GoBlueFlare
Statement of Work Draft for Discussion v0.1 July 2026

AI Inference Platform Evaluation & Bounded Optimization

Prepared for GoBlueFlare Prepared by [Our Team Name]
Abstract

GoBlueFlare operates data-centre and energy infrastructure and is evaluating unconventional and older-generation AI accelerators for commercial LLM inference distributed through OpenRouter. Our value is not hardware advice. It is the serving software that makes an unconventional accelerator viable. This engagement answers one question: which hardware platform produces the strongest risk-adjusted economics after a commercially sensible amount of targeted software engineering, measured against a real service envelope, not theoretical FLOPS. We establish a baseline, apply bounded, profiling-driven optimization, measure the response to effort, and stop when the selection becomes decision-stable.

Explore the Decision Model → Review Scope
Bounded optimization SLA-qualified economics Hardware–software co-design

Tiered engagement, From $24,000 CAD, stop at any tier. See the tiers ↓

Ph(e) = performance of hardware h after engineering effort e
Economic value = inference revenue − hardware cost − power & operating cost − engineering cost − ongoing maintenance
We evaluate each candidate as a function of engineering effort and stop when the selection becomes decision-stable, not when the hardware is theoretically maxed out.

Executive Thesis

Three ways to compare platforms. Only one produces a defensible decision.

Unconventional accelerators are routinely mis-judged because the comparison method is wrong before a single benchmark runs. Our differentiator is the method itself.

✕ Naive

Out-of-box shootout

Run each platform on stock, unoptimized software and pick the apparent winner.

Unfairly penalizes unconventional hardware: mature CUDA tooling wins on software readiness, not on the silicon's real potential.
✕ Overengineered

Optimize everything first

Fully optimize every candidate to its ceiling, then compare and choose.

Spends heavily before knowing whether the extra performance changes the procurement decision. Slow, expensive, and mostly wasted.
✓ Proposed

Bounded optimization

Baseline, profile, apply capped high-value engineering, measure the response, and stop when the selection is decision-stable.

Measures Ph(e), performance as a function of effort, and selects on SLA-qualified economics. This is the engagement's core method.

The positioning: we will determine which unconventional hardware becomes competitive after a commercially sensible amount of targeted software engineering, quantify the engineering required, and identify the point beyond which further optimization is unlikely to change the procurement decision.

Interactive Decision Lab

Model the decision before you buy the hardware.

Every input below is editable and every result is computed live in your browser. Values are clearly tagged by provenance. Measured benchmark fields are intentionally blank until real evaluation produces them, it shows Input required rather than inventing a number.

Illustrative assumptions, not benchmark results assumption user-provided default illustrative calc calculated measured benchmark n/a not yet available

Scenario Configuration