The Software Layer

IThe ranking

Below the model giants sits a second tier of AI businesses defined not by the size of their models but by the revenue they clear. ClickHouse, an analytical database, leads at $15bn after a January round; Harvey, which builds legal-workflow agents, follows at $11bn. Glean (enterprise search) and Modal (AI compute) come next. Two younger names — Exa and Parallel Web Systems, both building web infrastructure for AI agents — already carry marks around $2bn. A third primitive, OpenRouter, routes inference across models for a fee of roughly 5% of spend; it was valued at $500m in an April 2025 round and, in May 2026, raised a $113m Series B (led by CapitalG) that more than doubled its mark to ≈$1.3bn.

IIWhat the marks rest on

Where revenue is reported on a roughly comparable basis, the implied multiple is the clearest lens. It runs from about fifteen times at Modal to about sixty at ClickHouse — a four-fold spread inside one cohort, with Harvey now near the middle at ≈37× after reported revenue caught up to ≈$300m. These figures are illustrative: ARR is reported or estimated, the bases differ, and a multiple is not a forecast. But the dispersion is the point. The market is not paying a single price for a dollar of AI-software revenue; it is paying for growth rate, gross margin and defensibility, and it is paying very differently for each.

IIIRevenue, accelerating

What distinguishes this cohort from the frontier hardware tier is that the revenue is real, and growing fast. ClickHouse tripled annualised revenue year-on-year to $250m; Glean tripled to $300m in fifteen months; Harvey has climbed to a reported $300m; Modal’s run-rate rose roughly five-fold to an estimated $300m. These are among the steepest revenue ramps in private software. They also explain, in part, why the marks have moved as quickly as they have.

IVThe price and the basis

Set mark against revenue directly and the dispersion sharpens. At comparable annualised revenue — roughly $250m to $300m — ClickHouse and Harvey are marked two to three times higher than Glean and Modal. The revenue is similar; the price is not. The difference is the market’s read on durability: a database and a legal-agent platform with deep enterprise lock-in are being underwritten differently from a search tool or a compute layer, however fast each is growing.

Two cautions close the note. First, the bases are not clean. Consumption-priced run-rate is not contracted ARR; an annualised month is not an audited year. We have used reported figures and flagged estimates, but the multiples above should be read as orders of magnitude, not decimals. Second, the no-revenue edge — Exa, Parallel, Sycamore — is being priced on the same agentic thesis as the leaders, without the income statement to test it. Where public software multiples compress, the question is how quickly these private marks follow. We observe the exposure; we draw no conclusion the reader has not earned for themselves.