Ep. 027 - OpenAI Jalapeño: Better Than Nvidia Blackwell (Accelerators)
Спикеры
The podcast covers SemiAnalysis's article on OpenAI's self-designed Jalapeño accelerator, which beats Nvidia's GB300 and July Vera Rubin results on tokens per megawatt and inference TCO. The speakers discuss benchmark caveats, CUDA moat erosion, AI-assisted chip design and kernel bring-up, and the architecture that gives Jalapeño its efficiency. They also detail why Samsung's HBM4 is winning and what the custom ASIC shift means for Nvidia, Broadcom, and AMD.
- OpenAI Jalapeño is a credible ASIC that beats Nvidia GB300 and early Vera Rubin inference results on efficiency.
- The performance debate centers on tokens per megawatt, TCO, and interactive versus throughput tokens.
- CUDA moat is eroding as AI-assisted programming and custom ASICs lower software lock-in.
- Samsung HBM4 leads with advanced 1C DRAM and SF4 base die, ahead of SK hynix and Micron.
- The custom ASIC economics benefit Broadcom and reflect a broader vertical integration trend.
- Anthropic is using AMD as AI-assisted porting reduces software-stack barriers.
- Meta and Microsoft's silicon teams look weakest after the OpenAI reveal.
- The episode ends with system scaling challenges and Doom running on Jalapeño.
Идеи
Custom ASIC shift benefits Broadcom's design margins.
Custom in-house AI accelerators can lower hyperscaler chip costs because the buyer pays Broadcom's design margin only, rather than NVIDIA's or Broadcom-plus-Google-TPU margins. OpenAI's Jalapeño TCO advantage on tokens per dollar versus GB300 and Vera Rubin supports this custom ASIC economics, making Broadcom a key beneficiary of hyperscaler silicon programs.
OpenAI Jalapeño threatens Nvidia's CUDA moat.
OpenAI's Jalapeño is the first non-NVIDIA, non-AMD accelerator to beat NVIDIA's GB300 and Vera Rubin on public inference benchmarks, winning on both throughput per megawatt and interactivity. This signals that the CUDA moat is eroding as AI-assisted programming and vertically integrated ASICs reduce NVIDIA's software lock-in, though NVIDIA's broader ecosystem and supply chain still make it valuable.
AI-assisted porting reduces AMD software disadvantage.
Anthropic is bringing in AMD as a hardware provider because agentic AI programming lets labs bypass AMD's software stack challenges. This suggests AI-assisted kernel development is eroding NVIDIA's software moat and expanding the addressable market for AMD accelerators.
Samsung leads HBM4; SK hynix, Micron lag.
Samsung has emerged as the HBM4 leader because its DRAM dies are built on a more advanced 1C process and its base die uses Samsung Foundry SF4 4nm, while SK hynix and Micron are on older processes and have faced HBM4 delays or redesigns. Samsung's HBM4 supplies OpenAI Jalapeño/Broadcom with 15.4 TB/s bandwidth, higher than expected NVIDIA Rubin HBM4, so Samsung is the HBM4 winner while SK hynix and Micron lag.
Samsung leads HBM4; SK hynix, Micron lag.
Samsung has emerged as the HBM4 leader because its DRAM dies are built on a more advanced 1C process and its base die uses Samsung Foundry SF4 4nm, while SK hynix and Micron are on older processes and have faced HBM4 delays or redesigns. Samsung's HBM4 supplies OpenAI Jalapeño/Broadcom with 15.4 TB/s bandwidth, higher than expected NVIDIA Rubin HBM4, so Samsung is the HBM4 winner while SK hynix and Micron lag.
Далее
This SemiAnalysis video, published August 29, 2026, features Jordan Nanos, Myron discussing AVGO, NVDA, AMD, 005930.KS, MU, 000660.KS. 5 trade ideas extracted by AI with direction and confidence scoring.
Speakers: Jordan Nanos, Myron · Tickers: AVGO, NVDA, AMD, 005930.KS, MU, 000660.KS