盖购置成本八成，余下出资来自此前可转债融资所得，设备预计年底前交付。
有客户、有匹配期限的合同，运营商同时承担部分出资，这笔采购因而有了具体资金安排。9％的固定利率也提醒市场，筹得到钱与资金便宜是两回事。新云需要比较的不只是GPU租金和设备价格，还包括借款成本，以及这些成本如何影响下一批采购能够承受的规模。
结语
芯片的大买主们并没有一起陷入资金枯竭，但持续扩大采购，正在考验各自的筹资能力。成熟业务收入、长期客户合同，以及供应商提供的支持，决定了不同买家还能以什么条件继续投入。对芯片产业而言，需求有多大之外，还要看谁能按计划付款、完成采购，以及这种投入能够持续多久。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The semiconductor industry.
On 28 October, Alphabet, the Google parent company, published a three-quarter financial paper: operations brought about $42.3 billion in cash, while equipment and property were purchased at about $51.4 billion and free cash flows at around $9.1 billion. Search, advertising and cloud operations continue to grow, but the cash generated during the season is no longer sufficient to cover capital inputs.
For new cloud service providers, such as CoreWeave and Nebius, which specialize in the procurement of GPUs, and renting out computing power, the financial resources more directly determine the pace of expansion. On 22 October, Hoang In-hoon, referring to this group, stated that there was an extremely high level of computing needs and capacity to provide services, and that what was lacking was operating funds; the multiple areas of AI were simultaneously expanding their inputs and the demand for funds exceeded the available capital.
Three days later, Mask made a more pessimistic judgment. He stated that in the past few weeks the debt market had almost stopped funding companies that lacked large revenues outside the AI data centre, and predicted that more than 200 new clouds would experience mass collapse and integration, with only a few surviving. While Hoang In-hoon believed that the funding problem could be resolved, Mask believed that many enterprises might not be able to wait for that day.
On the one hand, there is a constant demand for growth, on the other hand, there is a growing focus on investment and debt-servicing pressures. Are the big buyers of chips really running out of money?
The order went up fast and the cash didn't keep up.
In the past few years, new clouds, such as CoreWeave and Lambda, have expanded their clusters through GPU mortgage loans and finance providers have expanded from private lending institutions to large banks. Operators rely on equipment and customer contracts as a basis for financing, obtaining purchase funds in advance and then servicing their debts with computing rents. This arrangement supports rapid expansion and makes the next round of procurement increasingly dependent on how lenders assess the value of equipment and customer credit.
At the beginning of October, some banks and regulators required stronger guarantees than had been the case under the original programme in the Calculus Financing Project promoted by Weida. The lenders are concerned not only with whether the GPU can be rented today, but also with how much will be earned in a few years, how long customer payments will last, and how much will be sold after default.
The judgement of GPU values has been divided. Tony Trzcinka, Senior Portfolio Manager, Impax Asset Management, stated that banks usually assessed GPU under a depreciation arrangement of three to four years, while Inverda believed that a high-end GPU could generate income for up to 10 years. It is not the same thing as whether the equipment will continue to be used for long enough to generate sufficient revenue. The absence of new clouds of income from other operations, more conservative mortgage valuations and higher collateral requirements may mean more self-financing or more expensive loans.
CoreWeave was also under pressure in the bond market. A group of bonds issued by the company in June had a yield of about 9.7 per cent, and by mid-October the market rate had risen to about 15 per cent as transaction prices fell. Investors require higher returns to cover their credit risk, and companies may also need to accept higher financing costs.
Its second quarter accounts showed a net cash flow from operating activities of approximately $679 million, while cash expenditure on the acquisition of equipment and property amounted to $6,422 million. During the same quarter, the net interest cost was approximately $640 million, equivalent to a quarter of the amount collected. Companies have to buy equipment for their customers' calculus and bear interest on existing loans, and continued expansion cannot be based solely on current seasons of business.
On 20 October, CoreWeave responded to the debt doubt, emphasizing that most of the debt was supported by long-term customer contracts. The company obtains a commitment to purchase from its more creditworthy clients and then finances and builds infrastructure accordingly; the contractual revenues from these contracts are sufficient to cover the underlying interest and do not have to rely on the customer for renewal. Corresponding income from contracts to be completed at the end of the second quarter amounted to $104 billion, with an additional $25 billion in commitments at the beginning of the third quarter.
Contracts already exist to provide a source of repayment for the related obligations, but additional procurement still requires continued funding. If lenders require higher returns and more guarantees, operators recalculate the cost-effectiveness of expanding clusters, even if they receive new orders. The pressure on the new cloud further extends from the size of the input to the costs and conditions of each round of financing.
The sustainability of these procurement plans has also been factored into the capacity assessment of the build-up. At the meeting on 15 October, the Wei Cheolers indicated that companies would consider the capital expenditure of cloud manufacturers, the financial situation, etc., and interact with end-users. The direct client of the Crystal Circle-generation plant is the chip designer, but advanced production and containment capacity need to be prearranged, and the designer's demand will ultimately have to be supported by input from the end buyer. The verification of the financial position of these buyers will help to determine how long the demand for power is sustainable.
The chip manufacturers are also starting to be involved in raising money.
Weeda has supported several projects. The disclosure of transaction information on 28 October indicates that Zankore, Indonesia's Calculator Platform, is introducing more lending banks for $3.1 billion of GPU loans. It had a six-year agreement with Weida, which included a share of income and credit support, covering the unsold computing capacity of the first 100 MW facilities. Related financing had been made public in September, and recent disclosures had made the manner in which suppliers were involved more specific.
Following the general model described earlier by Weida, companies provide a minimum income commitment for part of their capacity to provide financial institutions with evidence to support project construction. New Clouds can still sell their calculations to third parties, and when the conditions are met, Young Weidar participates in the revenue-sharing. Operators are supported beyond their own creditworthiness, while British Wida is given the opportunity to share the proceeds of subsequent leases in addition to hardware sales, with corresponding support responsibilities.
Model companies are also required to raise funds for equipment. At the end of the customized chip, Chase is already involved in the funding arrangements for Anthropic. On 26 October, Chase announced with its financial partners a $63 billion financing deal for the purchase by the new project company of its customized AI accelerator and network equipment, to be leased to Anthropic and expected to start delivery in 2027. Banks and institutions such as Blackstone provide funds for prior purchases, and model companies obtain equipment by lease and access supplies according to schedule.
In order to put this arrangement in place, Chase provided residual value support for the equipment. The regulatory documents submitted on 27 October indicated that the maximum potential liability for all aircraft hangers, after deployment, would be approximately $43 billion, based on an open calibre. This increased liability as a result of random deliveries, reduced by the payment of rent by Anthropic; if the customer defaults, Chase may bear the difference between the agreed outstanding amount and the disposal value of the equipment.
The $43 billion is the maximum amount of potential liability, not the cash already paid by Chase. The change is specific: after the vendor has completed the sale of the equipment, it will still need to look at whether the customer can pay the subsequent rent. The availability of chips and financing arrangements are further linked by funding from financial institutions with clients bearing the cost of the equipment in instalments, while the Chases offer partial risk-taking in exchange for procurement push.
It supports the value of equipment, and in the above-mentioned model, it supports partial computing income. Both approaches provide an additional layer of security for lenders and an opportunity to expand procurement for clients who have difficulty bearing large amounts of inputs on their own. For suppliers, the acquisition of such orders involves, in addition to product performance, delivery and price, the degree of support they are willing to provide to clients.
But lenders still have their own judgment. In early October, banking and asset management agencies involved in some of the financing projects in Weida required stronger safeguards than the original programme. The involvement of the supplier did not eliminate the risk, but changed who took the risk. The continued expansion of the transaction depends on the acceptance of these conditions by financial institutions and the willingness of the chip manufacturer to reserve a subsequent responsibility for the additional sales.
Funding is still in, but conditions are different.
On the second day of Mask ' s look at the new cloud financing, SpaceX announced the acquisition of approximately $40 billion of the GPU financing lease commitments. The entities managed by Apollo purchase GPU and associated equipment and lease it to SpaceX, which is expected to be delivered in batches in the first half of 2027. Financing still needs to meet the agreed terms of delivery, but the scale of commitments suggests that large-scale GPU purchases can still attract financial institutions.
In explaining the financing advantages of SpaceX, Mask emphasized investment-grade credit, as well as revenues from the chain and launch operations. It is not the same as the new cloud, which relies mainly on GPU rents to service debts, and the source of funding that can be shown to lenders. It was not contradictory to look at the financing prospects of some of the failing operators, while financing their own expanded procurement; the difference was who was in a position to bear the continuous input.
Cash from mature operations also leaves greater scope for cloud manufacturers. Microsoft disclosed a net cash flow of $58 billion in operating cash by the end of September, $39.4 billion in cash expenditure on the acquisition of equipment and property and $18.6 billion in free cash flows. The company projected capital expenditure of about $170 billion for the natural year 2026 and expected a positive free cash flow for the financial year 2027, while indicating that additional financing would be used when conditions were right.
Alphabet, which was previously mentioned, received some $41.1 billion in net financing during the same quarter through common shares and senior unsecured instruments. Operating cash flows did not cover current capital inputs and did not prevent it from continuing to access capital markets. For such buyers, the raising of money is part of the arrangement of huge inputs, and their procurement capacity cannot be judged to be declining solely by financing actions.
New clouds may not be able to mobilize such business income, but some of them have already been supported by large client contracts. On 22 October, Nebius and Citibank signed $3.5 billion in financing quota for data centre construction, GPU procurement, etc. The company indicated that long-term contracts of over $40 billion from investment-level clients could support subsequent asset financing, including Microsoft and Meta.
It also presents another layer of relationship between a cloud manufacturer and a new cloud. Microsoft and Meta invest their own infrastructure and buy power from external operators. New Clouds carries out hardware purchases and operations for its clients, who in turn use their multi-year payment commitments for financing. The lenders see not only the size of the new cloud itself, but also who is on the other side of the lease and whether the revenue from the contract is reliable.
The Nebius financing period is 364 days and will be used in conjunction with client advances and operating cash flows to support the recent expansion. Long-term funding for 2027 and beyond will need to be continued. Large client contracts did not address all funding needs at one time, but provided conditions for phased procurement and differentiated their fund-raising capacity from operators lacking stable tenants.
On 27 October, a three-year loan of about $129 million at a fixed annual interest rate of 9 per cent was also announced by BUZZ Alpha Cloud under the HIV/AIDS flag for the purchase of the GB200 cluster, which is planned to be repaid from the three-year contract income of Cohere. The loan covers approximately 80 per cent of the acquisition cost, with the remaining funding coming from prior reversible financing, and the equipment is expected to be delivered by the end of the year.
Specific funding arrangements were made for the purchase of a customer, a contract with a matching duration and a partial contribution from the operator. The 9 per cent fixed interest rate also serves as a reminder to the market that raising money is not the same as making it cheaper. The new cloud needs to be compared not only with the GPU rent and equipment prices, but also with the borrowing costs and how they affect the scale of the next purchase.
Concluding remarks
Large buyers of chips are not all caught up in financial depletion, but are testing their capacity to raise funds through continued procurement expansion. Promising business income, long-term customer contracts and support from suppliers determine the conditions under which different buyers can continue to invest. In addition to the size of the demand for the chip industry, it depends on who is able to make the planned payments, complete the procurement and how long this input will last.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.