Nvidia CEO Jensen Huang announces new AI chips: ‘We need bigger GPUs’

Original Article ↗ Hacker News Discussion ↗

Market reaction & valuation

  • Stock was flat-to-down around the keynote; some expected a “big pop” from major news.
  • Several argue near‑term price moves say little about product strength, but many see Nvidia as “hilariously overvalued” or “priced for perfection.”
  • Others counter that forward P/E is in line with recent years and that Nvidia currently controls a critical resource with extraordinary revenue and margins.
  • Historical tech bubbles (Cisco, dot‑com, crypto, Tesla) are cited as analogies: AI can be real and huge while Nvidia still crashes from bubble pricing.

Moat, competition, and CUDA

  • Many question whether Nvidia’s AI lead is durable, since NN hardware is simpler than graphics; others argue the real moat is:
    • CUDA and its ecosystem (cuDNN, TensorRT, libraries).
    • Mature drivers and software, where competitors repeatedly failed.
    • Massive R&D and TSMC capacity, and ownership of networking (Mellanox/Infiniband).
  • AMD, Google TPUs, and various startups are seen as inevitable challengers, but AMD’s underinvestment in software and buggy ROCm stack are repeatedly criticized.
  • Some predict future antitrust pressure to “open” CUDA, though others think that still wouldn’t fix AMD’s software gap.

Blackwell architecture, FP4, and performance claims

  • Raw uplift of ~2.5× FP8 vs Hopper is seen by some as underwhelming, especially since it’s effectively 1.25× per die in a dual‑die package.
  • Nvidia’s headline “30× inference” improvement is broadly viewed as narrow/marketing:
    • Relies on FP4, sparsity, packaging, networking, and a specific giant MoE model.
    • Not representative of general workloads; still impressive for hyperscale inference.
  • Debate over FP4:
    • Confirmed as 4‑bit float with new “precision‑aware” engine.
    • Seen as promising for inference; training at 4‑bit still niche/experimental.
    • Some question real‑world uptake; others note strong research momentum.

Platform strategy and NIM

  • Many see Nvidia “moving up the stack” from chips to a full platform:
    • Hardware: GPUs, NVLink, CPUs, NICs, switches, complete racks.
    • Software: CUDA, Triton, TensorRT‑LLM, and new NIM inference microservices.
  • NIM is described as “Docker for LLMs” with OpenAI‑compatible APIs, making it easy to swap from hosted APIs to on‑prem Nvidia stacks with minimal code change.
  • This threatens thin AI startups that just wrap models with simple UIs, and pushes more lock‑in for smaller customers, while hyperscalers may continue building their own accelerators.

Keynote style and “platform company” framing

  • Several viewers found the keynote awkward and under‑rehearsed; jokes often missed and pacing felt off.
  • Others prefer this to over‑polished, pre‑recorded tech keynotes and argue Nvidia is “selling water in a desert” so presentation polish matters little.
  • Confusion around the phrase “platform company”:
    • Many clarify Nvidia is not becoming a general cloud provider, but offering an integrated hardware+software stack.
    • A smaller cloud evaluation environment and NIM services do push them closer to a platform role, but not a full AWS‑style competitor.

AI scaling, use cases, and societal angle

  • Some worry that ever‑larger GPUs encourage unsustainable scaling of LLMs, with massive resource use and memorization/privacy issues.
  • Others emphasize that current incentive structures reward scaling over optimization, and that software is already heavily tuned to existing hardware.
  • There’s tension between optimistic visions (AGI, robotics, fusion design) and pessimistic ones (ad targeting, hype‑driven spending, bubble dynamics).

Ecosystem, acquisitions, and hardware packaging

  • Speculation on Nvidia acquiring more of the stack (e.g., Canonical, Run:AI, AI labs like Anthropic/Mistral), though cultural mismatches and existing cloud relationships are noted.
  • Some argue Nvidia should stick to GPUs; others point to DGX systems and analogies to Apple’s vertical integration.
  • Suggestions for “GPU stations” or tower‑like appliances arise, but many note integration is lower‑margin and enterprises prefer their own OS/software stacks.