Artificial intelligence platform pricing: buy a seat or run a GPU

An artificial intelligence platform is sold on one of three meters, and which meter you’re on decides more about your bill than which model you picked. You either pay per person per month, or per unit of work the software does, or per hour of the hardware it runs on. Microsoft, Google and Anthropic sell you the first. Google’s API and Salesforce sell you the second. Renting a GPU and running open weights on it is the third. The word “platform” sits on all three, which is why the quotes look incomparable when they land on your desk.

So the buying question isn’t which product is best. It’s which meter matches the way the work arrives at your company. Everything below comes from a vendor pricing page we opened in August 2026, and every derived figure is labelled as our calculation. Where a vendor doesn’t publish a number, we say that instead of estimating it.

Three meters hide behind one word

A seat licence charges for access. The person gets an assistant inside software they already use, and the bill is the same whether they open it forty times a day or never. That’s the model behind Microsoft 365 Copilot, Google Workspace with Gemini, and Claude Team.

A consumption meter charges for output. Tokens on an API, or credits on an agent platform, so the bill tracks how much work actually happened. Nobody pays for a quiet week, which cuts both ways, because nobody caps a loud one either.

The third meter charges for time on a machine. You rent a GPU by the hour, put an open-weight model on it, and the meter runs at the same rate whether the model is saturated or idle. That’s a fixed cost wearing a variable cost’s clothes, and it’s the one people most often mis-price. We walked through the layers that sit on top of it in the self-hosted AI stack, and this piece is about the layer above that: what you buy instead.

A seat bills whether or not the person opens the app. A token bill only appears when work happens. That difference settles more of these purchases than any benchmark does.

Rundowns AI

Seat pricing bills the person, not the work

Microsoft’s business Copilot page lists the Microsoft 365 Copilot Business add-on at “Originally starting from $21.00 now starting from $18.00 user/month, paid yearly”. The bundles that include the productivity suite cost more: Business Standard with Copilot at $23.50 user/month paid yearly, Business Premium with Copilot at $32.00. Those are the yearly-billing figures displayed in August 2026, and only the add-on carries a struck-through original price. It’s also the one to hold onto, because it’s what the AI itself costs on top of software you’re probably already buying.

Google doesn’t separate the two. Its Workspace pricing page, quoted in pounds on the UK site, lists Starter at £5.90, Standard at £11.80 and Plus at £18.40 per user per month, with Enterprise on request. Gemini comes inside every tier rather than as a line item: Starter gets the “Gemini AI assistant in Gmail”, Standard widens that to “Gmail, Docs, Meet, and more”. That’s a different commercial posture from Microsoft’s, and it makes a straight comparison impossible, which is presumably the intent.

Anthropic splits the difference. Claude’s pricing page puts a Team standard seat at “$20 Per seat / month if billed annually. $25 if billed monthly”, a premium seat at $100 annually or $125 monthly, and Team itself at “For teams of 2 to 150”. Enterprise is quoted as “Seat price + usage at API rates”, which is the honest version of what a seat licence really is once agents start doing heavy work in the background.

Product	Unit charged	List price	What the price includes
Microsoft 365 Copilot Business	User seat	From $18.00 user/month, paid yearly	Copilot as an add-on to existing Microsoft 365
Microsoft 365 Business Standard with Copilot	User seat	$23.50 user/month, paid yearly	Productivity suite and Copilot together
Google Workspace Standard	User seat	£11.80 user/month	Gemini in Gmail, Docs, Meet and more
Claude Team, standard seat	User seat	$20 seat/month billed annually	Team workspace, 2 to 150 people
Claude Team, premium seat	User seat	$100 seat/month billed annually	Higher usage allowance per seat
Agentforce add-on (Sales, Service, Field Service)	User seat	$125 user/month	“Unmetered Agentforce usage for employees”

Consumption pricing bills the work, and hides the volume

Token pricing is the cleanest version of a consumption meter. Google’s Gemini API pricing page lists Gemini 3.7 Flash at $0.75 per million input tokens and $3.75 per million output tokens on the standard paid tier, with a note that the price rises on 1 January 2027. Cheaper tiers exist: Gemini 2.5 Flash-Lite runs at $0.10 input and $0.40 output per million. There’s a free tier too, with limited model access and content used to improve the product, which is a cost of a different kind.

Agent platforms wrap the same idea in a currency of their own. Salesforce charges for Agentforce in Flex Credits, which its page describes as “fungible across Actions, Prompts, Translations, and Voice Actions”. Its worked example prices two actions at 40 Flex Credits and US$0.20, then scales that to 24,000 credits and US$120 a month. Divide it out and a credit costs half a cent, so an action costs about $0.10 at 20 credits each. That’s our arithmetic on Salesforce’s own example, not a figure they print.

The catch with credits is that you can’t convert them into anything you can compare. A token is a unit of text, so a token price can be checked against another token price. An “action” is whatever the vendor defines it to be this quarter. Salesforce also sells an Agentforce User License at “$5 USD/User/Month (Requires Flex Credits)”, which is a seat charge that grants you the right to start a consumption meter, and Agentforce 1 Editions “from $550 user / month” bundling 2.5M Flex Credits per org per year. Three meters, one product line.

Running it yourself buys the machine and nothing else

The self-hosted route is more viable than it was, because the open-weight models got small enough to fit a single card. The gpt-oss-120b model card describes “117B parameters with 5.1B active parameters” under a “Permissive Apache 2.0 license”, and says quantisation of the mixture-of-experts weights lets it fit “into a single 80GB GPU (like NVIDIA H100 or AMD MI300X)”. Only 5.1B parameters are active per token, which is why a 117B model runs at all on one card.

Ollama’s library page for the same family puts real numbers on the download: 14GB for the 20b, 65GB for the 120b, both quantised to “4.25 bits per parameter” in MXFP4. It states the smaller model runs “on systems with as little as 16GB memory” and the larger one fits a single 80GB GPU. So the 20b is a laptop proposition and the 120b is a data centre one, and conflating them is the most common error in self-hosting plans.

Serving many people from one card is a separate problem, and it’s the one vLLM exists to solve. Its documentation describes “state-of-the-art serving throughput” from “efficient management of attention key and value memory with PagedAttention” plus “continuous batching of incoming requests, chunked prefill, prefix caching”. The project came out of the Sky Computing Lab at UC Berkeley and now lists over 2,000 contributors. That software is free, which matters less than it sounds, because the hardware under it isn’t.

Amazon’s P5 instance page gives the shape of that hardware: p5.4xlarge with one H100 and “80 GB HBM3”, p5.48xlarge with eight H100s and “640 GB HBM3”, and the P5e and P5en sizes carrying eight H200s at “1128 GB HBM3e”. The page doesn’t publish an hourly rate, so we took a published one from a vendor that does. Lambda’s GPU cloud pricing lists H100 SXM at $3.99 per GPU per hour on an 8x node, $4.29 on a single GPU, and H100 PCIe at $3.29 for one, all “plus applicable sales tax/VAT/GST”. B200 nodes run $6.69 per GPU per hour.

One workload, priced four ways

Prices only compare on a task, so here’s the task. Forty people at one company use an assistant on a working day. Each asks about 30 questions, each answer runs roughly 400 output tokens against 2,000 tokens of input, over 22 working days. That’s 10.56 million output tokens and 52.8 million input tokens a month, and 26,400 assistant turns. The arithmetic below is ours, applied to the list prices above.

How you buy it	The sum	Monthly cost	What you don’t get
Claude Team seats, billed annually	40 seats x $20	$800	Usage above the seat’s allowance
Microsoft 365 Copilot Business add-on	40 seats x $18	$720	Anything outside Microsoft 365
Gemini 3.7 Flash API	52.8M in x $0.75, 10.56M out x $3.75	$79.20	The interface, admin, permissions, audit
Agentforce Flex Credits	26,400 actions x ~$0.10	$2,640	Nothing, if the actions are the CRM work
One rented H100 PCIe, always on	720 hours x $3.29	$2,368.80	Everything above the model weights

The spread is 33 to 1 between the cheapest and the dearest row, which tells you the meter matters more than the vendor. It also flags the trap. That $79.20 is the model, not the product: no interface, no single sign-on, no admin console, no audit trail, no data controls. Building those is a project, and we costed one end to end in what an AI feature costs to run.

Where self-hosting loses, and where the seat does

At 40 daily users, the rented GPU costs three times the dearest seat plan on the list, so it loses and it isn’t close. Work out where it stops losing and the number is brutal. At $2,368.80 a month for one H100 PCIe, and $3.75 per million output tokens on Gemini 3.7 Flash, the GPU only breaks even at about 632 million output tokens a month. That’s our calculation, and at the per-person rate defined above it implies roughly 2,400 daily users of that intensity before the machine pays for itself.

There’s a second condition, and it’s the one that sinks most plans. 632 million output tokens a month is about 244 output tokens every second, sustained, day and night. Whether one H100 running gpt-oss-120b holds that rate depends on batch sizes, context lengths and how bursty the traffic is, and we didn’t benchmark it. Continuous batching is exactly the thing vLLM’s docs claim to do well, so the honest answer is that this is the number to measure before signing anything. We compared the serving options themselves in Ollama vs llama.cpp vs vLLM.

Now the counter-case, in its strongest form. The seat price isn’t a markup on tokens, it’s the price of the things a token doesn’t include. Salesforce sells its $5 user licence with “Access to limited Salesforce CRM objects” attached, which is a plain statement that what you’re buying is the wiring, not the model. That wiring is expensive, dull and hard to keep correct, and paying $18 or $20 a head to skip it is often the cheapest line in the whole comparison. Anyone pricing a build against an API rate is comparing a finished product to a raw ingredient.

Where the seat loses is utilisation. It bills every licensed head whether or not they log in, so a rollout to 500 people that 80 people actually use costs the same as one that all 500 use. A consumption meter can’t fail that way, because an unused agent is a bill of zero. That’s why the metered options tend to win on pilots and lose on mature, heavy, everyday use, while the seat does the reverse.

What the pricing pages can’t tell you

Utilisation is the variable nobody publishes, and the adoption data suggests it’s low. The US Census Bureau’s Business Trends and Outlook Survey found overall AI use among businesses hovering between 17% and 20% from December 2025 to May 2026. It splits hard by size: 37% among firms with at least 250 employees, 32% at 100 to 249, and under 20% at the smallest. By sector, Information ran at 39.7% and Finance and Insurance at 33.9%, against about 14% in Retail Trade.

Those are firm-level rates, not per-employee ones, so they don’t tell you what share of a licensed workforce opens the assistant. Even so, they set the prior. If a fifth of businesses are using AI at all, assuming full seat uptake in your own rollout is optimistic, and a seat plan priced on that assumption is the most likely way to overpay. Which is a discipline question rather than a procurement one, and we looked at what survives contact with production in enterprise artificial intelligence.

Three things would change the conclusion here. The first is a change in the token price, and Google has already flagged one: its 3.x pricing rises on 1 January 2027, which moves the self-hosting break-even in favour of the GPU. The second is a serving benchmark. If one H100 sustains far below 244 output tokens a second on a 120B model, the break-even user count rises and self-hosting gets narrower still.

The third is the quietest. Every seat licence quoted above has a usage allowance behind it that the vendor doesn’t publish, which Anthropic admits by pricing Enterprise as “Seat price + usage at API rates”. If agents keep getting more expensive per task, the unmetered seat becomes hard to sell, and the seat meter quietly turns into a consumption meter with a subscription attached. Watch the words “unmetered” and “unlimited” on those pages. They’re the first thing that moves. For the model access layer underneath all of this, we compared the channels separately in generative AI platforms compared.