News Reporter |
News Editor | Wen Shuqi
After doubling its revenue, Nvidia has, for the first time ever, offered investors a peek at next year’s performance.
On August 26, after the US market closed, Nvidia reported its Q2 FY2027 earnings: quarterly revenue hit $96.221 billion, up 106% year-over-year; data center revenue reached $89 billion, up 117%; and adjusted EPS (non-GAAP) came in at $2.22, a 120% jump from a year ago.
Nvidia founder and CEO Jensen Huang said, “AI has reached a turning point. It’s doing real work. AI technology has become valuable and profitable. Now, compute has become a revenue stream.”
During the subsequent conference call, CFO Colette Kress dropped two key numbers: one, preliminary expectations for FY2028 revenue growth of about 70% year-over-year; two, gross margin will gradually slide from the current level to a trough of 71% to 72% by Q4 of this year.
This marks the first time in Nvidia’s history that it has given revenue guidance a full year in advance. Kress emphasized that the 70% figure is “supply-constrained.”
Explaining the unprecedented early guidance, Huang said a big reason is that the company now has far better visibility into its upstream and downstream partners than ever before. Nvidia has not only locked in resources early with upstream suppliers like memory and manufacturing, but has also gotten involved in downstream infrastructure like power, land, and data centers. That gives the company a very clear picture of future compute delivery capacity.
“Our actual demand is much higher than 70%, but with current supply, we’re confident we can deliver that 70% growth,” Huang said. He added that Nvidia will keep working with the supply chain to squeeze out more capacity. The reason for issuing guidance this early, he explained, is to align customers, shareholders, and suppliers around the same expectations so everyone can invest early. “Next year is going to be a truly extraordinary year.”
On the supply gap everyone’s asking about, he said during the call: “Right now, the supply we have can support 70% growth or a bit more. Demand is way higher. We have to keep hustling, or we’ll disappoint our customers.” When asked what’s the tightest link—wafers, memory, power, or data centers—he didn’t rank them, only noting that “the entire supply chain is running at full capacity.”
The demand beyond that 70% is coming first from the rapid rise of AI agents. Kress pointed out that a single agent completing a task requires roughly 15 to 100 times more compute than a human doing the same task, because the models are bigger, reasoning and planning need many more rounds, and tools get called repeatedly.
Asked whether AGI could push demand even higher, Huang said that in some scenarios, a broad form of AGI is already here, and demand will keep climbing. He sees AI shifting from “humans actively issuing prompts” to agents running continuously. In the future, he predicts, a company could have far more agents than employees, running around the clock in the background and collaborating with each other.
But rather than obsessing over whether some AGI or recursive self-improvement milestone has been crossed, Huang said those labels matter less now. What really matters are three things: AI is starting to do genuinely useful work, AI-generated tokens are already producing profit, and if you add more compute, you can generate even more profitable tokens.
Huang also shared another set of numbers to illustrate how much revenue opportunity each gigawatt of data center capacity represents for Nvidia. He said revenue per gigawatt stood at $18 billion in the Hopper era, rose to $25 billion with Grace Blackwell, jumped to $40 billion with Vera Rubin, and will go even higher after that.
“I heard the other day that payback periods are now under a year—and we’re talking about data centers on the order of $50 billion,” he said.
Looking at the data center customer mix this quarter, hyperscale cloud providers contributed $49 billion in revenue, up 13% quarter-over-quarter; the ACIE segment—comprising sovereign AI, NeoCloud, industrial, and enterprise customers—contributed $40 billion, up 25% quarter-over-quarter and 138% year-over-year.
Huang made a point of highlighting the latter. “Most people only see the hyperscalers—that’s just half the picture. The other half is what we call ACIE.”
He explained that these customers typically don’t develop their own custom chips like the big cloud providers do, and they’re not just buying chips piecemeal either—they need a whole “AI factory” platform. This segment now accounts for roughly half of data center revenue and is growing at about 100% annually.
The next-generation product riding this wave of demand is Vera Rubin. Nvidia said the platform began volume shipments this month and has secured orders from major hyperscalers, AI cloud providers, and system OEMs. The company expects Vera Rubin to contribute about 20% of data center revenue in Q3.
Gross margin is one of the few metrics trending downward in this report. Kress used the word “extreme” to describe the current memory pricing environment, saying price increases have exceeded expectations and will be even higher next year.
As a result, Nvidia has directly reset its gross margin expectations: Q3 is projected at 74%, dropping to a low of 71% to 72% in Q4; with already-executed product price increases taking effect in Q1 FY2028, gross margin for FY2028 is expected to stabilize at 72% to 73%.
Kress also pointed out that this memory price surge isn’t some exogenous cost unrelated to AI demand—it’s precisely AI infrastructure buildout that’s driving memory demand higher.
TrendForce data from June shows that conventional DRAM contract prices actually rose about 93% to 98% quarter-over-quarter in Q1 2026, with another 58% to 63% increase expected in Q2. The firm further projects Q3 of this year will see gains narrow to 13% to 18%.
Beyond chip supply and demand, Nvidia’s deepening involvement in financing downstream AI companies became another focal point of the call. Kress disclosed that Nvidia has invested nearly $50 billion in frontier AI labs and has partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build financing platforms aimed at mobilizing over $500 billion in third-party capital.
That raises an obvious question: Nvidia is investing in AI labs like OpenAI and Anthropic—customers that are also developing their own chips. The two sides are both partners and potential competitors.
Huang responded by saying that customer-developed XPUs are mostly inference-specific chips tailored to a single cloud or a single service, whereas Nvidia offers a platform that can be deployed across different clouds and covers the entire AI lifecycle—from data processing and training to post-training and agentic inference.
As for the investments in these AI labs, he said: “If there’s any regret, my only regret is that we didn’t invest more and earlier.”
In this quarter, China’s direct contribution to Nvidia’s performance was quite small. Kress mentioned on the call that Nvidia delivered H200 products to Chinese customers under US government licenses this quarter, but the volume was less than 1% of data center revenue, and those shipments diluted overall gross margin. Given geopolitical uncertainties, the Q3 guidance excludes any data center compute revenue from China.
Based on Huang’s previous public statements, China once contributed 20% to 25% of Nvidia’s data center revenue. But in an interview at the end of April, he said Nvidia’s share of China’s AI accelerator market had dropped to 0%, essentially ceding that market to Huawei.
That said, Chinese companies came up quite a bit during the call. When Kress listed mainstream open-source models running on Nvidia’s platform, she named Qwen, Kimi, GLM, DeepSeek, and MiniMax.
Huang then expanded on open-source models. He said, “The world needs both closed-source and open-source models,” noting that usage of both is growing rapidly.
According to him, almost all open-source models run on Nvidia’s platform, and most frontier models are built on Nvidia’s platform as well. Whether closed or open source, as long as model usage keeps growing, it will generate new compute demand.
At the end of the call, Huang circled back to supply. “We’re not even at next year yet—we still have plenty of time.” Going forward, Nvidia will continue pushing the supply chain for more capacity.
