SproutVest Field Notes · No. 1 · July 2026

Build, Buy, or Rent: An Operator's Guide to AI Compute Decisions

By Erick Watson, Managing Director, SproutVest Advisers

Most AI startups treat compute as a budget line item. It is not. Compute is a strategic decision that sets a company's margin structure, shapes its fundraising narrative, and in some cases decides whether the business survives first contact with production traffic.

This report gives operators and investors the cost model that makes compute conversations worth having. It is about economics and decision structure, not silicon benchmarking. It will not tell you whether your workload prefers an H100 or a B200. It will tell you which of the three paths your numbers actually support.

The numbers justify the attention. CloudZero's May 2026 analysis found a 23x spread in per-GPU-hour pricing across providers for the same H100. Very few companies would tolerate that spread when buying laptops. Most tolerate it when buying compute, because nobody truly owns the number.

Inside the report

1 · Why compute decisions silently kill AI companies

Three failure modes: paying on-demand rates for around-the-clock workloads, buying hardware on naive break-even math, and GPU hoarding as fundraising theater. All three share a root cause, and it is not the rate card.

2 · The real cost stack

Every component from the GPU itself to power, cooling, networking, and staffing, and why utilization multiplies all of them. Includes a worked example pricing one 8x H100 server's worth of capacity across on-demand, reserved, and owned paths at three utilization levels.

3 · Build, buy, or rent, and when each is right

A decision framework mapping workload type against demand predictability. Rent when demand is unproven or spiky. Reserve when a baseline becomes visible. Colocate when the baseline is large, stable, and long-lived. Build only with scale, a decade-long horizon, and above all, power.

4 · What investors should actually diligence

The four questions that separate founders who run the business on numbers from founders who have committed to forecasts they never made. None require a deep technical background.

5 · The strategic layer

Why a well-modeled compute plan is one of the cheapest credibility signals available to a technical founder at the fundraising table, and why the reverse signal is just as strong.

Pricing figures were gathered in July 2026 and are cited to their sources in the appendix. GPU pricing moves monthly. The numbers will change, but the framework will not.

About the author

Erick Watson has sat on several sides of this decision. He led product for an AI valuation platform that ran inference against all 105 million US households, led the buildout of a one megawatt, Tier 3 datacenter in Washington State, and through Chainlodge, an AI datacenter infrastructure venture, sees today's compute demand from the supply side.

Get the report

Enter your email and the download starts immediately. The report is a five-page PDF with the full worked example, a glossary, and every figure cited to its source.