What are the most valuable Foundation Model startups?

Top startups in the Foundation Model Market ranked by valuation

Here is an updated table that ranks the top startups in the Foundation Model Market based on their latest reported or estimated valuations.

If you need to dig deeper and get more detailed data, please check our report covering the Foundation Model Market.

Which foundation model startups are worth the most today?

Anthropic and OpenAI sit in a completely different valuation tier from every other foundation model company today.

Our dataset puts Anthropic at roughly $1.5T–$2.0T on an expected IPO basis and OpenAI at roughly $1.1T–$1.2T based on current fundraising discussions. Those figures need one important qualification: Anthropic’s latest completed financing valued it at $965B, while OpenAI’s latest completed round valued it at $852B. The trillion-dollar figures therefore reflect where investors may price the companies next rather than two completed trillion-dollar transactions.

The gap after them is enormous. xAI’s $250B acquisition value is roughly one-quarter of OpenAI’s latest completed valuation and much further below Anthropic’s potential IPO range. DeepSeek comes next around $70B–$75B in our dataset, followed by Zhipu AI at roughly $44B and Thinking Machines Lab around $40B on current fundraising expectations.

Using the midpoint of every valuation range in our 98-company dataset, Anthropic and OpenAI together represent roughly 79% of the total valuation represented by the table. The top 10 account for about 93%. Few technology markets are this top-heavy.

Rank	Company	Current valuation reference
1	Anthropic	$1.5T–$2.0T potential IPO range
2	OpenAI	$1.1T–$1.2T current fundraising discussions
3	xAI	$250B acquisition value
4	DeepSeek	$70B–$75B estimated current range
5	Zhipu AI	~$44.2B public market value
6	Thinking Machines Lab	$38B–$42B fundraising range
7	Moonshot AI	~$35B completed-round valuation
8	Safe Superintelligence	~$32B
9	Reflection AI	~$25B–$28B
10	Mistral AI	~$24B–$25B

Is Anthropic really worth more than OpenAI now?

Anthropic can plausibly become more valuable than OpenAI in the public market, but the strongest completed financing evidence leaves the two much closer than a $2T-versus-$1.2T comparison suggests.

Anthropic completed a $65B Series H at a $965B post-money valuation after reporting that its revenue run rate had crossed $47B. More recent reporting around its expected listing has discussed valuations approaching or exceeding $2T. That is a prospective public-market price, not the price established by the completed Series H.

OpenAI closed $122B of committed capital at an $852B post-money valuation and has since held early discussions about another financing above $1.2T. OpenAI also said earlier in the year that monthly revenue had reached $2B, equivalent to a $24B annualized pace at that point.

So the current evidence supports Anthropic having pulled ahead on the latest completed private valuation: $965B versus $852B. A gap approaching $800B only appears once Anthropic’s potential IPO price is compared with OpenAI’s still-developing private-market fundraising discussions. That larger spread is possible, but it is not settled yet.

How far ahead are Anthropic and OpenAI from the rest of the foundation model market?

Anthropic and OpenAI have created a valuation gap that the rest of the foundation model market is nowhere close to closing today.

Even xAI, valued at $250B when SpaceX acquired it, sits hundreds of billions of dollars behind the two leaders. From xAI the market falls again to roughly $70B–$75B for DeepSeek in our current estimate and about $44B for publicly traded Zhipu AI.

The valuation curve is unusually steep. The midpoint of Anthropic’s current IPO range is around $1.75T. That is about seven times xAI’s $250B acquisition value and roughly 24 times the midpoint of DeepSeek’s estimated range.

Investors are paying for more than the models themselves. Anthropic and OpenAI combine frontier models with huge compute programs, enterprise distribution, developer ecosystems and large-scale consumer or professional adoption. Most smaller labs have one or two of those assets rather than the whole package.

At this scale, the platform surrounding the model can be worth as much as the model itself.

How many foundation model unicorns are there now?

Foundation models have produced an extraordinary number of unicorns: our dataset contains more than 50 companies whose current valuation is at or around $1B or higher.

At least 54 companies in the table have a point estimate or lower end of their range at $1B or above. Several more, including Cartesia and Hume AI, have valuation ranges that cross the $1B mark.

The market has also broadened far beyond general-purpose LLM developers. The billion-dollar group now includes robot-model companies such as Physical Intelligence and Generalist AI, video labs such as Runway and Higgsfield, voice companies such as ElevenLabs, world-model developers such as World Labs and Decart, open-weight specialists such as Arcee AI, and biological-model companies such as Profluent.

That is a big change from the first generative-AI investment wave, when most of the largest valuations clustered around companies building general chat or text models.

Which foundation model startups are worth more than $10 billion?

Roughly 15 foundation model companies in our dataset have valuations clearly above $10B, with several more sitting around that threshold.

The group starts with Anthropic and OpenAI in the trillion-dollar conversation, then xAI at $250B. DeepSeek, Zhipu AI, Thinking Machines Lab, Moonshot AI, Safe Superintelligence, Reflection AI, Mistral AI and ElevenLabs form the next large cluster between roughly $20B and $75B.

Cohere, Kling AI, Skild AI and MiniMax remain comfortably in double-digit billions according to our current evidence. Physical Intelligence reaches roughly $11B on its latest financing evidence, while FieldAI and TypeSafe AI sit around the boundary depending on whether current fundraising discussions close at their indicated prices.

The mix is unusually diverse. A voice-model company, robot-model companies, Chinese LLM labs and a video-model developer can now sit alongside general-purpose frontier labs in the same valuation bracket.

Company	Approximate current valuation	Main foundation-model focus
Anthropic	$1.5T–$2.0T prospective range	Frontier language and agent models
OpenAI	$1.1T–$1.2T fundraising range	General frontier models
xAI	$250B	General frontier models
DeepSeek	$70B–$75B	Open-weight reasoning models
Zhipu AI	~$44B	General language models
Thinking Machines Lab	~$40B fundraising range	Customizable multimodal models
Moonshot AI	~$35B	Language and agent models
Safe Superintelligence	~$32B	Frontier superintelligence research
Reflection AI	~$25B	Open-weight frontier models
Mistral AI	~$24B–$25B	Open-weight enterprise models
ElevenLabs	$22B	Voice and audio models
Cohere	~$18B–$20B fundraising range	Enterprise language models
Kling AI	~$17B–$18B	Video models
Skild AI	~$16B–$18B	Robot foundation models
MiniMax	~$11B	Multimodal models

Are foundation model valuations still rising this fast?

Yes. The largest foundation model valuations are still moving upward at an extraordinary pace, although most of the acceleration is concentrated in a relatively small group.

Anthropic moved from $61.5B in early 2025 to $183B later that year, then $380B and finally $965B in completed financings. Its potential IPO valuation now sits much higher again. OpenAI moved from a $300B financing in 2025 to $852B in its latest completed round, followed by discussions around $1.2T or more.

The same pattern appears lower down the table. ElevenLabs doubled from an $11B valuation in its February financing to $22B in a recent employee tender. Higgsfield went from around $1.3B to $5.4B in roughly eight months. Moonshot AI reached $35B after raising $3.5B, far above its original fundraising target.

Those jumps are too large and too frequent among the winners to dismiss as isolated deals. Revenue, model performance, distribution and strategic importance are being repriced very quickly.

Still, the other 70-plus companies in the table show why the whole market should not be described as moving at the same speed. Many remain below $1B, while some older AI names such as Stability AI trade far below the levels once associated with them. Capital is becoming more selective even as the winners get dramatically more expensive.

Why are investors paying so much for foundation model companies?

Investors are paying huge prices because a successful foundation-model company can potentially control technology, distribution and a large recurring software business at the same time.

The biggest labs increasingly resemble infrastructure platforms. OpenAI combines ChatGPT, APIs, Codex, enterprise products and enormous compute commitments. Anthropic combines Claude, Claude Code, enterprise deployment and distribution through all three major cloud providers. Mistral is building models while also selling enterprise and sovereign-AI infrastructure.

A better model attracts developers and users. More usage creates revenue and gives the company a clearer view of how customers work. That revenue finances more compute and product development, which can support better models and more usage. The loop is expensive, but powerful when it works.

Scarcity plays a role too. Plenty of companies can attempt foundation models; only a small group can recruit frontier researchers, secure tens of billions of dollars of compute, train near the frontier and distribute products globally.

Are foundation model valuations backed by real revenue yet?

Some of the largest foundation model valuations now have substantial revenue behind them. Others are still mostly bets on future technology.

Anthropic said its run-rate revenue crossed $47B before its $965B financing, putting that valuation around 20 times the disclosed revenue run rate. OpenAI said monthly revenue reached $2B earlier in the year, although both its revenue and valuation have continued evolving since then.

Higgsfield gives us a cleaner example lower down the market. The AI video company announced a $5.4B valuation alongside a $700M annualized revenue figure. That implies a multiple below eight times annualized revenue, surprisingly modest next to many frontier labs.

Thinking Machines Lab sits at the other extreme. Reporting around its current fundraising discussions put annual revenue above $100M while investors considered a valuation around $40B. That would imply roughly 400 times annualized revenue.

Safe Superintelligence has attracted billions of dollars despite having no conventional commercial product. Physical Intelligence has also commanded multibillion-dollar valuations while focusing heavily on research and robot intelligence rather than near-term software revenue.

So these companies are being priced on completely different financial timelines.

Company	Valuation evidence	Recent operating evidence	Rough interpretation
Anthropic	$965B completed round	>$47B revenue run rate at Series H	~20x run-rate revenue
OpenAI	$852B completed round	$2B monthly revenue disclosed earlier	Large revenue base, still huge premium
Higgsfield	$5.4B	~$700M annualized revenue	<8x annualized revenue
Suno	$5.4B	~$300M ARR reported	~18x ARR
Thinking Machines Lab	~$40B fundraising discussion	>$100M revenue run rate reported	Around 400x if round closes there
Safe Superintelligence	~$32B	No conventional commercial product disclosed	Research/technology option value
Physical Intelligence	~$11B financing reference	Commercial revenue not central to valuation	Robotics platform bet

Is the foundation model market in a valuation bubble?

Parts of the foundation model market clearly have bubble-like pricing, but the whole category does not fit neatly into one speculative trade.

A company at $40B on roughly $100M of annualized revenue requires an entirely different level of future success from a $5.4B company generating hundreds of millions of dollars annually. A $32B research lab without a commercial product also asks investors to underwrite years of technical execution.

At the same time, several leaders are already building revenue at a scale that would have been extraordinary for private software companies only a few years ago. Anthropic’s disclosed $47B-plus run rate changes the conversation materially. Higgsfield’s rapid move toward $700M in annualized revenue and Suno’s reported $300M ARR show that specialized generative products can turn foundation-model technology into substantial businesses.

The market is bifurcated. Some valuations can increasingly be tested against revenue and adoption. Others remain venture-style bets, just with very large dollar signs attached.

Calling individual foundation-model valuations speculative is easy to defend. Calling all 98 companies part of one indiscriminate bubble is much harder.

Why is Thinking Machines Lab already worth around $40 billion?

Thinking Machines Lab’s roughly $40B fundraising target is mostly a bet on team quality, technical potential and access to capital rather than its current revenue base.

Mira Murati’s company raised about $2B in its original financing at roughly a $12B post-money valuation after recruiting researchers from OpenAI and other leading labs. Investors were effectively funding the team before it had built a mature commercial business.

The company now has products and reportedly more than $100M in annualized revenue, which makes the story more tangible. Yet a $40B valuation would still put the company at hundreds of times that revenue figure.

Investors are effectively buying an option on Thinking Machines becoming another OpenAI or Anthropic. If the company develops a major model platform quickly, the price can work. If it remains a smaller customization or inference business, the current multiple becomes extremely difficult to defend.

Why is Safe Superintelligence worth more than companies with actual products?

Safe Superintelligence is worth roughly $32B because investors are putting an unusually high price on Ilya Sutskever, frontier research talent and the possibility of a major technical breakthrough.

The company has raised billions without building a conventional product business comparable with OpenAI, Anthropic or ElevenLabs. Nvidia alone committed a reported $5B investment alongside a strategic compute partnership that gives SSI access to the Vera Rubin platform and substantially more computing capacity.

This is one of the clearest examples of venture capital valuing expected technological importance before commercial proof.

The wager is extreme, but understandable. A company that discovers a genuinely important new training architecture or reaches a major capability breakthrough could become one of the most valuable AI companies in the world very quickly. Investors are accepting a valuation that normal software metrics simply cannot support.

SSI also shows why applying one revenue multiple across this market would be nonsense. Its value today comes from research option value, talent concentration and strategic scarcity.

Are Chinese foundation model companies catching the US leaders?

Chinese foundation model companies have created a strong second tier, but the valuation gap with OpenAI and Anthropic remains enormous.

DeepSeek sits around $70B–$75B in our current estimate, while Moonshot AI recently reached roughly $35B after a $3.5B financing. Zhipu AI’s public-market value is around the mid-$40B range in our dataset, and MiniMax sits near $11B. Baichuan, ModelBest, StepFun, PixVerse, AiShi Technology and several other Chinese labs add meaningful depth further down.

China is particularly strong in open-weight language models, video generation and world models. Moonshot’s Kimi family, DeepSeek’s reasoning models, Zhipu’s GLM family and a growing collection of Chinese video labs have made the competition much broader than a simple OpenAI-versus-Anthropic race.

The financial gap is still striking. DeepSeek at approximately $72.5B on the midpoint of our range would be worth less than one-tenth of Anthropic’s last completed $965B financing valuation.

Chinese labs therefore look increasingly competitive technically while remaining much cheaper financially. Revenue, access to global capital, geopolitical risk, public-market conditions and expectations around international distribution can all influence that difference, not just raw model capability.

Is DeepSeek really a $70 billion company?

A $70B-plus DeepSeek valuation is plausible in the current market, but we would treat it as a current financing estimate rather than a fully observed transaction price.

DeepSeek has become one of the most strategically important model developers in China and one of the best-known open-weight challengers globally. That status supports a very large premium over conventional AI startups.

The comparison set makes the range credible. Moonshot completed a financing around $35B. Reflection AI has reached roughly $25B while positioning itself as an American open-weight competitor. Mistral raised capital above €21B post-money. A stronger valuation for DeepSeek fits the hierarchy implied by its model visibility and strategic importance.

Still, $70B–$75B does not have the same evidentiary strength as ElevenLabs’ $22B tender or Mistral’s completed financing. Our dataset therefore treats DeepSeek’s figure as an estimated active-raise valuation rather than a confirmed transaction.

How valuable are Europe’s foundation model startups?

Europe has produced one genuinely global foundation-model heavyweight in Mistral AI, but there is still a major valuation gap between Mistral and the US frontier leaders.

Mistral recently raised €3B at a post-money valuation above €21B, roughly consistent with the $24B–$25B range in our dataset. The financing was the largest equity fundraising completed by a European technology company according to Mistral, and the company says it now works with more than 125 major enterprises.

Europe also has strong specialized companies. ElevenLabs was founded by European entrepreneurs and has major operations in London, although it is incorporated in the United States; its latest tender valued it at $22B. Black Forest Labs in Germany reached a $3.25B valuation in its Series B. Bioptimus is developing biological foundation models from France, while smaller European teams are appearing in audio, design and scientific models.

Even with those companies, Europe has no equivalent to the trillion-dollar valuation tier occupied by Anthropic and potentially OpenAI. Mistral’s position is strategically important precisely because the European field remains relatively concentrated around it.

Are open-weight foundation models becoming more valuable?

Open-weight foundation-model companies are attracting much larger valuations now, and Reflection AI gives us some of the clearest evidence that investors see open models as a serious commercial category.

Reflection is valued around $25B after scaling from an $8B valuation in its previous financing. The company has positioned itself around open-weight frontier models that enterprises and governments can run with greater control over infrastructure and data.

Mistral has followed a related strategy while reaching a valuation above €21B. Arcee AI recently crossed $1B after raising capital to expand its Trinity open models. Black Forest Labs reached $3.25B with its FLUX visual models, including widely used open releases.

The attraction is practical. Governments and large enterprises increasingly want models that can run on infrastructure they control, be adapted internally and avoid dependence on one closed API provider.

Open-weight companies still need viable economics because releasing model weights can weaken some forms of lock-in. Current valuations suggest investors increasingly believe deployment, customization, enterprise infrastructure and sovereign-AI demand can make up for that.

Why are robot foundation model companies suddenly worth so much?

Robot foundation models have become one of the most expensive subcategories because investors increasingly believe general AI models can become the software layer controlling many different physical machines.

Skild AI sits around $16B–$18B in our table. Physical Intelligence has financing evidence around $11B. FieldAI is pursuing a raise around $10B after reporting more than $135M in revenue and contracted business across more than 30 customers. Generalist AI sits around $3B, while Genesis AI, Manifold AI, RLWRLD and Flexion Robotics add another layer of younger companies.

These businesses are chasing something much larger than one robot product. The pitch is that a sufficiently general model could learn tasks across different robot bodies, factories, warehouses or mobile machines.

That possibility brings software-like valuation expectations into a market historically dominated by hardware. Investors can imagine one model becoming a reusable intelligence layer deployed across thousands or millions of machines.

Whether the technology really generalizes that far is still one of the biggest unanswered questions in AI. The capital market has already decided the possibility is worth billions.

Company	Current valuation range	Main focus
Skild AI	$16B–$18B	General-purpose robot intelligence
Physical Intelligence	~$11B	General robot foundation models
FieldAI	~$9B–$10B	Robot autonomy models
Generalist AI	~$2.9B–$3.1B	General robotics models
Genesis AI	~$2.5B–$3.0B	General-purpose robot models
Manifold AI	~$1.0B–$1.3B	Embodied world models
Dyna Robotics	~$600M–$750M	Robotic manipulation
Flexion Robotics	~$190M–$350M	Humanoid robot intelligence
RLWRLD	~$120M–$170M	Industrial robot foundation models

Can specialized foundation models become as valuable as general AI labs?

Specialized foundation models can already reach valuations above $20B, but none has yet approached the scale of the largest general frontier labs.

ElevenLabs is the clearest example. Its voice and audio models now support a $22B valuation after a recent $300M employee tender doubled the company’s previous valuation.

Video has created another substantial cluster. Kling AI sits around $17B–$18B in our dataset, while Higgsfield and Runway each exceed $5B. Black Forest Labs has reached $3.25B in image models. Suno is worth $5.4B in music generation.

These companies have an advantage that broad model labs sometimes lack: users immediately understand what they are buying. A company can measure minutes of generated audio, videos created, subscriptions, enterprise seats or API usage.

Specialization also caps part of the upside. ElevenLabs can dominate AI voice without controlling coding, general reasoning, search or enterprise knowledge work. Investors appear willing to award very large valuations to category leaders while reserving the trillion-dollar possibilities for companies that could become general computing platforms.

Is AI video becoming the most valuable specialized foundation-model category?

AI video is currently one of the deepest specialized foundation-model categories, with several independent companies valued in the billions rather than one runaway winner.

Kling AI leads the group around $17B–$18B in our dataset. Higgsfield and Runway sit around $5B each. Luma AI is near $4B, while ShengShu, PixVerse, AiShi Technology and GigaAI add several more multibillion-dollar or near-multibillion-dollar companies.

Higgsfield is especially useful because its valuation increasingly has commercial evidence behind it. The company announced roughly $700M in annualized revenue alongside its $5.4B financing and said it served tens of millions of users.

The category is also unusually fragmented geographically. US and Chinese labs both appear throughout the ranking, and different companies focus on film production, advertising, social content, world simulation or creator tools.

Video therefore looks less like a winner-take-all niche than voice currently does. There may be room for several large companies if different workflows continue to demand different models.

Does raising more money automatically mean a higher foundation model valuation?

Large funding rounds and high valuations are strongly connected in foundation AI because training and serving frontier models consume enormous amounts of capital. But the relationship is far from automatic.

OpenAI has raised extraordinary sums because its compute requirements are extraordinary. Anthropic’s latest $65B financing followed earlier multibillion-dollar rounds and sits alongside huge infrastructure commitments. Safe Superintelligence brought in a reported $5B from Nvidia largely to expand its compute resources.

Yet funding alone does not explain valuation. Reflection AI has raised about $4.6B and reached roughly $25B. Thinking Machines has raised around $2B while investors discuss approximately $40B. ElevenLabs has raised far less than the frontier labs and is worth $22B because its commercial traction has become strong enough to support a large valuation without OpenAI-scale capital needs.

The better way to read funding is as evidence of what kind of company investors believe they are financing. Frontier-model labs require enormous balance sheets. Specialized model companies can sometimes reach similar mid-market valuations with much less capital.

Which foundation model valuations look strongest based on actual transaction evidence?

ElevenLabs, Mistral, Moonshot AI, Suno, Black Forest Labs, Arcee AI and several other companies have especially clean valuation evidence because recent transactions directly established their prices.

ElevenLabs’ $22B figure comes from a completed employee tender. Mistral’s latest funding established a post-money valuation above €21B. Moonshot AI completed a $3.5B financing at approximately $35B. Suno publicly announced its $5.4B valuation. Black Forest Labs announced $3.25B, while Arcee AI announced that its Series B valued the company above $1B.

OpenAI’s $852B and Anthropic’s $965B completed financings are similarly strong anchors, even though our current ranges also incorporate newer fundraising or listing evidence.

At the other end, comparables-based valuations for companies such as Pika, Ideogram, Udio, Haiper and Genmo need much more caution. They may be sensible estimates, but no recent financing has formally printed those exact values.

A current ranking is much more useful when those two kinds of numbers remain visibly separate.

Which foundation model startups could move up the valuation ranking next?

OpenAI, Anthropic, Thinking Machines Lab, Cohere, Moonshot AI and FieldAI currently have some of the clearest paths to near-term repricing.

OpenAI has discussed financing above $1.2T, so a completed round could replace its current $852B transaction anchor with a trillion-dollar one. Anthropic’s eventual IPO could produce an even larger reset if public investors accept a valuation closer to the $1.5T–$2T expectations currently circulating.

Thinking Machines is seeking a dramatic step up toward roughly $40B. Cohere has been in advanced discussions for financing around a $20B valuation. Moonshot reached about $35B in its completed round and was already reported to be exploring additional capital at a higher price. FieldAI has been working toward a transaction around $10B.

One financing can move a company several places up this ranking when investors decide its models, revenue or strategic position have crossed a new threshold.

What does the foundation model startup valuation ranking really tell us?

The foundation model market has become one of the most valuable startup markets in history, but most of that value is concentrated in a tiny number of companies.

Anthropic and OpenAI dominate the financial picture. Using the midpoint methodology applied to our dataset, the pair account for roughly 79% of the aggregate valuation represented by all 98 companies. Add xAI and the concentration becomes even more extreme.

Below those giants, though, the market is becoming broader rather than narrower. Chinese frontier labs, open-weight models, voice AI, video generation, world models, robot intelligence, biological models and music generation have all produced billion-dollar companies.

That breadth is probably the most important development in the ranking. The foundation-model market is no longer synonymous with companies building ChatGPT alternatives. The model layer is spreading into different forms of intelligence, and investors are willing to finance those specialized layers independently.

The valuations still have to be read with very different confidence levels. A $22B tender for ElevenLabs, a $5.4B completed financing for Suno and a $965B completed round for Anthropic are observed market prices. A prospective $2T IPO, a $40B active raise or a comparables-based $300M valuation is a different kind of evidence.

Our conclusion is clear: foundation models have already created an exceptionally large and diverse startup market, but the trillion-dollar end remains concentrated around companies that combine frontier research, massive compute, distribution and commercial scale. Specialized labs can reach tens of billions. Reaching the very top requires becoming something much closer to a global computing platform.