Archimedes · AI analyst · 48 of 100 filed · 49 reported

Archimedes: an analyst that sources every number and grades its own calls.

An AI analyst on your side of the table, covering the top 100 public and private AI companies: audited filings for the public names, tiered disclosures for the private ones. Archimedes grades every source it cites, audits itself nightly, and refuses claims the evidence cannot support. The work product, the AI Leverage Index, is below.

No hedged answers. If a number can’t be sourced, Archimedes says so.

Put your company on Archimedes’ desk — the calculator →

+399.3%
NVDA revenue-per-employee move, FY23→FY26
+51.5%
hyperscalers’ median move — the platform owners
+11.9%
AI adopters’ median — the buyers, so far
+21.3%
whole-set median — the honest baseline
answers with evidence tiers· audits itself nightly· adjudicates disputed figures· predicts, then grades itself· works inside your own AI (MCP)
Archimedes’ work product — the AI Leverage Index

Forty-eight companies, four ways to read them

Every score is a call. Archimedes publishes the grade. Click any row for the full picture: how the leverage was made (revenue growth against headcount), what the employee number is sourced from, and the go-to-market economics underneath. The tag says how the leverage was achieved — growing revenue, cutting people, or neither decisively.

The map

Level against trajectory

One composite hides two dials, so here they are separated: how much revenue each employee carries today (across), and how far that has moved since FY2022 (up). Top-right is where AI leverage is compounding; bottom-right earns a lot per head but isn’t moving; bottom-left is where AI, so far, is a line item.

The analyst

Ask anything Archimedes can source

Answers run on the index’s own stack (gpt-5-mini over the bounded tools), every exchange is logged, and the log is audited nightly by the same grounding check that gates the index.

Try: “Compare NVDA and CRM on GTM efficiency” · “What’s Anthropic’s revenue per employee — and how good is the evidence?” · “Did the market reward revenue-led companies?”
Method

Built like an instrument, not an opinion

The design rule throughout: code binds, models persuade. Every number is computed deterministically; the model is granted authority only where code can bound it. This is how Archimedes is held to account. The metrics investors screen on — growth, Rule of 40, capital efficiency — are computed only where the evidence supports them, and labeled where it doesn’t.

Source

Audited filings only

Revenue from SEC EDGAR XBRL. Employee counts extracted from the 10-K text itself, each with its accession number and supporting quote.

Verification

Agents check, code decides

147 employee-count rows collected or verified by a bounded agent. A number without a source stays MISSING — and a company with missing data is excluded with its reason, never guessed.

Computation

Deterministic engine

Same inputs, same ranking, every run. Method pre-registered before the first run; hand-worked gold checks gate every change.

Assurance

The judge is audited too

Every chat answer is checked nightly against the artifacts — and the checker itself must first catch a planted fabrication before it is allowed to audit anything.

Market

The market lens

Total shareholder return computed over each company’s exact measurement window, from free daily price history. Presented as correlation only — the index never claims the market moved because of leverage.

Reported-index evidence tiers: A filed/registry · B named outlet · C company statement, including a named executive on the record · S specialist data provider · E analyst estimate — shown, never ranks · U unvetted — shown, never ranks · CONTESTED — sources disagree.

Provenance

Where every number comes from

99 from 10-K text, quote + accession 32 sourced with URL 14 verified against draft 2 corrected by the agent IBM excluded — no disclosed total headcount ON excluded — extraction unresolved

Baselines are each company’s last fiscal year ending on or before March 2023 — the world before the AI buildout. Latest is the newest annual filing with a sourced headcount. Mixed fiscal calendars are flagged per company, not silently normalised. Company logos are the property of their respective owners and imply no affiliation or endorsement.