Damnang
· Damnang’s Substack
· April 17, 2026 at 09:15
· ⏱ 10 min read
| Read on Substack ↗
Summary
The article argues that asking 'Can AMD beat NVIDIA?' is the wrong question. Instead, Lisa Su is executing a multi-pronged strategy—securing long-term hyperscaler partnerships via equity warrants, building integrated rack systems (Helios), and leveraging CPU strength for agentic AI—that positions AMD for sustainable revenue growth even if it never overtakes NVIDIA's market share. For markets, this implies AMD's revenue trajectory is becoming more predictable from multi-year contracts, but NVIDIA's ecosystem moat remains deep.
•AMD signed 6GW GPU supply contracts with OpenAI and Meta, each granting 160 million warrants at $0.01, structured as joint investments rather than discounts.
•The MI455X GPU carries 432GB of HBM4 with 19.6 TB/s bandwidth, roughly 1.5x ahead of NVIDIA Vera Rubin on memory capacity by AMD's claims.
•In MLPerf Inference 6.0, MI355X tied B200 on Llama 2 70B Offline and beat B200 on GPT-OSS-120B (111% Offline, 115% Server), but share remains ~80% NVIDIA vs ~10% AMD.
•Rumors include MI455X potential delay, NVIDIA Vera Rubin pulled forward, Microsoft developing a CUDA-to-ROCm conversion toolkit, and Anthropic possibly being the next hyperscaler to sign a deal.
•The author interprets the 20% warrant equity as a joint investment mechanism that locks in hyperscaler demand, not a bearish signal of weak organic demand.
The article details AMD's closed spec gap (MI455X 1.5x memory bandwidth ahead, MLPerf wins) and a strategic shift to long-term hyperscaler contracts with warrants as joint investments, positioning AMD
The article details AMD's closed spec gap (MI455X 1.5x memory bandwidth ahead, MLPerf wins) and a strategic shift to long-term hyperscaler contracts with warrants as joint investments, positioning AMD for predictable multi-year revenue growth. Author states 'The gap closed on specs' and highlights the Helios rack as an integrated system that goes head-to-head with NVL72.
Risk: AMD's market share remains stagnant at ~10% despite spec parity; NVIDIA's ecosystem moat (CUDA, TensorRT, vendor lock-in) could limit adoption; MI455X delay or Vera Rubin pull-forward could erode the window of opportunity.
The article quantifies AMD's competitive hardware performance (MI455X memory capacity 1.5x ahead, MLPerf inference benchmarks beating B200 on GPT-OSS-120B) and describes a strategic hyperscaler deal s
The article quantifies AMD's competitive hardware performance (MI455X memory capacity 1.5x ahead, MLPerf inference benchmarks beating B200 on GPT-OSS-120B) and describes a strategic hyperscaler deal structure that could lock in significant compute demand for AMD. While the article notes NVIDIA still commands ~80% share, the gap in performance is narrowing, potentially threatening NVIDIA's pricing power and growth trajectory.
Risk: NVIDIA's 20-year ecosystem moat (CUDA, TensorRT-LLM, NIM containers) remains intact for now; if Vera Rubin lands early with superior performance, AMD's window may close quickly.