Interview, Fireside Chat
Dylan Patel on the AI Chip Race - NVIDIA, Intel & the US Government vs. China
- Intel faces potential shareholder dilution during future capital raises unless it secures major investment announcements from entities like Apple or Trump-backed investors to reach an estimated $50 billion target; the company may also abandon internal graphics and AI efforts like the Gaudi program due to high-end competitiveness issues.
- The NVIDIA-Intel partnership is projected to create an optimal x86 laptop market product while potentially undermining ARM's positioning and creating significant headwinds for AMD by combining their respective strengths.
- Huawei is expected to reach logic chip production capacity replacing TSMC within the next couple of years, though it faces bottlenecks in custom HBM manufacturing due to limited domestic capacity and etching equipment constraints despite recent import increases.
- China may experience a temporary supply gap in domestic AI chips, potentially leading to a reversal of NVIDIA bans and continued low-to-moderate volume smuggling of restricted chips from other countries.
- Hyperscaler capital expenditures are projected to reach $360 billion annually with a speaker estimate of $450–$500 billion next year, potentially growing to multiple trillions annually in the long term, driven by demand for AI infrastructure.
- OpenAI's annual recurring revenue is forecast to reach $35–$45 billion by the end of next year, though the company is expected to burn $15–$25 billion in cash annually and remain unprofitable until 2029.
- NVIDIA is likely to adopt a demand-betting strategy involving pre-orders of production capacity and feature cuts to ensure rapid time-to-market, risking cumulative inventory write-downs of many billions of dollars historically.
- The GPU market is expected to remain tight for large volume purchases similar to the 2023 era, with Hopper prices rising as Blackwell deployment faces capacity crunches and reliability learning curves involving single-point failures in 72-GPU systems.
- Inference workloads may become significantly cheaper and more efficient through the new CPX chip architecture, which strips out expensive HBM, while Hopper GPU prices are anticipated to have bottomed months ago.
- Amazon is projected to re-accelerate revenue growth to over 20 percent next year, leveraging spare data center capacity and Traneum GPUs, though its internal AI chips remain difficult to use for inference compared to NVIDIA's ecosystem.
- Oracle is expected to capture the AI compute market by remaining hardware-agnostic and is projected to sign an over $80 billion annual deal with OpenAI, necessitating significant debt financing starting in 2027–2029.
- Elon Musk is expected to lead in gigawatt-scale data center construction by leveraging cross-border regulations to move facilities, while Huawei remains a formidable competitor in foreign markets such as the Middle East and Southeast Asia.
- NVIDIA may backstop the AI ecosystem by investing directly in startups like Anthropic or Xai and shifting focus to the data center and energy layers, which are becoming primary bottlenecks, while the speaker cannot predict market dynamics beyond a five-year horizon.
- The consensus view suggests China will backtrack on NVIDIA bans if domestic supply cannot meet demand, and ByteDance continues to prioritize NVIDIA chips for performance over domestic alternatives, while the US government may relax export restrictions as a negotiating tactic.