NVIDIA 积累 5300 亿美元财务担保以垄断 AI 市场

毫无疑问,NVIDIA 在 AI 硬件市场的主导地位,不仅源于其 AI 加速器的性能与庞大的软件生态,也源于其激进的采购策略——这一策略实际上使竞争对手无法从台积电(TSMC)及其他代工方获得足够的产能。据 SemiAnalysis 称,这些交易的结果是,供应与产能承诺在单个季度内增加了近 2500 亿美元,主要来自存储采购。不过,根据《华尔街日报》和 SemiAnalysis 的报道,就连 NVIDIA 前首席执行官黄仁勋在离职前,显然也已开始质疑这一策略的可持续性。
过去几年间,NVIDIA 不仅拿下了 AI 加速器市场的绝大部分份额,还锁定了全球相当大比例的先进逻辑芯片、高带宽内存(HBM)以及构建 AI 加速器和系统所需各类其他元器件的产能,这限制了竞争对手提升处理器出货量的能力。为支撑这些交易,黄仁勋亲自与台湾和韩国的主要供应商谈判达成协议。
NVIDIA 采购承诺的庞大规模,也让该公司获得了比竞争对手更优惠的报价。SemiAnalysis 估计,2027 年 SK hynix 向 NVIDIA 收取的 HBM4 内存价格可能在每吉比特 3.00 至 3.30 美元之间,而 AMD、Amazon、Broadcom、Google 等其他一线客户则需支付每吉比特 3.70 至 4.10 美元。分析师认为,作为两家公司之间非正式安排的一部分,NVIDIA 会给予 SK 集团优先获取其 AI GPU 的权利。
尽管 NVIDIA 锁定庞大产能的能力是其取得成功的关键因素,但此类承诺也伴随着相当大的风险。据《华尔街日报》报道,在 8 月与高管举行的一次会议上,黄仁勋对 NVIDIA 承诺采购的巨量晶圆和内存表示担忧,因为主要云服务提供商正寻求减少对该公司硬件的依赖,转而采用其他 AI 加速器平台。
为了分散客户基础并为其产品创造额外需求,NVIDIA 向那些原本无力大规模采购其硬件的新兴云服务提供商和数据中心开发商提供了财务担保。据 SemiAnalysis 称,截至第二财季末,NVIDIA 的表外担保总额达 5300 亿美元,高于上一季度的 1840 亿美元,此外还持有价值 990 亿美元的股权投资。5000 亿美元这一数字看似极为庞大,以下是基于提交给美国证券交易委员会(SEC)的文件所做的拆解:
此类承诺的规模甚至在 NVIDIA 内部也引发了担忧。时任首席财务官 Colette Kress 质疑了挑选财务支持对象的标准以及此类安排的潜在回报。据报道,她还将 NVIDIA 对 OpenAI 位于俄亥俄州的大型数据中心项目的初始财务承诺削减了一半,此后两家公司于 8 月 17 日签署了协议。该设施预计将容纳价值约 3000 亿美元的 NVIDIA 处理器。
到了某个时点,黄仁勋本人也开始质疑:一旦他离任,Nvidia 这套激进的融资策略究竟还有多少价值。他实质上把现任临时 CEO Colette Kress 称为仅仅是一个「负责汇报的 CFO」,而非能够做出战略决策的高管。他还对巨额的晶圆和存储采购承诺表示担忧,因为此时主要云服务商正寻求减少对 Nvidia GPU 的依赖。
尽管这些承诺存疑,但如果没有黄仁勋本人以及他与主要半导体公司高管的私人关系,Nvidia 可能更难从供应商那里获得优待。SemiAnalysis 认为,到 2028 年,Nvidia 以外的客户可能占据台积电超过 70% 的产能。这虽然可能让 Nvidia 的竞争对手大幅提升其 AI 处理器的出货量,但也可能影响台湾和韩国供应商积极投资新增产能的意愿——毕竟,他们不再能得到黄仁勋过去提供的那种保证。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Nvidia amassed $530 billion in financial guarantees to corner the AI market)
Without a doubt, Nvidia's dominance in the AI hardware market stems not only from the capabilities of its AI accelerators and its vast software ecosystem, but also from its aggressive procurement strategy that effectively denies its rivals access to sufficient production capacity at TSMC and other contractors. The result of his deals is that supply and capacity commitments rose by almost a quarter of a trillion dollars in a single quarter, primarily due to memory procurement, according to SemiAnalysis. However, even Nvidia's former chief executive Jensen Huang had apparently begun to question the sustainability of this strategy before his departure, based on reports from The Wall Street Journal and SemiAnalysis.
Over the past several years, Nvidia has not only captured the vast majority of the AI accelerator market, but has also secured a substantial proportion of the world's production capacity for advanced logic, high-bandwidth memory (HBM), and various other components required to build AI accelerators and systems, which has limited the ability of its competitors to ramp up shipments of their processors. To support the deals, Jensen Huang personally negotiated agreements with major suppliers in Taiwan and South Korea.
The massive scale of Nvidia's procurement commitments also enabled the company to secure more favorable quotes than its rivals. SemiAnalysis estimates that SK hynix could charge Nvidia between $3.00 and $3.30 per gigabit of HBM4 memory in 2027, whereas other first-tier customers, such as AMD, Amazon, Broadcom, and Google, would pay between $3.70 and $4.10 per gigabit. The analysts believe that as part of an informal arrangement between the two companies, Nvidia gives SK Group preferential access to its AI GPUs.
While Nvidia's ability to secure vast production capacities has been instrumental in its success, such commitments also carry considerable risks. At a meeting with senior executives in August, Huang reportedly expressed concerns about the massive volumes of wafers and memory that Nvidia had committed to purchase as major cloud service providers sought to reduce their reliance on the company's hardware in favor of alternative AI accelerator platforms, according to The Wall Street Journal.
In a bid to diversify its customer base and create additional demand for its products, Nvidia offered financial guarantees to emerging cloud service providers and data center developers that otherwise could not afford to procure its hardware at scale. As of the end of its fiscal second quarter, Nvidia had $530 billion in gross off-balance-sheet guarantees, up from $184 billion a quarter earlier, as well as $99 billion worth of equity investments, according to SemiAnalysis. As half a trillion dollars seems gargantuan, here is a breakdown based on a filing with the Securities and Exchange Commission:
- $366 billion in future commitments (of which $279 billion was supply and capacity commitments, mainly for memory and manufacturing);
- $56 billion in additional commitments (including $36 billion in AI cloud agreements/take-or-pay arrangements and $20 billion in data-center leases expected to be reassigned);
- $108.5 billion in maximum gross exposure under guarantees (of which $105 billion was a lease/credit-support guarantee related to SB Energy’s PORTS Technology Campus in Ohio for an OpenAI affiliate, plus $3.5 billion for other AI-cloud partner land/power/shell guarantees).
The scale of such commitments raised concerns even within Nvidia. Then-CFO Colette Kress questioned the criteria used to select recipients of financial support and the potential returns on such arrangements. She also reportedly reduced Nvidia's initial financial commitment to OpenAI's massive data center project in Ohio by half before the two companies signed an agreement on August 17. The facility is expected to house some $300 billion worth of Nvidia processors.
Massive outcome?
At some point, Huang himself began to question the merits of Nvidia's aggressive financing strategy in the event of his departure, essentially calling the current interim CEO, Colette Kress, a mere 'reporting CFO' rather than an executive capable of making strategic decisions. He also expressed concerns about massive wafer and memory procurement commitments as major cloud service providers sought to reduce their reliance on Nvidia’s GPUs.
Despite questionable commitments, without Huang and his personal relationships with executives at major semiconductor companies, Nvidia may find it harder to secure preferential treatment from its suppliers. SemiAnalysis believes that customers other than Nvidia could account for over 70% of TSMC production capacity in 2028. While this could enable Nvidia’s competitors to significantly increase shipments of their AI processors, it could also affect the willingness of Taiwanese and South Korean suppliers to invest aggressively in additional production capacity without the kind of assurances that Huang used to provide.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 Tom's Hardware(Nvidia amassed $530 billion in financial guarantees to corner the AI market),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。