About Capital & Compute: Coverage & Method
What Capital and Compute covers, how every number gets sourced and dated, who writes it, and how the site pays for itself.
Capital & Compute is an independent publication with two questions behind everything on it: what does this actually cost to run? and where does the money end up? Vendors sell AI like software and bill it like infrastructure. The useful reporting lives in that gap.
What this site covers
The beat spans 6 connected areas: 143 published analyses, supported by33 free tools and trackers.
- AI costs: token rates, subscription floors, cache pricing, and what it costs to finish a task rather than start one. A model at half the price per million tokens often costs more per completed job.
- Coding agents: harness design and workflow. The model matters. How it reads a repository, what it retries, and where the loop gives up usually matters more.
- Models and benchmarks: which leaderboard scores hold up on a real codebase, which benchmarks have saturated, and where contamination inflates a result.
- Local AI and hardware: whether to buy the machine or rent the instance. The answer moves with memory capacity, bandwidth, utilization, and a DRAM shortage that will not sit still.
- Compute infrastructure: data centers, training runs, GPU and memory supply, and the debt and depreciation assumptions paying for the buildout.
- AI markets: adoption, measured productivity, publishing economics, and national compute strategy. Where AI creates value, and where it only relocates it.
Who it is for
You are probably here because you have a decision with a budget attached. Which model goes behind the product. Whether to buy a machine or rent an instance. Whether a benchmark score in a launch post is worth switching providers over. Whether a memory price means order now or wait a quarter.
That means engineers, technical founders, and the people who approve their infrastructure spend. The writing assumes you can read a table and would rather see the arithmetic than a verdict.
It is not written for investors hunting positions, and nothing here is advice to buy or sell anything. When a public company shows up, it is because it makes something the supply chain depends on, not because the stock looks interesting. The disclaimer has the full statement.
How the research works
Every factual, statistical, or research claim links to its source at the point it is made, and names who produced it and when. Labels stay accurate. A preprint is called a preprint, not a peer-reviewed study. Documentation written by a vendor is called documentation, not an independent finding. Coverage from another outlet says whose reporting it is. Primary sources come first, and a secondary one is used only where no primary exists, labeled as such.
Anything time-sensitive carries the date it was verified. A price confirmed in one quarter proves nothing in the next, and leaderboards reshuffle faster than any page can keep up. When a number is estimated rather than measured, the assumptions are laid out so you can redo the math, or argue with it. Posts built on research end with a full source list.
The trackers
About half the site is data rather than prose. Model prices, GPU and memory street prices, benchmark definitions, inference providers, and cost-per-task comparisons are kept as dated datasets and read again at the source as the market moves. None of it is captured once and left to rot.
Each tracker is a dated snapshot, not a live feed, so confirm anything time-critical at its source before quoting it. The underlying data is served as JSON at stable URLs, so you can cite it or rebuild it instead of scraping the page.
Who writes it
Capital & Compute publishes under the brand, not under individual bylines. That is a deliberate trade. The work is meant to stand or fall on whether its sources hold up, which is why every claim is linked and dated, and why the method sits next to the number instead of in a footnote. Posts are researched, sourced, and reviewed against the editorial standards, which also spell out how and where AI tools are used in the process.
When something turns out to be wrong, the correction gets published. It does not get edited in quietly. Posts that are materially revised carry an updated date.
Independence
No recommendation on this site is for sale. Nothing here is written, ranked, or reordered because a company asked for it, and no vendor gets a preview before publication. If that ever changes for a given piece, it will be disclosed inline and marked in the link itself. The disclaimer and disclosures page has the full policy.
Get in touch
Corrections, questions, and tips are welcome at contact@capitalandcompute.net, or through the contact page. Corrections get read first. New posts go out over RSS, and start here is the guided way into the archive.