Nearly half of Nvidia's revenue comes from four mystery whales each buying $3B+

Original Article ↗ Hacker News Discussion ↗

Nearly half of Nvidia’s revenue now comes from just four huge customers buying billions of dollars’ worth of AI chips each quarter, likely major cloud and social platforms racing to build and rent out AI capacity. Commenters debate whether this concentration signals an unsustainable AI bubble or a durable shift toward GPU‑driven computing, comparing it to past manias like dot‑coms, railroads, and crypto. They also examine Nvidia’s CUDA software moat, the slow rise of in‑house and custom accelerators, and what a future glut of data‑center GPUs could mean for open‑source AI, hobbyists, and gamers.

Debate over an AI bubble

  • Many see a “massive AI bubble,” likening it to dot-com, 1990s rail/telecom overbuild, or 80s AI winter: real tech, but overhyped and overfunded, followed by funding collapse and consolidation.
  • Others argue this isn’t like pure-speculation bubbles (NFTs/crypto/tulips) because Nvidia and many AI products already generate substantial revenue and real usage.
  • Several expect a pop in valuations and capex growth, not in the underlying tech, similar to how the internet thrived after the dot-com crash.

Nvidia, GPU demand, and future glut

  • Nvidia’s profits and margins are viewed as “insanely high”; some expect competition or a capex slowdown to compress them.
  • Commenters anticipate that any severe shortage will eventually be followed by a glut of used datacenter GPUs (like post-crypto), though quality and reliability of ex-datacenter cards are debated.
  • Gamers hope for cheaper GPUs, but many doubt much “trickle down” due to segmentation and higher-margin datacenter priority.

Who are the “mystery whales”?

  • Most assume the big four buyers are hyperscalers: Microsoft, Meta, Google, Amazon; some articles cited in-thread say exactly that.
  • Other large buyers mentioned: Oracle, CoreWeave, Lambda, Chinese cloud companies, Tesla/xAI.
  • Some speculate about indirect government/NSA/DoE demand, but note such purchases would likely be routed through intermediaries.

Custom silicon, CUDA moat, and competition

  • Cloud and big-tech efforts: Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia, Tesla D1, plus specialized players (Groq, Cerebras, etc.).
  • CUDA is seen as a major moat; AMD/Intel and others have struggled to attract training workloads despite hardware.
  • Ideas discussed: CUDA-compatibility layers and open alternatives to gradually weaken Nvidia lock-in; skepticism remains about difficulty and incentives.

Open vs closed and self-hosted AI

  • Some companies are commissioning ~$20k in-house AI servers running open-source models, citing flexibility and richer APIs than proprietary services.
  • There’s uncertainty whether proprietary “frontier” models or a diverse open-source ecosystem will dominate long term.

Use cases, productivity, and limits

  • Reported uses: search/lookup, translation, moderation, coding assistance, writing, design, data analysis, medical and legal document work, recommendation systems.
  • Individual experiences vary: some feel AI is a “new power” and increasingly indispensable; others see only modest productivity gains and fading novelty (e.g., image generation).
  • Open questions raised:
    • Whether LLM quality is plateauing and hallucinations can be tamed enough for broad deployment.
    • How big non-LLM markets (robotics, autonomy, scientific computing, drug discovery) will be, and whether they sustain current GPU growth.

Concentration and systemic concerns

  • Worry that AI progress and infrastructure are consolidating into a handful of tech giants and clouds; some advocate open source and Linux as partial counterweights.
  • Others note that even if this is a bubble, like railroads or dark fiber, the overbuilt infrastructure could still provide long-term economic benefit.