The ROIC tier framework is the most useful lens here. It maps almost perfectly to moat durability.
Several of those same names sit at the top of the Soji QualityScore for exactly this reason. ROIC is the single metric that most reliably separates businesses worth owning from ones just riding a cycle.
Denis, your four-vendor analysis is clean. I am contemplating whether this analysis signals a change of dynamics in my industry — Telecoms. In B2B tech the entity which has the current pricing power is not the same as the entity that captures durable margin. Hardware is a commodity and receives the highest pricing pressure. The concentration we see today in my view generates a short window of competitive advantage until the market gets normalized. The durable margins will migrate to whoever owns the use-case outcomes — that is where the real business value exists. Curious whether you see the same dynamic in the AI stack, or whether you think NVIDIA's position is structurally different because of CUDA-style lock-in plus the pace of model evolution.
You are right! I worked for quite a lot in the european telco space as an M&A banker and i can see the parallel. It is true on the inference level whre price per token is the only thing that matters - Nvidia is already under pressure. But on the Traning/Invoation layer I think Nvidia transformed hardware into an integrated system with serious swtiching costs that are indeed CUDA related. Also for frontier model training and custom kernel development (FlashAttention 3), CUDA remains the exclusive language. All in all, in my view the real threat to NVIDIA isn't a better GPU from AMD etc.. it’s the vertical integration of the hyperscalers.
Denis, your refinement is precise. The integrated stack is genuinely the value-creation layer for now — it has moved away from the hardware commodity layer, and that is where the AI models are produced. CUDA + cuDNN + the custom-kernel ecosystem is an engineering-years lock-in, and hyperscaler vertical integration is the disruption vector to watch — not a better GPU.
Profits dominate at this layer short-to-medium term, one to three years. As the model supply normalises and inference cost-per-token collapses — already happening — the AI model itself becomes a commodity input, and the integrated stack commoditises with it. The differentiator then moves to the services layer — usage and outcomes wrapped around the model, vertical applications, regulated-industry deployments, outcome-based commercial models. The services layer captures the durable premium long term, seven to ten years out. The entire compute layer, Nvidia and hyperscalers combined, prices like infrastructure. Utility margins at the compute layer. Premium margins at the services layer where the outcome is sold.
The telecom precedent maps almost line for line. I will come back to it properly in a piece I am developing.
This mapping of the infrastructure stack is incredibly useful. The breakdown of the custom silicon challengers attempting to bypass the CUDA software moat clarifies the exact battle lines of the next five years. We often talk about compute as an abstract resource, but this piece grounds it in the physical supply chains that actually build it. I write fiction about the reality of these very constraints, and your analysis is a fantastic grounding point.
No disrespect, but the reason you won’t say which moat and what extraction is purely because of the many ways in which bubbles swirl.
This breakdown, of the smell of individual notes-of-shit on the same shovel, is truly nought but a smell enhancer.
This AI long-game, it ends in a short, with any moat surrounding any castle left standing filled to-brimming with drowned kingpins and eye-pecking birds.
As human, why bet against yourself right now?
Win this bet, you know you’ll be bedding down for the rest of your life with losers.
The ROIC tier framework is the most useful lens here. It maps almost perfectly to moat durability.
Several of those same names sit at the top of the Soji QualityScore for exactly this reason. ROIC is the single metric that most reliably separates businesses worth owning from ones just riding a cycle.
Denis, your four-vendor analysis is clean. I am contemplating whether this analysis signals a change of dynamics in my industry — Telecoms. In B2B tech the entity which has the current pricing power is not the same as the entity that captures durable margin. Hardware is a commodity and receives the highest pricing pressure. The concentration we see today in my view generates a short window of competitive advantage until the market gets normalized. The durable margins will migrate to whoever owns the use-case outcomes — that is where the real business value exists. Curious whether you see the same dynamic in the AI stack, or whether you think NVIDIA's position is structurally different because of CUDA-style lock-in plus the pace of model evolution.
You are right! I worked for quite a lot in the european telco space as an M&A banker and i can see the parallel. It is true on the inference level whre price per token is the only thing that matters - Nvidia is already under pressure. But on the Traning/Invoation layer I think Nvidia transformed hardware into an integrated system with serious swtiching costs that are indeed CUDA related. Also for frontier model training and custom kernel development (FlashAttention 3), CUDA remains the exclusive language. All in all, in my view the real threat to NVIDIA isn't a better GPU from AMD etc.. it’s the vertical integration of the hyperscalers.
Denis, your refinement is precise. The integrated stack is genuinely the value-creation layer for now — it has moved away from the hardware commodity layer, and that is where the AI models are produced. CUDA + cuDNN + the custom-kernel ecosystem is an engineering-years lock-in, and hyperscaler vertical integration is the disruption vector to watch — not a better GPU.
Profits dominate at this layer short-to-medium term, one to three years. As the model supply normalises and inference cost-per-token collapses — already happening — the AI model itself becomes a commodity input, and the integrated stack commoditises with it. The differentiator then moves to the services layer — usage and outcomes wrapped around the model, vertical applications, regulated-industry deployments, outcome-based commercial models. The services layer captures the durable premium long term, seven to ten years out. The entire compute layer, Nvidia and hyperscalers combined, prices like infrastructure. Utility margins at the compute layer. Premium margins at the services layer where the outcome is sold.
The telecom precedent maps almost line for line. I will come back to it properly in a piece I am developing.
The AI supply chain is complex, and you’ve communicated it very clearly and effectively. I learnt a lot - great piece
This mapping of the infrastructure stack is incredibly useful. The breakdown of the custom silicon challengers attempting to bypass the CUDA software moat clarifies the exact battle lines of the next five years. We often talk about compute as an abstract resource, but this piece grounds it in the physical supply chains that actually build it. I write fiction about the reality of these very constraints, and your analysis is a fantastic grounding point.
No disrespect, but the reason you won’t say which moat and what extraction is purely because of the many ways in which bubbles swirl.
This breakdown, of the smell of individual notes-of-shit on the same shovel, is truly nought but a smell enhancer.
This AI long-game, it ends in a short, with any moat surrounding any castle left standing filled to-brimming with drowned kingpins and eye-pecking birds.
As human, why bet against yourself right now?
Win this bet, you know you’ll be bedding down for the rest of your life with losers.
Take the short.
Only Chuck Norris can short this market