The impact of competition and DeepSeek on Nvidia

Original Article ↗ Hacker News Discussion ↗

Nvidia’s towering AI valuation is being re‑examined in light of new models like DeepSeek, which claim GPT‑4‑class performance at a fraction of the training and inference cost. Commenters weigh whether such efficiency gains will compress Nvidia’s margins and reduce demand for its highest‑end GPUs, or instead unleash Jevons‑style growth where cheaper AI leads to vastly more usage and ultimately more chips sold. The thread also explores Nvidia’s broader moat in software (CUDA, drivers, ecosystem), emerging competition from AMD, Apple and Chinese players, and how quickly large labs can absorb DeepSeek-style optimizations.

Thread reception and meta

  • Many readers praise the article as one of the clearest, most comprehensive breakdowns of the GPU/AI landscape, though some feel the title under‑sells the breadth of content.
  • A few note technical quibbles (e.g., precision history, driver quality), but generally see it as informed and nuanced rather than typical finance-guy hot take.

Nvidia valuation and investment debate

  • Broad agreement that Nvidia is “priced for perfection” and highly sensitive to any slowdown in growth, margin compression, or loss of share.
  • Disagreement on overvaluation: some argue current P/E isn’t obviously excessive given growth; others say eventual commoditization and physical/power limits make current prices unsustainable.
  • Comparisons with AMD: some see AMD as the better risk/reward; others warn simple P/E comparisons are naive and expectations matter more.

DeepSeek’s impact on demand and economics

  • One camp: DeepSeek’s claimed ~45× training efficiency and much cheaper inference show massive over‑provisioning; future AI workloads may need far fewer top-end Nvidia GPUs, threatening margins.
  • Opposing camp: cites Jevons paradox – cheaper, more efficient models expand use cases and total consumption; efficiency will increase, not reduce, aggregate AI compute demand, likely helping Nvidia/TSMC over time.
  • Some stress that DeepSeek still uses Nvidia GPUs and that its main breakthrough is better algorithms and distillation, not non‑GPU hardware.

Competition and moats

  • Nvidia’s moat is seen as multi‑layered: CUDA, mature tooling, Linux drivers (for compute), software stack, and high-speed interconnect/Mellanox.
  • Counterpoints: higher-level frameworks and compilers (MLX, Triton, JAX, etc.) could erode CUDA lock‑in; cloud and hyperscaler custom silicon (TPUs, Apple, Huawei/China) may slowly chip away at Nvidia over years.
  • AMD is viewed as real but lagging competition; ROCm and drivers draw mixed reviews, with anecdotes ranging from “hilariously bad” to “works great for desktop/gaming.”

Technical debates around DeepSeek

  • Clarifications that mixed-precision and sub‑FP32 training have been used for years; DeepSeek pushes further (e.g., FP8 training, MoE routing, multi-token prediction, RL without labeled supervision).
  • MoE discussion: generally agreed it saves per-token compute and bandwidth, not total VRAM (experts still loaded across GPUs; batching and routing matter).
  • Some argue DeepSeek bundles many existing efficiency tricks (also seen in Llama) more aggressively rather than inventing something wholly new.

Infrastructure and physical limits

  • Disagreement on whether electricity and cooling/water will be real constraints for Nvidia’s projected growth.
  • Some argue current valuations implicitly assume AI datacenter power use can scale orders of magnitude, which skeptics doubt; others point to rapidly falling solar costs and geographic flexibility of training as mitigating factors.

Broader implications and sentiment

  • Several see DeepSeek as accelerating commoditization of frontier models and compressing model-provider margins more than harming chipmakers.
  • Others worry about Chinese strategic advantages, possible state backing, and potential propaganda/astroturf around DeepSeek’s narrative.
  • A recurring theme is unease at AI’s pace and capital intensity, contrasted with excitement that efficiency gains might democratize model training and enable more players beyond mega‑caps.