Pop quiz, applied
AI Lab Build-vs-Buy Simulator

Annual total cost vs utilization

Hover or tap the chart to see which strategy leads at each point.

Every formula, inspectable

Training compute

FLOPs = 6 × params × tokens × (1 + 0.15 × extra experiments). Fine-tuning uses ~2% of pretraining compute. Ranges shown as ×0.7 to ×1.4 around the point estimate.

Training time and cost

Days = FLOPs ÷ (cluster × effective TFLOP/s × utilization) ÷ 86400. Rented cost = GPU-hours × hourly rate. Owned cost = amortized capex + energy (0.7 kW/accelerator) + operations.

Inference cost per million tokens

Illustrative serving rates: API $3-8, fine-tuned self-host $0.60, own pretrained on rented $0.50, custom silicon $0.22, each reduced by half the annual model-efficiency gain. These are estimates, not quotes.

Breakeven volume

Breakeven B tokens/mo = one-time capital ÷ (API rate − self rate) ÷ 1000 ÷ (12 × horizon). If self-serving is not cheaper per token, breakeven never occurs.

Risk and optionality rubric

Schedule risk reflects integration surface (API low, custom silicon high, scaled by tape-out risk). Optionality scores 1-10 across data/IP control, roadmap independence, capacity assurance, and workload optimization. Qualitative and transparent, not measured facts.

Nothing here claims building is automatically cheaper. The point is to see when it might be, and what you pay for control.

Pretraining vs fine-tuning

Pretraining builds a model from raw tokens and consumes roughly six FLOPs per parameter per token; fine-tuning adapts existing weights with a small fraction of that compute. Fine-tuning buys capability cheaply but inherits the base model's ceiling, license, and roadmap.

Training vs inference economics

Training is a large one-time (and repeated-experiment) cost; inference is a recurring cost that scales with demand. High-volume products are often dominated by inference, which is why serving efficiency and hardware ownership matter more as token volume grows.

Utilization and total cost of ownership

A cluster at 30% utilization costs the same to own as one at 60% but delivers half the work. TCO includes hardware, power, cooling, networking, engineers, and the opportunity cost of capital, not just the sticker price of accelerators.

Custom silicon, trust, and control

Custom chips can improve performance-per-watt and secure capacity, but demand enormous fixed costs, compiler complexity, and fab dependence. Labs also integrate for reasons beyond unit economics: supply assurance, data and IP control, recruiting signal, and iteration speed. The honest quiz answer is usually 'all of the above, weighted by your scale.'