Edited by Brian Birnbaum and an update of my Palantir, AMD and Nvidia deep dives.
We are on a trajectory towards exponentially more parameters per dollar. This is what truly matters for Palantir, AMD and Nvidia shareholders.
Also, the battle for chip supremacy between the US and the former USSR teaches us that regimes cannot innovate as sustainably as free economies can.
DeepSeek-R1 has 671 billion parameters and allegedly cost $5.58M, while ChatGPT has around 170B parameters and cost orders of magnitude more than that. If you choose to trust the information coming out of China, which historically hasn’t been a good policy, this actually means that Palantir is set to grow way quicker. More parameters for less dollars means that we are about see the end applications of AI explode and with it, the utility of Palantir’s platform for end customers. This also means Palantir’s platform gets easier to deploy and use, driving more business.
If veracious, DeepSeek is merely a continuation of the curve you see below, originally popularised by futurist Ray Kurzweil. Over time humanity has constantly found ways to deliver more compute per dollar (in constant terms). In terms of LLMs, this means we are going to see models with more parameters being trained for less money. This is ultimately how trillion dollar parameter models become a commodity and AI goes mainstream. This is the path along which Palantir’s digital twins become the epicenter of (Western) capitalism and AMD, Nvidia and others come to sell more chips than we can imagine.
It’s true that such a leap in parameters trained per dollar makes the Nasdaq’s AI CapEx look a bit silly, but what will actually happen is the deployed infrastructure will get far more efficient practically overnight. Even if DeepSeek’s training cost is much higher than the alleged $5.58M, the trend in the graph above still applies. Any given installed compute capacity becomes 10X more efficient almost instantly the moment you slash the training cost per parameter by 10X. We are going to see bigger, higher performing and cheaper models emerge with a relatively regular cadence and that’s likely to make us all far richer.
LLMs are clearly on a path towards being commoditised as they become increasingly undistinguishable one from the other, driving downward pricing pressure. The value for companies is in being able to use LLMs to drive more revenue and lower costs and for that, Palantir’s software is essential. Together with the semiconductor companies, Palantir stands to benefit the most from the progress and commoditisation of LLMs, as it makes their digital twins exponentially more powerful with zero incremental R&D dollars.
Nvidia and AMD are down 12.47% and 4% respectively in the premarket today, but this is a short sighted move on behalf of the market. As previously mentioned, more parameters per dollar makes the existing infrastructure more efficient. In the short term this makes additional GPU purchases redundant, but the demand for compute naturally doesn’t stop there. Cheaper models mean more AI workloads across the board, which will still drive exponentially higher levels of training and inference workloads.
For AMD specifically, DeepSeek’s progress (at least in terms of performance) is great news. DeepSeek-V3 was launched before the R1 model, on the 26th of December of 2024. AMD announced the integration of the V3 model in their MI300X chip just yesterday and the relatively rapid integration reveals two important facts:
The MI300X’s larger memory is ideally suited to support the rapidly increasing size of LLMs. V3 has 671B parameters, as does R1, but it’s still larger than Meta’s Llama 405B which has 405B parameters.
Their software is getting better.
As LLMs get larger, it becomes increasingly crucial to be able to store them on-chip in order to minimise latency. In theory, AMD’s chiplet platform enables them to add higher memory capacity to their chips than competitors, which ultimately enables them to perform cheaper and faster inferences. This is well evidenced by Meta’s decision to run inferences exclusively on the MI300X, but until yesterday we weren’t seeing signs of traction in the market beyond that. Much of this slow progress has to do with AMD’s software not being good enough to scale beyond deployments were they provide tailored support, as they have done with Meta.
The fact that AMD could integrate DeepSeek-V3 briskly likely indicates that whatever software they used to get Llama’s inferences running on the MI300X is likely being productised. This suggests AMD is making good progress with its software and that we will likely see it scaling beyond manual support in the near future. Since AMD has competitive AI hardware now, especially on the inference side, the progress on the software side is likely to be quite accretive for the company’s earning power. As better models drive more applications and thus inference workloads, this seems like the perfect storm for AMD.
Lastly, DeepSeek poses questions about America’s ability to beat China in the AI race. While history doesn’t repeat itself, it does rhyme. This is why I believe there is much to be learned by studying the battle between the US and the former USSR for semiconductor supremacy during the Cold War. In essence, what happened is that people in Russia couldn’t innovate sustainably because at some point their work would be inconvenient to the communist party. At times it seemed that the USSR would take off, but over time ideology would reign over innovation.
Innovation is structurally inconvenient for regimes because it involves the free flow of information. At some point regimes are forced to intervene information flow in order to preserve their interests. This is why long term it’s hard for a regime to compete with a free economy and why betting on the latter is statistically a sounder bet. For a deep dive on this topic, I recommend reading Chip War by Chris Miller. It’s a fascinating read and it will equip you with the philosophical framework to better understand the AI race that’s unfolding as we speak.
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