What are the top AI infrastructure startups by revenue?

The ranking of top startups in the AI infrastructure market by revenue

Below is a table ranking all the companies in this market by their current revenue scale. You can find our methodology at the end of this page.

If you want a deeper understanding of the market and its current dynamics, get our report on AI infrastructure funding, valuations and who's ahead.

What actually counts as an AI infrastructure startup?

For this ranking, an AI infrastructure startup is an independent, venture-backed company whose core business provides the compute, chips, inference, storage, data systems or software infrastructure used to train and run AI models.

That definition immediately removes companies such as NVIDIA, AMD, Dell, HPE, Broadcom and Arista. They are central to AI infrastructure, but they are mature public companies rather than startups.

We also separate former startups that have already gone public. CoreWeave is the obvious example. Its 2025 revenue reached $5.13 billion, which would put it far above every private company in this ranking, but CoreWeave is now a public company.

Late-stage private companies preparing an IPO can still qualify. Cerebras and Nscale are useful examples because they remain independently operated venture-backed businesses and their filings give us unusually good financial information.

The result is a narrower group: companies such as Fireworks AI, Together AI, Lambda, Baseten, Cerebras, fal, Modal, VAST Data, Fluidstack, Nscale, Runpod, GMI Cloud, Crusoe, Rebellions and FuriosaAI.

Which AI infrastructure startups make the most revenue today?

Fireworks AI currently has the strongest claim to the top spot, while Together AI, Lambda, Baseten and Cerebras make up the rest of the leading group.

The numbers need to be read carefully because private AI infrastructure companies rarely publish the same financial metric. Some disclose annual revenue. Others give ARR, annualized revenue, bookings or contracted revenue. In a few cases, the best figure comes from a credible external estimate.

Fireworks has directly disclosed more than $1 billion in annualized revenue run-rate. Sacra estimates Together AI at roughly $1 billion in annualized revenue and Lambda at around $760 million. Baseten is estimated at about $600 million annualized. Fireworks, Together AI and Lambda sit close together, and we compare who is ahead and who is falling behind in our AI infrastructure report.

Cerebras gives us something cleaner: its SEC filing shows $510 million of recognized 2025 revenue.

Fal is around $400 million in annualized revenue according to Sacra. Modal directly disclosed more than $300 million of annualized revenue. VAST Data and Fluidstack have older ARR estimates around $200 million and $180 million respectively. Nscale reported $140.6 million of revenue in the first half of 2026, while Runpod has directly disclosed more than $120 million ARR.

Rank	Startup	Best current figure	Metric	Evidence
1	Fireworks AI	>$1B	Annualized revenue run-rate	Company disclosed
2	Together AI	~$1B	Annualized revenue estimate	Third-party estimate
3	Lambda	~$760M	Annualized revenue	Third-party estimate
4	Baseten	~$600M	Annualized revenue	Third-party estimate
5	Cerebras	$510M	2025 revenue	SEC filing
6	fal	~$400M	Annualized revenue	Third-party estimate
7	Modal	>$300M	Annualized revenue	Company disclosed
8	VAST Data	~$200M	ARR	Third-party estimate
9	Fluidstack	~$180M	ARR	Third-party estimate, older
10	Nscale	$141M	Six-month revenue	IPO reporting
11	Runpod	>$120M	ARR	Company disclosed

Is Fireworks AI really the biggest private AI infrastructure startup by revenue?

Fireworks AI currently has the cleanest case for number one because the company itself says it has passed $1 billion in annualized revenue run-rate.

Fireworks disclosed the milestone alongside its $1.505 billion Series D. The same announcement said the platform was serving more than 40 trillion tokens per day, with more than 95% of them coming from models specialized for customer workloads.

The wording around the revenue figure is important. Fireworks said annualized revenue run-rate, which tells us what recent revenue would look like over a full year at the same pace. It is more current than a completed fiscal year, but it is also easier for a fast-growing company to reach than $1 billion of revenue already recognized during the previous twelve months.

Even with that caveat, the disclosure is unusually strong for a private AI infrastructure company. It comes directly from Fireworks, it is recent, and it refers to revenue rather than bookings or contract value.

Together AI is close enough to make the top genuinely competitive, but its cleanest company-disclosed billion-dollar figure is bookings. For now, Fireworks has the stronger revenue case.

How much revenue does Together AI make?

Together AI appears to be running at roughly $1 billion in annualized revenue today, although the company itself has publicly emphasized $1.15 billion of annual bookings rather than recognized revenue.

Sacra estimates Together reached approximately $1 billion of annualized revenue after ending 2025 at around $618 million. That would be an extraordinary increase in only a few months.

The company has separately said annual bookings crossed $1.15 billion. Reuters reported the figure when Together announced an $800 million financing round at an $8.3 billion valuation.

Bookings show how much customers have committed to spend, but they can run ahead of revenue when contracts cover future periods or infrastructure that has not yet been delivered.

Together's commercial scale is hard to dispute even if the exact revenue number is less certain than Cerebras' SEC figure. The company now serves thousands of paying customers and has named Cursor, Cognition and Decagon among them.

For the ranking, we use the roughly $1 billion revenue estimate rather than quietly treating $1.15 billion of bookings as revenue.

How big is Lambda now?

Lambda is already around a three-quarter-billion-dollar annualized business and has become one of the largest private GPU-cloud companies still outside public markets.

Sacra estimates Lambda reached approximately $760 million in annualized revenue by the end of 2025, up from about $425 million one year earlier. It also estimates around $700 million of revenue for the full calendar year.

Lambda has changed a lot as it has grown. The company once sold physical GPU workstations and servers, but scarce accelerators became more valuable inside its own cloud. Lambda eventually shut down that hardware business and concentrated more of its capacity on large cloud deployments.

It also moved toward much bigger customers. Sacra documented a four-year $1.5 billion agreement involving NVIDIA and a separate multi-billion-dollar Microsoft deployment covering tens of thousands of GPUs.

The limitation is straightforward: the $760 million figure is an external estimate rather than a filed result. Lambda clearly belongs near the top, but the gap between it and another company a few hundred million dollars away should not be read as accounting-grade precision.

Is Baseten really making around $600 million a year?

Baseten is estimated at roughly $600 million in annualized revenue, putting the inference platform surprisingly close to the largest private GPU clouds.

Sacra estimates Baseten reached that level after sitting around $200 million at the end of 2025. If the estimate is accurate, the company roughly tripled its annualized revenue in a matter of months.

Baseten makes money from production AI workloads rather than conventional software seats. Customers use the platform to deploy models while Baseten handles GPU infrastructure, autoscaling, observability and the surrounding deployment layer. Baseten charges for production workloads rather than software seats, so we weigh which business models are winning in our AI infrastructure market report.

That usage-based model can move very fast when a customer application suddenly takes off. More requests create more inference consumption without waiting for another annual seat-renewal cycle.

We give Baseten's $600 million figure less weight than a filed number because it remains an estimate and represents an annualized pace. Still, the order of magnitude is large enough that Baseten belongs in the top tier.

How much real revenue does Cerebras make?

Cerebras generated $510 million of actual revenue in 2025, giving it the strongest conventional financial disclosure among the largest private AI infrastructure startups.

The company's SEC registration statement shows revenue rising from $290.3 million in 2024 to $510 million in 2025, a 76% increase.

About $358 million came from hardware and roughly $152 million from cloud and other services. That mix makes Cerebras quite different from Fireworks, Baseten or Modal, where more of the business sits in software and cloud services around inference.

Cerebras also reported a 39% gross margin for 2025. The company gives us far more financial detail than most private names in this market.

Its position at number five reflects current scale, not weak evidence. If we ranked only by the quality of reported financial data, Cerebras would look much stronger.

How fast is fal growing?

Fal is currently around a $400 million annualized revenue business after growing more than tenfold from late 2024.

Sacra estimates fal reached roughly $400 million in annualized revenue in early 2026, up from about $285 million at the end of 2025 and only $25 million at the end of 2024.

Fal focuses on generative-media inference. Developers use its infrastructure to run image, video, audio and 3D models without setting up their own GPU stack. The platform says it now supports hundreds of models and millions of developers.

That specialization has become valuable as image and especially video models consume much more compute than simple text requests. Fal captures that spending through API calls and GPU usage instead of selling a fixed software subscription.

The trajectory is one of the steepest in the ranking. The $400 million figure is still an estimate, though, so we treat it as evidence of current scale rather than an audited annual total.

How did Modal get above $300 million in revenue so quickly?

Modal has directly disclosed more than $300 million in annualized revenue after growing fivefold in less than a year.

The company announced the figure when it raised $355 million at a $4.65 billion post-money valuation. Modal raised at a $4.65 billion valuation, and we lay out what each company is worth, then and now, in our report on top AI infrastructure companies.

Modal provides a serverless cloud for AI workloads. Developers can run inference, reinforcement learning, batch jobs and other compute-heavy tasks without manually managing the underlying machines.

The business effect is simple: customers can use a lot of expensive compute while Modal handles when and where the capacity turns on. Revenue can scale with workloads rather than with employee seats.

The fivefold increase also shows how quickly private-company revenue figures go stale in AI infrastructure. A number that looked substantial one year ago can already be badly out of date.

How much revenue does Nscale actually make?

Nscale reported $140.6 million of revenue in the first half of 2026, while its future contracted business is already measured in tens of billions of dollars.

Financial Times reporting on Nscale's IPO filing showed revenue rising from $10.4 million in the comparable six-month period to $140.6 million. That is more than a thirteenfold increase.

At the same time, Nscale had around 55 MW of capacity operating or leased, roughly 1.3 GW under development and about $101 billion of contracts tied to future infrastructure.

Its signed business is therefore far ahead of the infrastructure it has already brought online.

We do not double the $140.6 million six-month figure to invent an annual revenue number. Nscale could finish the year much higher or lower than a simple extrapolation, particularly when large clusters can become operational in chunks.

The revenue figure looks modest beside the backlog, but it is the cleaner measure of what the company has actually delivered so far.

Nscale metric	Reported scale
Prior comparable six-month revenue	$10.4M
Latest six-month revenue	$140.6M
Growth between those periods	>13x
Capacity operating or leased	~55 MW
Capacity under development	~1.3 GW
Contracted value	~$101B
Revenue figure used in our ranking	$140.6M

How big are VAST Data, Fluidstack and Runpod?

VAST Data, Fluidstack and Runpod sit around the $100–200 million range on their best available revenue metrics, but Runpod currently gives us the freshest direct disclosure.

Sacra estimates VAST Data reached about $200 million ARR in early 2025. VAST operates further down the infrastructure stack than the neoclouds, selling data systems designed to keep large AI clusters fed with storage and data.

Fluidstack reached an estimated $180 million ARR at the end of 2024, according to Sacra, after growing 620% during that year. Around $112 million was attributed to its higher-value private-cloud activity and roughly $68 million to the older platform model.

That Fluidstack figure is now old enough that we treat its exact ranking cautiously. The company has expanded materially since then, but replacing the $180 million with a synthetic newer number would just create false precision.

Runpod is easier to pin down. The company has directly said it passed $120 million ARR and now serves more than one million developers. It repeated the $120 million milestone recently, so that remains the strongest public basis for its position.

Is Crusoe much bigger than this revenue ranking makes it look?

Crusoe is clearly much larger commercially than its position in a revenue-only ranking suggests, but its $140 billion-plus contracted value cannot be treated as annual revenue.

Crusoe recently disclosed more than $140 billion in total contracted value across its vertically integrated infrastructure platform. It also said it had more than 6 GW of gross contracted capacity, with about 1 GW already delivered and operational. Crusoe has 6 GW contracted but only about 1 GW delivered, and we dig into the biggest unsolved problems in our report on AI infrastructure.

Those figures put Crusoe in a different league when we look at infrastructure commitments.

The catch is timing. Total contracted value can cover many years of future services, buildings and capacity. A contract signed now may generate revenue gradually as power, data-center space and GPUs become available.

Crusoe has also said bookings in its cloud business increased more than twentyfold year over year and that its newer managed-inference product already has more than $100 million of contracted ARR.

We therefore treat Crusoe as one of the largest private AI infrastructure companies commercially while leaving it outside the numerical revenue ranking until there is a sufficiently current, comparable revenue figure.

How big is GMI Cloud right now?

GMI Cloud has passed $500 million in contracted ARR, but the company has not disclosed the absolute live ARR figure needed to place it cleanly against the revenue leaders.

GMI Cloud said its live ARR increased 2.4 times during the first half of 2026 and that contracted ARR had surpassed $500 million, more than eight times its year-end 2025 level.

It also said its inference business grew 34 times in four weeks and was processing around 2.5 trillion tokens per week.

The contracted number tells us GMI has already sold a large amount of capacity. The live figure would tell us how much of that business is actually operating and generating revenue now.

Without the absolute live ARR number, placing GMI at $500 million beside Cerebras would give the wrong impression. We keep it outside the strict ranking while recognizing that it could move high on the table once more contracted infrastructure goes live.

Are AI chip startups actually generating serious revenue?

Cerebras has already built a half-billion-dollar business, while Rebellions and FuriosaAI show that the next group of AI chip startups is finally moving from pilots into real sales.

Cerebras sits far ahead with $510 million of 2025 revenue.

Rebellions generated roughly ₩35 billion in 2025, according to Korean financial reporting, after about ₩10.3 billion the previous year and roughly ₩2.7 billion in 2023. Its first-generation AI accelerators have been deployed by customers including SK Telecom, KT Cloud and LG Electronics.

FuriosaAI was earlier in commercialization. Seoul Economic Daily reported ₩5.7 billion of 2025 revenue, while the company was targeting ₩80–100 billion as production of its second-generation RNGD accelerator increased.

The gap shows how difficult AI-chip revenue is to scale. Cloud and inference companies can add customer workloads quickly when capacity exists. Chip startups still have to finish designs, manufacture them, qualify them with customers and then move into volume shipments.

AI chip startup	Commercial figure	Metric
Cerebras	$510M	2025 revenue
Cerebras one year earlier	$290M	2024 revenue
Rebellions	~₩35B	2025 revenue
Rebellions one year earlier	~₩10.3B	2024 revenue
Rebellions in 2023	~₩2.7B	2023 revenue
FuriosaAI	₩5.7B	2025 revenue
FuriosaAI current-year target	₩80–100B	Company target

Are AI inference startups now as big as GPU clouds?

AI inference startups have now reached the same broad revenue league as some private GPU clouds, one of the biggest changes in this market.

Baseten is estimated around $600 million annualized. Fal sits around $400 million. Modal has passed $300 million.

Lambda, by comparison, is estimated around $760 million and Runpod has passed $120 million ARR.

A few years ago, most of the infrastructure economics appeared to sit with the companies controlling physical GPU capacity. These days, substantial businesses have formed one layer higher, where companies handle serving, routing, autoscaling, model optimization and developer APIs.

The physical infrastructure is still underneath every one of those services. Developers are simply willing to pay another company to make that compute easier and more efficient to use.

The revenue numbers make the change pretty hard to miss. Inference infrastructure is already a major market in its own right.

How many AI infrastructure startups have crossed $100 million in revenue?

At least eleven private AI infrastructure startups in our research have credible evidence of $100 million-plus revenue, ARR or annualized revenue.

Five have reached roughly $500 million or more on their strongest current operating metric: Fireworks, Together AI, Lambda, Baseten and Cerebras. Five companies at $500 million or more make a short list; we cover the top companies today in our report on AI infrastructure.

Fal and Modal are already well above $300 million.

VAST Data and Fluidstack have crossed the $100 million mark on older ARR estimates. Nscale generated more than that in only six months, and Runpod has directly disclosed $120 million-plus ARR.

The important qualification is that this is a $100 million commercial-scale club rather than eleven companies with identical GAAP revenue. The underlying metrics still differ.

Still, having this many venture-backed infrastructure companies above nine figures shows how far the market has moved beyond experimental spending.

Startup	$100M+ evidence
Fireworks AI	>$1B annualized run-rate
Together AI	~$1B annualized estimate
Lambda	~$760M annualized
Baseten	~$600M annualized
Cerebras	$510M annual revenue
fal	~$400M annualized
Modal	>$300M annualized
VAST Data	~$200M ARR
Fluidstack	~$180M ARR
Nscale	$141M six-month revenue
Runpod	>$120M ARR

How much of this AI infrastructure ranking is based on real revenue?

Only a minority of the largest private AI infrastructure companies currently give us conventional full-year revenue, so the confidence behind each ranking position varies quite a lot.

Cerebras provides the strongest large private-company number because its $510 million is filed annual revenue.

Nscale also gives us hard financial data, although only for a six-month period.

Fireworks and Modal provide direct company disclosures, but their metrics are annualized revenue. Runpod reports ARR.

Together AI, Lambda, Baseten, fal, VAST Data and Fluidstack depend partly or entirely on outside estimates for the figures we use.

That changes how precisely the table should be read. We can be highly confident that the top group consists of very large businesses. We should be much less confident that a $600 million estimated run-rate is precisely $90 million larger than $510 million of filed annual revenue.

The ranking works better when that uncertainty stays visible instead of being hidden behind a clean-looking number.

Are bookings making AI infrastructure startups look bigger than they really are?

Bookings and long-term contracts can make an AI infrastructure company look much bigger than the revenue it is generating today, especially when new data-center capacity still has to be built.

Together AI has reported more than $1.15 billion in annual bookings. That sits close to its roughly $1 billion estimated annualized revenue, so the gap is manageable.

GMI Cloud is harder to compare. More than $500 million of contracted ARR has been disclosed, but the absolute live ARR figure remains private.

Nscale reported roughly $141 million of six-month revenue while carrying about $101 billion of contracts.

Crusoe goes further with more than $140 billion in total contracted value.

These figures are valuable because they show customers are willing to commit serious money. They also help infrastructure companies finance projects that require billions of dollars before the first server starts billing.

For a ranking of today's revenue, though, money that may arrive five or ten years from now cannot carry the same weight as revenue already recognized or a live recurring run-rate.

How concentrated is AI infrastructure startup revenue?

The private AI infrastructure revenue we can measure is heavily concentrated in a small group of companies.

Using the ten companies for which we have the most usable revenue, ARR or annualized-revenue figures, the selected metrics add up to roughly $5.1 billion. Revenue sits with a handful of firms, so we give our read on the likely future winners in our AI infrastructure market report.

The first five account for around $3.9 billion, or roughly three-quarters of that pool.

The top three alone contribute more than half.

These percentages are directional because the underlying figures are not perfectly comparable. The shape of the market is still clear: hundreds of AI infrastructure startups may have raised funding, but only a small number have reached several hundred million dollars of commercial scale.

Group	Combined selected metric	Approx. share
Top 2	~$2.0B	~39%
Top 3	~$2.76B	~54%
Top 5	~$3.87B	~76%
Top 7	~$4.57B	~90%
Remaining three in selected top 10	~$500M	~10%
Selected top 10	~$5.07B	100%
Companies around $500M or more	5	-

Why are AI infrastructure startups growing revenue so fast?

AI infrastructure startups can grow unusually fast because they sell an expensive product that customers consume continuously.

A SaaS startup might need thousands of customers paying $20,000 a year to reach serious scale. One AI infrastructure contract can already be worth tens or hundreds of millions of dollars.

Usage-based inference has a similar effect from another direction. When an AI product suddenly becomes popular, its token, image or video volume can multiply almost immediately. The infrastructure provider grows with that traffic.

Fal went from roughly $25 million annualized at the end of 2024 to around $400 million today according to Sacra's estimates. Modal says its revenue pace increased fivefold before passing $300 million annualized. Baseten's estimated annualized revenue moved from around $200 million to $600 million in a very short period.

Hardware scarcity and large AI budgets have amplified the effect. Customers are willing to make large commitments to secure GPUs, power, data-center space and fast inference before they need all of it.

The result is a set of revenue curves that look almost abnormal compared with conventional enterprise software.

Does a bigger AI infrastructure revenue number mean a better business?

Higher revenue tells us which AI infrastructure companies have reached the most commercial scale, but it does not tell us which company has the strongest economics.

Cerebras sells hardware and cloud services. Lambda increasingly runs large GPU clusters. Fireworks, Baseten and fal sit higher in the inference stack. Modal abstracts cloud infrastructure through serverless compute. VAST Data sells the data layer. These companies sell chips, GPU clusters, inference and data layers; we look at which technologies are gaining traction in our report covering AI infrastructure.

Two companies with $500 million of revenue can therefore have completely different gross margins, capital requirements, customer concentration and financing needs.

CoreWeave provides a useful public benchmark. The company reached $5.13 billion of 2025 revenue, yet it also spent aggressively on infrastructure and reported a large net loss.

Cerebras offers another example from inside the private ranking. Its SEC filing shows a 39% gross margin in 2025, information we simply do not have at the same level for most private rivals.

This article answers a narrower question: who is generating the most revenue? Profitability, capital efficiency and long-term business quality would need separate comparisons.

How small are these startups compared with NVIDIA and the public AI infrastructure giants?

Even today's billion-dollar AI infrastructure startups are still tiny beside the largest public companies supplying the same market.

NVIDIA generated $193.7 billion of Data Center revenue in FY2026.

Supermicro reported $39.1 billion of FY2026 net sales.

CoreWeave, which only recently graduated from the startup category, generated $5.13 billion of revenue in 2025 after producing $1.92 billion one year earlier.

Those comparisons put the private-company numbers in the right order of magnitude. A startup reaching $500 million or $1 billion this quickly is extraordinary, but it still has a long way to go before approaching companies already selling AI infrastructure at industrial scale.

The more interesting development is how many independent businesses are reaching meaningful scale in layers that barely existed a few years ago: specialized inference, serverless GPU compute, open-model clouds and AI-native data systems.

So which AI infrastructure startups are actually the biggest by revenue right now?

Fireworks AI currently has the strongest number-one claim, followed by Together AI, Lambda, Baseten and Cerebras, but the confidence behind those positions is not identical.

Fireworks has the best combination of scale, freshness and direct company disclosure at the top.

Together AI appears to be in the same broad billion-dollar tier, although part of the public evidence around the company comes from bookings and external revenue estimates.

Lambda and Baseten look larger than Cerebras on their latest annualized estimates. Cerebras gives us the cleaner financial proof because its revenue is filed and covers a completed year.

Below them, fal and Modal have already created several-hundred-million-dollar businesses around inference and developer compute. VAST Data, Fluidstack, Nscale and Runpod form the next visible group, while companies such as Crusoe and GMI Cloud could move much higher once more contracted infrastructure turns into live revenue and comparable financial disclosures become available.

Private AI infrastructure now has a genuine billion-dollar top tier, a growing group in the several-hundred-million-dollar range, and at least eleven companies with credible evidence of more than $100 million in commercial scale.

The gaps just are not as clean as the table makes them look. A $510 million SEC revenue figure and a $600 million estimated annualized run-rate belong near each other in a market-scale analysis, but they are different financial facts.