Blog · Global · · 5 min read

Inference platforms rent their GPUs. Here is how they earn.

How do inference platforms make money? They charge per token or per GPU-hour, and most rent the GPUs. What that means for GPU providers.

How do inference platforms make money? They sell the output of a model, or the use of a GPU, and they aim to charge more for it than the GPUs cost them. Most of them do not own those GPUs. They rent them from clouds and data centres. That choice shapes what a GPU provider should expect from one as a customer.

What an inference platform does

A company that trains or fine-tunes a model needs somewhere to run it. Running a model for users is called inference. It is a different job from training. The load is steady, it must answer fast, and it has to stay up.

Inference platforms take that job on. Together AI, Fireworks AI, Baseten and Modal do it. So does SiliconFlow in Beijing. A customer sends a request over an API. The platform finds a GPU, runs the model and sends back the answer. The customer never touches a GPU.

Platforms offer two ways to buy. Serverless means you pay for what you use. Dedicated means you reserve a GPU or a group of them for your own model.

How they earn

Serverless is priced per million tokens. A token is a small piece of text. Input and output tokens have separate prices. Dedicated service is priced per GPU-hour, or per second or minute.

All of these prices are public. On 3 November 2026 the same open models cost the same at two of the largest platforms.

Identical prices suggest a market where the model is the same and the buyer can compare. The hour prices are wider. On the same day Modal listed an H100 at $0.001097 a second, about $3.95 an hour. Together AI listed an H100 at $3.99 an hour in a cluster and $5.49 an hour for dedicated inference. Fireworks AI listed $8.00. These are the platforms' own list prices. Large customers rarely pay list. The list price is not the price explains why.

The gap between what a platform pays for a GPU and what it charges for the output is where it earns. That gap is not public. We can only see the revenue side.

  • Fireworks AI. Its chief executive said on X that it passed $800 million of annualised revenue (Newcomer, 29 May 2026).
  • Modal. It said in its funding announcement that it crossed $300 million of annual recurring revenue (Newcomer, 29 May 2026).
  • Baseten. Newcomer reported its figure had reached $600 million. That is reporting, not a company statement.

All three are claims or reports, not audited figures.

Why most of them rent

Newcomer wrote on 29 May 2026 that Baseten, Fireworks AI and Modal lease their capacity. It set them against Lambda and Crusoe, which own the chip stack and also sell inference. The article does not say why the three rent. It notes that leasing raises questions about the gap between cost and price, and that it is unclear whether such firms can sustain growth without owning compute.

The platforms give some reasons of their own.

  • Baseten says its cloud runs thousands of GPUs across several cloud providers and regions. Its engineering book lists four reasons to pool providers: more capacity, protection from outages, lower latency near users, and rules on where data may sit.
  • Modal says it runs a pooled multi-cloud capacity, and that this avoids the usual hyperscaler quota requests. Sacra, an analyst, says the set-up lets Modal grow into new regions without owning data centres. Sacra names AWS, Google Cloud and Oracle as the sources. Modal does not name its hosts on the page we read.
  • SiliconFlow gives the most detail, because it filed to list in Hong Kong on 30 June 2026. KrAsia reports from the prospectus that it mainly leases computing resources through partners. Its five largest suppliers, all compute or cloud providers, were 70.8% of 2025 purchases. The largest alone was 20.4%.

The same report gives SiliconFlow's 2025 revenue as RMB 55.33 million and its net loss as RMB 345 million. Its cost of sales was RMB 68.63 million, above revenue. Computing resources were 86.9% of that cost. These are the company's own filing figures as reported by KrAsia.

Why rent? The sources point to speed and flexibility. A platform with no data centre can add a region or a GPU type by signing a contract. It can also move work when one provider fails. Owning would tie it to one place and one chip. This is our reading of the sources, not a statement from the companies.

What this means for a GPU provider

Selling to an inference platform is not the same as selling to a lab that trains one model for a year. Three things stand out. They are our reading of the sources above.

  • Term. A platform earns by usage, and its usage can change fast. It may prefer shorter or flexible terms. A provider that asks only for long, fixed commitments may lose to one that does not. Price the term, not just the hour.
  • Scale. Platforms buy in blocks and from many sources. SiliconFlow's top five suppliers took 70.8% of its purchases, so a few suppliers can hold a large share. A new provider may start small and grow if it delivers.
  • Reliability. Baseten lists outage protection as a reason to pool providers. A platform that can move work in minutes can also move it away from you. Uptime, fast replacement of failed GPUs and clear service levels are what it will test. How to buy GPU capacity lists what buyers ask on these points.

KrAsia reports that two of SiliconFlow's main partners are both customers and suppliers. A provider should know whether a buyer is also a rival.

What to take from it

Inference platforms are resellers of compute with software on top. Their prices are public. Their costs are not. Their growth is large, but the figures are their own claims. For a GPU provider, they are a demanding customer: they compare, they spread their buying and they leave when service slips.

Sources

Have GPU capacity? Tell us what you hold.

GPUs, site, power and term. Send what you have and what you want for it.

Submit capacity

Read next. More from the compute desk.

Ask an assistant about Casitel.

ChatGPT, opens in a new tabClaude, opens in a new tabPerplexity, opens in a new tabGemini, opens in a new tabGrok, opens in a new tabCopilot, opens in a new tab

The question: What is Casitel (https://casitel.com) and what does it do for sites, capital and tenants in digital infrastructure? Use casitel.com as the source.