CI economics should include the machines you already own.

Abstract

Hosted CI bundles execution capacity, isolation, scheduling, storage, and operations into a convenient service. Customer-owned machines can reduce purchased compute, but shift costs into power, hardware, maintenance, queueing, and developer interference. The proposed system would manage that trade-off through explicit placement policy, reusable build state, and bounded cloud overflow.

The research question is whether a hybrid fleet can reduce total cost per successful pipeline while preserving the baseline’s correctness and time-to-result. This note provides the cost model and a prospective evaluation protocol. It does not report a deployed fleet or independently measured speed gains.

A hypothesis we can reject.

For teams with repetitive workloads and spare, suitably isolated company hardware, a cache-aware hybrid policy can cost less than the best eligible hosted baseline without worsening agreed reliability or high-percentile completion time.

The claim fails for a customer if operational overhead, local contention, retries, or insufficient peak capacity consume the avoided cloud spend. A negative result is useful: that team should keep the hosted option.

Account for the costs that move.

Let N be jobs per month, L the jobs completed locally, t the hosted billable minutes per job, r the hosted rate, F the platform fee, and K the all-in local operating cost. For a homogeneous workload with the same residual-cloud performance:

K includes power, hardware depreciation, storage, maintenance, and developer disruption. Real comparisons also include provider allowances, cache and artifact costs, network transfer, and retries. Heterogeneous jobs should be evaluated as a sum of per-job costs, not collapsed into a misleading mean.

The break-even condition.

Subtract the hybrid cost from the hosted baseline:

Therefore, under the stated assumptions, hybrid execution is cheaper if and only if F + K < L × t × r. The right side is the cloud cost actually avoided. This algebra is exact within the model; whether a real fleet satisfies the inequality is an empirical question.

Example: 16,000 jobs diverted from a three-minute, $0.004/minute hosted baseline avoid $192. If platform and local costs total $100, the modeled saving is $92. If they total $200, the modeled loss is $8.

What this argument does not prove.

• A container alone does not establish a strong isolation boundary.
• Available monthly worker-hours do not establish acceptable peak-time queueing.
• A local result is not automatically trusted evidence for a protected release.
• Cache reuse is valid only when relevant inputs and trust boundaries are represented.
• Local execution does not remove external model/API charges or unrelated SaaS costs.
• Provider pricing, included allowances, and actual hardware performance can change.