What are the most valuable AI infrastructure startups?

Top startups in the AI infrastructure market ranked by valuation

Here is an updated table that ranks the top startups in the AI infrastructure market based on their latest reported or estimated valuations.

If you need to dig deeper and get more detailed data, please check our report on AI infrastructure funding, valuations and who's ahead.

Which AI infrastructure companies are worth the most today?

Cambricon is currently the most valuable AI infrastructure company in our dataset at roughly $105 billion, while Nebius, Cerebras and CoreWeave form a second group worth around $48 billion to $65 billion each.

The first surprise is how mixed the top of the ranking has become. Cambricon and Cerebras sell AI processors. Nebius and CoreWeave run AI clouds. MetaX, Enflame and Moore Threads are chip companies. Crusoe combines data centers, power and cloud infrastructure. VAST Data reaches $30 billion from the data layer rather than from owning giant GPU fleets.

Private companies are also much closer to the public leaders than they were a couple of years ago. Crusoe has just raised $3.9 billion at a $30.9 billion post-money valuation. VAST Data closed its latest round at $30 billion after being valued at $9.1 billion in late 2023. Etched raised $700 million at $21 billion after shipping its first inference rack to Jane Street. VAST went from a $9.1 billion valuation to $30 billion, and we cover what each company is worth, then and now, in our AI infrastructure report.

Nscale needs slightly different treatment. The company was valued at $14.6 billion in a private round earlier this year, while reports around its newly filed IPO point to a possible public valuation around $30 billion to $35 billion. Until that listing is priced, $30 billion-plus is a prospective IPO value rather than a completed valuation.

Company	What it mainly sells	Current valuation reference
Cambricon	AI processors	≈$105B public market cap
Nebius	Full-stack AI cloud	≈$64B public market cap
Cerebras Systems	Wafer-scale AI systems	≈$45B–$50B public market cap
CoreWeave	GPU cloud	≈$48B–$50B public market cap
MetaX	GPU accelerators	≈$35B public market cap
Enflame Technology	AI training processors	≈$33B public market cap
Moore Threads	GPU computing	≈$31B public market cap
Crusoe	Data centers + AI cloud	$30.9B
VAST Data	AI data infrastructure	$30.0B
Nscale	Full-stack AI cloud	$14.6B last private round; ≈$30B–$35B possible IPO

Is Cambricon really worth twice as much as CoreWeave?

Yes, Cambricon's public value is currently around twice CoreWeave's, even though CoreWeave already has a much larger operating business in revenue terms.

CoreWeave generated $2.58 billion of revenue in the second quarter alone, according to its latest SEC filing. Revenue for the first six months reached $4.65 billion, more than double the same period one year earlier. The company also said recently that contracted power had reached about 4.2 GW and that it had added more than $25 billion of new customer commitments early in the third quarter.

Cambricon's valuation reflects something different. Investors are placing a huge premium on an independent Chinese AI-chip supplier at a time when domestic alternatives to Nvidia have become strategically important. Its public value has risen by an order of magnitude from where it stood only a few years ago.

Cerebras makes the comparison even more striking. Its recent market capitalization has been around $45 billion to $50 billion, close to CoreWeave, despite generating just $180 million of revenue in the latest quarter. Cerebras' cloud revenue did grow 281% year over year, so investors are clearly paying for the expected scale of fast inference rather than today's revenue base alone.

The ranking tells us who the market values most, not who currently sells the most infrastructure.

How many AI infrastructure unicorns are there now?

Our dataset contains at least 44 AI infrastructure companies that are clearly worth $1 billion or more, and the number moves into the high 40s if we include companies whose estimated ranges cross the $1 billion line.

That depth is remarkable. The category has already produced more than ten companies around or above $20 billion, roughly 17 at $10 billion or more and more than 30 around the $2 billion mark.

Below the obvious unicorns, several companies sit directly on the boundary. MinIO is estimated at $900 million to $1.4 billion, Moffett AI at $800 million to $1.2 billion, Axiado at $700 million to $1.2 billion and Enfabrica around $900 million to $1 billion. Those should not be counted with the same confidence as a company that has actually completed a financing at $1 billion.

So the cleanest answer is 44 confirmed or strongly anchored billion-dollar companies in our dataset, with several more plausible candidates.

Valuation threshold	Number of companies
$30B+	About 9–10 depending on Nscale's treatment
$20B+	About 11
$10B+	About 17
$5B+	About 20
$2B+	More than 30
Clearly $1B+	At least 44
Potentially $1B+ including valuation ranges	High 40s
Companies analyzed	99

Is most AI infrastructure valuation concentrated in a few companies?

Very much so: roughly 70% of the combined value in our 99-company dataset sits in the top 10 companies.

Using the midpoint when a company has a valuation range, the dataset comes to roughly $660 billion in aggregate value. The first 10 account for around $460 billion. Extend the calculation to the top 20 and we reach roughly $580 billion, close to 88% of the total. Ten companies hold most of the value, and we cover who the top companies are today in our top AI infrastructure companies report.

That leaves 79 companies sharing only about 12% of the value.

This helps explain why AI infrastructure can feel both crowded and concentrated at the same time. There are dozens of GPU clouds, chip designers, storage companies and networking startups, but investors have placed most of their money on a very small number of perceived control points.

A company worth $300 million may be technically impressive. Financially, it barely moves the total when Cambricon alone is worth more than $100 billion.

Group	Approximate combined value	Share of dataset
Top 5	≈$303B	≈46%
Top 10	≈$460B	≈70%
Top 20	≈$580B	≈88%
Remaining 79 companies	≈$80B	≈12%
All 99 companies	≈$660B	100%

Which private AI infrastructure startups are worth the most?

Crusoe and VAST Data currently lead the private AI infrastructure market at around $31 billion and $30 billion, followed by Etched at $21 billion and Fluidstack at roughly $18 billion.

Crusoe has the cleanest very recent anchor. The company announced a $3.9 billion Series F at a $30.9 billion post-money valuation and said its vertically integrated infrastructure platform now has more than $140 billion of total contracted value.

VAST Data reached almost the same valuation from a very different business. Its latest Series F closed at $30 billion, more than triple the $9.1 billion valuation from late 2023. VAST says its systems already support AI environments spanning millions of GPUs.

Etched has moved extraordinarily quickly. The inference-chip company announced a $700 million round at $21 billion after Jane Street tested its hardware and became both investor and customer. Fluidstack has also jumped sharply: after a $7.5 billion Series A earlier this year, top-tier reporting around its latest financing puts the company at roughly $18 billion.

Baseten and SambaNova make up another important tier. Baseten raised $1.5 billion at $13 billion after reporting 20x revenue growth in a year, while SambaNova completed the first close of a $1 billion Series F at an $11 billion post-money valuation.

These are substantial companies by any technology-market standard, even before reaching the public market.

Are neoclouds taking over the AI infrastructure market?

Neoclouds have become one of the most valuable parts of AI infrastructure, but chipmakers and data platforms are still too large for the market to be described as a neocloud takeover.

Nebius is worth roughly $64 billion. CoreWeave sits around $50 billion. Crusoe is worth $30.9 billion, Fluidstack around $18 billion and Lambda remains in the low-to-mid-teens range in our dataset. Nscale's last private round valued it at $14.6 billion, while its planned listing could push that figure much higher. Neoclouds are worth tens of billions, and we weigh which business models are winning in our AI infrastructure market report.

The operating scale is also catching up with the valuations. CoreWeave reported $4.65 billion of revenue in the first half of this year and recently raised contracted power to about 4.2 GW. Nebius reported 479% revenue growth for 2025 and continues to add capacity in the United States, Europe and Asia.

What has changed lately is the definition of a serious neocloud. The largest players increasingly control power, data-center development, networking, financing and software alongside the GPU capacity itself. Crusoe describes itself as vertically integrated from energy generation through AI cloud services. Nscale combines power, data centers, compute and a software platform. Its proposed acquisition of Anyscale would push that model further into software.

A company that simply rents GPUs has a harder story to tell today. The market is paying much more for infrastructure businesses that control several expensive constraints at once.

Why are AI inference startups suddenly worth so much?

Inference has become one of the hottest parts of AI infrastructure because running models for millions of real users can create a larger and more recurring compute workload than training a model once.

Etched is the clearest example. Its valuation rose to $21 billion after it shipped its first rack to Jane Street, which had already tested the company's hardware. Etched is designing chips, racks and software specifically around transformer inference rather than trying to reproduce every workload handled by a general-purpose GPU.

Baseten attacks the same problem from the software side. The company says revenue grew 20x over the last year while inference volume grew 40x. That growth helped support a $13 billion valuation after its latest $1.5 billion financing.

SambaNova reached $11 billion while increasingly positioning its hardware around enterprise inference. Positron went from a valuation above $1 billion earlier this year to $5 billion after raising $875 million for its next-generation inference silicon.

Cerebras gives us a public-market version of the same trend. Its second-quarter cloud revenue rose 281% year over year, much faster than hardware revenue, and the company now has 600 MW of data-center capacity under contract.

Different products, same money flow: investors are paying heavily for businesses that can make production AI faster, cheaper or easier to deploy.

Are AI chip startups or AI cloud startups worth more?

AI chips currently produce the single largest valuation in the market, while AI clouds produce more companies in the $10 billion-to-$60 billion range.

Cambricon sits above $100 billion. Cerebras is around $45 billion to $50 billion. MetaX, Enflame and Moore Threads are all worth tens of billions in our dataset. Etched has reached $21 billion and SambaNova $11 billion.

The cloud side is broader at the upper end. Nebius is around $64 billion, CoreWeave around $50 billion, Crusoe $30.9 billion, Fluidstack about $18 billion and several others sit in or around the $10 billion tier.

Lower down, chips become much more fragmented. Rebellions, DEEPX, d-Matrix and EnCharge AI sit around $2 billion. FuriosaAI remains below $1 billion in our estimate, while Mythic, MemryX and Rain AI are much smaller.

Designing a new accelerator can begin with a relatively small engineering company, which creates many venture-backed attempts. Building a serious AI cloud quickly becomes a multibillion-dollar financing problem.

For now, scarce chips and scarce compute capacity are both receiving enormous premiums.

Which parts of AI infrastructure get the highest valuations?

Compute still attracts the largest valuations, but the money is spreading into data, inference software and networking as clusters become bigger.

Processors have Cambricon above $100 billion and Cerebras around $50 billion. AI clouds have Nebius and CoreWeave in a similar range. Vertically integrated infrastructure has Crusoe above $30 billion, while VAST Data proves that the data layer itself can also support a $30 billion company.

Inference software has now produced Baseten at $13 billion. Optical interconnects have produced Ayar Labs around $3.75 billion from its latest disclosed round, Celestial AI at a $3.5 billion acquisition value and Lightmatter in the multibillion-dollar range. Inference software and optical interconnects now carry multibillion-dollar valuations, and we cover which technologies are gaining traction in our report on AI infrastructure.

The pattern follows the physical constraints of AI systems. More accelerators require more electricity. Bigger clusters need faster networks. More inference creates heavier pressure on memory, storage and data movement. Software then has to make those expensive resources usable.

Investors keep moving toward whichever bottleneck looks hardest to replace next.

Is photonics really becoming a big AI infrastructure market?

Yes, optical interconnects have quietly become a multibillion-dollar AI infrastructure category rather than a collection of laboratory-stage photonics projects.

Ayar Labs closed a $500 million Series E earlier this year at a $3.75 billion valuation, then added another $150 million of capital. The company is moving co-packaged optical technology toward high-volume manufacturing.

Celestial AI gives us an even clearer transaction benchmark. Marvell completed its acquisition of the company for $3.5 billion, according to its latest SEC filing. Marvell explicitly linked the deal to next-generation AI and cloud data-center connectivity.

Lightmatter also sits in the multibillion-dollar range in our dataset, while Lightelligence, AttoTude, Avicena, Mesh, Xscape Photonics, Salience Labs, Oriole Networks and others are attacking different parts of the same problem.

The technical reason is simple enough: electrical connections consume more power and become harder to scale as enormous numbers of accelerators exchange data. AI clusters therefore need better ways to move information across racks and eventually across much larger systems.

There are now enough companies and enough multibillion-dollar outcomes to treat optical connectivity as a real AI infrastructure market of its own.

Which AI infrastructure startups have increased their valuations the fastest?

Etched, Baseten, Positron, VAST Data and Fluidstack stand out because their valuations have multiplied within a short period rather than simply drifting upward with the wider AI boom.

Etched moved from roughly $5 billion to $21 billion through successive financing steps. Positron announced a valuation above $1 billion earlier this year and reached $5 billion with its latest round. VAST Data went from $9.1 billion in late 2023 to $30 billion.

Fluidstack's increase is especially abrupt. The company confirmed an $830 million Series A at a $7.5 billion valuation earlier this year; more recent reporting around its next financing puts the valuation near $18 billion.

Baseten has had an even more dramatic multi-year move. Its earlier rounds valued the company in the hundreds of millions, while the latest Series F values it at $13 billion. The company's own explanation is backed by unusually strong operating growth: 20x revenue and 40x inference volume over one year.

The recurring theme is more useful than the exact percentages. Most of these companies sit in inference, AI data infrastructure or physical compute deployment, areas where demand has recently turned into much larger production workloads.

Company	Earlier valuation anchor	Latest valuation	Rough increase
Etched	≈$5B	$21B	≈4x
Positron	$1B+	$5B	≈5x
VAST Data	$9.1B	$30B	≈3.3x
Fluidstack	$7.5B	≈$18B	≈2.4x
Baseten	Hundreds of millions in earlier rounds	$13B	More than an order of magnitude
Ayar Labs	≈$1B in an earlier round	$3.75B latest disclosed round	≈3.8x

Does raising more money actually produce a higher AI infrastructure valuation?

No. AI infrastructure companies often need huge amounts of capital, but our dataset contains too many counterexamples for total funding to be a reliable predictor of valuation.

CoreWeave and Crusoe fit the intuitive pattern. Both have raised billions because data centers, GPUs, power infrastructure and networking are extremely expensive. Lambda has also added another $1 billion credit facility and a $926 million term loan this year to finance GPU deployments.

VAST Data looks very different. Its tracked funding is below $1 billion in our dataset, yet the latest completed round values the company at $30 billion. VAST sells a data infrastructure layer rather than funding entire GPU fleets itself. VAST is valued at $30 billion on under $1 billion raised, and we trace who has raised the most, and from whom, in our AI infrastructure report.

Past failures show the relationship can work in the opposite direction too. Graphcore raised roughly $684 million but was eventually acquired after losing much of its earlier private-market value. Mythic has raised close to $300 million while our current valuation estimate sits below the amount invested. Untether AI shows an even more extreme version of that problem.

The companies worth more than $10 billion usually have access to a lot of capital, but they reach that valuation for different reasons. Asset-heavy clouds need financing to build capacity. Software and data companies can potentially create far more equity value per dollar raised.

Funding tells us how expensive the journey has been. It does not tell us what the company is worth now.

Is CoreWeave still the best benchmark for AI cloud startups?

Yes, CoreWeave is still the most useful benchmark for neocloud valuations because it gives us both public-market pricing and billions of dollars of actual revenue.

The latest SEC filing makes the scale easy to see. CoreWeave generated $2.58 billion in second-quarter revenue and $4.65 billion in the first half. Active power reached 1.5 GW by the end of the quarter, contracted power reached 3.7 GW, and the company later said that figure had risen to about 4.2 GW.

That gives us a much stronger operating reference than a private startup announcing future capacity. CoreWeave's current market value around $50 billion can be compared against real quarterly revenue, reported expenses and signed customer commitments.

Nebius is particularly interesting beside it. Nebius is currently worth more despite having a much smaller revenue base, which means public investors are paying a much richer price for expected future growth.

Nscale shows the other extreme. Its IPO filing reported just $140.6 million of revenue for the first half of this year, yet the company has more than $103 billion of total contracted value and could seek a valuation around $30 billion.

CoreWeave remains the benchmark because it has already crossed the gap many neoclouds are still trying to cross: turning enormous infrastructure plans into billions of dollars of recognized revenue.

Can Nscale really be worth $30 billion to $35 billion?

Nscale could reach a $30 billion-plus public valuation, but that price would be based overwhelmingly on its future contract book rather than on the business it operates today.

The company's SEC filing shows $140.6 million of revenue in the first half of this year. Reuters reports that Nscale's largest customer accounted for 52% of that revenue. The filing also shows very heavy losses as the company builds data centers and buys infrastructure ahead of revenue.

Now compare that with the forward pipeline. Nscale says its total contracted value has risen above $103 billion from only $100 million about two and a half years ago. The company operates across 14 regions and says its power pipeline exceeds 10 GW. A recent Anthropic agreement alone is worth tens of billions of dollars over its life.

Those numbers explain why investors could entertain a $30 billion valuation. Contracted value spread across multi-year agreements is still very different from annual revenue already being recognized.

Nscale's last completed private round valued the company at $14.6 billion. Reports of $30 billion to $35 billion refer to the possible IPO price. Until shares are actually priced, the lower number remains the firmer observed valuation and the higher number is a public-market target.

Why is VAST Data worth $30 billion without owning a giant GPU cloud?

VAST Data has reached $30 billion because investors are betting that the data layer can capture a large part of AI infrastructure spending without financing every GPU and data center underneath it.

The latest round gives us a clean valuation anchor. VAST closed its Series F at $30 billion, more than triple its $9.1 billion valuation from late 2023. The transaction included about $1 billion of primary and secondary capital.

The company says thousands of organizations use its infrastructure and that its systems support environments spanning millions of GPUs. Customers and users mentioned by VAST include CoreWeave, Cursor, Lowe's and the U.S. Air Force.

That positioning gives VAST a different economic profile from a neocloud. CoreWeave must continuously finance physical capacity. Crusoe builds data-center infrastructure. VAST sells a layer that can sit across different accelerators, clouds and enterprise environments.

As AI systems grow, those customers have to store, move and process huge quantities of data regardless of whether the underlying accelerator comes from Nvidia, AMD or someone else.

A $30 billion valuation for VAST shows that investors no longer see the valuable part of AI infrastructure as compute alone.

Is Baseten really worth $13 billion?

Baseten's $13 billion valuation is aggressive, but the company's recent growth is strong enough to explain why investors are paying far more for inference infrastructure than they did a year or two ago.

Baseten announced a $1.5 billion Series F at a $13 billion valuation. The company says revenue grew 20x over the preceding year while inference volume grew 40x. Revenue up twentyfold in a year is hard to match; we follow who is growing the fastest in our report covering AI infrastructure.

Its customer base also gives those numbers more weight. Baseten works with companies including Cursor, Notion, Harvey, HubSpot, OpenEvidence, Abridge and Decagon: businesses that need models running continuously in production rather than occasionally during experimentation.

Baseten's bet is that running models remains difficult enough for customers to pay an independent platform to handle deployment, optimization and infrastructure across underlying hardware.

The obvious risk comes from the companies surrounding it. Cloud providers, model developers and chip companies can all add more inference tooling themselves. Baseten therefore needs to stay useful even as the basic infrastructure becomes easier to access.

At $13 billion, investors are already assuming it will.

Are smaller AI chip startups being left behind?

Yes, AI-chip valuations are splitting into a small group of very expensive winners and a long tail of companies that remain below $1 billion despite years of development.

At one end we have Cambricon above $100 billion, Cerebras near $50 billion, Etched at $21 billion, SambaNova at $11 billion and Positron at $5 billion. Rebellions, DEEPX, d-Matrix and EnCharge AI form another tier around $2 billion.

Then valuations fall quickly. FuriosaAI is estimated below $1 billion in our dataset. Mythic sits around $145 million to $175 million. MemryX is around $130 million to $180 million. Rain AI is estimated around $100 million to $180 million.

Graphcore is the clearest warning from the previous cycle. The company once had a multibillion-dollar valuation and raised hundreds of millions of dollars before eventually being acquired by SoftBank at a much weaker implied outcome.

The lower end still contains technically interesting companies, including photonics, neuromorphic processors, compute-in-memory chips and edge accelerators. Some could become strategically valuable acquisition targets.

But investors are no longer paying a huge premium simply because a startup has designed a novel AI processor. The company increasingly needs customers, working hardware, manufacturing capacity and a believable path to large production volumes.

Is AI infrastructure becoming more about power and data centers than software?

Power and physical infrastructure have become much more important, but software and data are getting more valuable at the same time.

CoreWeave now has about 4.2 GW of contracted power. Nscale says its development pipeline exceeds 10 GW. Crusoe is building vertically integrated infrastructure from energy through cloud services and reports more than $140 billion in contracted value.

Fluidstack has become another striking example. It has moved from a relatively obscure cloud provider to an approximately $18 billion company while helping develop large AI data-center projects for major customers. The physical buildout itself has become a technology-market category.

Yet VAST Data is worth $30 billion and Baseten $13 billion without following the same asset-heavy model. Their value comes from helping customers use the physical infrastructure more effectively.

Optical networking adds another layer. Marvell paid $3.5 billion for Celestial AI because large accelerator systems need better connectivity. Ayar Labs has raised hundreds of millions to push co-packaged optics into volume production.

The AI stack is becoming more capital intensive from end to end. More software demand creates more compute demand, which creates more demand for power, networking, storage and deployment software.

Who's actually winning the AI infrastructure valuation race right now?

Cambricon currently leads the AI infrastructure valuation ranking, but the broader winners are companies that control one of the bottlenecks created by the massive AI buildout.

Our dataset of 99 companies represents roughly $660 billion of combined value using midpoint estimates. Around 70% sits in the top 10 companies and close to 88% sits in the top 20. At least 44 companies have reached a defensible $1 billion valuation, but the real financial power remains highly concentrated.

Chip companies still occupy some of the largest positions. Cambricon is above $100 billion and Cerebras remains around $50 billion. Neoclouds have become equally important, with Nebius and CoreWeave around the same scale and Crusoe above $30 billion.

The newest shift is happening around inference. Etched is worth $21 billion, Baseten $13 billion, SambaNova $11 billion and Positron $5 billion. Cerebras' cloud revenue is growing far faster than its hardware revenue. The market has moved beyond the first phase of the AI boom, when training capacity attracted most of the attention.

Data and networking are also producing unusually large outcomes. VAST Data is worth $30 billion. Celestial AI sold for $3.5 billion. Ayar Labs reached $3.75 billion in its latest disclosed round. The next valuable bottleneck does not necessarily have to be another GPU.

AI infrastructure is already a huge market, but investors are selective about where they place the biggest valuations. The companies pulling away are the ones that control something difficult to replace: chips, power, deployed compute, inference performance, data movement or high-speed connectivity.