The AI Bull Market Continues
Update
Edited by Brian Birnbaum and an update of my original AMD and Nvidia deep dives.
The Nvidia Q4 2024 earnings call is bullish for AI companies. They suggest that Palantir’s operating margins are set to continue growing exponentially.
And they’ve given me further clarity on AMD’s competitive advantage, which is AI workload specialisation.
The Nvidia Q4 2024 earnings report gave me additional clarity on two things: the continuity of the AI bull market and AMD’s competitive advantage. Starting with the former, Jensen explained that we now have two additional scaling laws which are driving a 100X increase in the demand for compute. These are the post-training and reasoning scaling laws: the former consists in improving models over time after they’ve been deployed and the latter is about getting models to think deeply, versus just perform one-shot inferences.
Jensen’s remarks about this during the Q4 2024 earnings call were fascinating:
The traditional scaling laws of AI remains intact. Foundation models are being enhanced with multimodality, and pre-training is still growing. But it's no longer enough.
We have two additional scaling dimensions.
Post-training skilling, where reinforcement learning, fine-tuning, model distillation require orders of magnitude more compute than pre-training alone.
Inference time scaling and reasoning where a single query and demand 100x more compute. We defined Blackwell for this moment, a single platform that can easily transition from pre-trading, post training and test time scaling.
The AI bull market has thus far been propelled by the pre-training scaling law. As we’ve added more parameters to the models and thrown more computation at them, they’ve been getting exponentially smarter. This has ultimately driven positive returns on AI CapEx, which has kept the bull market alive for AI companies across the value chain. Per Jensen’s words, these two additional scaling laws are thus set to accelerate the progress of AI further, which will hypothetically drive higher returns.
As Jensen explained during the call, Nvidia customers make more money per every increment in the aforementioned scaling laws:
And the third thing I would say is that our performance in our rhythm is so incredibly fast. Remember that these data centers are always fixed in size. They're fixed in size or they're fixing power. And if our performance per watt is anywhere from 2x to 4x to 8x, which is not unusual, it translates directly to revenues.
And so if you have a 100-megawatt data center, if the performance or the throughput in that 100-megawatt or the gigawatt data center is 4 times or 8 times higher, your revenues for that gigawatt data center is 8 times higher.
During the call Jensen also said that they see the next AI coming up: Agentic and Physical AI. The world has been focused on LLMs for some time now, but Jensen says that agents and AIs that understand the physical world are on the horizon and that they will be much bigger than what we have seen to date.
What this means for Palantir shareholders is that Palantir’s operating margins and free cash flow per share are likely going to continue growing exponentially. The main driver of Palantir’s financial outperformance over the past year has been AIP, which in turn has been enabled by the rapid advancement of LLMs. As I have explained recently, Palantir’s future is about creating autonomous enterprises and both Agentic and Physical AI are key pillars of that future.
To the degree that LLMs have increased Palantir’s margins, by making digital twins easier to deploy and use, the next two AI waves will do so considerably more. This is because these two types of AI will enable autonomous employees to emerge from Palantir’s platform, that will be capable of performing actions both in the virtual and in the physical space. This would exponentiate the value delivered to Palantir’s end customers at a marginal cost, since all Palantir has to do is plug these new AI models to their customers’ digital twins.
Palantir stock is down over 30%, which is fairly typical of growth stocks of this sort. As I’ve discussed recently, the market is attempting to price in the company’s exponential fundamental progress. The volatility is the result of essentially no one understanding how to price a company evolving in this manner. Nvidia is perhaps the first instance in history of this phenomenon and it’s ironic to see that investors were actually paying just 19 times earnings for the company just a year ago.
Nvidia teaches us that in winner-takes-all scenarios, it’s more important to make sure you’re betting on the actual winner than making precise calculations of the company’s intrinsic value at a given point in time. Nvidia investors have won because they spotted exponential growth early. While there is no guarantee of success for Palantir, I see a similar pattern in play.
Further, when asked about his view of ASICs, Jensen explained in the Q4 2024 earnings call that Nvidia’s ecosystem is generalist. ASICs are chips that are built for a specific neural network, in the case of AI. Thus ASICs tend to outperform general GPUs when running the specific AI model they’ve been built for. Jensen’s reply gave me great clarity on AMD’s competitive advantage in the AI field, which is specialisation at a marginal cost.
See Jensen’s comments about this topic, during the Q4 2024 earnings call:
Well, we built very different things than ASICs, in some ways, completely different in some areas we intercept. We're different in several ways.
One, NVIDIA's architecture is general whether you're -- you've optimized for unaggressive models or diffusion-based models or vision-based models or multimodal models or text models. We're great in all of it.
We're great on all of it because our software stack is so -- our architecture is sensible, our software stack ecosystem is so rich that were the initial target of most exciting innovations and algorithms. And so by definition, we're much, much more general than narrow. We're also really good from the end-to-end from data processing, the curation of the training data, to the training of the data, of course, to reinforcement learning used in post training, all the way to inference with tough time scaling.
While I don’t expect this approach to yield anything but exponentially growing revenues over the long term, Nvidia’s generalist approach leaves a large gap in the market for AMD to fill. While in the past I’ve been vocal about AMD’s inference advantage, the implications of AMD’s chiplet platform are broader. Meta’s Llama 405B running inferences exclusively on AMD’s MI300X chip is at a higher level the result of AMD being able to personalise chips at a marginal cost, with a clear total cost of ownership advantage.
In her most recent interview, Lisa Su explains that the traction on the inference side is the result of tactics - implying that what they’ve done is leverage their highly versatile platform to bring Meta (among other customers) a chip that AMD felt would best address a specific niche. Despite it’s extraordinary success, Nvidia can’t do that because its platform is geared for a generalist approach. AMD’s platform has enabled them to add an unusually high memory capacity, which brought Meta onboard.
I believe that as AI continues to mature, with the aforementioned new scaling laws evolving and Agentic and Physical AI coming into the scene, the variety of AI workloads in the world will explode. AMD’s ability to produce specialised chips at a marginal cost will likely be its competitive advantage over the next decade or so. Further, what this means is that AMD’s ROCm doesn’t have to catch up to Nvidia’s at face value: it just has to get good for whatever specialised chips AMD brings to the market, that Nvidia’s cant quite compete with.
Lisa explains this concept elegantly in her latest interview:
Until next time!
⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.
You can also reach me at:
Twitter: @alc2022
LinkedIn: antoniolinaresc