AI Infrastructure: what are the top startups now?

Why are AI infrastructure startups attracting so much money right now?

AI infrastructure startups are attracting billions because companies need far more computing power to train AI models, run applications, and handle the growing volume of AI-generated work.

The money involved has reached remarkable levels. Crusoe raised $3.9 billion at a $30.9 billion valuation in September 2026. Fireworks AI secured $1.505 billion at a $17.5 billion valuation in July. Databricks reached approximately $190 billion in August, making it one of the world's most valuable private technology companies.

These investments cover quite different businesses. Crusoe develops computing facilities and AI cloud services, Fireworks specializes in running AI models, and Databricks helps companies manage the data behind their AI applications.

Some startups already generate hundreds of millions or billions in recurring revenue. Others have signed enormous contracts but still need to build the data centers required to fulfill them. And several software companies are growing quickly without financing anything close to the same amount of physical infrastructure.

What should count as an AI infrastructure startup?

AI infrastructure includes the companies providing the computing power, data systems, and software that developers need to build and operate AI applications.

Crusoe, Lambda, and Nscale supply computing capacity through specialized cloud services and data centers. Fireworks AI, Together AI, and Baseten help companies deploy models and process requests efficiently. Databricks and VAST Data handle the large datasets that AI systems need to access.

Developer platforms such as Modal, Runpod, and Hugging Face also belong in this comparison. They provide the tools, computing environments, or model ecosystems that make AI development practical.

For this ranking, we concentrate on privately held companies with meaningful commercial activity, a strong technical position, or major customer commitments. We exclude publicly traded companies such as CoreWeave and Nebius from the startup ranking, although both are useful benchmarks.

Databricks is the largest by private valuation and disclosed revenue. Crusoe has one of the strongest positions in privately held physical AI infrastructure. Fireworks AI stands out in inference, where companies pay to run trained models.

Which AI infrastructure startups are the biggest today?

Databricks is currently the most valuable private AI infrastructure company in our comparison, followed by a group of specialists led by Crusoe, VAST Data, and Fireworks AI.

Databricks reached approximately $190 billion in its August 2026 financing, far ahead of Crusoe's $30.9 billion and VAST Data's $30 billion. Fireworks AI followed at $17.5 billion, while Nscale and Baseten reached $14.6 billion and $13 billion respectively.

Databricks earns money from enterprise software, while Crusoe operates capital-intensive computing infrastructure. Fireworks, Together AI, and Baseten compete over the growing demand for AI inference.

The table uses completed funding valuations rather than reported negotiations over future rounds.

Startup	Main business	Latest confirmed valuation	Commercial evidence
Databricks	Enterprise data and AI	~$190B	$7B+ annualized revenue
Crusoe	AI data centers and cloud	$30.9B	$140B+ contracted value
VAST Data	AI storage and data platform	$30B	Positive free cash flow reported
Fireworks AI	Model inference	$17.5B	$1B+ annualized revenue
Nscale	AI cloud and data centers	$14.6B	Major Microsoft contracts
Baseten	AI inference	$13B	20x annual revenue growth
Together AI	Open-model AI cloud	$8.3B	$1.15B+ annual bookings
Modal	Serverless AI infrastructure	$4.65B	~$300M annualized revenue
Groq	AI inference cloud	$3.5B	6M+ developers reported
Runpod	Developer GPU cloud	$1B	$325M+ ARR

Is Databricks the strongest AI infrastructure startup overall?

Databricks is currently our strongest overall AI infrastructure startup because its commercial scale is far ahead of most private competitors, while its business continues growing exceptionally quickly.

According to its August 2026 financing announcement, Databricks surpassed $7 billion in annualized revenue and recorded more than 80% year-over-year growth in the second quarter. Just six months earlier, its annualized revenue had exceeded $5.4 billion.

That represents at least $1.6 billion in additional annualized revenue over approximately six months.

Databricks has also expanded beyond its original data analytics business. Lakebase, its database platform built for AI applications and agents, passed a $100 million revenue run rate. Its data warehousing business exceeded $1.5 billion, while products such as Genie and Unity AI Gateway help businesses use AI with their own data.

The company benefits from having established enterprise customers already storing and processing information on its platform. Selling them AI products is potentially easier than convincing those same customers to move their data to a completely new provider.

Its $190 billion valuation is demanding, at roughly 27 times disclosed annualized revenue. But among private infrastructure startups, Databricks has the clearest combination of multibillion-dollar sales, rapid growth, and a broad enterprise customer base.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Has Crusoe become the leading private AI computing startup?

Crusoe is now one of the strongest private AI computing companies, with more than $140 billion in contracted value and substantial operational capacity already in place.

In its September 2026 financing announcement, Crusoe reported over 6 gigawatts of gross contracted capacity, including approximately 1 gigawatt already delivered and operational.

Roughly one-sixth of the company's reported contracted capacity was operational, leaving a large pipeline still to be delivered.

Crusoe also reported that Crusoe Cloud bookings had grown more than twentyfold year over year. Its managed inference business had contracted over $100 million in annual recurring revenue.

The company combines energy infrastructure, data centers, GPU computing, and inference services. That makes it unusually integrated compared with cloud startups that depend more heavily on outside data-center operators.

We rank Crusoe first among private, vertically integrated AI computing providers. But there is a lot of construction between its current capacity and its ambitions, and that execution risk is substantial.

Can Lambda and Nscale catch Crusoe?

Lambda and Nscale are credible challengers to Crusoe, but Nscale's recent construction problems show how difficult it is to turn major AI cloud contracts into working infrastructure.

Nscale raised $2 billion at a $14.6 billion valuation in March 2026. Its Microsoft agreement announced in October 2025 covered approximately 200,000 NVIDIA GB300 GPUs across several countries. The company also announced plans to deploy more than 30,000 NVIDIA Rubin GPUs in Norway.

Reuters reported on October 19, 2026, that Microsoft had withdrawn from a planned Nscale data center in Loughton, England. Grid connection delays meant electricity might not be available until the 2030s. Microsoft said the cancellation concerned one site and that its broader relationship with Nscale remained intact.

The Financial Times also reported that Nscale recorded approximately $141 million in first-half revenue against a net loss of roughly $1 billion. The company had more than $100 billion in contracted revenue, according to the report.

Lambda has taken another route. It secured a $926 million financing facility in August and approximately $1 billion in fixed-rate financing in October, expanding its ability to fund GPU deployments.

For now, Crusoe has the stronger disclosed combination of operational capacity and contracted demand. Nscale has enormous commercial ambitions, while Lambda has demonstrated access to major infrastructure financing. Neither has conclusively caught Crusoe on operating scale.

Are AI inference startups becoming the most interesting infrastructure companies?

AI inference startups are currently among the fastest-growing infrastructure businesses because they benefit whenever developers and companies use AI models in production.

Fireworks AI, Together AI, and Baseten provide good examples. Their customers pay to run models, customize them for specific tasks, and manage large volumes of requests without building all the infrastructure themselves.

Inference demand is closely tied to actual application activity. A coding assistant generating suggestions throughout the day, a customer-support agent handling thousands of conversations, or a search service answering questions continuously can create recurring computing demand.

Fireworks reported that the volume processed on its platform rose from 15 trillion to more than 40 trillion tokens daily over one year. That is approximately 167% growth in daily processing volume, although its annualized revenue grew about fivefold during the same period.

Revenue grew much faster than raw token throughput. Customer mix, pricing, specialized workloads, and higher-value services may have contributed, although Fireworks has not disclosed enough detail to quantify each factor.

Inference also gives startups opportunities to build software around deployment, customization, and monitoring. Those capabilities can help them retain customers even as GPU prices change.

Is Fireworks AI winning the inference startup race?

Fireworks AI currently has one of the strongest commercial positions among independent inference startups, with more than $1 billion in annualized revenue and over 40 trillion tokens processed each day.

Its July 2026 announcement included an especially interesting detail: more than 95% of its token volume came from models specialized for customer-specific tasks.

That means over 38 trillion daily tokens were associated with specialized models. This is already a large-scale business in customized production workloads, not simply a place where developers experiment with standard models.

Customers include Uber, Shopify, and Doximity, while Fireworks also works with AI-native companies developing specialized applications.

Its valuation has climbed remarkably quickly. Fireworks was valued at approximately $552 million in July 2024, $4 billion in October 2025, and $17.5 billion in July 2026. Over those two years, its valuation increased more than thirtyfold.

According to Sacra, Fireworks has gross margins of around 50%, with management targeting approximately 60% through better GPU utilization and optimization. That is below the margins typical of many software subscription businesses, but it also suggests the company can earn meaningful margins beyond raw compute resale.

Subsequent reports have suggested discussions around a possible $30 billion to $40 billion valuation, although no completed transaction at those prices has been established.

We currently give Fireworks the lead in specialized inference. The revenue, token volume, and proportion of customized workloads make a stronger case than valuation momentum alone.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Can Together AI and Baseten overtake Fireworks?

Together AI and Baseten have become serious competitors to Fireworks AI, and Baseten's reported growth is especially aggressive.

Together AI raised $800 million at an $8.3 billion valuation in July 2026. Its value had risen from $3.3 billion in February 2025, an increase of approximately 152% over seventeen months.

Reuters reported that Together AI had exceeded $1.15 billion in annual bookings, with customers including Cursor, Cognition, and Decagon. Annual bookings are useful evidence of sales activity, although they cannot be compared directly with Fireworks' recognized revenue run rate.

Baseten's valuation rose from $5 billion in January 2026 to $13 billion in June, increasing 160% in roughly five months. The company also reported twentyfold revenue growth over the preceding year.

The companies have overlapping products but somewhat different strengths. Together AI combines open-model training and inference, while Baseten concentrates on deploying and managing production models. Fireworks has placed particular emphasis on specialized models and high-volume serving.

Reported negotiations suggest investors are still willing to pay much more for these businesses. Bloomberg reported discussions around a potential $26 billion valuation for Baseten, twice its last completed funding valuation.

We currently favor Fireworks for demonstrated revenue and usage scale, but Baseten is advancing quickly. Together AI has a substantial position in the open-model ecosystem, although its disclosed bookings make precise revenue comparisons difficult.

Is Modal becoming a major AI infrastructure startup?

Modal is rapidly becoming one of the most interesting AI infrastructure startups because developers increasingly need somewhere to execute AI-generated code and run computational tasks automatically.

In May 2026, Modal raised $355 million at a $4.65 billion valuation. Reuters reported approximately $300 million in annualized revenue, up from $60 million in September 2025.

That is fivefold growth in roughly eight months.

Modal's CEO Erik Bernhardsson told Reuters that coding-related demand was driving much of the expansion. The company serves customers across industries, including biotechnology, finance, and weather forecasting.

Another revealing detail is how Modal sources its computing capacity. It worked with 13 cloud providers in May, compared with five the previous year. Rather than relying entirely on its own data centers, Modal can access capacity from different suppliers and focus on the software that manages workloads.

More recently, reports indicated that Modal was discussing a funding round at approximately $15.75 billion, over three times its May valuation. That remains a proposed financing rather than a confirmed new price.

Modal already has substantial revenue and a clear connection to growing AI-agent workloads. Its ability to scale across multiple cloud suppliers makes the business particularly interesting.

Is Runpod the most underrated AI infrastructure startup?

Runpod may be one of the most interesting AI infrastructure companies relative to its last funding valuation, especially after its latest revenue announcement.

On October 13, 2026, Runpod reported more than $325 million in annual recurring revenue, up from $120 million in January.

That is roughly 171% growth in nine months, equivalent to adding more than $22 million to its annual recurring revenue run rate per month on average.

Developer adoption has also accelerated. Runpod reported more than 1.2 million developers, compared with approximately 495,000 a year earlier. September alone brought more than 95,000 new signups.

Runpod's June 2026 financing valued it at $1 billion. Comparing that valuation with its subsequently reported $325 million ARR gives a multiple of approximately 3.1 times recurring revenue.

The comparison uses figures from different dates, so it should not be treated as a conventional transaction multiple. Still, Runpod looks unusually inexpensive on this rough measure compared with several private inference companies trading at double-digit revenue multiples.

Its challenge is keeping developers as they graduate from experiments to demanding enterprise applications. Runpod has already been adding enterprise governance tools and security capabilities to support that transition.

If the reported ARR is durable and margins are reasonable, Runpod's last valuation looks surprisingly low for a company growing this quickly. That is a more interesting starting point than simply chasing the startups with the biggest funding rounds.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Can Groq still compete with NVIDIA in AI inference?

Groq remains an important inference infrastructure company, but its relationship with NVIDIA has changed the way we should evaluate its future.

In December 2025, Groq announced a non-exclusive agreement to license its inference technology to NVIDIA. Founder Jonathan Ross and several other team members joined NVIDIA, while Groq continued operating independently.

The company subsequently expanded its financing and infrastructure business. It announced a $350 million Series A at a $3.5 billion valuation in August 2026 and reported more than six million developers using its platform.

Groq also became an NVIDIA Cloud Partner and expanded its inference availability across multiple data centers.

Its original appeal centered on specialized language-processing hardware designed to run AI models efficiently. Following the NVIDIA transaction, its competitive position increasingly involves providing inference services within a broader ecosystem.

The developer audience is substantial, but Groq has disclosed less comparable revenue information than Fireworks AI or Runpod.

We still include Groq among notable infrastructure startups because its technology and distribution remain important. Its commercial standing, however, is harder to establish than its developer adoption suggests.

Why is VAST Data worth $30 billion?

VAST Data has reached a $30 billion valuation because its storage and data software is increasingly used in large AI computing environments, including projects run by major GPU-cloud companies.

In April 2026, VAST completed a financing round that valued the company at $30 billion, up from $9.1 billion in late 2023. That represents approximately 230% valuation growth.

The company helps customers store and process large datasets used to train and operate AI systems. Its customers include CoreWeave, xAI, and the U.S. Air Force.

One particularly concrete example is its agreement with CoreWeave. In November 2025, the companies announced an expanded commercial relationship worth approximately $1.17 billion.

That contract illustrates how AI infrastructure spending spreads beyond GPUs. Large computing clusters also require fast, reliable access to enormous datasets, creating business opportunities for specialized storage software.

VAST has reported positive free cash flow, an encouraging distinction from infrastructure startups still investing heavily to build their initial capacity.

Its recent revenue is less transparent. Sacra previously estimated approximately $200 million in annual recurring revenue in early 2025, but that older estimate should not be used to calculate a reliable current valuation multiple.

Is Hugging Face still one of the most important AI infrastructure startups?

Hugging Face remains one of the most influential AI infrastructure startups because developers depend on its ecosystem to discover, share, and work with open AI models.

The company was last widely valued at $4.5 billion during its 2023 financing. Unlike many competitors, its influence comes mainly from software distribution, community adoption, and developer workflows.

Hugging Face provides model repositories, datasets, libraries, and tools used throughout the open-model ecosystem. This gives it a presence across multiple stages of AI development without requiring it to own enormous GPU campuses.

Its commercial business has been growing as well. Sacra estimated that Hugging Face reached approximately $150 million in annual recurring revenue in August 2026, compared with around $81 million at the end of 2025.

That would represent approximately 85% growth over eight months, although both figures are third-party estimates rather than audited company disclosures.

The challenge is monetization. A platform can be central to how developers work while capturing only a fraction of the money spent on the resulting infrastructure.

We regard Hugging Face as a strategic infrastructure leader, particularly in open AI models. Its importance to developers is arguably greater than its last valuation suggests. In a market obsessed with GPU capacity, that distribution advantage is easy to overlook.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Which AI infrastructure startups are growing the fastest?

Baseten, Fireworks AI, and Modal have reported some of the fastest revenue growth among major AI infrastructure startups, while Databricks stands out for growing rapidly from a much larger base.

Baseten reported twentyfold revenue growth over a year. Fireworks AI and Modal each reported approximately fivefold growth, while Runpod increased its annual recurring revenue from $120 million to more than $325 million during the first nine months of 2026.

The absolute amounts are just as revealing. Databricks added at least $1.6 billion in annualized revenue between February and August, even though its percentage growth was lower than that of smaller competitors.

We should also separate actual revenue from bookings and estimates. Crusoe's twentyfold cloud bookings growth indicates strong sales activity, but bookings are not equivalent to revenue generated during the period.

Company	Reported growth	Period	Measurement
Baseten	20x	Approximately one year	Revenue growth
Fireworks AI	~5x	One year	Annualized revenue
Modal	~5x	September 2025–May 2026	Annualized revenue
Runpod	~2.71x	January–October 2026	ARR
Databricks	>80%	Year over year, Q2 2026	Revenue growth
Together AI	~$618M to ~$1B	Late 2025–February 2026	Sacra revenue estimates
Hugging Face	~$81M to ~$150M	End-2025–August 2026	Sacra ARR estimates
Crusoe Cloud	>20x	Reported year-over-year comparison	Bookings

Are AI infrastructure startups actually making money from their giant contracts?

The largest AI infrastructure startups have secured substantial customer commitments, but the financial results of publicly traded competitors show how expensive delivering those contracts can be.

Crusoe reported over $140 billion in total contracted value. Nscale has also announced enormous customer commitments, particularly with Microsoft and Anthropic.

For comparison, CoreWeave disclosed approximately $104 billion in revenue backlog at the end of June 2026, excluding more than $25 billion of additional commitments signed early in the next quarter.

Despite that backlog, CoreWeave recorded approximately $2.575 billion in second-quarter revenue and a net loss of around $626 million.

Even companies already operating large-scale AI computing businesses can lose substantial amounts while financing equipment and expanding capacity.

Contract value also depends on timing. A customer agreement running for several years may generate revenue gradually, while the supplier has to spend heavily on GPUs, buildings, and infrastructure in advance.

Nscale's recently reported first-half losses reinforce the same concern. It has secured major commitments, but turning those promises into operational projects with attractive financial returns remains a difficult and expensive job.

Is electricity now more important than GPUs for AI infrastructure startups?

Reliable electricity has become one of the biggest competitive advantages for AI data-center startups because power shortages can delay entire projects, regardless of customer demand.

The recent Nscale-Microsoft cancellation in England provides a particularly clear example. Microsoft withdrew from a planned site after delays pushed the expected electricity connection into the 2030s.

That happened despite the two companies having a broader commercial relationship involving large GPU deployments.

Crusoe has placed energy infrastructure at the center of its strategy, combining power access with data-center development and cloud services. Publicly listed Nebius has also been securing large amounts of contracted power capacity, with guidance exceeding 4 gigawatts.

As seen above, Crusoe has reported approximately 1 gigawatt of operational capacity against more than 6 gigawatts contracted. Delivering the remaining capacity requires coordination across grid connections, cooling, construction, equipment supply, and financing.

GPUs can be purchased when supply is available. A suitable site with enough reliable electricity can take years to secure. For the biggest AI computing projects, that difference is becoming crucial.

Can AWS, Microsoft, and Google push AI infrastructure startups out of the market?

AWS, Microsoft, and Google Cloud are formidable competitors, but independent AI infrastructure startups are already supplying services that the hyperscalers themselves need.

Microsoft's contracts with specialized providers such as Nscale and its relationships with other AI computing companies show that hyperscalers sometimes prefer buying outside capacity to building everything internally.

For inference, Fireworks AI and Together AI offer specialized deployment environments and support for customizable models. Modal and Runpod provide developer workflows designed around AI computing tasks rather than the broader needs of traditional cloud customers.

The risk is that major cloud providers can copy attractive features, negotiate better hardware prices, and offer discounts to customers already using their platforms.

Startups are therefore in a stronger position when they provide something difficult to reproduce quickly. For Crusoe, that includes access to physical sites and power. For Fireworks, it involves specialized inference software and customer workloads. For Hugging Face, the advantage is its established developer ecosystem.

Are AI infrastructure startup valuations becoming ridiculous?

Some AI infrastructure startup valuations are already extremely aggressive, especially when investors are paying for growth that has not yet translated into strong profits.

Databricks's $190 billion valuation represents approximately 27 times its disclosed $7 billion annualized revenue. Fireworks AI's $17.5 billion valuation corresponds to about 17.5 times its $1 billion run rate, while Modal's $4.65 billion valuation is around 15.5 times its $300 million annualized revenue.

Those multiples are high, although the businesses have reported exceptionally fast growth.

Their economics also differ. Databricks sells enterprise software, while Fireworks and Modal incur significant costs associated with running computing workloads. A company with 50% gross margins needs much more revenue to generate the same gross profit as a software provider with 80% margins.

The valuation increases have sometimes been faster than the periods over which companies could demonstrate sustained profitability. Baseten's valuation climbed 160% in approximately five months, and Fireworks moved from $4 billion to $17.5 billion in nine months.

Runpod offers an interesting contrast. Its $1 billion June valuation looks relatively modest against the $325 million ARR it subsequently disclosed.

The infrastructure boom is commercially real. But several valuations already assume years of exceptional growth, and not every company can keep multiplying revenue while preserving attractive margins. Some of these prices will be difficult to justify if growth slows even moderately.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

So, what are the top AI infrastructure startups now?

The top AI infrastructure startups right now are Databricks, Crusoe, Fireworks AI, VAST Data, Together AI, Baseten, Nscale, Lambda, Modal, and Runpod, with Hugging Face and Groq also holding important positions in the market.

Databricks is our overall number one. Its multibillion-dollar revenue base, strong growth, and established enterprise customers put it ahead of the private field in commercial scale.

Crusoe leads among privately held, vertically integrated AI computing providers. Its operational capacity and contracted infrastructure pipeline give it a strong position, although a large share of that capacity still needs to be delivered.

Fireworks AI is our leading specialized inference startup based on disclosed revenue and usage volume. Together AI and Baseten are credible challengers, particularly as more businesses deploy customized and open models.

VAST Data is a major name in AI data infrastructure, while Nscale and Lambda are important competitors in specialized computing capacity. Modal and Runpod deserve particular attention for their rapid growth and exposure to developer and AI-agent workloads.

Hugging Face remains central to the open-model ecosystem, and Groq still has a meaningful inference business following its technology agreement with NVIDIA.

The most important distinction is between current commercial strength and future infrastructure potential. Databricks already generates revenue at a scale few private competitors can match. Crusoe has secured an enormous physical infrastructure pipeline. Fireworks has demonstrated that running AI applications can support a billion-dollar revenue business.

If we choose one overall private AI infrastructure leader today, it is Databricks. For physical AI computing infrastructure, we choose Crusoe. And for specialized AI inference, Fireworks AI currently has the strongest disclosed commercial case.