OpenAI and Nvidia announce partnership to deploy 10GW of Nvidia systems

Original Article ↗ Hacker News Discussion ↗

Positioning of Major Players (Apple, Microsoft, Oracle)

  • Some wonder where Apple is in this capex arms race; replies note Apple spends similar sums on buybacks and focuses on efficient on-device AI instead of massive datacenters.
  • The Nvidia–OpenAI deal is read as OpenAI diversifying away from exclusive dependence on Microsoft/Azure; others note Microsoft already hedged with Anthropic and that OpenAI is also tied up with Oracle.
  • Several comments see the whole ecosystem as increasingly incestuous: cloud vendors, model labs, and Nvidia all cross‑investing and reselling to each other.

What 10GW Actually Means

  • Estimates for GPUs range from ~2–5 million accelerators depending on per‑GPU/system power (1–5 kW+) and cooling.
  • Comparisons offered:
    • Roughly the average power use of the Netherlands, or NYC+Chicago combined.
    • About 40% of Bitcoin’s electrical draw, but a tiny fraction of its hash power due to ASICs.
    • Equivalent to multiple large nuclear plants or over 100 nuclear submarines’ reactors.
  • Many emphasize that at this scale, power (not chip count) is the binding constraint.

Why Use Power (GW) as the Metric?

  • Datacenters are planned, permitted, and financed around megawatts/gigawatts, since compute per watt changes but grid capacity and cooling do not.
  • Some view the GW framing as honest and alarming given fragile grids and rising residential prices; others see it as marketing theater to impress investors.
  • There is confusion over whether 10GW is nameplate capacity vs typical utilization.

Nvidia’s “Up to $100B” Investment in OpenAI

  • Interpreted as in‑kind or circular: Nvidia “invests” and OpenAI uses the money to buy Nvidia hardware.
  • Many call this a form of vendor financing or “round tripping” that inflates revenue and valuations: Nvidia sells chips funded by its own capital and gets OpenAI equity back.
  • Others argue it’s legitimate strategic investing: real hardware is built and used, and Nvidia simply trades margin today for equity in a key customer.

Bubble, Accounting, and Systemic Risk Concerns

  • Large minority sees this as classic late‑stage bubble behavior, likening it to the 1990s telco boom and dot‑com era vendor financing.
    • Story: debt and stock-fueled capex, circular deals, and valuations dependent on unrealistically high future AI profits.
  • Counterarguments:
    • Inference revenues are already large; some claim each major model has recouped its training cost.
    • Even without AGI, AI is already deeply useful (coding, search, enterprise features), so massive investment may be rational.
  • Disagreement over legality: some call it “Enron‑like”, others note it’s disclosed, equity-based, and thus unlikely to be prosecuted as fraud.

Grid, Bills, Water, and Climate

  • Strong anxiety that AI datacenters will drive higher consumer electricity bills and grid upgrades paid by ratepayers, while hyperscalers secure favorable industrial pricing.
  • Technical commenters note:
    • 10GW requires huge new generation and transmission; power lines, transformers, and cooling are long‑lead bottlenecks.
    • Co‑locating datacenters near generation (hydro, solar, nuclear, gas) and using waste heat (e.g. district heating) are possible mitigations but not trivial.
  • Water use for cooling sparks debate: some consider it overblown, others highlight local drought impacts and water pollution around DCs.

Value vs Waste: What Are We Buying?

  • Skeptics: 10GW likely yields marginal gains—better chatbots, influencers, ads—rather than cures for cancer; they see misallocation of capital that could go to medicine, clean energy, or education.
  • Supporters:
    • Argue compute is the new foundational infrastructure (like fiber in 2000), and AI will eventually underpin huge productivity gains.
    • Emphasize that post‑bubble, society may inherit abundant compute and power infrastructure, as happened with dark fiber after the telco crash.

Future Overhang and Secondary Effects

  • If AI demand disappoints, commenters expect:
    • A glut of aging GPUs and overbuilt datacenters; possible crash in AI infra prices; pressure on power generators that expanded for AI.
    • But also cheaper compute for research and other industries, similar to cheap broadband post‑dot‑com.
  • Some foresee intensifying vertical integration: AI firms or their backers directly investing in new generation (including nuclear and large solar) and power trading to secure their own supply.