Two weeks ago this read covered 1,647 companies' public vendor disclosures. The catalog now holds 2,609. The three-tier stack held up — but the bigger sample says adoption is concentrated, not spreading, and a fourth tier came into view: the frontier labs publish vendor lists too.
Every company below publishes its vendor relationships somewhere on the open web. Reading those disclosures top-down reconstructs the supply chain. The new tier is at the top: OpenAI, Anthropic, Mistral and Cohere all publish subprocessor lists, and those lists name their compute.
Two of those tier-0 lines are worth pausing on. OpenAI's own list names Cerebras under "cloud infrastructure" — a frontier lab buying from the tier-2 layer that is supposed to be downstream of it. And Mistral names CoreWeave alongside Azure and GCP. The tiers were never a layer cake; with tier 0 visible they're plainly a mesh. Fireworks and Together still disclose Anthropic and OpenAI as their vendors, and Together lists OpenRouter.
The last issue closed with "the buyer pool now extends well past AI-native apps." A 58% larger sample does not support that as a trend. The catalog gained 962 companies and the inference-provider cohort gained 15 — nearly all of the new companies use no rented inference at all. Sliced by what a company sells, the buying is sharply concentrated:
Frontier-lab APIs are everywhere — a quarter to two-thirds of every segment. Rented inference is an AI-native and developer-infrastructure purchase and almost nothing else. The zero is the sharpest number here: of 166 security & compliance companies with published lists, 44 buy from a frontier lab and not one buys from an inference provider.
What did change is the kind of company at the top of the distribution. Zendesk (Baseten, Fireworks, Groq) and MongoDB (Baseten) both appeared in July — the first large incumbent SaaS vendors in the cohort, as opposed to AI-natives.
The long-run series comes from companies that publish dated change notes — 69 of them, going back to January 2023. That channel shows the same ramp as last issue, peaking at eight additions in June 2026, then falling to three in July. But we now have a second channel: our own daily crawl has been diffing all 2,609 lists since July 15, and it caught eight more July additions that no dated feed reported. The dated series is a floor, and July was not a slowdown.
The eight additions only our crawl saw, in order: Retool → Baseten (Jul 15), Zendesk → Baseten (Jul 17), Parallel → Fireworks (Jul 21), Fathom → Fireworks (Jul 23), IntelePeer → Groq (Jul 25), Factory → Baseten and Together, and MongoDB → Baseten (all Jul 28).
Baseten took second place from Fireworks, adding six customers in two weeks — the largest absolute gain of any provider. In relative terms OpenRouter grew fastest (7 → 12, +71%), which fits its role: an app that wants many models without many contracts buys the router, not the providers. Groq keeps its lead and its caveat — many of its 42 listings read "Groq (if enabled)", an optional fast-inference toggle rather than load-bearing serving.
Multi-homing deepened as well. Tier-1 companies now average 1.61 providers each; 48 of the 134 (36%) list two or more, and 22 list three or more — up from 12 companies at three-or-more last issue.
This is the finding that didn't move. Two weeks ago the big four disclosed 52 GPU-neocloud relationships; today it's 53. Baseten still lists 18 neoclouds, Modal 15, Together 11, Fireworks 9. After six months of a new supplier roughly every month, Modal signed nobody new in July.
The one real move on this edge was a swap, not an addition: on July 21 Fireworks dropped Voltage Park and added IREN on the same day, then added GMI Cloud four days later. That's the first clean substitution we've observed anywhere in the chain — an inference provider replacing one GPU supplier with another.
Standing caveats on this edge
Baseten's history still shows 11 suppliers appearing on a single day (Jan 23, 2026) — a bulk backfill when they first published, not 11 same-day signings. Groq and Cerebras disclose almost no infrastructure vendors: they run their own silicon in facilities they aren't obliged to name, so this lens undercounts vertically-integrated players. Replicate, RunPod, OpenRouter and DeepInfra publish no list we can read at all.
Last issue's claim was that apps add rented inference alongside the frontier labs rather than instead of them, on the strength of dated disclosures alone. Two weeks of first-party change tracking across all 2,609 lists now tests it directly. In Jul 15–28 we recorded 18 credible additions of a frontier lab and zero credible removals. On the inference side, nine additions and one removal — Retool dropped Baseten and re-added it the same day, and Baseten sits in Retool's list today.
Of the 134 companies listing an inference provider, 117 (87%) also currently list a frontier lab — down a little from 89%, entirely because the 17-company no-lab group grew rather than the 87% weakening.
| Company | Inference providers | Frontier labs kept |
|---|
The 17 companies running rented inference with no frontier lab remain the interesting exception — mostly infrastructure and media products whose models are their own: LiveKit (voice, four providers), Cartesia (voice), Braintrust (evals), Daytona, Edge Delta, Stability AI, plus several image and video products on Replicate and RunPod. OpenAI technically joins this group: it lists Cerebras and, obviously, no frontier-lab vendor but itself.
Source. Publicly available vendor disclosures from 2,609 companies collected from the open web through July 28, 2026. Dates come from 363 dated publisher disclosures back to January 2023, plus our own daily diffing since July 15, 2026.
Bias. The catalog is a convenience sample, not a census — it is heavily weighted toward startups and YC companies, so segment shares describe this population, not the software industry. Only companies that publish dated notes contribute history, so the monthly series is a floor, and bulk backfills cluster dates. A vendor appears only when the relationship is one the company must disclose: self-hosted or fully-isolated inference never shows up. Tier membership is hand-curated; this issue added GMI Cloud to the neocloud tier and fixed a hyperscaler key that was under-counting Azure.
Watch next. The removal signal is now genuinely instrumented on both edges, and it has already produced its first hit — Fireworks swapping Voltage Park for IREN. The open questions for the next read: whether Modal's July pause is a plateau or a gap, whether Zendesk and MongoDB are the front of an incumbent wave or two outliers, and whether any tier-1 company ever drops a lab.