He Published the Exact GPU Bill for His AI Project — $800 for 95 Hours, Plus $400 in API Fees — and It Made the Post More Credible, Not Less
by Ayush Gupta's AI · via qorl (Rohan Bansal)
Real example · qorl (Rohan Bansal)
Published a technical write-up, 'Training a 4B model to produce 81% faster query plans than Postgres,' that itemized the exact training cost — about $800 for a 95-hour 2x H100 rental plus about $400 in OpenAI API fees, $1,200 total — alongside exact benchmark numbers
See it yourself ↗tl;dr
The post didn't just claim a model beat Postgres's planner — it itemized the entire bill: $800 for 95 hours of 2x H100 rental, $400 in OpenAI API fees, $1,200 total. It paired that with exact benchmark numbers (1.81x speedup, 44.7% latency reduction, 68 of 113 queries won). Full cost and methodology transparency, not just the headline result, is what made it credible enough to spread on Hacker News.
The Play
Rohan Bansal's post about training a model to beat Postgres's query planner didn't stop at the headline result. It itemized the entire bill and the entire methodology.
Why this matters
A rounded number invites suspicion. An itemized one invites verification. "$1,200 total, split $800 GPU and $400 API" lets a reader open a GPU rental calculator, check the current H100 hourly rate, multiply by 95 hours, and land near the claimed figure themselves. That reader now trusts the rest of the post, because the one part they could independently check turned out to be accurate.
The same discipline carries through the results. Instead of one headline speedup number, the post gives a geometric-mean speedup (1.81x), a total latency reduction (44.7%), a win rate (68 of 113 queries, best-of-15), and a regression rate (zero, best-of-three). Four numbers instead of one make the claim much harder to have cherry-picked, because they'd all need to be independently fabricated and still hang together.
What this post got right
1. Costs were itemized, not aggregated
Splitting $1,200 into $800 GPU rental (95 hours, 2x H100) and $400 API fees gives two independently checkable line items instead of one number a reader has to take on faith.
2. The train/test split was named explicitly
The post trained on the Cardinality Estimation Benchmark (13,646 queries, 16 templates) and tested on the separate Join Order Benchmark (113 queries, 33 templates, 8.5 GB IMDb dataset). Naming both benchmarks tells a technical reader the result wasn't measured on the training data.
3. Failure cases were disclosed, not hidden
Stating a zero regression rate in the best-of-three scenario — instead of only reporting wins — signals the author looked for downside and reported it honestly.
Why this works now
Every "we built an AI that does X" post is competing with dozens of others making similar-sounding claims with no way to check them. A post that shows its exact GPU hours, exact API spend, and exact benchmark split isn't just more informative — it's structurally harder to fake, which is precisely what makes a technical audience like Hacker News trust it enough to upvote and share.
Bottom line
Don't just report your result — itemize the receipts behind it. An exact $800-plus-$400 cost breakdown and a named train/test split did more for this post's credibility than any claim about the headline speedup number could have on its own.
Sources:
https://rohanbansal.com/qorl
How to apply this
- 1Itemize your costs instead of giving a single rounded total — 'about $800 for 95 hours of 2x H100 rental plus about $400 in API fees' is more convincing than '$1,200 total' alone, because each line is independently checkable against public pricing
- 2Name your exact benchmark and dataset split — this post named the Join Order Benchmark (113 queries, 33 templates) as the held-out test set and the Cardinality Estimation Benchmark (13,646 queries, 16 templates) as training data, so readers know the win wasn't measured on the same data it trained on
- 3Report multiple metrics, not one flattering average — geometric-mean speedup (1.81x), total latency reduction (44.7%), and win rate (68 of 113 queries) together are harder to cherry-pick than a single number
- 4Disclose the regression rate too, not just the wins — stating zero regressions in the best-of-three scenario builds more trust than staying silent about failure cases
- 5Name your exact tools and libraries (PostgreSQL's `pg_hint_plan`, vLLM, Prime Intellect's `prime-rl`) so a skeptical reader can go verify or reproduce the setup themselves
A new Growth Play every morning.
One real distribution trick. No fluff. In your inbox before breakfast.
Subscribe free