저자: Alex · 외국인
요약 보기
엔비디아 실적은 매출·EPS 급성장 및 마진 일시 하락에도 견조하며, 주가 PER 42는 5년 저점 수준이라고 주장. 추론(inference) 연산량 증가와 AI 스케일링 법칙이 GPU 수요를 지속시켜 엔비디아 우위가 유지된다고 봄. 향후 GTC, Computex, 차세대 칩 발표를 촉매로 제시하고 저점 분할매수를 권유.
다듬은 글
I'm convinced that Wall Street still doesn't understand Nvidia. Nvidia stock is down more than 10% after one of its best earnings calls ever, and I think I've figured out why. So in this video I'll break down everything you need to know about Nvidia's latest earnings, the three big things investors seem to be missing right now, and what that means for Nvidia stock in 2025 and beyond.
Up front, let me quickly highlight Nvidia's excellent earnings results. Then I'll walk you through three key points that most investors are missing right now, three big catalysts that could be huge for Nvidia stock in the short term, and what all of this means for Nvidia shareholders. There's a lot to cover, so let's dive right into the earnings call.
Nvidia posted record revenue of $130 billion for 2024, which is its fiscal year 2025. Top-line revenue is up a whopping 114% year over year, and earnings per share came in at $2.94, up 147% from a year ago after accounting for the 10-for-1 stock split last summer. Nvidia's data center segment is its biggest business unit by far, accounting for 88% of total revenue. Those revenues came in at $115 billion, up an insane 142% year over year.
For reference, Amazon Web Services revenue came in at $108 billion in 2024, and AWS powers around a third of the entire internet. So Nvidia's data center revenue coming in higher than AWS's is a pretty big deal.
One thing that's scaring Wall Street is that Nvidia's gross margin fell to 73% in the fourth quarter, a three-point year-over-year decline. CFO Colette Kress said the lower margins are due to speeding up Blackwell manufacturing, and Nvidia expects gross margins to return to the mid-70s later this year. But nobody has pointed out that Nvidia's gross margin for the last 12 months came in at 75%, which is actually up 2.3 points from last year, even after accounting for the expedited Blackwell production. That's the first thing Wall Street is missing, and I'll come back to it.
A 73% gross margin means Nvidia is as profitable as most software companies. On top of that, gaming revenue was up 9%, professional visualization revenue was up 21%, and automotive revenue was up 55%, all year over year. I won't spend much time on these segments because they account for just 11% of Nvidia's revenue combined. But at this rate, Nvidia's total revenue could hit around $190 billion for calendar year 2025, which would mean more revenue than Tesla, AMD, and Palantir combined. That's pretty crazy.
If you've been watching this channel for a while, you know the key to finding great stocks is understanding a company's products, not just its profits. The best investments are usually in companies with platforms that other businesses pay to build on top of. Nvidia is worth around $3 trillion because every major tech company relies on its accelerated computing platform. Nvidia currently holds around a 90% share of the data center GPU market, and that's before Blackwell is fully ramped. Nvidia's hardware and software platforms are the picks and shovels of this entire AI gold rush, and how big it gets depends on how fast Nvidia can deploy its solutions around the world. It wants to go as fast as possible, which is why it's willing to accept lower margins in the short term.
Many of the companies building next-generation AI applications are not publicly traded. Think about the '90s and early 2000s: companies like Amazon and Google went public very early in their growth cycles, but today companies wait an average of 10 years or longer to go public. That means investors like us can miss most of the returns from the next Amazon, the next Google, the next Nvidia. This is where Fundrise's Innovation Fund, the sponsor of this video, can help. It gives you access to invest in some of the best tech companies before they go public. Venture capital has usually been reserved for the ultra-wealthy, but Fundrise's Innovation Fund gives everyday investors access to top private, pre-IPO companies, with an entry point starting at $10. They've already invested over $110 million into some of the largest, most in-demand AI and data infrastructure companies. If you want access to late-stage companies before their IPO, check out the Fundrise Innovation Fund with my link below.
Alright, back to Nvidia. Annual revenue and earnings per share both grew by triple digits year over year, and Nvidia is ramping Blackwell even faster to maintain its massive market share. That's why margins are down quarter over quarter even though they're still up for the full fiscal year. This is the first thing Wall Street seems to be missing. As a result, Nvidia is trading at a price-to-earnings ratio of 42, close to its five-year low. In fact, the last time it traded at P/E ratios this low, Nvidia stock was around $11 per share.
The second thing most investors missed was a pretty big bombshell that Nvidia's CEO Jensen Huang dropped during the Q&A portion of the earnings call. Timothy Arcuri from UBS asked Jensen how some of Nvidia's biggest customers are balancing ASICs and GPUs. For context, Amazon, Microsoft, and Google all have their own application-specific integrated circuits, or ASICs, which some investors worry compete directly with Nvidia's GPUs. Now we're seeing some super clusters use both kinds of chips together instead of relying only on Nvidia's GPUs.
Jensen gave a thorough answer, so I'll paraphrase a bit. Nvidia builds very different things than ASICs, even though sometimes they overlap. Nvidia's architecture is general, which means it's great for diffusion-based models, vision-based models, multimodal models, or text, without needing to specialize in any one of those. Nvidia's ecosystem is so big and feature-rich that most innovations and algorithms come out on CUDA first. Nvidia is also really good end to end, from data processing and curating training data to training, post-training, and inference themselves. Unlike ASICs, which are tied to one cloud, whether AWS, Azure, or Google, Nvidia is in every cloud and can also be on premises or in a robot. And Nvidia's upgrade cycle is every year, which tends to be faster than ASICs get upgraded. This includes multiple chips, GPUs, CPUs, and networking chips to connect them at scale.
As a result, Nvidia's total performance per watt is anywhere from 2x to 4x, or even 8x, better than these ASICs. That translates directly into revenue. AI applications charge based on the number of tokens generated, and most data centers have either a fixed amount of space or a fixed amount of power. Being able to generate eight times more tokens in the same space and power budget translates directly into an eightfold return on investment for those data centers.
I don't think Wall Street realizes that Amazon, Microsoft, and Google's custom chips don't compete with Nvidia at all. They're slower to upgrade, much more limited in the workloads they can support, and have two to eight times lower performance per watt than Nvidia's chips, depending on the innovations their massive install base has come up with and where Nvidia's chips are in their annual upgrade cycle.
Here's a clear example. Nvidia released an open-source software package called TensorRT-LLM, which literally doubled the inference performance for large language models running on Nvidia GPUs. That means the hundreds of thousands of H100s bought by companies like Amazon, Google, Microsoft, Meta, OpenAI, and xAI got twice as good at running large language models overnight. TensorRT-LLM also works on Nvidia's older GPUs, their Ampere A100s, which predate the H100s, and on Blackwell GPUs too.
This leads nicely into the third thing Wall Street seems to be missing, which is pretty ironic considering Jensen said it on CNBC for the whole world to hear. Basically, it's AI teaching AI how to be better AI. The post-training process is where an enormous amount of innovation is happening right now. A lot of it happened with these reasoning models, and that compute load could be 100 times more than pre-training. Then comes inference, the reasoning process. Instead of just spitting out an answer when prompted, the model reasons about the question. It thinks about how best to answer, breaks it down step by step, might even reflect on it, comes up with several versions, picks the best one, and then presents it to you. The amount of computation we have to do at inference time is now 100 times more than what we did when ChatGPT first came out.
That works out to roughly a tenfold increase in inference compute costs every year, thanks to the new reasoning approaches in OpenAI's GPT-4o and o-series, xAI's Grok 3, and DeepSeek's R1 model. That's huge for Nvidia, and I expect this trend to continue.
Let me explain why. Until recently, inference didn't need much compute, because answers were generated in a single shot. The user prompts the model, and the model responds with whatever tokens it predicts come next. Inference costs were basically a function of how many prompts a model received and the average difficulty of those prompts. Today we have techniques like step-by-step verification, chain-of-thought reasoning, and mixture-of-experts models, all of which generate many more tokens on the way to solving a problem. As a result, AI models have a whole new way to scale performance: spend more tokens thinking about a problem in exchange for a better final output. For a human, that would be like getting more time, more money, and more access to the right tools and experts to do a job. For AI models, it means more GPU hours and better algorithms.
So now there are three ways to scale AI models. The first is pre-training scaling, where the foundation model improves with more training data and more parameters, both of which increase the compute needed. The second is post-training scaling, which is reinforcement learning on more specialized data and prompts, refining outputs with different kinds of feedback, having the model practice millions of times on synthetic data, and so on. Think of this as the fine-tuning step, where the model keeps learning over time, which also requires a significant amount of compute. The third is test-time scaling, where models produce better outputs by spending more time, energy, and tokens on the inference step itself. Should GPT-4o provide the answer in one shot, or do some retrieval-augmented generation first? Should Grok 3 break its chain of thought into three steps, or five, or ten? Which expert models should DeepSeek R1 consult, or should it take a majority vote from all of them?
This is a huge area of research that will produce many different algorithms for all kinds of problems across a wide variety of markets. What they all have in common is that they benefit from more compute. Letting models generate more and more tokens on the way to solving a problem leads to better and better outputs, and as long as that's true, AI research and development teams will keep finding new ways to generate better answers at the cost of more tokens.
That brings us full circle. This is exactly why hyperscalers, supercomputers, and other companies with big data centers will keep buying Nvidia's chips, which are anywhere from 2 to 8 times more power-efficient than the specialized chips built for specific workloads on specific clouds. That's why Nvidia is pushing the pace of Blackwell's production and deployment, even at the cost of margins in the short term. More chips lead to more innovation, which leads to more AI adoption, which leads to more chips. Nvidia understands this feedback loop better than anyone else.
Now that I've covered Nvidia's earnings and the three key points Wall Street has been overlooking, let me highlight the three catalysts I think will be huge in the near term.
The first is Nvidia GTC, which is right around the corner. That's Nvidia's massive developer conference, and I'll be there live to cover Jensen's keynote, check out Nvidia's newest products and prototypes, and attend as many sessions as I can on topics like robotics, self-driving cars, and even quantum computing. I'll leave a link to the sessions I plan to attend in the description below. They're online and absolutely free, and a great way to learn more about the science behind this stock. The key thing I expect from Jensen's GTC keynote is talk about the B300 Blackwell Ultra GPUs, which are set to ship in the second half of 2025. Blackwell is an accelerated computing platform, so he should be discussing upgrades not just to the GPUs but also the Grace CPUs, the BlueField DPUs, NVLink switch chips, InfiniBand, and Spectrum-X networking solutions. And it's not just about data centers. I want to understand how Blackwell performs at the edge by touring the entire exhibition floor at GTC and seeing Blackwell chips in everything from robots and self-driving cars to desktop PCs and even Project DIGITS. Of course, I'll be interviewing as many Nvidia executives as I can and sharing everything I learn with you.
The second catalyst is Computex 2025, which takes place in May. At last year's Computex, Jensen revealed the architecture after Blackwell, called Rubin, slated for 2026. Just like Blackwell, Rubin is a platform, so we should expect to hear about all six next-generation chips involved, not just the GPUs themselves.
The third and final catalyst is the architecture after Rubin. In an article published just yesterday, Tom's Hardware confirmed that post-Rubin GPUs are already in the works, and Jensen said he'll announce them at GTC. It's not just about these next three chips. It's about the overall performance gains from chip to chip, and the trend in performance we can expect over the long term. The bigger the boost each chip provides to data center throughput, the further Nvidia extends its lead over every other AI chipmaker, and the faster AI costs come down, which means more AI adoption across the board. Both of those would obviously be great for Nvidia stock.
So between Nvidia's strong earnings call, the three big things most Wall Street analysts are missing, the upcoming catalysts at GTC and Computex, and Nvidia's P/E ratio sitting near a five-year low, my plan is to keep dollar-cost averaging in for the long term. That means buying more shares as the price goes lower while keeping a healthy amount of cash on the side. I'll be at GTC, so stay tuned, and I'll see you there before Nvidia GTC.
Comments 0