Spend intelligencedeliberately.
AI products don’t need the smartest model for every request.
You can find a cheaper model in an afternoon. Figuring out where you can actually use it without hurting the product takes real work.
Inference economics and AI infrastructure
Which parts of your product actually need frontier intelligence?
01 · Economics
$1 of inference
× 1,000,000 users
= $1,000,000
A two-cent model call looks irrelevant.
At scale, request-level decisions become infrastructure decisions.
Companies need to understand:
- 01what each product feature costs to serve
- 02which workloads actually require frontier models
- 03whether repeated context is being cached effectively
- 04whether smaller or open-weight models preserve acceptable quality
- 05when batching changes the economics
- 06when self-hosting beats API pricing at realistic utilization
- 07what latency and throughput look like under real concurrency
02 · Three dimensions
Economics
Where inference spend actually goes and which interventions produce measurable savings.
Performance
How model, runtime, hardware, batching, caching, concurrency, and utilization change the operating point.
Quality
Whether a cheaper execution path remains good enough for the workload it actually performs.
The cheapest model is not necessarily the cheapest system.
03 · Direction
From measurement to execution.
Today, Enough focuses on measuring and understanding inference workloads.
The larger question is:
Given this workload and these constraints, where should it execute?
- Measure
- Evaluate
- Choose
- Execute
- Observe
Subject to
- cost
- latency
- quality
- privacy
- hardware availability
- reliability
Right now I’m focused on measurement: cost, latency, throughput, and quality. The execution layer comes later.
04 · Writing
Research notes
No. 001 · August 2026
Inference Economics
$1 of inference to 1 million people is $1 million.
Read the note →
No. 002 · In progress
The Knight
More experiments and benchmark results coming.
05 · Tools
Tools
Tool · 01
Inference cost calculator
Price a workload against published rates, then see what caching, batching, a smaller model, or your own hardware does to the number.
Open the calculator →
Tool · 02
Throughput bench
Measured tokens per second and cost per million at real concurrency.
06 · Work with me
Have a meaningful inference bill?
I’m looking for real production workloads to analyze.
If your company is spending meaningfully on OpenAI, Anthropic, Gemini, hosted open-weight models, or self-hosted inference, I’d like to understand where the money is going and which questions are worth investigating.
I’m particularly interested in anonymized usage exports.
Send me the workload and the bill. I’ll dig into the engineering.