Cloud marketplaces as the enterprise billing rail — and now as the price-disclosure surface of last resort
Infrastructure, data and dev-tooling vendors increasingly let enterprises buy through the AWS, GCP and Azure Marketplaces — so the spend draws down an existing cloud commitment instead of triggering a fresh procurement cycle. About 38 of the 338 corpus vendors (~11%) offer it, concentrated in the infra/data/tooling layer; seat-based app verticals still sell direct.
What's happening — and why
What's happening: rather than signing a new vendor contract, an enterprise can buy an AI vendor's service through its hyperscaler marketplace listing. The charges flow through AWS, GCP or Azure billing and count against the cloud commitment the company has already negotiated. Roughly 38 of the 338 corpus vendors (~11%) now offer this — up from a dozen when the pattern was first observed — spanning GPU compute, inference APIs, vector databases, FinOps and billing tools, observability, and the frontier labs that route through hyperscaler model marketplaces (Bedrock, Vertex AI, Azure AI).
Why: large enterprises sit on committed-use agreements — an AWS EDP, a GCP or Azure commit — worth millions. Routing AI spend through a marketplace lets them draw that balance down, skip a new procurement cycle, and consolidate invoices, so the marketplace becomes the default enterprise distribution rail for the metered-compute and infra cohort. The hyperscaler takes a margin, so the rate is usually a touch above the vendor's direct price.
How it works
Evidence over time
25 supporting · 4 counter — hover or tap a point for detail, click to jump to the row.
Evidence
| Company | Date | What happened |
|---|---|---|
| Cohere | Apr 2024 | Command R+ launched day-one on Cohere API, Azure AI Studio and AWS Bedrock; charges route through Bedrock / Azure / Vertex marketplace billing — a multi-cloud-first GTM. |
| Anyscale | Nov 2025 | AWS Marketplace and GCP Marketplace billing added for BYOC, letting enterprises consume Anyscale spend through existing cloud-provider commit agreements. |
| Modal | Feb 2026 | AWS and GCP Marketplace billing added so enterprises can pay through existing cloud commitments. |
| Anthropic | Jul 2026 | Added Microsoft Foundry alongside AWS as a Claude Consumption Unit (CCU) marketplace billing platform — token usage rated in USD, converted to CCU at $0.01 each (100 CCU = $1.00), reported hourly and billed in arrears via the hyperscaler marketplace with no prepaid credits. A frontier model API's enterprise billing now routes through both AWS and Azure marketplaces via a dedicated consumption unit. |
| Baseten | Sep 2025 | Self-host / BYOC for enterprises; GCP Marketplace billing on Enterprise contracts. |
| Fireworks AI | Aug 2025 | Azure Marketplace billing offered on Enterprise. |
| Lightning AI | Aug 2025 | Multi-cloud GPU Marketplace plus AWS Marketplace billing on Enterprise; routes spend across seven providers. |
| RunPod | Feb 2026 | GCP Marketplace billing available on Enterprise commits alongside Secure / Community cloud. |
| Together AI | Sep 2024 | Enterprise GPU clusters with GCP Marketplace billing on Enterprise contracts. |
| Groq | Jan 2026 | GCP Marketplace billing on Enterprise tier. |
| Replicate | Sep 2025 | GCP Marketplace billing on Enterprise alongside dedicated deployments. |
| AssemblyAI | Apr 2024 | AWS Marketplace listing for enterprise procurement of pay-per-second transcription. |
| Qdrant | Mar 2026 | Cloud billing available through AWS/GCP/Azure marketplaces, converting at exactly $0.01 per Resource Usage Unit — an $85 month shows up as 8,500 units on the cloud bill. |
| Dify | Jun 2026 | Same product resold as 'Dify Premium' on AWS and Azure marketplaces for teams wanting custom branding without a Cloud subscription — marketplace as an alternate SKU, not just a billing rail. |
| Mistral AI | May 2025 | Le Chat Enterprise distributed via Google Cloud Marketplace. |
| Weaviate | Jun 2026 | Vector DB: AWS and GCP Marketplace billing available on Plus/Premium/Enterprise tiers — the pattern extends from GPU-cloud into managed vector database infra. |
| Pinecone | Jun 2026 | AWS Marketplace PAYG listing available for Standard/Enterprise alongside direct billing — vector database adopts marketplace billing rail. |
| Observe.AI | Jul 2026 | Breaks the app-layer-absent clause, and does it in the strongest possible form: a customer-service vertical whose own observe.ai/pricing returns 404 publishes a fully metered AWS Marketplace contract for VoiceAI Agents on two dimensions — $4.80 per Minutes-per-Month unit and $12.00 per Interactions unit for 12 months ($9.60 and $24.00 for 24 months, exactly double, so no posted commitment discount) — with NO per-agent seat dimension, even though its direct motion is reported to sell per-agent seats. The listing spells out entitlement mechanics: the contract 'entitles you to a specified quantity of use for the contract duration,' access expires if not renewed, no refunds mid-subscription, with a 'Request private offer' path for negotiated terms. Source: changes/observe-ai-2026-07-22-packaging.md. |
| Usage AI | Jul 2026 | The marketplace as price-disclosure surface of last resort. Usage AI's mid-2026 redesign removed the inline fee calculator (a worked ~15% example: $1,200 fee on $8,000/mo of savings) and the FAQ's '~20% fee' reference from usage.ai entirely — the FAQ now reads only 'a percentage of the savings we realize for you, billed monthly.' The ONLY published rate anywhere is the AWS Marketplace listing: 20% of EC2 savings, 35% on RDS/ElastiCache/Redshift/OpenSearch. Its whole model is marketplace-native ('the fee is billed through your existing AWS, Azure, or GCP marketplace contract'), so the reseller channel is now more transparent than the vendor's own site. Source: changes/usageai-2026-07-23-packaging.md. |
| Hugging Face | Jun 2026 | Uncounted until this review: "you also have the option to link your Hugging Face organization to your AWS account via AWS Marketplace. Hugging Face compute service usage will then be included in your AWS bill" — an org-level account link rather than a per-SKU listing, which is the lightest-weight form of the rail in the corpus. Source: monetization_signals billing stack entry. |
| SambaNova | Jun 2026 | Uncounted until this review, and the clearest statement of the procurement motive: "SambaNova is available through the AWS Marketplace, enabling enterprises to streamline procurement and consolidate billing through their existing AWS account." Procurement friction, not the commit drawdown, is the stated benefit. Source: monetization_signals billing stack entry. |
| Vantage | Jun 2026 | Uncounted until this review, and unusual in offering the rail as a self-serve SWITCH rather than an enterprise-only option: the pricing-page FAQ asks "can I change my payment method from Stripe to AWS or Azure Marketplace?" — marketplace billing presented to any customer as an alternative to the card, from a cloud-cost vendor whose buyers are by definition already on a cloud commit. Source: monetization_signals billing stack entries (AWS + Azure). |
| Upstash | Oct 2024 | The rail generalizing past the hyperscalers: Upstash bills through the AWS Marketplace as pay-as-you-go AND through the VERCEL Marketplace, which offers "integrated billing, automated account provisioning, and direct access to the Upstash console from directly within the Vercel Dashboard." A platform vendor rather than a cloud provider acting as billing intermediary — the same mechanic one layer up the stack. Source: monetization_signals billing stack entries. |
| Qodo | Jun 2026 | Uncounted until this review, and notable because it is an AI-CODING vendor rather than infra: "you will pay recurring monthly usage fees through your AWS bill, while AWS handles deployment and infrastructure management" — the rail reaching a developer-tool app alongside its pooled-credit direct model. Source: monetization_signals billing stack entry. |
| Poolside | Aug 2026 | The third price-disclosure-of-last-resort case, and the first where a price CHANGE — not just a listing — was executed entirely on someone else's marketplace. Poolside cut Laguna S 2.1 from $0.10/$0.20 to $0.09/$0.18 per 1M input/output tokens (a '10% off' badge on OpenRouter as of 2026-08-04, its first move since the model's 2026-07-21 launch), while Laguna XS 2.1 held its 40%-off $0.06/$0.12 rate. Poolside's own site carries no rate card for either model: the price lives on OpenRouter and Vercel AI Gateway, the third-party marketplaces it uses instead of publishing pricing itself. OpenRouter's trailing 7-day weighted averages even expose post-cache effective rates ($0.011/$0.179 per 1M for Laguna S 2.1 at a 97.7% cache-hit rate; $0.036/$0.119 for XS) — a level of disclosure the vendor publishes nowhere. This also extends the rail past hyperscalers AND past platform vendors (Upstash's Vercel Marketplace listing) to model-routing marketplaces, a third marketplace type. |
Counterexamples
- Groq · Aug 2026 — The bound on the disclosure-of-last-resort reading: the marketplace does not automatically catch what the vendor drops. Groq carries the rail (GCP Marketplace billing on Enterprise) and nonetheless deleted its own public pricing page on 2026-08-11 — groq.com/pricing now 308-redirects to a homepage with no pricing content and no Pricing nav link. Per-token, per-hour and per-character rates survive unchanged in the GroqCloud developer docs, but the three built-in agentic tool prices (web search, website visits, code execution) have no live public source anywhere, because their docs defer to the dead page. Two more first-party surfaces vanished in the same window: Hyperbolic's Serverless Inference page returned HTTP 404 with per-model rates surviving only in Mintlify docs and most of the catalog flagged 'Sunset' (2026-08-11), and Moonshot pulled Kimi Enterprise's $599/yr/seat card for a sales-only 'Contact us' (2026-08-04). Vendors are deleting price surfaces faster than marketplaces are backfilling them — check the listing, but do not assume it is there.
- Harvey · May 2026 — Fully sales-led per-seat legal vertical — direct enterprise contract, no cloud-marketplace rail.
- Glean · May 2026 — Enterprise Flex sold direct (seats + FlexCredits); not transacted through a hyperscaler marketplace.
- Suno · May 2026 — Consumer credit app — no enterprise marketplace listing at all.
Trivia
-
Cohere went multi-cloud-marketplace-first from day one for Command R+ (April 2024) — listing simultaneously on AWS Bedrock, Azure AI Studio, and GCP Vertex — making it the earliest corpus example of an AI vendor treating hyperscaler distribution as the primary enterprise go-to-market channel rather than a supplementary one.
-
The cloud marketplace channel has a hidden cost that vendor pricing pages rarely disclose: hyperscalers typically take a percentage of marketplace transactions, meaning the per-token or per-GPU rate charged through a marketplace is higher than the vendor's direct price. Buyers who use marketplace to draw down cloud commits are trading the commit benefit for a slightly elevated unit rate — a tradeoff that is almost never made explicit in pricing comparisons.
-
The marketplace-as-billing-rail pattern is almost entirely absent from the app and seat-based segment of the corpus: Harvey, Glean, Suno, and the consumer-facing vendors all sell direct or self-serve with no hyperscaler intermediary. The channel correlates precisely with metered compute sold to cloud-native enterprise buyers who already carry AWS EDP or GCP committed-use agreements, not with "enterprise" as a general category. (Amended 2026-07-30: Observe.AI and Qodo now break this — the correlation is with the METER, not the layer. Once an app-layer vendor prices in minutes and interactions rather than seats, the marketplace becomes a natural rail for it too.)
-
Anthropic and Qdrant landed on the exact same marketplace conversion rate — $0.01 per consumption unit (Claude CCU; Qdrant Resource Usage Unit) — despite selling completely different products (frontier LLM tokens vs vector-DB compute). The $0.01-per-unit marketplace primitive is emerging as a de-facto convention for wrapping metered usage into a single hyperscaler line item; Anthropic's 2026-07-06 Microsoft Foundry addition made Claude billable through both AWS and Azure at that rate.
-
For two corpus companies the hyperscaler marketplace is now the only place a price exists at all. Observe.AI's own observe.ai/pricing returned a 404 on capture while its AWS Marketplace listing published $4.80 per Minutes-per-Month unit and $12.00 per Interactions unit (2026-07-22); Usage AI's mid-2026 redesign deleted its inline fee calculator and the ~20% figure from usage.ai, leaving 20% of EC2 savings and 35% on RDS-class published only in its AWS listing (2026-07-23). The marketplace has become a price-disclosure surface of last resort, which is the opposite of how vendors usually treat a reseller channel.
-
Observe.AI's marketplace listing prices 24-month contracts at exactly double the 12-month rates — $9.60 vs $4.80 per Minutes-per-Month unit and $24.00 vs $12.00 per Interactions unit — so doubling the commitment buys precisely zero discount. It is also the corpus's clearest seats-to-outcomes substitution: a vendor reported to sell per-agent seats directly publishes minutes-and-interactions pricing with NO seat dimension on the one surface buyers can actually see.
-
The rail has started generalizing past the big three. Upstash bills through the Vercel Marketplace as well as AWS, with "integrated billing, automated account provisioning, and direct access to the Upstash console from directly within the Vercel Dashboard" — a platform vendor, not a hyperscaler, acting as the billing intermediary. And the count itself was badly understated: 25 of 353 corpus companies carry the rail, against the 12-of-158 the trend's prose had claimed since the 158-company era.
For buyers
If you already carry an AWS EDP, GCP or Azure commit, check whether the vendor has a marketplace listing — buying through it can apply your committed-use discount and skip procurement. Confirm two things first: that the marketplace rate matches the vendor's direct rate (the hyperscaler's cut can make it higher), and that the spend actually counts toward your commit.
For vendors
To run this play you need a private-offer flow on at least one marketplace, metering that reconciles to the hyperscaler's billing API, and a margin model that absorbs the marketplace fee. It's an enterprise-distribution lever, not a self-serve one — list where your enterprise buyers already hold commits. It's now table stakes across the infra/data/tooling layer: GPU compute, inference APIs, vector databases, FinOps tools and the frontier labs all support it.
Outlook — what to watch
At ~38 of 338 vendors and rising, marketplace billing is already table stakes for the infra, data and dev-tooling layer and the frontier labs; the next move is its appearance on the enterprise tiers of app-layer vendors chasing committed-use budgets. It stays absent from seat-based verticals (Harvey, Glean) that sell direct — if one of those lists for marketplace billing, the rail will have jumped layers, and the status would sharpen toward a cross-segment default.
Bottom line
About 38 of 338 corpus vendors (~11%) — concentrated in infrastructure, data, dev-tooling and the frontier labs — let enterprises buy through a hyperscaler marketplace to draw down existing cloud commits. It's the default enterprise rail for that layer, and still almost absent from seat-based app verticals above it.
FAQ
What is cloud-marketplace billing for AI?
Buying an AI vendor's service through the AWS, GCP or Azure Marketplace instead of via a direct contract, so the charges run through your cloud bill and count against an existing committed-use agreement.
Why do enterprises buy AI through a cloud marketplace?
To draw down a committed cloud-spend agreement (an AWS EDP, GCP or Azure commit), consolidate invoices, and skip a separate procurement cycle. The trade-off is a marketplace margin that can make the unit rate slightly higher than buying direct.
Which AI vendors offer cloud-marketplace billing?
About 38 of the 338 corpus vendors (~11%), concentrated in the infra/data/tooling layer — GPU compute and inference (Cohere, Anyscale, Modal, Baseten, Fireworks AI, Lightning AI, RunPod, Together AI, Groq, Replicate, SambaNova), vector databases (Pinecone, Weaviate, Qdrant, Milvus, LanceDB), FinOps and billing tools (Finout, Vantage, Usage.ai, Lago, Metronome, Kill Bill), observability (Langfuse), and frontier labs routing through model marketplaces (Anthropic, Mistral, AI21). Seat-based app verticals like Harvey and Glean sell direct.
Is the marketplace price the same as buying direct?
Not always. The hyperscaler takes a cut, so the per-unit rate is often a touch higher than the vendor's direct price — but the ability to spend down an existing commit can more than offset that. Always compare both.