962亿美元。8月27日美股盘后，英伟达公布2027财年第二季度（截至7月26日）财报：营收962亿美元，同比增长106%、环比增长18%；数据中心业务单季890亿美元，同比增长117%——一年前，总营收还只有467亿。数字本身没有悬念，悬念在同一天晚些时候到来：《华尔街日报》报道称，这家公司上周悄悄暂停了一项面向AI云服务商的收入分成计划。史上最猛的财报，和一次最敏感的商业模式急刹车，发生在同一周。
成绩单：千亿门槛前的一个季度
先把硬数字过一遍：GAAP与非GAAP毛利率均为75.0%；GAAP净利润597亿美元，GAAP摊薄每股收益2.46美元（non-GAAP每股2.22美元）。管理层给出的下一季度营收指引是1080亿美元——按这个斜率，单季千亿美元营收的门槛，就站在下个季度门口。本季度公司还向股东返还了260亿美元，回购加分红的力度同样在刷新自己的纪录。算一笔结构账：数据中心已占总营收的九成以上，这家公司的命运曲线，和全球AI算力开支几乎是同一条线。
黄仁勋的两句话
财报声明里，黄仁勋留了两句值得抄录的话。第一句：“Compute is revenue”（算力即收入）——把算力采购直接等同于收入创造的叙事，既说给客户听，也说给华尔街听。第二句是对Blackwell需求的形容：“really extraordinary”（真的非同寻常）。产品面上的注脚更硬：下一代Vera Rubin平台已进入全面投产（full production），配套的Vera CPU被定位为“首款为AI agent打造的CPU”。注意这个措辞——芯片公司的产品定义已经开始围着agent转，AI agent正在从软件叙事变成硬件设计目标，这是比任何一句口号都诚实的行业信号。
B面：被按下暂停键的分成实验
然后是这条真正值得展开的新闻。《华尔街日报》8月27日援引知情人士报道称，英伟达已于上周暂停一项面向AI云服务商的融资安排：原方案是由英伟达向云厂商提供信用支持，换取其云业务的营收分成。报道披露的条款细节颇具想象力——芯片只租给“经批准的客户”，并要求把算力分散给多家小型AI公司使用。促使暂停的原因，据称是内部员工担忧该模式可能招致反垄断审查，对公司“向客户施加经营控制的边界”也存在顾虑。必须强调：以上均为WSJ报道口径，并非公司官方公告；据知情人士说法，英伟达不排除调整方案后重启，或将其并入其他计划。
“资产负债表即服务”的质疑
同周还有一条容易被忽略的动态：据新浪财经转引的报道，摩根士丹利首次启动对英伟达的信用研究覆盖，给出中性评级，并警告这家公司正在把资产负债表转化为战略性的AI融资工具——也就是被质疑为“资产负债表即服务”的模式。把两条新闻放在一起，画面就完整了：当最大的芯片供应商同时扮演自己客户的融资方，收入的“成色”问题——有多少需求是被自家的信用支持催生出来的——就成了损益表之外必须追问的变量。历史上，电信设备与光伏行业都上演过“供应商加杠杆撑起客户需求”的剧本，结局参半。而这次主动暂停，至少说明公司自己也在重新掂量这条边界。
为什么B面比A面更值得盯
A面无需多虑：需求真实、指引清晰、客户排队，1080亿的指引摆在那里。值得盯住的是B面的两个长期变量。其一，厂商融资的闭环到底有多大——分成计划是暂停不是终止，换一种更收敛的形态卷土重来的概率不低；届时真正该看的不是营收第一行，而是应收账款、合同条款和现金流的附注。其二，反垄断的边界划在哪里——“只租给经批准的客户”已经不是销售行为，而是产业链控制；监管对这类安排的容忍度，将决定算力巨头商业模式的半径。千亿门槛很亮，但英伟达的下一阶段故事，多半写在财报附注里，而不是新闻标题里。
明天见。
$96.2 billion. After hours on August 27, NVIDIA reported its fiscal Q2 2027 results (quarter ended July 26): revenue of $96.2B, up 106% year over year and 18% sequentially; the data center business alone booked $89B, up 117% — a year ago, total revenue was $46.7B. The numbers held no suspense; the suspense arrived later that day, when the Wall Street Journal reported the company had quietly paused a revenue-sharing program aimed at AI cloud providers. The most ferocious earnings report in tech history, and the most sensitive business-model brake job, landed in the same week.
The Report Card: One Quarter From the Trillion-Run-Rate Gate
The hard numbers first: GAAP and non-GAAP gross margins both at 75.0%; GAAP net income of $59.7B; GAAP diluted EPS of $2.46 (non-GAAP: $2.22). Management guided next-quarter revenue to $108B — at this slope, a hundred-billion-dollar quarter is standing at next quarter's door. The company also returned $26B to shareholders this quarter, outbidding even its own buyback records. One structural note: data center now exceeds ninety percent of total revenue — this company's destiny curve and the world's AI capex line are, for better or worse, the same line.
Two Sentences From Jensen Huang
Two lines from the earnings statement deserve transcription. First: “Compute is revenue” — a narrative that equates compute purchases with revenue creation, pitched simultaneously to customers and to Wall Street. Second, on Blackwell demand: “really extraordinary.” The product footnote is harder: the next-gen Vera Rubin platform has entered full production, and its companion Vera CPU is positioned as “the first CPU built for AI agents.” Note the wording — chip product definitions are now being drawn around agents. AI agents have graduated from software storyline to hardware design target; that is a more honest industry signal than any slogan.
The B-Side: A Revenue-Sharing Experiment Hits Pause
Then comes the story actually worth unpacking. Per the Wall Street Journal on August 27, citing people familiar with the matter, NVIDIA paused last week a financing arrangement for AI cloud providers: the original design had NVIDIA extending credit support to cloud providers in exchange for a share of their cloud revenue. The reported terms were imaginative — chips leased only to “approved customers,” with requirements to spread compute across multiple smaller AI companies. What triggered the pause, reportedly, was internal worry that the model could invite antitrust scrutiny, plus discomfort over the boundary of “exerting operating control over customers.” To be crystal clear: all of this is WSJ-reported, not an official company announcement; per those sources, NVIDIA hasn't ruled out reviving the program in adjusted form or folding it into other plans.
The “Balance Sheet as a Service” Question
Another quiet data point landed the same week: per a report carried by Sina Finance, Morgan Stanley initiated credit coverage of NVIDIA with a neutral rating, warning that the company is turning its balance sheet into a strategic AI-financing vehicle — the setup critics have tagged “balance sheet as a service.” Place the two stories side by side and the picture completes itself: when the largest chip supplier doubles as financier to its own customers, the “quality” of revenue — how much demand was midwifed by one's own credit support — becomes a variable to interrogate beyond the income statement. Telecom equipment and solar both staged the “vendor leverage propping up customer demand” script before, with mixed endings. That this pause was proactive at least suggests the company itself is re-weighing that boundary.
Why the B-Side Matters More Than the A-Side
The A-side needs little worrying: demand is real, guidance is clear, customers are queueing — the $108B number sits right there. The long-term variables live on the B-side. First, how big the vendor-financing loop actually gets — the program is paused, not terminated, and a comeback in a more restrained form is far from unlikely; when it returns, read the receivables, contract clauses and cash-flow footnotes rather than the top line. Second, where the antitrust boundary gets drawn — “lease only to approved customers” is no longer a sales act but supply-chain control; regulators' tolerance for such arrangements will define the radius of compute giants' business models. The hundred-billion gate glitters, but NVIDIA's next chapter will most likely be written in the earnings footnotes, not the headlines.
See you tomorrow.
千亿门槛很亮，但英伟达的下一阶段故事，多半写在财报附注里。
—— Dawn Vision编辑部
The hundred-billion gate glitters, but NVIDIA's next chapter will be written in the earnings footnotes.
— The Dawn Vision Editorial Desk
英伟达Q2 FY2027（截至2026-07-26） · 营收962亿美元(+106% YoY, +18% QoQ) · 数据中心890亿(+117%) · GAAP/非GAAP毛利率均75.0% · GAAP净利597亿 · GAAP EPS 2.46美元/非GAAP 2.22 · Q3指引1080亿 · 本季股东返还260亿 · 黄仁勋'Compute is revenue' · Blackwell需求'really extraordinary' · Vera Rubin全面投产 · Vera CPU'首款为AI agent打造的CPU' · WSJ报道(8/27)：上周暂停收入分成计划，条款含'只租给经批准客户/算力分散给小型AI公司'，内部担心反垄断风险，不排除调整后重启 · 大摩首次信用覆盖：中性评级，'资产负债表即服务'质疑
NVIDIA Q2 FY2027 (quarter ended 2026-07-26) · Revenue $96.2B (+106% YoY, +18% QoQ) · Data center $89B (+117%) · GAAP and non-GAAP gross margin both 75.0% · GAAP net income $59.7B · GAAP EPS $2.46 / non-GAAP $2.22 · Q3 guide $108B · $26B returned to shareholders this quarter · Jensen Huang 'Compute is revenue' · Blackwell demand 'really extraordinary' · Vera Rubin in full production · Vera CPU 'first CPU built for AI agents' · WSJ (Aug 27): revenue-sharing program paused last week; terms included 'lease only to approved customers' and 'spreading compute across smaller AI companies'; internal antitrust concerns; may restart in adjusted form · Morgan Stanley initiates credit coverage: neutral, 'balance sheet as a service' concerns
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎对当日信号的交叉验证生成。财务数据全部来自英伟达官方财报新闻稿（一手来源）；收入分成计划暂停为《华尔街日报》8月27日报道口径（经电子工程专辑、Channel NewsAsia等多家媒体转引确认），非公司官方公告，正文已明确标注；摩根士丹利信用覆盖相关表述来自新浪财经转引报道。全文未引用精确盘中股价数字。
Produced by the Dawn Vision cognition engine and cross-checked against same-day signals. All financial figures come from NVIDIA's official earnings press release (primary source). The revenue-sharing pause is as reported by the Wall Street Journal on August 27 (corroborated via EE China, Channel NewsAsia and others), not an official company announcement, and is labeled as such in the text; the Morgan Stanley credit-coverage note is via Sina Finance's report. No precise intraday share-price figures are cited.
