Can AMD break the CUDA Moat? AMD Advancing AI 2026

Bryan Shan · SemiAnalysis · 25 июля 2026, 00:33 · ⏱ 70 мин чтения  | Читать в Substack ↗
Резюме
AMD has a credible path to gain market share in AI accelerators due to hardware leadership (MI455X, Helios rack) and improving software, but faces two major risks: Helios production ramp challenges (backplane retimers, cable-based design) and a persistent shortage of stable GPU clusters for internal development and CI testing. The article argues that the CUDA moat is eroding with agentic tools and open-source contributions, and that both AMD and NVIDIA can win as the pie grows, but AMD's execution on distributed inference and stable infrastructure is critical.
  • AMD's MI455X delivers 20 PF FP8 vs Nvidia's Rubin at 17.5 FP8, with 432 GB HBM4 vs Nvidia's 288 GB, but AMD's GPU microarchitecture trails Nvidia (e.g., lacks 3-bit LUT tensor cores).
  • AMD's Helios rack requires up to 85% of backplane links to be retimed due to weak SerDes, needing over 550 Broadcom ethernet retimers per rack, adding cost and complexity.
  • AMD is giving Meta and OpenAI close to a 105% equity rebate discount via finance engineering (stock options tied to $600 AMD stock price), effectively making Helios 'practically negative cost'.
  • Anthropic has publicly committed to deploying 2 GW of AMD's chips, and Microsoft (for OpenAI) will deploy MI455X Helios after skipping MI325X and MI355X.
  • AMD's internal software development is hindered by a lack of stable GPU clusters; planned ETA for CI parity (vLLM gating, Kubernetes NIC testing) with Nvidia was missed, and clusters are frequently reallocated.
  • Meta's custom cut-down MI455 (half compute, half HBM) is criticized as a 'bad decision' that will push Meta toward Nvidia's Rubin, reducing AMD's volume at Meta.
  • AMD's software stack is improving: AITER/vLLM optimizations showed up to 18× improvement on Kimi K2.5, but distributed inference (WideEP, disaggregated prefill/decode) still lags behind Nvidia's ecosystem.
  • Nvidia's NIXL library accepted upstream contributions from AMD's RIXL fork, enabling AMD's KV cache transfer to work within Nvidia's framework, a positive step for interoperability.
Время чтения 70 мин
Объём 70,851 симв.
Категория finance
Идеи
Bryan Shan Автор Substack, SemiAnalysis
The article presents a cautiously bullish view on AMD's AI accelerator prospects, citing hardware leadership (MI455X), improving software, and aggressive financial engineering. However, it identifies
The article presents a cautiously bullish view on AMD's AI accelerator prospects, citing hardware leadership (MI455X), improving software, and aggressive financial engineering. However, it identifies two major risks (Helios production ramp and GPU cluster shortage) that could derail progress. The author states 'we strongly believe that AMD will be well positioned to do well and take market share as long as AMD solves the 2 major risks'. Risk: Execution on Helios rack manufacturing and provision of stable internal GPU clusters remain unresolved; any delays could widen the gap with Nvidia.
Bryan Shan Автор Substack, SemiAnalysis
The article acknowledges that 'NVIDIA will continue to massively grow revenue' and that 'the pie is getting massive', while noting that AMD's competition will pressure Nvidia to move faster. Comparati
The article acknowledges that 'NVIDIA will continue to massively grow revenue' and that 'the pie is getting massive', while noting that AMD's competition will pressure Nvidia to move faster. Comparative specs show Nvidia's Rubin still leads in 3-bit LUT tensor cores and system integration, and Nvidia's software moat remains strong, particularly in disaggregated inference where Nvidia has a multi-year head start. Risk: AMD's improved hardware and software could erode Nvidia's market share over time, especially if Nvidia's internal bureaucracy slows innovation.
Bryan Shan Автор Substack, SemiAnalysis
The article highlights that AMD is the first to ship 2nm datacenter silicon (N2) for MI455's compute tiles and Venice CPU, and uses TSMC's CoWoS-L with active LSI (first shipping implementation). TSMC
The article highlights that AMD is the first to ship 2nm datacenter silicon (N2) for MI455's compute tiles and Venice CPU, and uses TSMC's CoWoS-L with active LSI (first shipping implementation). TSMC is the exclusive supplier for these advanced packaging and process nodes, directly benefiting from AMD's ramp. Risk: No direct risk; TSMC benefits from both AMD and Nvidia demand.
Bryan Shan Автор Substack, SemiAnalysis
AMD's Helios rack uses 12 Tomahawk6 switches from Broadcom (102.4T each) and over 550 Broadcom ethernet retimers per rack for backplane signal integrity. Broadcom is the sole merchant supplier for the
AMD's Helios rack uses 12 Tomahawk6 switches from Broadcom (102.4T each) and over 550 Broadcom ethernet retimers per rack for backplane signal integrity. Broadcom is the sole merchant supplier for these high-speed switches and retimers, capturing significant content per rack. Risk: Dependence on Broadcom's ability to deliver sufficient volumes for AMD's ramp; any supply constraints could impact Helios production.
Bryan Shan Автор Substack, SemiAnalysis
The article notes 'tight memory supply' and that HBM4 suppliers had to rework their designs to meet Nvidia's higher pin speeds. AMD's MI455 uses 12 stacks of HBM4, and the industry-wide HBM ramp benef
The article notes 'tight memory supply' and that HBM4 suppliers had to rework their designs to meet Nvidia's higher pin speeds. AMD's MI455 uses 12 stacks of HBM4, and the industry-wide HBM ramp benefits memory manufacturers like Micron. The article implies HBM4 supply is a bottleneck that validates strong demand for HBM suppliers. Risk: Memory pricing and supply allocation between AMD and Nvidia could create volatility; Micron's dependency on a few large customers.
Ещё от SemiAnalysis

This newsletter, published July 25, 2026, features Bryan Shan discussing AMD, NVDA, TSM, AVGO, MU. 5 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Bryan Shan  · Tickers: AMD, NVDA, TSM, AVGO, MU