Video summary: this article is a detailed summary of the video “AMD Beats Nvidia On Every Spec And Still Loses”, from the channel The Stack.
Video URL: https://www.youtube.com/watch?v=bOHhUVSDmrw
On paper, there should not even be a contest: the accelerator AMD MI300X carries 192 GB of memory versus 80 GB on Nvidia's H100, moves 5,3 TB/s of bandwidth, and delivers 1,31 petaflop — winning line by line on the datasheet. AMD itself raised the stakes in an official announcement: the Instinct MI325X outperforms the H200, precisely the chip the industry was lining up to buy. Even so, Nvidia closed 2025 with 92% of the enterprise GPU market — of every 100 GPUs in server racks, 92 are green. The Stack's video dissects why better specifications lose to network effects — and the answer is not in the silicon, it is in the software.
The moat is called CUDA
Anyone buying a cluster worth tens of millions of dollars is not buying a chip — they are buying compatibility. Nvidia launched CUDA in 2007, when GPUs for general-purpose computing were still an odd bet. Almost twenty years later, that bet became the industry's most effective castle moat: PyTorch, TensorFlow, and JAX — the three frameworks supporting modern AI — assume CUDA by default. The result: an estimated 90% of all AI training workloads run on Nvidia.
AMD's answer, ROCm, only arrived in 2016 — nine years late. In hardware, missing a window costs one product cycle; in software, nine years is a geological era: enough time for the incumbent to dictate how an entire generation of engineers learns to program, how universities teach, and how the surrounding tooling is built. Then there is the gravity of the installed base: with half the world's installed AI chips on Nvidia silicon, every optimization trick and every weekend project on GitHub is born tested on the architecture the developer has at hand — making the platform even safer for the next buyer, who expands the base again. It is a self-reinforcing cycle.
The three-headed hydra
Analysts describe Nvidia's structural advantage as a three-headed hydra: the chip is only one head. The second is the network layer that stitches thousands of GPUs into a single cluster (NVLink, InfiniBand); the third is the software stack on top of it all. A competitor matching its computing power still needs to beat the other two heads — and that is where the math falls apart.
Five months of benchmarking, and a painful verdict
The most forceful part comes from SemiAnalysis, which spent five months testing AMD hardware with direct support from engineers at both manufacturers — in other words, giving AMD every chance to optimize. The conclusion: the silicon is as fast as advertised, but the software stack is so “riddled with bugs” that the time saved on hardware evaporated while chasing strange errors, turning extremely expensive engineers into full-time IT support. The published verdict was blunt: AMD has not yet crossed the CUDA moat. Buying the most powerful chip gets expensive when the real cost is skilled people's hours spent putting out fires.
Attacking the moat from below
If the problem is software, software can be fixed — but rewriting code solves only half the equation; the other half is twenty years of accumulated habit. Two efforts attack exactly that:
- SCALE (Spectral Compute) — a compatibility layer that takes your CUDA binaries without any modification and runs them on AMD GPUs. Translation happens at compile time: PyTorch believes it is talking to an Nvidia chip while an MI300X does the computation underneath. No migrating models, maintaining two codebases, or retraining the team
- CANN (Huawei) — the toolkit for Ascend accelerators, opened in August 2025 specifically to test whether giving the stack away beats developers' muscle memory
If “run CUDA anywhere” becomes reliably real, the moat stops protecting the castle — the same move Wine/Proton made with Windows games on Linux.
The market is already cracking — but not for AMD
Nvidia's dominance fell from its peak of 87% to a projection of 75% in 2026 — 12 points of the planet's most profitable hardware sector looking for a new home. The catch: AMD is not capturing the difference. Even with a quarter of US$ 5,8 billion in data centers, its volume does not explain the hole in the spreadsheet. What explains it is the in-house chips from Google (TPU), Meta, Amazon, and Microsoft, growing 44,6% per year versus 16,1% for off-the-shelf GPUs.
And the homegrown chip has an unfair software advantage: it only needs to run the company's own internal models. It does not need to support PyTorch for the world or convince the global market to change habits — its target audience is a single building, and internal engineers optimize the closed stack and ignore the ecosystem. The real threat to Nvidia is not its usual competitor — it is its biggest customers ceasing to be customers.
The lesson for builders
The video's lesson applies to any infrastructure layer: ecosystem beats specifications. A bug can be fixed with a patch; asking an entire department to switch the tools it has mastered is harder than signing the check for a platform it already trusts. True for GPUs, languages, clouds — anyone who has migrated distros knows: the cost is never the kernel, it is everything built around it. For those of us in the Linux world, there is an upside — ROCm, CANN, and SCALE are open bets against a proprietary monopoly, and this entire battle is being fought on the penguin: virtually all AI training on the planet runs on Linux, whether the chip is green, red, or homegrown.
The original video is worth its 17 minutes — and if you run local AI on your hardware, the post Running local AI on Linux with Ollama is the practical companion.