Baseten Revenue and Valuation: How It Grew 20x in a Year

Baseten did not start as the AI inference company it is today.

For its first few years, the startup built a broad platform for machine learning teams. Users could deploy models, add backend logic, connect data, and build interfaces. The product worked, but the business did not grow much. CEO Tuhin Srivastava later told Forbes that revenue from 2019 through the end of 2022 was "basically zero."

Then Baseten cut most of the product and focused on one hard problem: running AI models in production.

The timing was almost perfect.

Stable Diffusion, Whisper, ChatGPT, Llama, DeepSeek, Qwen, and other models made AI useful to far more companies. Those companies suddenly needed GPUs, low latency, autoscaling, monitoring, failover, and engineers who knew how to make models run fast.

Baseten had already spent years building that layer.

By June 2026, Baseten said revenue had grown 20x year over year, inference volume had grown 40x, and its platform was handling more than one billion inference calls per day. That same month, the company raised $1.5 billion in a financing that used valuation tranches of $11 billion and $13 billion.

The short version of Baseten's growth story is:

Baseten started broad, found a painful infrastructure wedge, focused almost entirely on it, and then expanded as the AI market moved toward custom and open models.

The rest of the story is about how it turned that timing into a business.

Baseten's growth in one table

Year	What happened
2019	Tuhin Srivastava, Amir Haghighat, Phil Howes, and Pankaj Gupta founded Baseten.
2021	Baseten opened a closed beta after roughly 18 months of building.
2022	Public launch. The product was still a broad platform for building ML applications. Baseten also open-sourced Truss.
2023	The company narrowed around inference as generative AI and open models created much more demand for model serving.
Mar. 2024	Baseten raised a $40M Series B at a valuation above $200M. Forbes reported revenue had reached the mid-single-digit millions after being close to zero through 2022.
Feb. 2025	Baseten raised a $75M Series C. CNBC reported more than 100 enterprise customers and roughly 6x fiscal-year revenue growth.
Sep. 2025	Baseten raised a $150M Series D at a $2.15B valuation. Srivastava said revenue had grown more than 10x over the previous 12 months.
Feb. 2026	Baseten raised a $300M Series E at a $5B valuation and said inference volume grew 100x during 2025.
Jun. 2026	Baseten raised $1.5B. It said revenue grew 20x YoY, inference volume grew 40x, and it was processing 1B+ inference calls per day.
Sep. 2026	Baseten acquired Blaxel, extending from model inference and training into infrastructure for long-running AI agents.

1. Baseten started with a real problem, but its first product was too broad

The idea for Baseten came from a problem its founders had seen firsthand.

Srivastava, Howes, and Haghighat had worked together at Gumroad. Srivastava and Howes had built a credit-card fraud model there. Training the model was only part of the work. To make it useful, they also needed to turn it into a service, deploy it, monitor it, keep latency low, and make sure it could handle real traffic.

That gap between a model that works in a notebook and a model that works inside a product became the starting point for Baseten.

The founders launched the company in 2019.

Their first product tried to solve most of the machine learning application stack at once. Baseten described four parts of the platform:

Serve: deploy models.
Integrate: add business logic and APIs.
Design: build interfaces with drag-and-drop components.
Ship: put the full application in users' hands.

It was ambitious. It was also a lot of product surface for a small startup.

Baseten was competing, in different parts of the product, with model-serving tools, backend platforms, internal-tool builders, cloud infrastructure, and custom engineering teams.

The product found users. TechCrunch reported customers including Patreon, Cockroach Labs, Stanford, and the University of Sydney when Baseten publicly launched in April 2022.

But usage did not turn into much revenue.

In 2024, Srivastava told Forbes that revenue from the company's founding through the end of 2022 was "basically zero."

That forced the company to ask a useful question:

Which part of the product did customers need badly enough to pay for?

The answer was inference.

2. The biggest growth decision was cutting the product down to inference

Baseten already had a model-serving layer because every application on its original platform needed one.

A production model needs more than a Python file. It needs packaging, dependencies, an API, hardware, autoscaling, monitoring, rollouts, and enough capacity to survive traffic spikes.

Baseten had built much of this infrastructure while selling a much broader application platform.

Then the market changed.

Stable Diffusion and Whisper made powerful open models easy to experiment with. ChatGPT showed that model output could become the product itself. Llama and later open-weight model families gave companies more reasons to run models they controlled.

That changed the value of inference.

Before generative AI, a model might score fraud, moderate content, or help an internal operations team. A little latency was often tolerable.

With AI chat, image generation, search, coding, voice, and agents, users were waiting directly for the model.

Slow inference now meant a slow product.

Baseten narrowed its scope around this problem. Srivastava told Forbes that 2023 made the company realize inference was the bigger opportunity, so it "chopped everything" and focused on it.

That decision gave Baseten a much clearer pitch:

Bring us your model. We will make it fast, reliable, scalable, and easier to run in production.

The market finally matched the infrastructure Baseten had spent years building.

3. Generative AI created the demand wave Baseten needed

Baseten's pivot mattered because of when it happened.

The AI market moved from a world where a relatively small number of ML teams trained specialized models to one where thousands of companies wanted models inside customer-facing products.

Each new model wave created another reason to use infrastructure like Baseten:

Stable Diffusion increased demand for image inference.
Whisper increased demand for speech workloads.
Llama made open-weight LLMs more practical.
DeepSeek, Qwen, and GPT-OSS pushed open models further.
Reasoning models increased the amount of compute used at inference time.
Companies began post-training smaller or specialized models for their own tasks.

Baseten did not create this demand.

It positioned itself at the layer that benefited when the demand arrived.

That distinction matters.

A large part of Baseten's growth came from a market shift: more AI products were being built, more of them used custom or open models, and inference became one of their largest infrastructure costs.

Baseten's job was to make itself the easiest place to run those workloads.

4. Truss lowered the cost of trying Baseten

Baseten also made adoption easier for developers.

In 2022, it open-sourced Truss, a framework that packages model code, dependencies, configuration, and serving logic into a reproducible format.

The idea is simple: a model should behave similarly on a developer's machine and in production.

Instead of manually creating a web server, Docker setup, deployment configuration, and GPU environment, a developer can package the model with Truss and push it to an inference environment.

This mattered for growth in two ways.

First, it reduced friction. A developer could get from model to API faster.

Second, open source created trust. Developers could inspect the format instead of committing their model to an opaque proprietary packaging system.

It is tempting to say open source was a major acquisition channel for Baseten. The public evidence is not strong enough to prove that. Truss had about 1,200 GitHub stars in September 2026, which is useful attention but not massive developer distribution by itself.

A safer conclusion is that Truss helped the developer experience and adoption loop:

build model → package it locally → deploy it → scale it on Baseten.

The commercial value came after the package was created. Baseten could then sell the difficult parts: GPU access, optimized runtimes, autoscaling, observability, failover, and engineering support.

5. Baseten grew before it built a normal sales and marketing machine

One of the strongest signals in Baseten's story came in early 2024.

Forbes reported that Baseten had reached mid-single-digit millions in revenue, tens of thousands of developers, and about 20 large customer logos. It had done much of this before building a conventional sales and marketing team.

That tells us something about the early go-to-market motion.

Baseten was not primarily creating demand through top-of-funnel marketing.

Developers already had the problem.

They could try the product, deploy a model, and start sending requests. As their AI product grew, their inference usage grew too.

That creates a natural expansion loop for usage-based infrastructure:

A developer starts with one model.
The model enters production.
Product traffic grows.
More inference is required.
The customer's Baseten usage grows.
New models or workloads can be moved onto the same platform.

This is one reason infrastructure companies can grow quickly once they land the right workloads. Customer growth and vendor growth can happen at the same time.

6. Forward-deployed engineers helped Baseten win harder accounts

Self-serve developer adoption can get an infrastructure company started.

It is not always enough to win the largest workloads.

Inference performance depends on the model, GPU, serving engine, batch size, quantization, caching strategy, traffic pattern, latency target, and many other details. A customer may know that its model is too slow or too expensive without knowing which part of the stack to change.

Baseten built a Forward Deployed Engineering, or FDE, team for this problem.

Its FDEs sit inside engineering and work directly with customers. They help with problems such as model selection, runtimes, quantization, batching, speculative decoding, autoscaling, and production reliability.

The important part is what happens next.

Baseten says its north star is to move about 70% of what the FDE team builds back into the main product.

That creates a strong product loop:

customer has a hard problem → FDE solves it → solution becomes product → future customers get the improvement too.

Two customer stories show how the motion works.

Gamma

Gamma uses Baseten to serve open-source image models.

Baseten says Gamma now generates more than 3 million images per day on the platform. Gamma reported model-specific image-generation latency improvements of roughly 30% to 80% while scaling without building its own ML infrastructure team.

The case is valuable because Baseten is not only selling compute. It is selling access to engineers who help Gamma keep improving the workload.

Zed

Zed came to Baseten shortly before launching its AI code-editing feature.

According to Baseten's case study, its engineers tested more than 75 configurations. Zed then reported 45% lower p90 latency and 3.6x higher throughput compared with its previous setup.

These are company-published customer cases, so they should not be treated as proof that every customer gets the same result.

But they show the sales motion clearly.

Baseten can land a difficult workload by doing engineering work that would otherwise sit on the customer's team.

7. Technical performance became part of the marketing

Baseten sells to technical buyers.

That changes what good marketing looks like.

A CTO or ML engineer is less likely to be convinced by a generic claim that a platform is "fast." They want to see the model, hardware, latency, throughput, cost, and benchmark setup.

Baseten publishes a large amount of technical content around:

new model launches,
serving benchmarks,
latency,
throughput,
quantization,
GPU performance,
speculative decoding,
autoscaling,
model runtimes,
and infrastructure design.

This content does two jobs.

It helps existing developers solve problems. It also gives new buyers proof that Baseten's team understands the work deeply.

New model releases create repeated marketing moments too.

When a model such as DeepSeek, Qwen, or GPT-OSS becomes popular, developers immediately search for ways to run it quickly and cheaply. An inference provider that supports and benchmarks the model early can enter that demand while interest is highest.

For a developer infrastructure company, shipping fast can become distribution.

Publishing your numbers

If your product wins on something measurable, like speed, cost or accuracy, put the number in public and update it as it improves. Okara's Writer Agent drafts benchmark and technical posts, and the X Agent puts them in front of engineers.

8. Multi-cloud infrastructure turned GPU supply into part of the product

Inference has another problem: getting enough GPUs in the right place at the right time.

Baseten does not manufacture GPUs. It works across cloud providers and regions, then adds software on top.

The company calls this system Multi-cloud Capacity Management, or MCM.

Baseten says MCM treats GPU capacity across more than 20 clouds and regions as one pool. A global control plane can decide where model replicas should run, route traffic, and move workloads when capacity or infrastructure fails.

The company says the infrastructure team spent more than six months building the system after starting the project in early 2024.

This made GPU access itself part of the value proposition.

A customer no longer needed to negotiate and operate every cloud relationship on its own. Baseten could offer access to scarce hardware while also handling placement, scaling, and failover.

This became more important as customer workloads grew.

In June 2026, Baseten said it was processing more than one billion inference calls a day across 87 clusters and 18 clouds.

The exact availability claims need some care. Baseten markets parts of MCM around 99.99% availability, while its public service-level agreement for hosted Dedicated Inference promises 99.9% monthly availability and includes exclusions.

Still, the broader growth point is clear:

Baseten turned fragmented GPU supply into a managed infrastructure product.

9. Better open models made Baseten more valuable

Baseten benefits when companies decide they do not want to send every request to a closed model API.

Open and custom models can give a company more control over:

cost,
latency,
privacy,
deployment location,
model behavior,
and specialization.

The tradeoff is that someone has to run them.

That is where Baseten sits.

As open models improved, the quality gap with closed APIs became smaller for more tasks. That made it reasonable for companies to ask a new question:

Could we get the quality we need from a model we control, while improving cost or speed?

Baseten became one of the companies that helped them test that question in production.

This does not mean open models are always cheaper. Costs depend on model size, utilization, token length, hardware, batching, traffic shape, and the alternative being compared.

But every improvement in open models increases the number of workloads that can move to infrastructure providers such as Baseten.

That market shift became one of Baseten's strongest tailwinds.

10. Customer logos reduced the perceived risk for the next customer

Infrastructure purchasing is partly a trust problem.

A startup can recover from a broken marketing tool. It has a harder time recovering when the model powering its product goes down.

That means customer proof matters.

Over time, Baseten added companies such as Abridge, Writer, Descript, Patreon, Gamma, Clay, Cursor, Notion, OpenEvidence, Mercor, Harvey, Lovable, and others to its public customer list.

Those names become distribution.

A developer evaluating Baseten can see that other high-growth AI products already trust it with production workloads.

By February 2025, CNBC reported that Baseten had more than 100 enterprise customers, plus hundreds of smaller customers.

By September 2025, Srivastava told Fortune that revenue had grown more than 10x over the previous 12 months.

Baseten's enterprise story became easier to sell because the next buyer did not have to be the first one taking the risk.

11. Funding followed usage, then helped Baseten buy more capacity

Baseten's funding history mirrors the growth of the inference market.

The company disclosed a little over $20 million across its seed and Series A when it launched publicly in 2022.

Then the rounds came much faster:

Date	Round	Amount	Reported valuation
Mar. 2024	Series B	$40M	More than $200M
Feb. 2025	Series C	$75M	About $825M
Sep. 2025	Series D	$150M	$2.15B
Feb. 2026	Series E	$300M	$5B
Jun. 2026	Series F	$1.5B	$11B-$13B tranches

The June 2026 financing needs one clarification.

Baseten's headline says it reached a $13 billion valuation. Reporting around the round says the financing included investors entering at both $11 billion and $13 billion valuations.

So "$13 billion valuation" is useful shorthand, but it does not describe every dollar in the round.

The bigger point is how quickly the financing accelerated.

Baseten said the Series F was its fourth fundraise in 18 months. It planned to use the capital for compute, software, and hiring.

For an inference company, capital can itself support growth because large workloads require access to large amounts of expensive GPU capacity.

12. Baseten's growth numbers are huge, but not all of them mean the same thing

Baseten reported several extraordinary growth figures in 2025 and 2026.

They should be kept separate.

In February 2026, the company said inference volume had grown 100x during 2025.

In June 2026, it said:

revenue had grown 20x year over year,
inference volume had grown 40x year over year,
and the platform handled more than 1 billion inference calls per day.

Reuters also reported the 20x revenue-growth claim when covering the Series F.

Separately, Sacra estimates that Baseten reached roughly $600 million in annualized revenue in March 2026. The Information also reported a roughly $600 million annualized revenue pace around the end of the first quarter.

That is not the same thing as Baseten reporting $600 million in annual revenue or ARR.

Annualized revenue usually takes a recent revenue period and extrapolates it over 12 months. Baseten has not publicly disclosed enough detail to calculate recognized annual revenue, gross margin, profitability, or customer concentration.

For that reason, the cleanest metrics to cite are the ones Baseten itself has reported: revenue growth rates, inference growth rates, calls per day, customer counts, and funding.

13. After winning inference, Baseten started expanding again

There is an interesting loop in Baseten's strategy.

The company grew by cutting a broad platform down to a narrow inference product.

Now it is becoming broader again.

The difference is the order.

This time, Baseten is expanding outward from a product with strong demand.

It has moved into training and post-training, partly through its acquisition of Parsed. It launched products that help model labs serve and distribute closed models. And on September 10, 2026, Baseten acquired Blaxel, which builds isolated sandboxes, persistent storage, and networking for autonomous agents.

The company now wants to cover more of the lifecycle around models:

train → improve → serve → distribute → run agents that use them.

That creates a new risk.

The old Baseten struggled partly because it tried to do too much.

The new Baseten has to become a larger platform without losing the depth that made its inference product valuable.

The five growth loops behind Baseten

Baseten's growth is easier to understand as five loops.

1. The market loop

Better open models create more useful AI applications.

More applications create more inference.

More inference creates more demand for Baseten.

2. The developer loop

A developer can package and deploy a model.

If the model reaches production, traffic creates more usage.

More usage creates more Baseten revenue.

3. The FDE loop

A large customer brings a hard production problem.

Baseten engineers solve it.

The best solutions move back into the core product.

The product gets better for the next customer.

4. The capacity loop

More customers create more predictable demand for GPUs.

More scale gives Baseten reasons to secure more cloud capacity.

More capacity lets it support larger customers and traffic spikes.

Those workloads create even more demand.

5. The proof loop

Large AI companies run production workloads on Baseten.

Those logos reduce perceived risk for new buyers.

New buyers become new reference customers.

What founders can learn from Baseten

Baseten's story is not simply "pick a hot market."

The company was early to a market that did not look large enough for years.

A few lessons stand out.

Start with the painful layer

Baseten's original product did many things.

The part customers eventually paid for was the difficult infrastructure underneath them.

Sometimes the strongest startup is not the full workflow. It is the bottleneck every workflow depends on.

Cut scope when the wedge becomes clear

Baseten did not keep adding features to rescue the original product.

It removed large parts of it.

That made the company easier to understand and let a small team go much deeper on one problem.

Let customer growth become your growth

Inference is usage-based.

If a customer's AI product succeeds, its model traffic grows. Baseten benefits without having to resell the account from zero every month.

This makes picking the right customers especially valuable.

For technical products, support can be product development

Baseten's FDE team is not only a services layer.

Its job is to learn from difficult customer problems and push reusable solutions back into the platform.

That turns high-touch enterprise work into product research.

Market timing matters, but preparation matters too

Baseten benefited enormously from the generative-AI boom.

But it could benefit because it had already spent years building serving infrastructure.

When the market changed, it did not need to invent an inference product from scratch. It needed to remove everything around it.

So, how did Baseten grow?

Baseten grew because it stopped trying to own the entire ML application stack and focused on the infrastructure bottleneck underneath it.

Then several forces compounded:

Generative AI created far more inference demand.
Open models gave companies more reasons to run their own models.
Truss made model deployment easier for developers.
Usage-based infrastructure let revenue grow with customer traffic.
Forward-deployed engineers helped Baseten win hard production workloads.
Engineering work from those customers fed back into the product.
Multi-cloud infrastructure gave Baseten access to more GPU capacity.
Strong customer logos made the next enterprise sale easier.
Technical content and fast support for new models captured developer demand.
Funding let Baseten secure more compute and expand the platform.

The key move came before all of them:

Baseten found the part of its product the market wanted most, deleted much of the rest, and went deep.

Frequently asked questions

What does Baseten do?

Baseten provides infrastructure for running AI models in production. Its products cover inference, model deployment, scaling, GPU orchestration, training and post-training, model distribution, and, after its Blaxel acquisition, infrastructure for autonomous agents.

When was Baseten founded?

Baseten was founded in 2019 by Tuhin Srivastava, Amir Haghighat, Phil Howes, and Pankaj Gupta.

What was Baseten's original product?

Baseten originally built a broad platform for complete machine learning applications. It combined model serving, backend logic, APIs, data integrations, and a drag-and-drop interface builder.

Why did Baseten pivot to inference?

The original platform generated little revenue. At the same time, generative AI and open models created much more demand for fast and reliable model serving. Baseten already had an inference layer inside its product, so it narrowed the company around that problem.

How fast is Baseten growing?

In June 2026, Baseten said revenue had grown 20x year over year and inference volume had grown 40x. It also said it was processing more than one billion inference calls per day.

Does Baseten have $600 million in ARR?

Baseten has not publicly reported $600 million in ARR. Sacra estimates roughly $600 million in annualized revenue as of March 2026, and The Information reported a similar annualized revenue pace. Annualized revenue is not the same as recognized annual revenue or company-reported ARR.

Is Baseten valued at $13 billion?

Baseten uses $13 billion as the headline valuation for its June 2026 Series F. Reporting on the financing says the $1.5 billion round included two valuation tranches, at $11 billion and $13 billion.

How much has Baseten raised?

Baseten's disclosed primary funding rounds total roughly $2.085 billion through its June 2026 Series F.

Who uses Baseten?

Publicly named customers include Cursor, Notion, Abridge, OpenEvidence, Writer, Gamma, Clay, Mercor, Harvey, Lovable, Descript, Patreon, and others.

What is Truss?

Truss is Baseten's open-source framework for packaging and serving machine learning models. It helps developers define model code, dependencies, configuration, and serving logic in a reproducible format.