Nvidia makes gains on rising inference demand
Welcome to Cautious Optimism, a newsletter on tech, business, and power.
📈 Trending Up: Chinese AI hype (FT) … Podcasting on YouTube … deficits … domestic vaccines … cutting off our nose to spite our face … never going public … new AI models … robot startups … Quantum computing (again) …
📉 Trending Down: African venture capital … active-duty military personnel numbers … paid research … automation … soft power … crypto rules …
Japan’s population, which saw 720,988 births in 2024 against 1.6 million deaths. In more positive news, marriages were up in Japan and South Korea, which could imply more births in coming years. Enough to arrest net population declines? No, but perhaps enough to slow them.
Today we’re taking a look at Nvidia, Salesforce and Snowflake earnings data. If you are a pure-play startup fan, don’t worry. We’ll get back to that shortly! — Alex
P.S. After CO was first sent to its editor, a lot happened. Quickly:
Trump’s next round of tariffs — 25% on Mexico and Canada, and other 10% on China — are now set to to into effect on March 4th. That’s next week. So much for more delays. Stocks fell on the news.
US initial jobless claims came in higher than expected (242,000 versus a 221,000 estimate). That’s not good.
Maybe rate cuts are still coming this year!
Small comforts.
Nvidia’s earnings are good news for more than just the chip business
Yesterday, the big news after the bell was Nvidia’s earnings report. The company’s $39.3 billion in Q4 F2025 revenue (roughly Q4 calendar 2024) beat expectations of $38.05 billion. Nvidia also generated more earnings per share on an adjusted basis than expected $0.89 compared to $0.84.
In short, it was a blowout quarter, again, for the chip giant, reporting 12% revenue growth from the sequentially preceding quarter, and 78% growth from the year-ago period. Total revenue from Nvidia during its fiscal 2025 (roughly calendar 2024) was “$130.5 billion, up 114% from a year ago,” the company reports.
Nvidia beat expectations by a large enough margin that its shares managed a roughly 3% gain in after-hours trading. That limited pump implies that if Nvidia had merely met expectations it would have had its head cut off.
Shares of AMD, Broadcom, TSMC and other chip companies are up a point or two apiece after Nvidia’s earnings.
Nvidia is a bellwether for the AI space; the more demand for its chips, the more demand we can infer that AI foundation model companies, inference providers, and traditional cloud giants are seeing from their customers. That’s the good news — Nvidia is still besting investor expectations because, downstream, there’s massive demand for AI-powered services.
Pulling from the company’s earnings call, a few excerpts underscore the point. Here’s Nvidia discussing its chip mix in its data center business as its new Blackwell chip begins its commercial ramp (emphasis added):
In the fourth quarter, Data Center revenue of $35.6 billion was a record, up 16% sequentially and 93% year on year as the Blackwell ramp commenced and Hopper 200 continued sequential growth. In Q4, Blackwell sales exceeded our expectations.
We delivered $11 billion of Blackwell revenue to meet strong demand. This is the fastest product ramp in our company's history, unprecedented in its speed and scale. […]
Customers are racing to scale infrastructure to train the next generation of cutting-edge models and unlock the next level of AI capabilities. With Blackwell, it will be common for these clusters to start with 100,000 GPUs or more.
But training is only so much of the game; for long-term investments in AI compute capacity to bear out financially, inference is what matters. Recall that model training is a huge compute suck, but running AI models against user queries (inference) should prove even bigger over time. The good news there is that Nvidia’s having a ball in the inference arena:
Our inference demand is accelerating, driven by test time scaling and new reasoning models like OpenAI's o3, DeepSeek-R1, and Grok 3. Long-thinking reasoning AI can require 100x more compute per task compared to one-shot inferences.
Blackwell was architected for reasoning AI inference. Blackwell supercharges reasoning AI models with up to 25x higher token throughput and 20x lower cost versus Hopper 100. […]
Blackwell has great demand for inference. Many of the early GB200 deployments are earmarked for inference, a first for a new architecture. Blackwell addresses the entire AI market from pretraining, post-training to inference across cloud, to on-premise, to enterprise.
Breaking that down into bullets:
New AI models that think more use more compute, implying that inference demands will not collapse on a marginal AI query basis, even taking into account DeepSeek-esque improvements to AI model training and inference efficiency. That’s good news for Nvidia’s ability to sell chips.
Rising inference demand is predicated on more than simply rising test-time compute-powered AI models over their predecessors. Overall inference demand is rising so much that, as noted above, Nvidia’s new GB200 system — replete with two Blackwell chips — is being stood up not merely for training new, better AI models but to handle AI query demand (inference).
The company remains very bullish on what comes next. Here’s Nvidia CEO Jensen Huang (Fool transcript, once again; emphasis added):
We have, of course, forecasts and plans from our top partners. And we also know that there are many innovative, really exciting start-ups that are still coming online as new opportunities for developing the next breakthroughs in AI, whether it's agentic AIs, reasoning AI, or physical AIs. The number of start-ups are still quite vibrant and each one of them needs a fair amount of computing infrastructure. […]
And then the long-term signals has to do with the fact that we know fundamentally software has changed from hand-coding that runs on CPUs to machine learning and AI-based software that runs on GPUs and accelerated computing systems. And so, we have a fairly good sense that this is the future of software. And then maybe as you roll it out, another way to think about that is we've really only tapped consumer AI and search and some amount of consumer generative AI, advertising, recommenders, kind of the early days of software.
The next wave is coming, agentic AI for enterprise, physical AI for robotics, and sovereign AI as different regions build out their AI for their own ecosystems. And so, each one of these are barely off the ground, and we can see them. We can see them because, obviously, we're in the center of much of this development and we can see great activity happening in all these different places and these will happen.
To sum: Nvidia’s new chips are selling better than expected, there’s good reason to expect them to keep selling well, and demand signals from use if AI products (rising inference compute needs) are positive.
We have had three signals lately regarding how much datacenter capacity the AI era is going to need:
Microsoft pulling back on some datacenter leases (bearish)
All told, I’d bet on horns over claws today.
What about Salesforce and Snowflake?
I am glad you asked. In brief, so that we can hit one or two more topics before running out of word count:
Snowflake’s stock is up 13% in pre-market trading. The company’s adjusted EPS beat expectations ($0.30 instead of $0.18), while its revenue also came in above the tape ($986.8 million instead of $957.6 million). The company’s guidance was also welcomed.
Not surprisingly, the data company had a lot to say about AI:
“Snowflake is the most consequential data and AI company in the world”
And from its earnings call (transcript source):
When it comes to AI, last year was foundational for us. We introduced AI, which is now being used by customers to seamlessly build data agents for both structured and unstructured data with state-of-the-art retrieval using Cortex Search and Analyst. We are supporting a range of market-leading models, including Anthropic's Claude, Meta Llama, and DeepSeek. And I'm sure, you saw that we just announced our expanded partnership with Microsoft that brings OpenAI's models into Cortex.
This makes us the only data platform to seamlessly host both Anthropic and OpenAI, world-leading models enabling our customers to build data agents while ensuring that their data remains secure in Snowflake. And just a few weeks ago, we introduced Cortex Agents, a world-class agent orchestration framework to enable seamless planning and execution of tasks across structured and unstructured data, all powered by leading models such as Anthropic's Claude.
And:
We have seen broad adoption of AI. We expect that to turn into real revenue over the coming quarters.
A beat, solid guidance, and lots of AI goodies for analysts to bake into their long-term models? Wall Street loved it.
Salesforce’s stock is having less fun. Shares of the CRM giant are off 3.5% in pre-market trading, after it whiffed revenue expectations ($9.99 billion against $10.04 billion anticipated). The company’s adjusted profitability beat expectations, but Salesforce’s guidance was lackluster.
Here’s CNBC: “The company called for $2.53 to $2.55 in adjusted earnings per share for the fiscal first quarter, with $9.71 billion to $9.76 billion in revenue. Analysts polled by LSEG had anticipated adjusted earnings of $2.61 per share, with $9.9 billion in revenue.”
The company’s staid financial results compared to expectations mask a pretty darn profitable company with record cash flows and record RPOs.
And CEO Marc Benioff had some pretty big words on Salesforce’s AI work. Noting that it has landed 3,000 “paying Agentforce customers who are experiencing unprecedented levels of productivity, efficiency, and cost savings” in the first 90 days of the product being for sale, Benioff pitched a future in which humans will be so much more productive that his company can skimp on net-new engineering hires (transcript source):
It's pretty awesome. And we're not going to hire any new engineers this year. We're seeing 30% productivity increase on engineering, and we're going to really continue to ride that up. And we're going to grow sales pretty dramatically this year.
The CEO is not kidding. Later in the same riff, he added (emphasis added):
We're not, you know, doing some of these kind of engineering efforts that may or may not have some kind of huge payoff but [will] take down all of our cash and all of our margin for the next several years. We're like augmenting our existing product line with artificial intelligence, taking advantage of these incredible investments that are being made, you know, in infrastructure by others. And we're going to deliver the digital labor revolution. This is our goal.
Our goal is to be the No. 1 provider of digital labor in the world. That's it. I don't think there really is another goal.
What does that look like in practice? Benioff gave an example to investors (emphasis added):
Agentforce is a consumption product. But it's going to be a mix. It's going to be a mix between what's going on with our customers with how many humans do they have and then how many agents are they deploying. I think in one example that I can personally tell you about in the quarter, you know, we did a large transaction with a large telecommunications company.
It's incredible. And when I was talking to their CEO, she was asking me how are we going to price the transaction and so forth. And I can't remember the exact number of the deal. I think it was -- Brian, maybe it was about $20 million in ACV, something like that for the year, maybe $60 million TCV transaction.
And then, as part of that transaction, it's a mix of, you know, we're rebuilding this telecommunications company. So, it's Sales Cloud, it's Service Cloud, it's Marketing Cloud, it's all of our core clouds. But then, also, it's Agentforce. And the Agentforce component, I think, was maybe $7 million in the transaction.
So, she was buying $7 million of Agentforce. She bought $13 million in our products for humans. And I think that was about 20 million in total. These are about approximate numbers.
I think that's kind of the idea that you're going to see us be able to deliver, you know, the right package for the right customer.
That’s interesting. I don’t know at this juncture if that revenue split between selling ‘products for humans’ and ‘products for computers,’ to paraphrase a different part of the same earnings call will hold, but it’s fascinating to consider. $7 million is more than a third of $20 million. Provided that AI revenues don’t cannibalize trad Salesforce top line, you can see quite a bit of potential growth ahead of the CRM concern.