What readers ask, and how far the evidence gets.

What does a token actually cost?

Answered

The posted range across comparable models is 111x, and the quality-adjusted spread is wider still — but effective cost is set by output volume, retries, and cache economics, not by the rate card.

• Inference economics — five verified claims
• If you build with AI APIs

Still open
An effective-cost calculator. Board-gated tooling, not blocked on research.

Should I run my own GPUs?

Answered, and modeled in the repository

It depends on utilization and procurement scope, and the crossover band is quantified per lane rather than argued.

• GPU ownership — three verified claims and a crossover model
• If you run ML infrastructure

Still open
A public-facing version of the crossover model, and a quarterly price refresh.

Which model should I use?

Answered

Route by measured effective cost per successful task at your quality bar; the savings are real and quality-conditioned, and the routing fee is thin.

• If you build with AI APIs
• The routing layer itself — thin fees, thin moats
• Inference economics — five verified claims

Still open
A routing decision tool.

Is fine-tuning worth it?

Answered

The compute is cheap and everything around it is not; the substitute-or-complement question with retrieval decides it before the cost model does.

• If you're evaluating fine-tuning vs retrieval

Still open
A payback calculator.

How much of my AI spend is subsidized?

Answered structurally, anchored on the cost side

Frontier list pricing runs below sustainable infrastructure return, so posted prices are effectively investor-subsidized — but the user-side subsidized fraction is unmeasured and is carried as an open boundary.

• If you budget AI spend
• Evidence boundaries

Still open
An aggregate fraction, if one is ever published. Zero-price channel caps rechecked quarterly.

Where are costs headed in 24 months?

Answered as six dated bets

Not as a curve. Six binary bets, each with a probability, an evidence grade, and a condition that annuls it.

• The 24-month horizon — six dated bets

Still open
Resolution scoring as the dates arrive.

Can I run inference on the edge?

Answered in research custody; carried onto this site only as a forecast bet

Quantization studies and current-generation NPU throughput baselines are held in the research notebooks. The public report carries the edge question as a dated bet rather than a decision surface.

• The 24-month horizon — six dated bets

Still open
An edge feasibility route on this site, and the tool behind it.

How do all these dimensions interact?

Partially answered

Interactions are named where evidence exists — spot risk against on-prem, subsidy against list price, efficiency against the hardware split — and left open where it does not.

• Evidence boundaries
• The 24-month horizon — six dated bets

Still open
A cross-dimensional interaction study. Open research, not a publication gap.