Google Cloud now has a dedicated cluster of Nvidia GPUs for YC startups

Original Article ↗ Hacker News Discussion ↗

GPUs vs TPUs and Vendor Lock-In

  • TPUs are described as cheaper per FLOP at cloud prices but with weaker library/tool support and higher engineering overhead, plus lock-in to one cloud.
  • Nvidia GPUs are seen as the default: nearly all ML frameworks work out-of-the-box, and CUDA’s dominance makes the choice straightforward despite vendor lock-in.
  • Some expect Nvidia GPU prices to fall faster than proprietary accelerators like TPUs, and doubt Google will compete aggressively on FLOPs-per-dollar for TPUs.
  • Others argue Google isn’t the incumbent in accelerators, so milking lock-in would be short-sighted.

GPU Pricing and Scarcity on Major Clouds

  • Hyperscalers are viewed as expensive for GPUs compared to smaller providers; GCP and AWS H100 clusters are cited around similar high hourly prices.
  • Experiences conflict: some say it’s very hard to get GPUs on GCP and AWS (especially A100/H100, GovCloud, or >24GB VRAM), others report decent availability for mid-range GPUs (e.g., T4, L4).
  • Workarounds like AWS capacity blocks are mentioned, but several note that credits are hard to actually use due to capacity limits.

VC, Accelerators, and Cloud/GPU Deals

  • There’s disagreement on how common it is for VCs to directly buy bulk compute.
    • Broad cloud credits are said to be standard, usually funded by cloud vendors as sales/marketing.
    • A minority of investors reportedly buy or control GPU clusters and swap discounted access for equity or to win deals.
  • This Google–YC deal is framed as:
    • A way for Google to hook startups early with priority access and credits.
    • Potentially anti-competitive by favoring YC over other accelerators, though others see it as normal competition since Google lacks a monopoly.

Role of Credits for Startups and Academia

  • Multiple accounts state that large cloud credit packages (hundreds of thousands of dollars) are crucial for early-stage iteration.
  • Some warn that building a business dependent on big-tech credits, quotas, or changing terms is risky.
  • Educational and research credits (including TPUs) already exist, but concerns are raised about lack of hard spend limits and potential large surprise bills for students.

Alternative Compute Providers and Perception of Ads

  • One founder describes a startup building AMD-based MI300x clusters with bare-metal access and consulting, aiming to create an Nvidia alternative and a “credits flywheel.”
  • Other commenters debate whether this adds useful context or reads like investor-oriented advertising, prompting a meta-discussion on promotional posts on HN.