Pillar
The AI Stack
This page covers the layers a production AI system actually sits on, and it is organised around an awkward fact: some of these layers can be scored on public evidence and some cannot. Where a category is mostly open source, issue-resolution rates are measurable and a ranking means something. Where it is closed and enterprise-sold, there is no public signal at all, and the honest artefact is a price table or a directory. We publish both, labelled, rather than pretending the second kind is the first.
Layers we score
9 categories, 78 tools. Each is placed on a grid from measured signals — repository activity, issue-resolution rate, search demand and reach — not from opinion.
AI Coding Assistants
5 tools scored
AI-powered tools that help developers write, review, and maintain code
AI SRE Tools
15 tools scored
AI-powered tools for site reliability engineering, incident management, and observability
Agent Orchestration Frameworks
10 tools scored
Code-level frameworks for coordinating multiple AI agents, state, tools, and human-in-the-loop across long-running workflows. Three architectural fami…
Backend-as-a-Service
6 tools scored
Platforms that hand an application its entire backend as a managed service: a database, authentication, file storage, realtime subscriptions and serve…
Data Pipeline Orchestration
7 tools scored
The layer that decides what runs, when, and what happens when it fails: DAG schedulers, asset-aware orchestrators and declarative pipeline engines for…
GEO Tools
11 tools scored
Platforms for tracking and optimizing brand visibility in AI-generated answers across ChatGPT, Perplexity, Google AI Mode, Claude, and other generativ…
LLM Observability
9 tools scored
Platforms for tracing, evaluating and monitoring LLM and agent applications in production: span-level traces of model calls and tool use, offline and …
Open-Source Auth & CIAM
9 tools scored
Identity you can run yourself: self-hostable authentication servers and framework-level auth libraries covering sign-in, sessions, passkeys, MFA, ente…
Open-Source Business Intelligence
6 tools scored
Dashboards, SQL exploration and metrics layers you can clone, self-host and read the source of. Two generations sit on this grid: the 2013-2015 self-h…
The layer we cannot score, and what we publish instead
GPU inference is the clearest example of a category that defeats a ranking. Every vendor is closed source at the product layer, so there is no repository to measure, and every review platform we would otherwise use returns 403 to the machine that builds this site. There is no satisfaction signal to be had, and a quadrant needs two axes.
What this category does have is something rarer: published prices. So the artefact is a normalised table rather than a grid.
2.1× spread on identical silicon. On-demand H100 runs from RunPod at $3.29 to Fireworks AI at $7.00 per GPU-hour. Only 2 of 11 vendors publish a reserved-discount rate card; 1 does not publish an H100 price at all.
| Vendor | H100 / GPU-hr | What the vendor actually prints |
|---|---|---|
| RunPod | $3.29 | per GPU-hour, Secure tier (Community: H100 $2.69, A100 $1.39) |
| Nebius | $3.85 | per GPU-hour |
| Crusoe | $3.90 | per GPU-hour (A100 $2.30 SXM / $2.00 PCIe) |
| Modal | $3.95 | per second ($0.001097/s H100), converted to per-hour here |
| Lambda | $3.99 | per GPU-hour ($3.99 in 8x config, $4.29 single) |
| Together AI | $3.99 | per GPU-hour, cluster ($5.49 dedicated endpoint) |
| Replicate (Cloudflare) | $5.49 | per second ($0.001525/s H100), converted to per-hour here |
| CoreWeave | $6.16 | per 8-GPU HGX instance-hour ($49.24 H100); divided by 8 here |
| Baseten | $6.50 | per minute ($0.10833/min H100), converted here; MIG tier $3.75 |
| Fireworks AI | $7.00 | per GPU-hour until 2026-08-31; rises to $8.00 H100 / $13.00 B200 on 2026-09-01 |
| Anyscale | gated | per GPU-hour, A100 only; H/B/GB families GATED |
Read the third column before comparing the second. Vendors do not quote in a common unit. CoreWeave prints $49.24 because it prices per eight-GPU instance; Baseten bills per minute; Modal and Replicate per second. Every figure above is normalised to one GPU for one hour, and the printed unit is shown so the arithmetic can be checked. Comparing headline numbers without converting is an eightfold error.
Prices read from vendor pricing pages on 2026-08-26. GPU pricing moves; treat anything older than a month as indicative.
The consolidation wave
Between mid-2024 and August 2026, eight vendors across these layers stopped being independent companies. Five of them were in AI agent security alone, which had roughly fourteen vendors to begin with. This is the single most important thing to know before choosing anything in that layer, and it is invisible on a comparison page that was written a year ago.
| Company | Acquired by | When | What happened to the product |
|---|---|---|---|
| Protect AI | Palo Alto Networks | Jul 2025 | Folded into Prisma AIRS. Its llm-guard and rebuff repos are now archived. |
| Lakera | Check Point | Oct 2025 | Brand retained as Check Point’s AI-security centre of excellence. |
| Robust Intelligence | Cisco | Sep 2024 | Became Cisco AI Defense. |
| Prompt Security | SentinelOne | Sep 2025 | Integrated into Singularity. Price reported as $180M and $250M by different sources. |
| Aim Security | Cato Networks | Sep 2025 | Folded into Cato SASE Cloud; aim.security now redirects to catonetworks.com. |
| Replicate | Cloudflare | Nov 2025 | Model hosting. Still shipping under its own name. |
| Anyscale | Nscale | Jul 2026 | $1.65B. The Ray company; Ray itself stays open source. |
| iMerit | EXL | Aug 2026 | Up to $310M, closed 3 August. Its homepage headline now reads “EXL ACQUIRES iMerit”. |
Each row confirmed on the acquirer's or the vendor's own site on 2026-08-26, not from a news roundup. The archived repositories are worth noting on their own: acquisition in this category has repeatedly meant the open-source component stops moving.
Why some categories are directories
A grid needs two measurable axes. Reach we can always measure. Satisfaction comes from one of two places: an open repository whose issue-resolution rate is public, or a review platform with enough reviews to mean something. Categories built from closed enterprise products sold through demos have neither, and no amount of effort produces the number.
The tempting move is to score them anyway on impressions. We do not, because a grid presents whatever it plots as measured, and an invented satisfaction score is worse than a missing one — it looks exactly like a real one. Where the signal is absent we say so, list the vendors, and write about what is actually happening in the category instead.
Excluded from the price table above and why: vLLM, SGLang, OpenRouter, Vercel AI Gateway, Fluidstack. Full reasons are in the data file, but the short version is that two are open-source inference engines rather than clouds, two are routers that own no GPUs, and one publishes no price at any tier.