AI Infrastructure Pricing 2026: Buy the Meter First

The short version

Pick the meter before the vendor. For bursty work choose per-second compute that stops when your container does, which favours Modal over Replicate on identical hardware. Reserve per-hour capacity only when a machine is never idle, and treat monthly floors as the real comparison.

An Nvidia H100 costs $3.95 an hour on Modal and $5.49 an hour on Together AI, and that gap is the least useful number in this guide.

Modal quotes the same card at $0.001097 per second and charges only while a container is awake. Together AI quotes $5.49 per GPU-hour for a dedicated endpoint, which is a reservation rather than a stopwatch. Two meters, one piece of silicon, and they respond to completely different things about your workload.

We published a map of this category in May 2026. It named thirty-six tools and printed not a single price, which made it useless for the only question a buyer has. This is the replacement: five tests you can run against any vendor at this layer, then the vendors themselves, priced as read on 6 August 2026.

Four meters cover almost everything sold at this layer

Ignore the category labels for a moment. Infrastructure for AI work bills in one of four ways, and the four behave very differently when your usage moves.

Per token. OpenRouter resells other providers' models and states in its FAQ that it takes "no markup on inference pricing", earning instead on a 5.5% fee when you top up with a card, or 5% in crypto. Together AI's serverless rates run from $0.14 per million input tokens on DeepSeek V4 Flash up to $3.00 in and $15.00 out on Kimi K3.

Per second of machine time. Modal, Replicate and E2B sell wall-clock seconds. The unit is small enough to look harmless. E2B's default two-vCPU sandbox is $0.000028 a second, roughly ten cents an hour before memory.

Per unit the vendor invents. Chatbase sells message credits, Exa sells search requests at $7 per thousand, and LangSmith bills something it calls a Storage Unit at $1.00 each.

Per seat. Tabnine is $39 per user per month for its code assistant tier and $59 for the agentic one, both quoted on annual terms. Headcount is the meter. Usage is not.

Test one: can you predict next month from a number you already track?

A meter is only usable if you can already count the thing it counts.

Tokens you can count, because you send them. Requests you can count. Seats you can count without opening a dashboard at all. Seconds of container time are where teams lose the plot, because the number they track is requests per day and the number they get billed for is how long a machine stayed awake between those requests.

Run the test on a real month. If you cannot multiply one figure from your own logs by a published rate and land within about twenty percent of the invoice, you are buying on a meter you do not yet understand.

That is not a reason to walk away. It is a reason to stay on the free tier until the arithmetic works, and every vendor here has one. Modal gives Starter accounts $30 a month in credits, Exa gives new accounts $20 and then $10 a month, and E2B hands over a one-time $100.

Test two: does the meter stop when you stop?

This is the expensive one, and the answer is in the documentation rather than on the pricing page.

Replicate's billing documentation is unusually direct about it. For private models and deployments, it says, you pay for all the time instances are online: setup time, idle time waiting for requests, and active time processing them. Public models scale down to a minimum instance count, zero by default. Fine-tunes with fast booting are billed only while active.

Read that again with a prototype in mind. A private model left running through a quiet Tuesday bills every second of that Tuesday. On an H100 at $0.001525 a second, a day of doing nothing is $131.

Modal's meter behaves differently: it prices CPU at $0.0000131 per physical core-second with a floor of 0.125 cores per container, and memory at $0.00000222 per GiB-second, all of it tied to a container being up rather than to a reservation. Replicate does not publish a minimum billing increment on its billing page, so the smallest chargeable slice there is not something we can state.

Test three: find the base fee sitting under the meter

Usage pricing is often two prices wearing one coat. There is the meter, and underneath it a monthly floor that bills in a month when you ship nothing at all.

E2B's Pro plan is $150 a month before a single sandbox second, and what the money buys is limits rather than compute: sandbox sessions up to 24 hours instead of one, 100 concurrent sandboxes instead of 20, 20 GiB of storage instead of 10. Modal's Team plan is $250 a month and returns $100 of that as credits, so the true floor is $150. Chatbase Pro is $400 a month for 15,000 message credits.

Those floors are the numbers to compare, not the per-second rates. A team that runs sandboxes two hours a week is paying $150 a month for a session-length limit it will never reach.

The one-hour cap on E2B's free tier is the honest version of this. It is not a crippled product. It is a statement about which workloads the vendor wants on the paid plan, and long-running agent sessions are the answer.

Test four: who gets to define the unit?

Tokens and seconds are units the world agrees on. Credits are not.

Chatbase sells 4,000 message credits for $120 on its Standard plan, which works out at three cents a credit, and sells extra credits at $40 per thousand, which is four cents. Same product, same month, a third more for the marginal message. Extra agents are $300 a year each, and removing Chatbase branding is $1,188 a year, both priced independently of how much anyone uses the thing.

LangSmith is the cleaner illustration, because of what its pricing page does not say. The Developer tier is $0 per seat with up to 5,000 base traces a month, Plus is $39 per seat with up to 10,000, and usage beyond that is metered in Storage Units at $1.00 each.

How many traces make one Storage Unit is not published there. So the cost of an eleven-thousandth trace cannot be worked out from the pricing page alone, and we are not going to guess a conversion rate.

A vendor-defined unit is not dishonest. It is simply repriceable, and a repricing does not look like a price rise on any page a buyer bookmarks.

Test five: price the exit before you sign

Ask what a migration costs on the day you sign, while you still have a choice.

Per-token vendors are the easiest to leave, which is why OpenRouter's own answer is worth reading: bring your own provider key and the first million requests a month cost nothing, after which the fee is 5% of what that model and provider would have cost through OpenRouter. The exit is priced, published, and small.

Tabnine's own-LLM option makes the same move from the other direction. Point it at a model you host and usage is unlimited, with the seat fee covering the product rather than the inference. That is a genuinely different shape of deal from $59 per user per month with the vendor's model behind it, and it only matters if you already run inference somewhere.

Per-second compute sits in the middle. Container images and a Python decorator are portable in a way a proprietary credit balance is not.

Modal against Replicate is the head to head most teams actually face

Both sell GPU seconds. Both have a free entry point. They are close enough that most teams shortlist exactly these two and then choose on brand recognition, which is the wrong tiebreaker.

On price for a single H100, Modal is $0.001097 a second against Replicate's $0.001525, so Replicate is 39% more for the same card. Across the rest of the fleet the gap holds: an A100 with 80GB is $0.000694 on Modal and $0.001400 on Replicate, a T4 is $0.000164 against $0.000225. Replicate's rate card is simpler, with prices quoted for one, two, four and eight card configurations up to $0.012200 a second for eight H100s.

The verdict for a team of two to fifteen with bursty traffic: Modal's Starter plan, because the meter stops. Replicate earns the pick when you are running a public model or a fast-booting fine-tune, where its own documentation says you are billed only for active processing, and when you want a model reachable by URL without writing deployment code.

Neither is the right answer for steady, predictable load. That is what Together AI's dedicated endpoints at $5.49 per H100-hour and its on-demand clusters at $3.99 are for, and per-hour reservations beat per-second billing the moment your machine is never idle.

Who should skip this layer completely

If you are calling a hosted model through an API and storing results in a database you already run, you do not need anything in the infrastructure and platform category yet. Buy the model access, skip the rest.

The trigger for coming back is specific: you need to run code you did not write, or run a model that nobody hosts for you. Executing agent-generated code on your own server is the failure mode that sandboxes exist to prevent, and that is a security decision rather than a cost one.

Teams shopping the code and development tools shelf have a simpler version of the same question. Seat pricing there is the norm, and five seats of Tabnine's agentic tier is $3,540 a year whether the team ships every day or takes August off.

The strongest case against choosing by meter

The obvious rebuttal is that all of this is rounding error. A five-person team spending $200 a month across these vendors could pick the worst option on every test and lose less than one day of an engineer's salary, and the vendor that goes out of business or ships a breaking API change will cost them far more than the meter ever will.

That is largely right, and it is why the tests are ordered the way they are. Test one and test two cost nothing to run and protect you from the surprise invoice. The rest start earning at the moment a base fee appears. Modal Team at $250, E2B Pro at $150 and Chatbase Pro at $400 are all charges that arrive in a month when you shipped nothing, and they arrive every month after that too.

Below a few hundred dollars, buy on whichever developer experience your team will actually use. Above it, the meter is the decision.

What to watch next

The credit is the unit worth watching. Chatbase, Exa and LangSmith each bill in something they define themselves, and a vendor can change what a credit buys without any published number moving, which makes the bookmark on the pricing page a weaker safeguard than it looks. The first vendor at this layer to publish a fixed conversion rate and commit to it in writing will deserve a look on that alone.

Frequently asked questions

How much does an Nvidia H100 cost per hour in 2026?

Modal lists an H100 at $0.001097 per second, which works out at $3.95 an hour, and Replicate lists the same card at $0.001525 per second, or $5.49 an hour. Together AI sells a dedicated HGX H100 endpoint at $5.49 per GPU-hour and an on-demand cluster at $3.99. The per-second vendors only bill while a container is up, so the hourly figures are not directly comparable to a reservation.

Does Replicate charge for idle time?

For private models and deployments, yes. Replicate's billing documentation says you pay for all the time instances are online, including setup time and idle time waiting for requests. Public models scale down to a minimum instance count that defaults to zero, and fine-tunes with fast booting are billed only while actively processing.

What does it cost to run untrusted AI-generated code safely?

E2B's default two-vCPU sandbox is $0.000028 per second plus $0.0000045 per GiB-second of memory, and its free Hobby tier includes a one-time $100 of credit with sessions capped at one hour. The Pro plan is $150 a month and raises that cap to 24 hours with 100 concurrent sandboxes. Modal is the alternative if you want the same isolation inside a general compute platform rather than a dedicated sandbox product.

Is per-seat pricing ever the right choice for AI tools?

It is the right choice when usage is steady and headcount is stable, because the bill is predictable and nobody has to police consumption. Tabnine charges $39 per user per month for its code assistant tier and $59 for the agentic tier on annual terms. The drawback is that a seat bill does not fall in a quiet month, so five agentic seats cost $3,540 a year regardless of output.

What does OpenRouter charge on top of model provider prices?

OpenRouter states in its FAQ that it takes no markup on inference pricing, so token rates match the underlying provider. It earns on credit purchases instead, charging 5.5% with a $0.80 minimum on card payments and 5% on cryptocurrency. Bringing your own provider key is free for the first million requests a month, after which the fee is 5% of the equivalent OpenRouter cost.

Covered in this guide

Modal: AI infrastructure that developers love—serverless compute for ML inference, training, and batch processing
Replicate: Run open-source AI models via API without managing any GPU infrastructure.
Chatbase: AI chatbot platform for customer support. Train on docs, PDFs, website, or Notion. Deploy to web, Slack, WhatsApp, Instagram. Free tier. $40/mo Hobby plan.
E2B: Enterprise sandbox infrastructure for AI agents to execute code securely. Firecracker microVMs, 80ms cold start, Python and JavaScript SDKs. Free with $100 credit, Pro $150/month. 94% Fortune 100 use E2B.
Exa: Neural search API for AI agents, serving 400,000+ developers at $7/1K requests, with a free tier of 1,000 searches/month.
LangSmith: Observe, evaluate, and deploy reliable AI agents
OpenRouter: Single API endpoint for 300+ AI models from OpenAI, Anthropic, Google, and others — one bill, no lock-in.
Tabnine: Enterprise AI coding assistant with privacy-first architecture, agentic workflows, and flexible deployment
Together AI: The AI Native Cloud—full-stack platform for training, fine-tuning, and deploying open-source AI models

Sources

Still deciding?

This guide covers a handful of options. Smart Match checks every listing in the directory against how you actually work and what you can spend, then hands you the shortlist and the reason behind each pick.

Start Smart Match

Related guides

All AI guides Browse the AI directory