We saved a customer $2.3M in GPU spend and they still almost churned

The first cancellation threat came in May 2025. Their procurement team had reclassified our contract from "software" to "managed services," which triggered a different approval flow. The change had nothing to do with our product. It was about a $14,000 line item buried in our invoice for "data egress reimbursement" that, under the new classification, needed quarterly board sign-off.

For six weeks I was on weekly calls with their procurement, finance, and AI engineering teams trying to get the line item moved. Our actual buyer — the head of AI platform — apologized on every call. The head of AI platform did not have the authority to fix it.

§Don't eat the line item

I almost made the mistake every founder makes in this situation. I almost offered to eat the $14k. Owen, our former Stripe Capital cofounder, stopped me. He said: "If you eat it, you've just told them every line item in your contract is negotiable. The next renewal will be a bloodbath." Instead we restructured the contract so egress reimbursement was a pre-paid annual minimum rolled into the base fee. Same dollars. Different procurement category. Approved within a week. I later passed this exact framing on to the team at Quorum Risk during their carrier deal.

The second cancellation threat, in August, was different. Their CISO had flagged that we held read access to their AWS billing API and decided we should not. The fix took eight engineering weeks because we had to rebuild our cost-attribution layer to work from a customer-pushed event stream instead of pulled API data.

The breakthrough wasn't a feature. It was a hire. In September 2025 we brought on a former enterprise procurement officer — a woman named Yelena who had spent fifteen years writing the kind of contracts our buyers were tripping over. Her job title was Head of Customer Operations. Her actual job was making sure our contracts could be approved by buyers like the Fortune 500 retailer's procurement department on the first try.

Within six months under Yelena, our average sales cycle for enterprise deals dropped from 84 days to 41 days. Our cancellation rate among enterprise customers dropped from 22% annualized to under 5%. Renewals at expansion (more workloads, larger contracts) climbed from 38% to 71%.

As of April 2026 we have 47 enterprise customers, $14.6M ARR, and a team of 31. The Fortune 500 retailer is still a customer; they just renewed for $720k. Yelena is now our COO.

Product quality got us in the door. Procurement fluency kept us there.

§Why the renewal almost died and how we saved it

The renewal conversation in February 2026 was the one I had been dreading for nine months. The customer — a Series C generative-AI company running on AWS p4d and CoreWeave H100 clusters — had saved $2.3M in annualized GPU spend with us. They were also about to churn because their new VP of Infrastructure had decided to "consolidate vendors" onto Vantage and CloudHealth. Logo churn at our ACV level kills your year.

What changed the call was a single dashboard. We shipped a CFO-facing weekly digest that translated GPU utilization into dollars-per-trained-token and dollars-per-deployed-inference. The VP could not justify killing us once his CFO started forwarding our digest to the CEO every Monday. Renewal closed two weeks later, ARR up 38%. We hired our first head of finance — Naomi, ex-Datadog — that quarter, specifically to own the buyer-facing financial story across all enterprise accounts. Her existence raised our average deal size from $84k to $190k inside six months. The pattern of turning the *finance buyer* into the actual champion shows up in the AI underwriting account in this cohort too — different industry, identical move.

1. 1
   Saving the customer money is not the same as keeping the customer.

   Once you cross into enterprise-procurement territory, the rules of the buying motion are detached from the value of the product. Learn the rules.

2. 2
   Don't eat the line item.

   The minute you discount to fix a procurement problem, you've made every future invoice negotiable. Restructure instead.

3. 3
   Hire procurement fluency before you hire more AEs.

   Our most leveraged enterprise hire wasn't a salesperson. It was someone who could read a contract redline like a mother tongue.

4. 4
   The CISO will eventually find your data permissions.

   Build for least-privilege from day one. Refactoring under a 30-day clock is among the worst engineering work I've done.

5. 5
   Procurement weirdness is not the customer being difficult.

   It's the customer's organization protecting itself. Treat it as information, not an obstacle.

What this looked like in practice.

The thing nobody tells you about *saving a customer money* is that procurement does not remember last year's savings — they renew based on this quarter's risk score. We learned to re-prove value on a rolling 90-day basis, not annually.

The thing nobody tells you about *the FinOps category* is that Vantage and CloudHealth look like competitors on a slide and like complements in production. Most of our largest accounts run all three. The right fight is for the highest-leverage workload, not for the logo.

The thing nobody tells you about *GPU optimization* is that the savings come from scheduling, not from negotiation. Our biggest single win was a Slurm preemption policy change that moved 31% of one customer's training jobs to spot-equivalent capacity. The negotiation lever — committed-use discounts from AWS — was a third the size.

The thing nobody tells you about *enterprise AI infra customers* is that the VP of Infra changes every fourteen months on average. Build relationships with the CFO and the staff engineers; the VPs are weather.

The thing nobody tells you about *near-churn* is that the saved renewal is the highest-LTV account you will ever have, because you have already proven you can survive a hostile internal buyer. We track "almost-churned" as a positive cohort signal now.