Is AI Profitable Yet?

Original Article ↗ Hacker News Discussion ↗

Methodology and What the Site Actually Shows

  • Many argue the site is more “meme” than analysis:
    • Uses CEOs’ stated future capex, not actual spend.
    • Mixes hardware, power, salaries, and R&D into undifferentiated “AI spend.”
    • Treats cumulative capex minus revenue as “PNL,” i.e., effectively expensing data centers and GPUs upfront instead of amortizing.
  • Several say this will make any fast‑growing infra-heavy business look bad; under GAAP many of these lines might look much healthier.

Capex, Depreciation, and Hardware Lifetimes

  • Debate over how long AI hardware is useful:
    • Some cite 1–3 year service lives; others note decade-old accelerators still sold/used.
    • If hardware is amortized over even a few years, a 195% cost/revenue ratio during build‑out is seen by some as acceptable, by others as alarming.

Winners: Nvidia and the “Shovels” Analogy

  • Strong consensus that Nvidia (and to a lesser extent other chip and datacenter vendors) are the clear current winners, analogous to “selling shovels in a gold rush.”
  • Some complain the site is misleading by including Nvidia but not RAM/SSD, power, or other infra vendors that are also profiting.

Frontier Labs, Cloud Providers, and Circular Deals

  • Distinction drawn between:
    • Hardware makers.
    • Cloud providers.
    • Pure AI labs (OpenAI, Anthropic, etc.).
  • Concern about circular financing:
    • Clouds give AI labs compute credits; labs give inference credits/equity back.
    • Both sides book this as revenue even though it’s largely internal barter backed by cloud cash.
  • Some think new “lean” labs (e.g., Chinese players) may be structurally more efficient than US “legacy” labs.

Inference Economics and Pricing

  • One camp: inference margins are “fantastic,” especially at scale, and training is becoming a smaller share of cost.
  • Other camp: doubtful these margins remain after correctly pricing GPU depreciation and competition; suspect many providers may be underpricing and effectively subsidizing customers.
  • Debate on whether clouds or marketplaces are currently selling inference at a loss to grab share.

Is This a Bubble, and How Dangerous?

  • Frequent analogies: dot‑com bubble, 2008 crisis, railroad panic of 1873.
  • Optimists:
    • Note that only ~50% of cumulative AI infra spend is “in the hole” during a massive buildout; see that as a strong sign.
    • Argue infra can be repurposed even if frontier labs fail.
  • Skeptics:
    • Emphasize unprecedented scale: over $1.6T infra, multiples of Apollo and the US interstate system.
    • Worry about stock market overvaluation, pension/retirement exposure, and knock‑on layoffs if expectations collapse.
    • Point out that stock “losses” and bankruptcies still translate into real job and demand shocks.

Adoption, Real Value, and Saturation

  • Some claim AI usage has plateaued outside niches like coding; others counter with rapidly growing lab ARR and huge token volumes (e.g., Google serving quadrillions of tokens/month).
  • Question raised: given how omnipresent AI is in media and corporate roadmaps, why so little clear profit so far?
  • Others note hidden gains:
    • Ad ranking, recommendation, and other internal productivity uses that don’t show up as “AI revenue” but likely drive profits at firms like Google and Meta.

Distribution of Risk and Social Externalities

  • Disagreement on whether “losses” matter:
    • One view: markets reallocate capital efficiently; investor losses are just transfers.
    • Counter‑view: crashes propagate via confidence, credit, and employment, harming ordinary people.
  • Some criticize vast AI spend versus underfunded public goods (healthcare, childcare, education), arguing this fuels public backlash and even hostility toward AI.
  • One commenter attributes inevitable AI profit capture to “Wall Street and Jewish capital,” reflecting a conspiratorial/ethnic framing rather than an economic argument.