The meter tracks the buyer: usage for developers, commitments for enterprise
Pricing structure tracks who's buying, not the compute underneath. Of the 113 pure-usage vendors in the corpus, 88 (78%) sell to developers; of the 141 vendors with annual commitments, 130 (92%) target enterprise. The same vendor often runs both meters at once — chosen by who signs.
What's happening — and why
What's happening: across the corpus, the billable structure lines up with the buyer's procurement habits. Developers get clean pay-as-you-go meters; enterprises get annual commitments with overage. It's so consistent that the structure on offer signals the segment a vendor expects you to be.
Why: developers want to start instantly and pay for what they use; enterprise procurement wants a committed contract with a negotiated discount and a predictable floor. So vendors fit the meter to the buyer — Together's on-demand H100 for developers ($5.49/hr) versus its reserved 7–30-day commit ($4.99/hr) for enterprise is the archetype: identical compute, two meters, picked by who's signing.
How it works
Evidence over time
18 supporting · 5 counter — hover or tap a point for detail, click to jump to the row.
Evidence
| Company | Date | What happened |
|---|---|---|
| DeepSeek | Mar 2025 | Pure per-token API sold to developers — no commitment, no seat; the canonical developer-buyer / pure-usage pairing. |
| Modal | Sep 2025 | Per-second compute, self-serve to developers; commitments and Marketplace billing appear only at the Enterprise tier. |
| Fireworks AI | Mar 2025 | Pure-usage per-token developer API; enterprise adds dedicated deployments and committed-use rates. |
| Together AI | Sep 2024 | On-demand H100 self-serve for developers ($5.49/hr) vs reserved 7–30 day commit ($4.99/hr) for enterprise — usage for devs, commit for enterprise on the same product. |
| Anyscale | Nov 2025 | Pure-usage ACU credits for self-serve developers; annual commits + BYOC + Marketplace billing for enterprise. |
| Glean | May 2026 | Enterprise-only buyer: seats + pooled FlexCredits sold on annual commitment — no developer self-serve PAYG surface. |
| Harvey | May 2026 | Enterprise legal buyer: per-seat annual contracts, sales-led — the opposite end of the spectrum from a developer PAYG meter. |
| Vercel | Sep 2025 | Self-serve usage metering for developers on Pro; Enterprise layers annual commitments on the same metered units. |
| Lambda Labs | Jun 2026 | Pure-usage on-demand pricing for developers ($3.99/GPU/hr H100) plus multi-year committed discounts for enterprise — the same physical GPU in two metering structures based on buyer procurement. |
| Vapi | Jun 2026 | Build tier is developer PAYG ($0.05/min, self-serve); Scale tier is an annual enterprise contract with committed volume — developer vs enterprise in a single voice-API product. |
| LiveKit | Jun 2026 | Build free for developers, Ship/Scale plans for growing teams, Enterprise annual for large deployments — a three-buyer-tier structure that tracks commitment level with buyer scale. |
| RunPod | Jul 2026 | The thesis published as a product page. RunPod launched runpod.io/enterprise naming four commercial components — reserved baseline capacity, usage-based burst, committed-use pricing, post-paid billing on one invoice — plus contractual SLAs, SOC 2 Type II / HIPAA / GDPR, and clusters from 200+ on-demand to 10,000 reserved. The self-serve pricing page keeps pure PAYG Pods, Serverless, Clusters and Public Endpoints rates. Same GPUs, two meters, split by who signs. No enterprise dollar figures published. |
| GitLab | Jun 2026 | The enterprise commit absorbing the usage meter: GitLab Flex replaces separate per-seat plans plus GitLab Credits plus usage add-ons with ONE annual dollar commitment spanning all three, reshapeable month-to-month without contract amendments. The commit is no longer a floor under a specific unit — it is a floor under the whole relationship. |
| CoreWeave | Jul 2026 | Commitment discounts moving the wrong way at an infra-cloud vendor: CoreWeave removed the footnote offering 'discounts for reserved storage capacity' on AI Object Storage and Distributed File Storage, and withdrew the $15,000/mo 100G Virtual Direct Connect tier (10G remains at $1,500/mo). Its two new Standard Memory CPU shapes are published at SPOT rates only ($3.94/hr AMD Turin 9655P, $2.86/hr Intel Emerald Rapids 8562Y+) with on-demand listed 'Contact Sales' — inverting the usual pattern in which spot is the sales-gated tier. |
| Modal | Aug 2026 | The meter itself gated by plan tier — the thesis applied to the pricing SURFACE rather than to the rate. Modal's July 29 announcement of an OpenAI-compatible, token-billed Shared API for Kimi K3 originally framed it as available on any plan, with Starter's $30/month free-compute credit covering ongoing usage. The same post has been edited to gate Shared API token-based pricing to Team ($250/mo + compute) and Enterprise customers only. Starter and lower tiers are not locked out of the underlying model — they reach Kimi K3 through Modal's per-second Auto Endpoint — but they lose access to the token-metered pricing surface. Same models, two meters, split by who is paying rather than by what is running. Modal has published no per-token input/output rates for the Shared API on its pricing page or billing docs. |
| Nomic | Aug 2026 | One product, a self-serve meter and a procurement meter, published side by side for the first time. Nomic's AEC Platform had sold only Business ($40/user/month, 25-seat minimum, an explicit $1,000/month platform commitment, $20 of pooled AI usage per seat) and custom Enterprise since November 2025 — no self-serve path existed on the flagship at all. The rebuilt grid adds Free (no credit card, limited agent requests) and Individual ($20/month, self-serve, all product surfaces except the Agent API, AI usage funded by on-demand top-ups), with the standalone $1,000/mo commitment line removed. The developer buys top-ups; the enterprise buys 25 seats and a pooled allowance. |
| Ambience Healthcare | Aug 2026 | A direct test of the trend's strong claim that the UNIT predicts the model: Ambience moved billing_units from active-users|seats to outcomes and taxonomy from seat-based+subscription to outcome-based in one change. active-users and seats are the two units this trend identifies as collapsing to a subscription (active-users 71% seat-based against a 21% base; seats 13% pure-usage), and the vendor exited both units and the subscription together. The unit changed first and the model followed, in the predicted direction. |
| Schematic | Aug 2026 | A machine-emitted unit replaced by a FINANCIAL one, which the trend's machine-vs-human framing does not anticipate. Schematic dropped active-users, events and transactions for a single billing-volume meter — billing on a percentage-shaped measure of the customer's own revenue rather than on anything either a machine emits or a human occupies — and doubled Growth to $400/mo. Worth watching as a third pole alongside machine-metered and human-occupied units. |
Counterexamples
- Chargebee · Aug 2026 — The commit-is-an-enterprise-artifact half dented again, and this time outside infrastructure. Chargebee retired its Starter and Performance tiers for a single Flow plan sold with a choice of rate structure rather than a choice of tier: Pay-as-you-go at 0.80% of monthly billing value with a $0 platform fee, or Commit monthly at $99/mo plus 0.65%. Both include 100M usage events/mo; only Enterprise Plus (annual commitment, 500M events) stays quoted. That is a 19% rate discount for a commitment a customer takes on self-serve, in billing software — the same shape as Hyperbolic's 1-week self-serve reserved GPU capacity, in a segment with no GPUs. The commitment is detaching from procurement in more than one place.
- Hyperbolic · Jul 2026 — The clearest dent in the commit-is-enterprise half. Hyperbolic moved Reserved capacity from sales-quoted to SELF-SERVE in-app at a discounted prepaid $/GPU/hour, with terms from 1 week to 1 month, paid in full up front and no early termination — only Private Cloud (single-tenant, off-platform, multi-month to multi-year) stays sales-led. Funded by prepaid credits with a $5 minimum that never expire and Auto Top-Up. A commitment sold to a developer with a card.
- Cursor (Anysphere) · Feb 2026 — Sells usage-style credit pools to individual developers AND prosumers via flat seat tiers — usage logic applied to a non-developer, seat-anchored buyer.
- Intercom · Mar 2024 — Enterprise-ish buyer, yet prices on per-resolution outcome rather than an annual commitment — outcome unit breaks the commit-for-enterprise rule.
- Anthropic · May 2026 — Pure-usage API but explicitly has_commits:false — a frontier vendor large enough to skip the enterprise-commit convention while still selling to enterprise.
Trivia
-
Together AI's on-demand H100 for developers ($5.49/hr) versus reserved commit for enterprise ($4.99/hr) is the corpus's cleanest single example of the buyer-determines-meter thesis: the same physical GPU, the same model, priced in two different structures purely based on whether the buyer is a developer with a card or an enterprise with a procurement cycle. The ~10% discount for commitment is less important than the structural point that both structures exist on the same product.
-
Of the 52 corpus vendors with annual commits, 49 (94%) target the enterprise segment — commitment discounts almost never appear for individual or SMB buyers without an enterprise tier alongside. The commit is an enterprise-procurement artifact, not a loyalty incentive, which is why it travels with the enterprise sales motion rather than with higher usage volumes.
-
Anthropic is the corpus's most notable exception to the buyer-meter rule: it sells to enterprise customers on a pure-usage API with no annual commitment (has_commits: false), a structure that would be atypical for any other enterprise-targeting vendor. It can do this because its brand and frontier-model position give it sufficient negotiating power to avoid the commitment-discount convention that governs the rest of the enterprise inference market.
-
Every machine-emitted billing unit in the 353-company corpus clears 70% pure-usage against a 32% base — gpu-hours 23/26 (88%), requests 34/45 (76%), characters 12/16 (75%), api-calls 39/54 (72%), tokens 52/73 (71%), cpu-hours 14/20 (70%) — while `seats` runs at 25/188 (13%), a 0.41 lift. Six independent units in the same direction is not a correlation you can explain by segment; the unit is doing the work.
-
`credits` (137 companies, the corpus's second most common unit) belongs to neither pole. It lands on freemium 92/137 (67% vs a 48% base) and hybrid 81/137 (59% vs 44%), and on pure-usage barely at all — because a credit is not a unit, it is an abstraction layer bolted over one. Vendors reach for credits precisely when they want to sell a subscription while consuming a machine meter underneath.
-
21 corpus companies bill in tokens without being pure-usage, and 8 of them are the same trade: augment-code, codeium, continue, cursor, tabnine, v0, windsurf and bolt-new buy tokens and resell seats. They are the cohort absorbing usage risk on their customers' behalf — which is also why they are the cohort most often caught reworking quotas (Windsurf's April 2026 credits-to-quotas rework, Codeium's 2026-07-21 fold into Devin's daily/weekly allowance).
-
Hyperbolic broke the enterprise half of this trend on 2026-07-21: reserved GPU capacity, which had been sales-quoted only, became self-serve and prepaid in terms from 1 week to 1 month — paid in full up front, no early termination — with only Private Cloud left behind sales. A commitment a developer can buy with a card in a week is no longer a procurement artifact.
-
Two vendors ran the thesis in opposite directions six weeks apart. RunPod (2026-07-22) published a dedicated Enterprise surface that splits the meter by buyer — reserved baseline at committed-use rates plus usage-based burst, post-paid on one invoice — while GitLab's Flex (2026-06-10) collapses the split, folding seats, GitLab Credits and eligible usage into a single annual dollar commitment that can be reshaped month-to-month without a contract amendment.
For buyers
The structure you're offered signals the segment a vendor has slotted you into. As a developer on a pure-usage rate, don't expect a committed discount until you cross into enterprise sales. As an enterprise, expect an annual commit with overage rather than a clean PAYG meter — and when a vendor offers both, the commit trades a lower unit rate for a volume floor, so model the breakeven before signing.
For vendors
Offer the meter your target buyer's procurement expects: self-serve PAYG to win developers, an annual commit with overage (and often marketplace billing) to win enterprise. Running both on one product lets you land developers and expand into procurement — but keep the unit consistent across them so the upgrade path stays legible.
Outlook — what to watch
Expect the split to harden as agentic products add usage meters for developers and enterprise tiers add commits on top. The exceptions point to where it bends: a frontier vendor large enough to skip commits (Anthropic, has_commits:false), a prosumer product applying usage logic to seats (Cursor), or an outcome unit that replaces the commit entirely (Intercom).
Bottom line
The meter tracks the buyer: 78% of pure-usage vendors sell to developers, 92% of vendors with commitments target enterprise. Structure fits procurement, not compute — so the price shape on offer tells you which buyer a vendor thinks you are.
FAQ
Does AI pricing structure depend on who the buyer is?
Strongly. In the corpus, 78% of pure-usage vendors (88 of 113) sell to developers and 92% of vendors with annual commitments (130 of 141) target enterprise — the billable structure fits the buyer's procurement habits more than the underlying compute.
Why do developers get usage pricing and enterprises get commitments?
Developers want to start instantly and pay for what they use; enterprise procurement wants a committed contract with a negotiated discount and a predictable spend floor. Vendors fit the meter to each.
Can one vendor offer both usage and commitment pricing?
Yes — it's common. Together AI sells on-demand H100s to developers and reserved committed rates to enterprise on the same product; Modal, Anyscale, Fireworks and Vercel do similar. The commit usually trades a lower unit rate for a volume floor.
What are the exceptions to the meter-tracks-buyer rule?
It's a correlation, not a law: Anthropic sells to enterprise on a pure-usage API with no commitments, Cursor applies usage-style credit pools to seat-anchored prosumers, and Intercom prices an enterprise-ish product per outcome instead of on a commit.