Briefing

OpenAI‑Oracle Deal Reveals 80% of AI Spend Goes to Data Prep

ai-dev
by David Aronchick · OpenAI

Patch your AI projects to budget 80% for data prep and 20% for training, and build intelligent pipelines to reduce waste.

What to do now

Patch your AI pipeline to pre‑filter, deduplicate, validate, and compress data before training to cut costs.

Summary

OpenAI’s $300 B Oracle contract equals $60 B per year, with 2 M H100 GPUs at $30 k each. Data collection costs $9 B/year, deduplication $12 B/year, quality filtering $15 B/year, and failed experiments consume $12 B/year, leaving only $12 B/year for actual training. Thus 80 % of AI spend goes to data prep and only 20 % to model training. Large AI projects spend $5 M on data lake setup, $1.5 M on cleaning, $1 M on POCs, and only $500 k on production AI. Intelligent pipelines can reduce data prep costs by pre‑filtering, deduplication, validation, and compression. The cost of training GPT‑4 ($100 M) is dwarfed by data prep and failures. Every organization should budget 80 % for data prep and 20 % for training.

The article reveals that the real cost of AI is in turning the internet’s garbage into gold, and that intelligent pipelines are the only way to keep costs under control.

Companies that ignore data prep will face billions in wasted spend and legal exposure.

Key changes

  • OpenAI’s $300 B Oracle contract equals $60 B/year, with 2 M H100 GPUs at $30 k each.
  • Data collection costs $9 B/year, deduplication $12 B/year, quality filtering $15 B/year.
  • Failed experiments consume $12 B/year, leaving only $12 B/year for actual training.
  • 80 % of AI spend goes to data prep; only 20 % to model training.
  • Large AI projects spend $5 M on data lake setup, $1.5 M on cleaning, $1 M on POCs, and only $500 k on production AI.
  • Intelligent pipelines can reduce data prep costs by pre‑filtering, deduplication, validation, and compression.
  • The cost of training GPT‑4 ($100 M) is dwarfed by data prep and failures.
  • Every organization should budget 80 % for data prep and 20 % for training.

Affects

enterprise ads-customers internal

Customer impact

Analyzing matches…

Ask about this story

Impact on an agency? Which customers? Compare historically Risks of waiting