Edited by Brian Birnbaum.
Inference is the technological trend of our lifetime and AMD is best positioned to capitalise on it.
Generational wealth is created, among other methods, by surgically betting big on the most inevitable trends of your time before everyone else does. Rockefeller did it with oil. I did it, in my own small way, when I bought AMD heavily in 2014, and then Palantir in 2020, way before it became the S&P 500’s top performing stock of 2024.
Similarly, the following disclosure is tantamount to being told in 1859 that oil would change the world.
The economy is becoming an inference machine. In the pursuit of training the best models, big tech has invested nearly $200B on AI infrastructure. Nvidia has been the major beneficiary of this trend. Until now, Nvidia’s chips have powered the training of all the AI models and LLMs that have sprung up over the past two to four years. But what are they being trained to do?
Make predictions–otherwise known as inferences. While training AI models is important, the money is now moving to inference.
The world’s economic competition consists of either taking action or calculating the next set of actions. The latter is another way to describe the act of predicting. Inference will allow machines to assume a large percentage of this cognitive-analytical activity.
Sixteen-plus decades after Rockefeller bet big on oil, the economy is now essentially a machine that burns fossil fuels to deliver products and services. Energy from fossil fuels is converted into value by exponentially accelerating humanity’s decisions. The information-sharing technologies used today still rely, to a massive degree, on these fossil fuels, and have proliferated global real GDP.
The same will happen with data and inference. As we collect more and higher quality data, we’ll delegate a growing volume of decision-making to computers until the economy is simply a giant inference machine that calculates the best moves while we enjoy life and creative pursuits.
By decreasing the cost of decisions to zero or near-zero, inference will again exponentiate world GDP, only much faster than anything before. Inference will drive digital twins and power robotics, themselves converting decisions into actions at a marginal cost. The outsized economic impact is tantamount to that of oil in the 1860s and beyond.
AMD presents me with the most obvious mechanism to bet on inference. As inference makes its way through the economy, it’ll tend to run on small devices with more stringent energy requirements. By mixing and matching different compute units at will, AMD’s chiplet platform better navigates the thermodynamic requirements of inference, particularly on these small devices.
Take for example Meta’s most recent bet on AMD’s technology. As of 2024, Meta’s running all Llama live traffic on AMD’s MI300X exclusively. In other words, Meta’s top AI model performs all inference on AMD’s technology, and the plan is to do the same with the upcoming MI350 and MI400 series. The two companies are now engaged in a long-term relationship in which AMD is the key inference technology provider for Meta. Listen to Meta’s Infrastructure VP Kevin Salvatori discuss this in my last Youtube video.
AMD’s chiplet platform has enabled the company to infuse the MI300X with more on-chip memory than competitors. More memory means you can fit larger models on the same chip, thus reducing the electrical distance between memory and compute engine, where inferences are performed. More memory means lower latency and thus faster and cheaper inferences.
Meta is arguably the world’s best AI capital allocator, which shows that AMD’s tech truly has an edge in the market. While the rest of the world is stuck in training mode, Meta has already leveraged inference to make its Family of Apps tremendously addictive, accurately predicting what content billions of people across the globe want to watch next. Meta has become an inference machine, and the rest of the economy will soon follow suit.
As AMD continues investing in its software capabilities, the modular nature of its platform versus the monolithic nature of Nvidia’s gives it a clear advantage in inference. Additionally, we will likely see a wide variety of AI workloads emerge over the coming decades, each with its own specific computational requirements. The Meta partnership demonstrates AMD’s ability to pounce on new AI workloads quickly and at a marginal cost.
Inference is likely just the start of AMD’s AI journey. People don’t want frameworks or models with lots of parameters; they want AI to do the heavy lifting. With inference top-of-funnel for the entire AI industry, including its application layer, earning power will naturally accrue to the inference providers. Though hard to visualize now, I believe this dynamic will give AMD pricing power over Nvidia in the future.
I believe the inference market will yield many specialized inference workloads and having a modular platform will give AMD the ability to better serve each one. I also believe inference will be a big business for Nvidia, but the monolithic approach will prove to be more limited in its ability to cater for the needs of customers in an individualised way. Tailored compute commands a premium.
While analysts continue to downgrade AMD stock, datacenter revenue has grown at a 64.91% CAGR since 2020. And still, the numbers shown in the graph do not begin to factor in AMD’s inference advantage. Over the coming years, I believe we will see Nvidia-like revenue growth. Additionally, the story doesn’t stop with their GPUs–AMD has been working on a relatively secretive technology that presents even further upside.
Xilinx is the world’s leading FPGA company, and by a wide margin. By acquiring Xilinx in 2022, AMD set itself up to dominate a new $200B business: AI at the edge. FPGA stands for field programmable gate array, which is basically a chip that’s able to change shape on the go. This kind of chip allows AMD to optimize AI workloads at the circuit level in a way that no other kind of chip can.
As AI makes its way from the datacenter out into the edge/devices, FPGAs will be the only chip capable of running inferences at the edge economically. For a device to run any kind of neural network efficiently on-demand, you need a chip that changes shape on the go to minimize the distance electrical signals travel and thus any energy wasted in the form of heat. AMD’s acquisition of Xilinx means only they have the tech to make this happen right now.
In Q3 2024, Lisa Su announced that AMD’s FPGA tech will be powering SpaceX’s next-generation satellites. Much like the Meta deal, I believe this is a sign of the great things to come for AMD’s adaptive technology. I believe this technology will further enhance AMD’s inference advantage, making it much harder to replicate for competitors.
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
Thank you for the article.
This may be a stupid question, but I don't understand why AMD and Nvidia are seen as having so much potential. They only design chips, they don't produce them. Where is the moat? Alphabet, Amazon etc. are still buying chips in large quantities. However, they are already working on their own chips, so they are turning from customers into competitors (at least in certain areas).
Chips will increasingly become commodities, especially when large companies can develop the chips themselves thanks to AI.
AMD may be better at inferences than Nvidia, but why shouldn't Nvidia be able to catch up with all the financial resources. We are still at the beginning of AI chip developments.
Where is the moat for AMD?
Thank you.
Gracias por tu análisis Antonio. Da gusto como profundizas en los temas. Enhorabuena por tu disciplina