Which AI infrastructure startups generate the most revenue today?

What actually counts as an AI infrastructure startup?

For this ranking, an AI infrastructure startup is a private company whose main business is supplying the compute, inference, deployment, data or storage layer that AI models need to run.

That puts companies such as Crusoe, Lambda, Fireworks AI, Together AI, Baseten, Modal, fal, Fluidstack, Nscale and VAST Data inside the useful comparison. OpenAI and Anthropic sit outside it because they primarily sell models. CoreWeave and Cerebras are useful benchmarks, but both are now public companies.

Databricks and Scale AI are trickier. Databricks now generates more than $7 billion in annualized revenue, but it built a huge data engineering and analytics business long before the generative-AI infrastructure boom. Scale AI supplies training data, evaluation and enterprise AI services, which makes it part of the wider infrastructure stack without looking much like an AI cloud.

We therefore use two definitions. The main ranking covers private AI-first infrastructure companies. Databricks and Scale AI appear later to show how dramatically the answer changes when we widen the category.

Why is AI infrastructure revenue so hard to compare?

AI infrastructure startups report revenue in ways that can make companies look far closer in size than they really are.

Fireworks AI says it has passed $1 billion in annualized revenue run rate. Together AI says annual bookings have crossed $1.15 billion. VAST Data reports more than $500 million of committed annual recurring revenue. Nscale recently showed prospective investors about $103 billion of total contracted revenue.

Those four numbers measure four different things.

Nscale is the clearest example. Its contracts run for about 5.7 years on average, so $103 billion works out to roughly $18 billion a year if we simply divide contract value by contract length. Yet documents reviewed by The Information put Nscale’s actual second-quarter revenue at only a little above $100 million. The $18 billion figure tells us about future contracted capacity, not what Nscale is selling today.

CoreWeave gives us a cleaner public-company benchmark. Its latest quarter produced $2.575 billion of recognized revenue. That is money already earned during the quarter. Its $104.2 billion backlog tells us something different: how much contracted business is still waiting to turn into revenue.

For the ranking below, we give the most weight to recognized revenue and explicit current revenue run rates. Forecast revenue comes next. Bookings, contracted ARR and long-term backlog help us judge future scale, but they do not get treated as current revenue.

Metric	What we can safely read from it	Example
Recognized revenue	Revenue already earned during the period	CoreWeave: $2.575B in latest quarter
Annualized revenue run rate	Current monthly or quarterly pace extrapolated for a year	Fireworks: >$1B
Annual bookings	Business signed during a period	Together AI: >$1.15B
Contracted revenue	Future revenue attached to signed contracts	Nscale: ~$103B

Which private AI infrastructure startups make the most revenue today?

Crusoe and Lambda appear to lead private AI infrastructure today, while Fireworks AI and Together AI have already reached roughly billion-dollar scale in inference.

Crusoe has told investors it expects around $2.2 billion of revenue this year. Lambda is reportedly expecting more than $1.5 billion. Fireworks directly says its annualized revenue run rate has passed $1 billion, and Together AI has crossed $1.15 billion in annual bookings while outside estimates place annualized revenue around the billion-dollar mark.

The next group is already surprisingly large. Fluidstack is projected to generate about $660 million this year. Baseten reached roughly $600 million in annualized revenue by the end of the first quarter, according to figures reported by The Information. VAST Data exited its previous fiscal year above $500 million of committed annual recurring revenue. Sacra estimates fal reached about $400 million in annualized revenue, while Modal directly reported more than $300 million.

The ordering becomes less reliable once we move beyond Fireworks because several figures are forecasts or third-party estimates. The broad conclusion is much firmer than the precise rank: the private AI-infrastructure market now has at least four companies operating around or above a $1 billion annual revenue pace, with another group already in the $300 million to $700 million range.

Company	Best current revenue evidence	How much we trust the comparison
Crusoe	~$2.2B expected revenue	Medium
Lambda	>$1.5B expected revenue	Medium-high
Fireworks AI	>$1B annualized revenue run rate	High
Together AI	>$1.15B annual bookings; revenue around ~$1B by outside estimates	Medium
Fluidstack	~$660M projected revenue	Medium
Baseten	~$600M annualized revenue	Medium-high
VAST Data	>$500M committed ARR	High on CARR, lower for direct revenue comparison
fal	~$400M estimated annualized revenue	Medium
Modal	>$300M annualized revenue	High

Is Crusoe really the biggest private AI infrastructure startup?

Crusoe is our best candidate for the private AI infrastructure revenue lead right now, although its roughly $2.2 billion figure is still a forecast rather than reported full-year revenue.

The scale-up has been unusually fast. Forbes reported about $276 million of revenue for Crusoe in 2024, up 82% in a year. The Information later reported that Crusoe was telling investors revenue could reach roughly $500 million the following year and $2.2 billion this year.

That would mean revenue increasing around eightfold in two years.

Crusoe has also changed what kind of company it is during that period. It sold its bitcoin-mining business and pushed much harder into AI cloud computing and large data-center projects. The company says its contracted footprint now includes several gigawatts of data-center and cloud capacity, while projects tied to OpenAI’s Stargate buildout, Oracle, Microsoft and other large customers are moving through construction.

The catch is that Crusoe mixes businesses with very different revenue timing. Renting GPUs can generate revenue as soon as capacity goes live, while developing a giant data center can involve years of construction before the associated contracts are fully reflected in revenue.

So we put Crusoe first among private pure-play AI infrastructure companies, but with noticeably less confidence than we have in Fireworks’ disclosed $1 billion run rate.

If you want more recent data on this point, please see our latest AI infrastructure market report.

How big is Lambda’s AI cloud business now?

Lambda is now a billion-dollar-scale AI cloud company and appears to be running second only to Crusoe among the large private neoclouds.

The latest reporting around Lambda’s fundraising says the company expects more than $1.5 billion of revenue this year. CB Insights records $520 million for 2025, while Sacra estimated Lambda had already reached a $760 million annualized pace by the end of that year.

That puts the current expectation close to three times the 2025 reported figure.

The customer evidence also looks much stronger than it did a year ago. Lambda has a multi-billion-dollar infrastructure agreement with Microsoft, and Reuters reported that Hudson River Trading is renting more than 1,000 Nvidia Blackwell systems through Lambda. The company has also arranged a $1 billion secured credit facility and another $926 million senior secured loan to finance large deployments.

Those financings tell us something useful about the revenue. Lambda increasingly buys infrastructure against large committed workloads rather than simply filling GPUs one developer at a time. That should make a greater share of future capacity easier to monetize, although it also pushes the company toward the same capital-heavy model that defines CoreWeave.

For now, Lambda looks comfortably above the smaller inference platforms in total revenue and close enough to Crusoe that the final 2026 numbers could still change the order.

Did Fireworks AI really reach $1 billion this fast?

Fireworks AI has genuinely passed $1 billion in annualized revenue run rate, and its growth is among the fastest we found anywhere in AI infrastructure.

The company reported the milestone when it raised $1.505 billion at a $17.5 billion valuation. Fireworks also said revenue was growing fivefold year over year.

The short-term trajectory is even more striking. Around seven weeks before announcing $1 billion, Fireworks said it had crossed an $800 million annualized pace. Independent estimates had put the company closer to $300 million around the end of 2025.

Usage has risen with the revenue. Fireworks now says it serves more than 40 trillion tokens a day, roughly eight times its year-earlier volume. More than 95% comes from models specialized around customers’ own data and tasks.

That helps explain why Fireworks has grown beyond experimental developer traffic. Customers such as Cursor, Harvey, Uber and Shopify are running production workloads through the platform, and those workloads expand as the underlying AI products attract more usage.

Crusoe and Lambda can generate more revenue by building enormous physical infrastructure projects. Fireworks has reached the billion-dollar tier through model serving and inference, which makes its growth especially unusual.

Is Together AI really as big as Fireworks AI?

Together AI is probably operating around the same broad revenue scale as Fireworks AI today, but we have cleaner evidence for Fireworks.

Together AI said its annual bookings had exceeded $1.15 billion when it raised $800 million at an $8.3 billion valuation. The Information has also reported annualized revenue above $1 billion. Those figures put Together firmly in the top private inference group.

Still, bookings can move ahead of actual usage. Together has disclosed several very large enterprise contracts, including one worth more than $1 billion, so signed business can rise much faster than revenue recognized in a particular month.

The underlying usage does support the idea that Together has reached real scale. The company says it serves thousands of paying customers and more than one million developers. Cursor, Cognition and Decagon are among the companies it names publicly, and its platform now covers inference, fine-tuning, training and dedicated GPU clusters.

Fireworks gets the cleaner current-revenue claim because the company explicitly reports more than $1 billion of annualized revenue. Together looks close enough that we would place both in the same commercial tier rather than pretend the available private-company data can reliably separate them by a few hundred million dollars.

If you want more recent data on this point, please see our latest AI infrastructure market report.

How big are Baseten, Modal and fal now?

Baseten, fal and Modal have all crossed into hundreds of millions of dollars in annualized revenue, showing that the AI inference market now supports several large independent platforms rather than one or two breakout winners.

Baseten is the biggest of the three. The Information reported that its annualized revenue jumped from around $200 million at the start of the first quarter to roughly $600 million by the end. A year earlier, it was running at only about one-twentieth of that level.

Fal has followed a similar curve in generative media. Sacra estimates annualized revenue reached about $400 million, up from roughly $285 million at the end of 2025 and about $25 million at the end of 2024. Adobe, Canva, Shopify, Perplexity and Quora are among its public customers, while its platform now offers hundreds of image, video, audio and 3D models.

Modal is smaller in revenue but has recently been growing even faster. When the company raised $355 million at a $4.65 billion valuation, it said annualized revenue had passed $300 million after increasing fivefold since the previous September.

Together these companies show how quickly usage-based AI infrastructure can scale. Baseten is strongest in production inference, fal has concentrated on extremely compute-heavy generative media, and Modal sells a more programmable serverless cloud for inference, agents, reinforcement learning and batch workloads.

The common feature is that customers spend more as their AI applications get busier. These startups do not need customer counts to grow at the same rate as revenue, because one successful application can suddenly consume several times more compute.

Are Fluidstack and Nscale already bigger than Fireworks and Together AI?

Fluidstack and Nscale have signed enough future AI infrastructure business to become much larger than today’s inference leaders, but neither company currently generates anything close to the headline contract values attached to it.

Fluidstack is the easier company to size. Forbes recently reported an investor memo projecting about $660 million of revenue this year, up from roughly $200 million last year. The same memo says Fluidstack could manage as much as 1.3 gigawatts across more than ten sites.

Nscale is where the numbers become wild. Documents shown to prospective investors put total contracted revenue around $103 billion after a $45 billion Anthropic deal. Before that agreement, the company had already been presenting roughly $51 billion of contracted revenue.

Those contracts average about 5.7 years. Dividing them mechanically produces around $18 billion of contracted value per year, which would dwarf every private company in our ranking.

Actual revenue tells a very different story. The Information reported roughly $37 million for Nscale in the first quarter and more than $100 million in the second. Even annualizing the second quarter puts Nscale only somewhere above $400 million at that moment.

Nscale could jump several places very quickly as its contracted capacity comes online. As of now, though, calling it an $18 billion revenue company would confuse a future contract pipeline with a current business.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Does VAST Data belong with AI cloud startups?

VAST Data belongs in the broader AI infrastructure market and is already operating above $500 million in committed annual recurring revenue, although its business looks quite different from a GPU cloud.

VAST sells the data and storage layer behind large AI systems. When the company raised capital at a $30 billion valuation, it said it had surpassed $4 billion in cumulative bookings and exited its previous fiscal year above $500 million in committed annual recurring revenue.

The financial profile is unusual for this group. VAST also said it was operating-margin positive and free-cash-flow positive.

Compare that with companies financing billions of dollars of GPUs and data centers before collecting the associated revenue. VAST can grow by selling more software and data infrastructure into those environments without owning most of the accelerators itself.

We still need to be careful with the number. Committed ARR includes recurring revenue represented by signed contracts, including business that may not yet be fully active. It therefore cannot be dropped directly beside Baseten’s estimated $600 million current annualized revenue and treated as precisely the same thing.

But VAST is clearly one of the biggest private businesses created around the modern AI infrastructure stack, and its profitability makes the revenue more interesting than the raw ranking alone suggests.

Do Databricks and Scale AI completely change the ranking?

Databricks completely changes the answer if we count established data platforms as AI infrastructure, while Scale AI now sits closer to the billion-dollar pure-play group.

Databricks says it has passed $7 billion in annualized revenue and is growing more than 80% year over year. That is several times Crusoe’s projected revenue and roughly seven times Fireworks’ current run rate.

The company is also cash-flow positive on an adjusted basis. Its data-warehousing business alone has crossed a $1.5 billion revenue run rate, while Lakebase has already passed $100 million.

Still, calling Databricks the largest “AI infrastructure startup” hides what readers usually want to know. Databricks was founded in 2013, and a huge share of its revenue comes from data engineering, analytics and warehousing businesses that existed before today’s generative-AI boom.

Scale AI creates a different classification problem. Recent reporting puts annual revenue around or above $1 billion, but the company makes money from data labeling, model evaluation, government work and enterprise AI projects rather than from supplying compute alone.

Our answer depends on the question being asked. Databricks wins easily among private companies that provide infrastructure used for AI. Among companies built specifically around the recent AI compute and inference wave, Crusoe, Lambda, Fireworks and Together give us the more meaningful ranking.

If you want more recent data on this point, please see our latest AI infrastructure market report.

How far ahead is CoreWeave compared with private AI infrastructure startups?

CoreWeave is still several times larger than every private pure-play AI infrastructure startup, and its latest results show that the gap is widening rather than closing.

The now-public AI cloud generated $2.575 billion of revenue in its latest quarter, up 112% from $1.212 billion a year earlier. CoreWeave has since raised its full-year revenue guidance to between $12.4 billion and $13.2 billion.

Take the midpoint and CoreWeave is heading toward roughly six times Crusoe’s projected revenue, more than eight times Lambda’s expected revenue and around thirteen times Fireworks’ current annualized run rate.

Its active power capacity has reached about 1.5 gigawatts, after CoreWeave added nearly 500 megawatts during the latest quarter alone. Revenue backlog stands at roughly $104.2 billion, with more than $25 billion of additional customer commitments signed shortly afterward.

Those numbers give us a useful ceiling for the private-company ranking. Fireworks reaching $1 billion is huge for an inference startup, yet CoreWeave is already producing more than twice that amount every quarter.

Backlog should not be mistaken for present revenue. But CoreWeave has already converted enough capacity into recognized sales to show what the neocloud model can look like once tens of billions of dollars of infrastructure are actually running.

Do the AI infrastructure startups with the most revenue also have the best businesses?

Revenue tells us who has reached commercial scale, but it can badly overstate the quality of an AI infrastructure business when every new dollar requires another large chunk of GPU, power and financing spending.

CoreWeave makes the trade-off easy to see. Its latest quarter produced $2.575 billion of revenue and $1.51 billion of adjusted EBITDA. Yet depreciation and amortization reached roughly $1.39 billion, interest expense hit $640 million, and the company ended the quarter with a $626 million net loss.

The private neoclouds face versions of the same problem. Lambda has arranged close to $2 billion through two recent secured credit facilities. Crusoe expects to keep spending billions as it expands data centers and cloud capacity. Fluidstack and Nscale are pursuing gigawatt-scale infrastructure that has to be financed and built before customers can use it.

Inference providers such as Fireworks, Together and Baseten avoid some of the real-estate burden, but they still pay heavily for underlying compute. A billion dollars of inference revenue therefore should not automatically receive the same economic value as a billion dollars of high-margin SaaS revenue.

VAST Data offers the interesting contrast. The company reports more than $500 million in committed recurring revenue while also saying it is operating-profit and free-cash-flow positive. Modal, meanwhile, has passed $300 million in annualized revenue without trying to own a multi-gigawatt data-center empire.

So when we move beyond the revenue ranking, the most attractive position may belong to companies that can keep growing usage without increasing capital requirements at the same speed.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Which AI infrastructure startups generate the most revenue today?

Crusoe is our best candidate for the largest private pure-play AI infrastructure startup by revenue today, with Lambda close behind and Fireworks AI and Together AI forming a second billion-dollar group.

Crusoe is heading toward roughly $2.2 billion based on figures it has presented to investors. Lambda expects more than $1.5 billion. Fireworks directly reports more than $1 billion in annualized revenue, while Together has crossed $1.15 billion in annual bookings and appears to be around the same broad revenue scale.

Below them, Fluidstack and Baseten are already around the $600 million range. VAST Data has passed $500 million in committed annual recurring revenue, fal is around $400 million by independent estimates, and Modal says it is above $300 million annualized.

Nscale deserves a special mention because its future contracted business is enormous. Its roughly $103 billion of contracts could eventually push it far above today’s private leaders, but its latest reported quarterly revenue was only a little above $100 million. For a ranking of companies making the most money now, the contracts cannot substitute for revenue that has not yet been earned.

If we broaden the definition, Databricks wins comfortably at more than $7 billion annualized. We would still keep it separate because much of Databricks is an established data platform rather than a company born from the current AI-compute boom.

And CoreWeave shows how much further this market can go. The former startup now expects $12.4 billion to $13.2 billion of revenue this year as a public company.

The clearest private-market hierarchy today is Crusoe and Lambda at the top of physical AI infrastructure, followed by Fireworks and Together in inference. Crusoe gets our overall number-one spot for now, but Fireworks has the strongest directly disclosed billion-dollar revenue figure among the newer inference-first companies.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.