NVDA Q3 FY26 finance report
Disclaimer: The information provided here is intended for general knowledge and informational purposes only, and does not constitute financial advice. Investment decisions should be based on your specific financial situation and needs, and after consultation with a qualified financial advisor.
免責聲明：網頁屬個人博客，一切言論純粹是表達本人的個人意見或經驗分享，無論在任何情況下，不應被視為投資建議，也不構成要約、招攬、邀請、誘使、建議或推薦，本人亦無法保證部落格內容的真實性和完整性。讀者務請運用個人獨立思考能力自行求證和分析，讀者一切的投資決定以及該投資決定引致的收益或損失，概與本人無涉。
NVDA Q3 FY26 finance reportNvidia's Q3 earnings hit $57 billion, surpassing estimates, with Q4 projected at $65 billion.
Nvidia's CEO Jensen Huang announced record sales of Blackwell GPUs during the earnings call, highlighting that cloud GPUs are sold out, reflecting strong demand in the AI infrastructure market.
Evercore raised Nvidia's price target to $352, citing a 95% upside from rising demand for Blackwell products, with revenue up nearly 50% to over $13 billion as supply improved.
Income statement:
Revenue jumped +22% Q/Q and 62% Y/Y to $57.0 billion ($1.9 billion beat).
⚙️ Data Center +25% Q/Q and +66% Y/Y to $51.2 billion.
🎮 Gaming -1% Q/Q and +30% Y/Y to $4.3 billion.
👁️ Professional Viz +26% Q/Q and +56% Y/Y to $0.8 billion.
🚘 Automotive +1% Q/Q and +32% Y/Y to $0.6 billion.
🏭 OEM & Other +1% Q/Q and +79% Y/Y to $0.2 billion.
Gross margin was 73% (-1pp Y/Y).
Operating margin was 63% (+1pp Y/Y).
Net Profit Margin was 56% ( (+1pp Y/Y).
Non-GAAP operating margin was 66% (flat Y/Y).
Non-GAAP EPS $1.30 ($0.04 beat).
Cash flow:
Operating cash flow was $23.8 billion (42% margin).
Free cash flow was $22.1 billion (39% margin).
Balance sheet:
Cash and cash equivalents: $60.6 billion.
Debt: $8.5 billion.
Q4 FY26 Guidance:
Revenue +14% Q/Q and +65% Y/Y to $65.0 billion ($3.2 billion beat).
Gross margin 75% (+1.4pp Q/Q).
WHAT to make all of this?
⚙️ Data Center hit 90% of total revenue, growing 25% Q/Q to $51.2 billion, a massive acceleration. The Blackwell ramp broadened again, with hyperscalers, sovereign customers, and large enterprises all increasing deployments. Management noted strong adoption across training, inference, and early agentic workloads. Supply remains the gating factor, not demand:
⚡ Compute revenue surged 27% Q/Q as Blackwell availability improved and large projects moved into production. The absence of China revenue is now fully baked into the baseline. The growth you’re seeing is entirely ex-China, an important signal that the AI cycle no longer depends on China at the margin.
🔌 Networking rose 13% Q/Q, reflecting the build-out of AI factories. CFO Colette Kress highlighted continued strength in Spectrum-X Ethernet, InfiniBand, and NVLink as clusters get denser and model complexity rises. Networking is becoming a structural growth engine.
🎮 Gaming was mostly flat Q/Q at $4.3 billion, lapping the Blackwell-powered GPU launch last quarter but still up 30% Y/Y. GeForce maintains strong demand, even as the company reallocates more supply toward Data Center.
👁️ Professional visualization grew 26% Q/Q, benefiting from workstation upgrades and AI-accelerated design workflows. This segment continues to recover as enterprises modernize their tooling.
🚘 Automotive was steady, up 1% Q/Q, reflecting gradual adoption of NVIDIA’s autonomous driving and digital cockpit platforms.
📉 Margins expanded again, with gross margin climbing sequentially to 73% and Q4 guided even higher to 75%. The mix continues to improve as networking scales and Blackwell availability normalizes.
🔮 The outlook is exceptional: Q4 revenue is guided up another 14% sequentially to $65 billion ($3.2 billion ahead of estimates) despite assuming zero shipments to China.
Big picture: NVIDIA is growing at breakneck speed, and the ramp is now powered by two engines: Blackwell compute and the networking fabric behind AI factories. The China reset is already absorbed, and the cycle is being driven by global AI infrastructure demand that continues to broaden, deepen, and compound.
2. Business highlights
🤝 So. Many. Partnerships.
The past quarter was a blur of mega-partnerships, each one expanding NVIDIA’s reach deeper into the AI stack.
OpenAI: Working with NVIDIA to build and deploy at least 10 GW of AI data centers. NVIDIA plans to take an equity stake and invest up to $100 billion over time as part of the multi-year buildout.
Anthropic: Signed a deep platform deal to run up to 1 gigawatt of Grace Blackwell and Rubin systems, alongside a planned $10 billion investment from NVIDIA. Anthropic will purchase $30 billion of compute from Azure and collaborate with NVIDIA on model training and hardware optimization, turning a prior non-customer into a full-stack NVIDIA partner.
xAI: xAI is building gigawatt-scale Colossus 2 AI factories anchored on Blackwell, including a 500 MW flagship site with Humane. AWS will supply up to 150,000 NVIDIA accelerators to power these workloads.
Saudi Arabia (KSA): Framework agreement for roughly 400,000 to 600,000 GPUs over three years.
Palantir: Bringing CUDA X into Ontology, with customers like Lowe’s already using it for supply chain and analytics workflows.
Fujitsu, Intel, and Arm: Announced NVLink integrations that wire their CPU roadmaps directly into NVIDIA’s ecosystem.
The combined commitments of Microsoft, OpenAI, and Anthropic effectively ensure multi-year visibility for NVIDIA’s systems and keep demand anchored in the hyperscaler ecosystem.
Jensen Huang touched on the circular financing of customers:
“No company has grown at the scale that we’re talking about and have the connection and the depth and the breadth of supply chain that NVIDIA has. The reason why our entire customer base can rely on us is because we’ve secured a really resilient supply chain, and we have the balance sheet to support them.”
NVIDIA now has a Berkshire-style challenge to put its money to work. With a fast-growing pile of $61 billion in cash, the company is investing in the most important players in AI to secure future offtake and expand the CUDA ecosystem. These deals ensure that the fastest-growing AI companies have the resources to scale, which reinforces NVIDIA’s long-term demand rather than propping it up.
Source: Fiscal.ai
🏗️ Customers build chips and still buy NVIDIA
Custom silicon is rising across cloud providers: TPUs at Google, Trainium at AWS, Maia at Microsoft, and Meta’s internal accelerators. These chips target specific workloads, not the frontier or the broad platform layer.
Blackwell remains the default for large-scale training, inference, and agentic systems. Even alternative accelerators often rely on NVIDIA’s networking stack, reinforcing the systems moat rather than weakening it.
Customer ASICs shift negotiating leverage at the edges, but they expand the total compute pie, and NVIDIA remains essential at the center.
🇨🇳 From Great Wall to Firewall
China has moved from a swing factor to a structural constraint. Q3 confirmed what the last two quarters hinted: H20 demand never materialized, B-series chips face fresh US scrutiny, and China’s regulators are directing state-backed data centers toward domestic accelerators. NVIDIA shipped just $50 million of H20 this quarter, effectively zero.
Management now assumes no China revenue in both Q4 and FY27 guidance. The real risk is long-term. China is racing to build a full domestic AI stack that bypasses CUDA entirely. If successful, a market Jensen once pegged at ~$50 billion becomes structurally closed. For now, the rest of the world is more than compensating.
3. Key quotes from the earnings call
Check out the earnings call transcript on Fiscal.ai here.
CFO Colette Kress:
On demand and long-term visibility:
“We currently have visibility to $500 billion in Blackwell and Rubin revenue from the start of this year through the end of calendar year 2026. [...] We believe NVIDIA will be the superior choice for the $3 trillion-$4 trillion in annual AI infrastructure build we estimate by the end of the decade. Demand for AI infrastructure continues to exceed our expectations.”
This anchors the growth looking forward and frames the $500 billion Blackwell–Rubin pipeline inside a multi-trillion-dollar decade-long build-out.
CEO Jensen Huang:
On power limits and performance per watt:
“In the end, you still only have one gigawatt of power, one gigawatt data centers, one gigawatt of power. Therefore, performance per watt, the efficiency of your architecture, is incredibly important. [...] Your performance per watt translates directly to your revenues, which is the reason why choosing the right architecture matters so much now.”
Power is the binding constraint, and that perf-per-watt is the real battleground for economics that could favor NVIDIA over the long haul.
On the ecosystem and running every model:
“NVIDIA’s architecture, NVIDIA’s platform is the singular platform in the world that runs every AI model.[...] We run OpenAI. We run Anthropic. We run xAI [...] We run Gemini [...] We run science models, biology models, DNA models, gene models, chemical models [...] AI is impacting every single industry.”
Huang’s ecosystem argument is simple: one architecture, every major model, across consumer apps, enterprises, and science. It reinforces CUDA as the default AI operating layer.
4. What to watch moving forward
Depreciation anxiety
A growing concern on Wall Street is the useful life of high-end AI hardware.
Every new GPU generation makes the previous one look older faster, raising fears that hyperscalers will be forced to accelerate depreciation, pressuring long-term margins and EPS projections.
Colette Kress pushed back. She emphasized that software updates extend the useful life of NVIDIA GPUs and that the shift from simple chatbots to agentic AI dramatically increases compute intensity, keeping older clusters fully utilized even as Blackwell and Rubin ramp. The worry is real, but NVIDIA’s argument is that the demand curve determines effective lifespan.
Who’s buying?
Many super investors dialed back their NVIDIA enthusiasm in the latest 13F filings covering Q3 2025. It’s not entirely surprising with the stock hitting new highs. NVIDIA remains one of the most widely held names, yet many funds are still underweight relative to its nearly 8% weight in the S&P 500.
At ~32x forward earnings, NVIDIA trades mostly in line with the rest of Big Tech. But with EPS surging 67% Y/Y, you could argue NVIDIA looks cheap. NVIDIA’s growth is supply-constrained. That means quarter-to-quarter noise matters less than understanding how long this cycle can run and what the business looks like when demand normalizes.
Source: Fiscal.ai
What to watch next
If you’re a regular reader, you already know what to watch:
Cycles boom and bust: Like every major tech cycle, demand will eventually plateau or reset, and it will inevitably create volatility. However, in the words of Arya Stark: “Not today.”
Networking vs. compute mix: Is networking going to make a larger share of Data Center as AI factories scale?
Sovereign AI momentum: Are more countries committing to large, multi-year infrastructure builds?
China visibility: Does NVIDIA maintain a zero-China baseline, or is there any sign of reopening?
Hyperscaler silicon adoption: How quickly are cloud providers shifting toward in-house chips, and does it eventually impact NVIDIA's margins?
Power constraints: Energy efficiency and performance-per-watt remain the next competitive frontier, with many moving pieces.
Rubin cadence: Rubin is already in the lab. Any clarity on sampling, production, or deployment timing for the next platform will move the stock.
NVIDIA sits at the center of the most important computing cycle in decades. The details will shift (China, custom silicon, financing), but the direction hasn’t changed. As long as AI factories keep scaling, NVIDIA remains the one building the rails.
--------Chinese edition ------------
英偉達2026財年第三季財報
英偉達2026財年第三季財報顯示，第三季營收達570億美元，超出預期，第四季營收將達650億美元。
英偉達執行長黃仁勳在財報電話會議上宣布，Blackwell GPU的銷售量創下歷史新高，並強調雲端GPU已售罄，反映出人工智慧基礎設施市場的強勁需求。
Evercore將英偉達的目標股價上調至352美元，理由是Blackwell產品需求成長，股價有95%的上漲空間；隨著供應改善，營收成長近50%，超過130億美元。
損益表：
營收季增22%，年增62%，達到570億美元（超出預期19億美元）。
⚙️ 資料中心業務營收季增25%，年增66%，達到512億美元。
🎮 遊戲業務季減 1%，年增 30%，營收達 43 億美元。
👁️ 專業視覺化業務較上季成長 26%，較去年同期成長 56%，營收達 8 億美元。
🚘 汽車業務季增 1%，年增 32%，營收達 6 億美元。
🏭 OEM 及其他業務季增 1%，年增 79%，營收達 2 億美元。
毛利率為 73%（年減 1 個百分點）。
營業利益率為 63%（年增 1 個百分點）。
淨利潤率為 56%（年成長 1 個百分點）。
非 GAAP 營業利潤率為 66%（與去年同期持平）。
非 GAAP 每股收益為 1.30 美元（超出預期 0.04 美元）。
現金流：
經營現金流為 238 億美元（利潤率為 42%）。
自由現金流為 221 億美元（利潤率為 39%）。
資產負債表：
現金及現金等價物：606 億美元。
債務：85 億美元。
2026 財年第四季業績展望：
營收季增 14%，年增 65%，達到 650 億美元（超出預期 32 億美元）。
毛利率為 75%（季增 1.4 個百分點）。
這一切意味著什麼？
⚙️ 資料中心業務佔總營收的 90%，較上季成長 25%，達到 512 億美元，這是一個巨大的成長。加速成長。 Blackwell 的部署規模再次擴大，超大規模資料中心、獨立客戶和大型企業都在增加部署。管理階層指出，訓練、推理和早期智能體工作負載的採用率都很高。供應仍是限制因素，而非需求：
⚡ 計算收入環比增長 27%，這得益於 Blackwell 可用性的提高和大型專案投入生產。中國市場收入的缺失已完全反映在基線中。您目前看到的成長完全來自中國以外的市場，這是一個重要的訊號，表明人工智慧週期不再僅依賴中國市場。
🔌 網路業務較上季成長 13%，反映了人工智慧工廠的建設。財務長 Colette Kress 強調，隨著群集密度增加和模型複雜性上升，Spectrum-X 乙太網路、InfiniBand 和 NVLink 的持續強勁表現。網路業務正成為結構性成長引擎。
🎮 遊戲業務環比基本持平，收入為 43 億美元，與上季度 Blackwell GPU 的發布相比略有增長，但仍同比增長 30%。 GeForce即使公司將更多產能重新分配到資料中心，需求仍然強勁。
👁️ 專業視覺化業務較上季成長 26%，受惠於工作站升級和 AI 加速的設計工作流程。隨著企業工具的現代化，此業務板塊持續復甦。
🚘 汽車業務保持穩定，季增 1%，反映出 NVIDIA 的自動駕駛和數位座艙平台正逐步被市場接受。
📉 利潤率再次提升，毛利率環比攀升至 73%，預計第四季將進一步提高至 75%。隨著網路規模的擴大和 Blackwell 可用性的恢復，業務結構持續改善。
🔮 前景非常樂觀：儘管假設對中國的出貨量為零，但預計第四季度營收環比仍將增長 14% 至 650 億美元（比預期高出 32 億美元）。
整體而言：NVIDIA 正以驚人的速度成長，而推動這一成長的兩大引擎是：Blackwell 運算和 AI 工廠背後的網路架構。中國市場重塑已然完成。市場已消化吸收，而推動這一周期發展的，正是全球人工智慧基礎設施需求的持續成長、深化和複合。
2. 業務亮點
🤝 合作專案眾多
上個季度，NVIDIA 達成了一系列重量級合作，每項合作都進一步拓展了其在人工智慧領域的覆蓋範圍。
OpenAI：與 NVIDIA 合作，建置並部署至少 10 吉瓦的人工智慧資料中心。 NVIDIA 計劃持有 OpenAI 的股權，並在未來幾年內投資高達 1000 億美元，作為這項多年建設計劃的一部分。
Anthropic：簽署了一項深度平台協議，將運行高達 1 吉瓦的 Grace Blackwell 和 Rubin 系統，同時 NVIDIA 計劃向其投資 100 億美元。 Anthropic 將從 Azure 購買價值 300 億美元的運算資源，並與 NVIDIA 在模型訓練和硬體優化方面展開合作，從而將這家此前並非 NVIDIA 客戶的公司轉變為 NVIDIA 的全端合作夥伴。
xAI：xAI 正在建造一個基於 Blackwell 的千兆瓦級 Colossus 2 人工智慧工廠，其中包括一個 500 兆瓦的工廠。與 Humane 合作的旗艦專案。 AWS 將提供多達 15 萬個 NVIDIA 加速器來支援這些工作負載。
沙烏地阿拉伯 (KSA)：簽署了一項框架協議，將在三年內提供約 40 萬至 60 萬個 GPU。
Palantir：將 CUDA X 整合到 Ontology 中，Lowe's 等客戶已將其用於供應鏈和分析工作流程。
富士通、英特爾和 Arm：宣布了 NVLink 集成，將各自的 CPU 產品路線圖直接連接到 NVIDIA 生態系統。
微軟、OpenAI 和 Anthropic 的聯合承諾有效地確保了 NVIDIA 系統多年的可見性，並維持了需求的穩定性。
在超大規模雲端生態系統中，NVIDIA 扮演著重要角色。
黃仁勳談到了客戶的循環融資：
“沒有任何一家公司能夠達到我們所談論的規模，並擁有像 NVIDIA 這樣深厚且廣博的供應鏈。我們所有客戶都能信賴我們的原因在於，我們擁有真正具有韌性的供應鏈，並且我們擁有足夠的資產負債表來支持他們。”
NVIDIA 現在面臨著類似伯克希爾·哈撒韋公司那樣的挑戰，需要有效地利用資金。憑藉著快速成長的 610 億美元現金儲備，該公司正在投資人工智慧領域最重要的參與者，以確保未來的市場需求並擴展 CUDA 生態系統。這些交易確保了成長最快的人工智慧公司擁有擴展所需的資源，從而鞏固了 NVIDIA 的長期需求，而不是僅僅支撐其成長。
來源：Fiscal.ai
🏗️ 客戶自行建構晶片，但仍購買 NVIDIA 的產品
客製化晶片正在雲端服務供應商中興起：Google的 TPU、AWS 的 Trainium、微軟的 Maia 以及 Meta 的內部加速器。這些晶片針對的是特定工作負載，而非尖端技術或廣泛的平台層。
Blackwell 仍然是大規模訓練、推理和智能體系統的預設選擇。即使是其他加速器也通常依賴 NVIDIA 的網路協定棧，反而增強了系統的競爭優勢，而不是削弱它。
客戶 ASIC 晶片雖然改變了邊緣運算的談判籌碼，但它們擴大了運算總量，而 NVIDIA 仍然在核心領域扮演著至關重要的角色。
🇨🇳 從長城到防火牆
中國市場已從一個搖擺因素轉變為結構性限制因素。第三季證實了前兩季的跡象：H2O 需求從未真正實現，B 系列晶片面臨美國的新一輪審查，而中國監管機構正在引導國有資料中心轉向國產加速器。 NVIDIA 本季 H2O 的出貨量僅 5,000 萬美元，實際上為零。
管理階層目前在第四季和 2027 財年的業績指引中均假設中國市場收入為零。真正的風險在於長期層面。中國正競相建構一套完全繞過CUDA的國產人工智慧技術棧。如果成功，Jensen曾估計規模約500億美元的市場將徹底封閉。目前，世界其他地區的市場規模已足以彌補這一缺口。
3. 財報電話會議要點
點擊此處查看Fiscal.ai上的財報電話會議記錄。
財務長Colette Kress：
關於需求和長期前景：
「我們目前預計，從今年年初到2026年底，Blackwell和Rubin的收入將達到5000億美元。[...] 我們相信，到本十年末，NVIDIA將是每年3萬億至4萬億美元人工智能基礎設施建設的最佳選擇。市場對人工智能基礎設施的需求持續超出我們的預期。」
這為未來的成長奠定了基礎，並將5000億美元的Blackwell-Rubin項目納入長達數兆美元的十年建設規劃中。
CEO 黃仁勳：
關於功耗限制和每瓦性能：
“最終，你只有 1 吉瓦的電力，1 吉瓦的數據中心，1 吉瓦的電力。因此，每瓦性能，也就是架構的效率，至關重要。[...] 每瓦性能直接影響你的收入，這就是為什麼選擇合適的架構現在如此重要的原因。”
功耗是根本的限制因素，而每瓦性能才是真正決定經濟效益的關鍵所在，從長遠來看，這可能對英偉達更有利。
關於生態系統和運行所有模型：
「英偉達的架構，英偉達的平台是世界上唯一能夠運行所有 AI 模型的平台。[...] 我們運行 OpenAI。我們運行 Anthropologie。我們運行 xAI。[...] 我們運行 Gemini。[...] 我們運行科學模型、生物模型、DNA 模型、基因模型、化學模型。[...] 正在影響每個產業正在影響著每一個產業。」人工智慧
黃仁勳的生態系統論點很簡單：一個架構，即可運行所有主流模型，涵蓋消費應用、企業和科學領域。它進一步鞏固了 CUDA 作為預設 AI 作業系統的地位。
4. 未來值得關注的面向
折舊焦慮
華爾街日益關注高端 AI 硬體的使用壽命。
每一代新的 GPU 都會讓上一代產品顯得更快過時，這引發了人們的擔憂：超大規模資料中心將被迫加速折舊，進而對長期利潤率和每股盈餘預期構成壓力。
科萊特·克雷斯 (Colette Kress) 對此提出了反駁。她強調，軟體更新可以延長 NVIDIA GPU 的使用壽命，而且從簡單的聊天機器人到智慧體 AI 的轉變會顯著提高運算強度，即使 Blackwell 和 Rubin 等新架構的架構不斷升級，也能確保舊叢集得到充分利用。這種擔憂不無道理，但 NVIDIA 的論點是，需求曲線決定了有效使用壽命。
誰在買入？
在最新的涵蓋 2025 年第三季的 13F 文件中，許多超級投資者降低了對 NVIDIA 的熱情。考慮到該股股價屢創新高，這並不完全出乎意料。英偉達仍是最受投資人青睞的股票之一，但許多基金的持股比例仍低於其在標普500指數中近8%的權重。
英偉達的預期本益比約為32倍，與其他大型科技股基本一致。但隨著每股收益的飆升，其估值有望進一步提高。
