AMD Is Becoming an AI Powerhouse Because of SemiAnalysis
The catalyst that made AI hardware a competitive market again
Happy Sunday and welcome to Investing in AI. Check out our new AI in NYC podcast, and new video shorts we are putting up on Youtube. Here is one of me explaining why I think most startup advice is crap. Also check out our new research at Neurometric showing that the AI jagged frontier exists in a technical sense as well as an applied AI sense.
Today I want to talk about how AMD’s GPU business has quietly become a real competitor to NVIDIA, and the stock might be undervalued because everyone is still stuck thinking about the market the way it looked 12 months ago.
The catalyst for turning around AMD’s business, from what I can tell, was a post from SemiAnalysis last December about how hard it was to work with the AMD software stack.
This caught the attention of Lisa Su who took it seriously and made it a top priority to fix. Now if you follow the AMD chatter online, there has been a big shift in the excitement of people working with AMD GPUs.
For the last few years the conventional wisdom was NVIDIA has a monopoly on AI compute, their CUDA moat is unbreakable, and AMD is just making noise on the sidelines. The numbers seemed to support this. NVIDIA has something like 90% market share in data center GPUs, and their H100s and H200s were the gold standard for training large language models.
But if you dig into what’s actually happening in AI infrastructure right now, a different picture emerges. AMD has fixed their software stack at a key time in the industry growth trajectory. The market is splitting into two distinct workloads: training and inference. Training is where you build the model, and NVIDIA still dominates here. But inference—running the models for actual users—is where the real volume is. And for inference, AMD’s MI300 series chips are extremely competitive.
The math is pretty straightforward. AMD’s MI300X delivers comparable performance for many inference workloads, but at a significantly cheaper cost. When you’re Meta or Microsoft running millions of inference requests per day, that cost difference adds up fast. It’s not about raw performance anymore. It’s about performance per dollar, and AMD wins that calculation for a lot of use cases.
What’s more interesting is that the software moat everyone talks about is eroding faster than people realize. AMD’s ROCm platform used to be a disaster, but it’s gotten genuinely good over the past year thanks to the call out by SemiAnalysis. More importantly, the major AI frameworks like PyTorch and the inference platforms like vLLM now support AMD chips pretty seamlessly. You’re not betting on AMD catching up to CUDA anymore. You’re betting on “good enough” being, well, good enough when there’s a 30-40% cost savings attached.
The customer wins are real too. Microsoft is deploying MI300 chips at scale. Meta has publicly committed to using them. Oracle Cloud offers them. These aren’t pilot programs or PR announcements. These are production deployments at companies that live and die by their infrastructure costs.
I’m not saying AMD is going to overtake NVIDIA anytime soon. NVIDIA is the better company with better execution and a bigger ecosystem. But the market is undervaluing what AMD could become. AMD’s data center GPU revenue is growing at triple digit rates off a small base, and the TAM for AI inference is massive and growing.
Sometimes the best opportunities in tech aren’t betting on the obvious winner.
Thanks for reading.
The inference versus training split is the underappreciated part of this story. Meta's commitment to AMD chips at scale makes perfect economic sense when you're running billions of inference requests where every marginal dollar compounds. The 30-40% cost savigs on comparable performance is real money at their scale. AMD's timing with the software fixes was impeccable, catching the wave right as inference workloads are exploding.
Meta's public commitment to deploying AMD MI300 chips at scale validates your thesis that the software moat is eroding faster than the market realizes. The cost arbitrage you've outlined becomes exponentially more valuable when you consider that inference workloads are growing faster than training and will likely represent 80%+ of compute demand within 18 months. What's particularly astute is recognizing that 'good enough' plus 30-40% cost savings beats perfection in infrastructure economics, especiay when companies like Meta are running hundreds of millions of requests daily where marginal costs compound quickly. The timing of AMD's software fixes coinciding with the inference workload explosion is either incredibly lucky or brilliant positioning, but either way it's creating a genuine competitive dynamic that should compress NVIDIA's margins over the next few quarters.