We deleted our GPU cluster and shipped ten times more

A short story about spending nine months building AI on the wrong stack, and one weekend rebuilding it on the right one.

The nine months we spent doing it wrong

Two H100s. A vector database SaaS bill I do not want to type out loud. Two engineers spent three months tuning cosine-similarity thresholds. Our answers were 70% right. Regulators would not sign off. Deployment kept slipping. Every review meeting was a debate about top-k and re-ranker weights.

Then one weekend we tried something stupid: what if we just stored the facts as text files and walked a graph? No embeddings. No vector DB. No re-ranker. Just an inverted index and typed relations.

What the weekend produced

Sub-5ms retrieval on 90,000 entries. On a Ryzen. In our office. We deployed the same day. Answers went from 70% right to auditable, cited, and correct. The regulator conversation flipped inside a week.

The bill this month vs the bill nine months ago

Why nobody tells you this

Because the entire AI industry is monetised around the assumption that you need the GPU stack. The GPU maker wants to sell GPUs. The vector DB vendor wants a subscription. The frontier lab wants an API you cannot leave. Their business model needs your architecture to look expensive.

Our business model needs your architecture to look cheap. Because that is when you buy sovereignty. Green Intelligence is not just an environmental argument — it is the fastest capex reduction in your organisation right now, and nobody in the AI industry is selling it to you.