Which AI startups are growing without burning billions?

Are AI startups really starting to grow without burning billions?

Yes. A small but increasingly convincing group of AI startups is now reaching hundreds of millions of dollars in revenue without anything close to the multibillion-dollar cash burn of the biggest AI labs.

That distinction has become much clearer lately because the numbers at the frontier have become so extreme. Financial documents reported by The Information showed OpenAI generating $5.7 billion of revenue in the first quarter of 2026 while burning $3.7 billion of cash during the same three months. Research and development alone cost $8.6 billion. xAI had already recorded a $6.4 billion operating loss on $3.2 billion of revenue in 2025.

At the other end of the market, Midjourney generated $300 million of revenue in 2024 while profitable and without outside funding, according to PitchBook figures reported by Forbes. ElevenLabs made an estimated $116 million of profit in 2025 and has since passed $500 million ARR. HeyGen says it has reached $200 million ARR after burning only about $25 million since launch. Gamma crossed $100 million ARR profitably with roughly 50 employees.

These companies are doing more than spending less because they are smaller. Some are already producing revenue at a scale that would qualify as substantial even for established software companies.

A real split has emerged inside AI. Frontier laboratories are financing giant research programs and compute infrastructure. A growing set of AI product companies is capturing the commercial value of those models without carrying the same cost structure.

What does “burning billions” actually mean for an AI startup?

For an AI startup, “burning billions” should mean consuming billions of dollars of cash, rather than simply raising billions from investors.

That distinction sounds basic, but it changes which companies look efficient. ElevenLabs has now raised around $800 million. Lovable has raised roughly $950 million. Neither figure tells us how much of that money has actually been spent.

A company could raise $1 billion, spend $100 million and leave the rest in the bank. Another could raise $500 million and consume nearly all of it. Funding totals alone would make the first company look less efficient even though the opposite is true.

So we put much more weight on reported profitability, free cash flow, cumulative cash burn and gross margins. Revenue relative to funding and headcount helps when those figures are unavailable, but it remains indirect evidence.

ARR also deserves some skepticism in this market. Fast-growing AI startups often annualize their latest month or quarter, meaning a “$500 million run rate” does not necessarily mean the company collected $500 million during the previous 12 months. TechCrunch raised this point recently with Glean, where part of the reported $300 million top line comes from consumption pricing and is better thought of as annualized revenue than traditional recurring subscription revenue.

Two companies with identical headline ARR can therefore have completely different economics.

Which AI startups look the most capital-efficient today?

Midjourney, ElevenLabs, HeyGen and Gamma currently have the strongest combination of serious revenue and hard evidence that the underlying business does not require enormous cash burn.

Midjourney remains the cleanest case because it has never raised outside capital. ElevenLabs stands out because it develops important proprietary models while still reporting substantial profits. HeyGen has disclosed remarkably low cumulative burn relative to revenue. Gamma has produced perhaps the neatest application-software example: $100 million ARR, profitability and a team of about 50.

Lovable is close behind. Its revenue per employee is extraordinary, and annualized revenue has now passed $500 million, but we still lack good public figures for cumulative burn or profitability.

Glean has also become harder to ignore after reaching $300 million in reported ARR or annualized top line, triple its level 15 months earlier. The company has raised considerably more money than Gamma or HeyGen, however, and we do not have comparable profit figures.

Cursor belongs further down this particular ranking despite having much more revenue. Its annualized revenue passed $2 billion earlier this year, but TechCrunch reported that the company had been running at negative gross margins until recently. Scale alone cannot compensate for that.

Company	Latest useful revenue signal	Efficiency evidence	Our view today
Midjourney	$300M 2024 revenue	Profitable, zero outside funding	Exceptional
ElevenLabs	$500M+ ARR	Profitable; estimated $116M 2025 profit	Exceptional
HeyGen	$200M+ ARR	Around break-even; ~$25M lifetime burn	Exceptional
Gamma	$100M+ ARR	Profitable since 2023; ~50 employees	Exceptional
Lovable	$500M+ annualized revenue	Extremely high revenue per employee	Very strong, burn unclear
Glean	$300M reported ARR/top line	Tripled in 15 months	Strong growth, profit unclear
Cursor	$2B+ annualized revenue	Gross margin only recently turned slightly positive	Huge growth, weaker economics

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

Is Midjourney still the clearest low-burn AI success story?

Yes. Midjourney is still probably the hardest company to dismiss when someone argues that major AI startups need gigantic funding rounds to succeed.

According to PitchBook figures reported by Forbes, Midjourney generated $300 million of revenue in 2024 and was profitable. Forbes still listed the company as profitable in its 2026 profile. More unusually, Midjourney has done this without taking outside venture funding.

Its team remains tiny for a company of this scale. Forbes currently lists roughly 60 employees. Even if we use the older $300 million revenue figure rather than less certain estimates of more recent revenue, that works out to about $5 million of revenue per employee.

The company has also started using its financial strength more aggressively. Midjourney expanded from images into video in 2025, licensed technology to Meta and has recently started making acquisitions. Those moves show a self-funded company using internally generated cash to broaden its position rather than depending on continuous fundraising.

Midjourney does carry obvious concentration risk. Image and video generation are becoming crowded, with Google, OpenAI, Adobe and open-source models improving quickly. That competitive threat does not change the financial record so far: Midjourney built one of the largest generative-AI consumer businesses with its own cash.

How is ElevenLabs growing so fast and still making money?

ElevenLabs is currently the strongest evidence that an AI startup can develop its own important models, grow past $500 million ARR and still avoid frontier-lab economics.

Forbes estimates that ElevenLabs made $116 million of profit in 2025. That is unusually strong for any startup growing at this speed, and especially unusual for one doing meaningful model research rather than simply calling external APIs.

Its revenue curve has kept accelerating. ElevenLabs needed roughly 20 months to reach $100 million ARR, another ten months to reach $200 million and around five months to move beyond $300 million. The company then said it passed $500 million ARR during the first part of 2026.

The business underneath that revenue has also changed. ElevenLabs started with voice cloning, dubbing and speech generation for creators. AI voice agents now account for more than half of the company, according to Forbes. That means more revenue is coming from businesses using voice AI for customer service and other recurring workflows rather than from one-off creative experimentation.

ElevenLabs now employs around 450 people and has raised roughly $800 million, so it is no longer a tiny operation. But the important number here is the profit. A company can raise a huge war chest while remaining economically sound.

Among the startups we reviewed, ElevenLabs has probably done the best job of combining proprietary AI research with actual software-like economics.

Is HeyGen the most capital-efficient venture-backed AI startup?

HeyGen has one of the strongest cases today: more than $200 million ARR, around break-even cash flow and only about $25 million of reported cumulative burn.

HeyGen said in June 2026 that ARR had doubled to more than $200 million in eight months. The company also reported more than 30 million users across 196 countries and usage within 85% of the Fortune 100.

Its historical growth makes the burn figure particularly striking. HeyGen finished 2024 around $57 million ARR, moved through roughly $100 million the following year and has now doubled again. That growth was achieved after raising only around $74 million of equity capital.

In rough terms, current ARR is therefore about 2.7 times all the equity HeyGen has ever raised. More importantly, management says only around $25 million has actually been burned.

HeyGen has also resisted the temptation to compete across every possible video use case. The company focuses heavily on avatars, localization, marketing, training and business communication. Those jobs already cost companies money when performed with presenters, studios, editors or localization teams, which gives HeyGen a straightforward way to demonstrate value.

That narrower approach may eventually face pressure as general-purpose video models improve. For now, though, HeyGen has turned AI video into one of the cleanest high-growth businesses in the sector.

HeyGen metric	Approximate figure
Current ARR	$200M+
Equity raised	~$74M
ARR / equity raised	~2.7x
Reported lifetime cash burn	~$25M
Cash-flow position	Around break-even

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

How did Gamma reach $100 million ARR with only about 50 employees?

Gamma reached $100 million ARR profitably by keeping the company unusually small and letting the product do most of the selling.

Gamma founder Grant Lee said the company passed $100 million ARR with about 50 employees after initially raising only $23 million. Gamma says it has been profitable since 2023. Its later $68 million Series B took total funding to roughly $90 million, but part of that round was a secondary transaction giving existing employees liquidity rather than financing operating losses.

The ratio is remarkable: roughly $2 million ARR per employee when Gamma crossed the milestone.

The company also reached that scale without building the kind of large sales organization common at enterprise SaaS businesses. Gamma spread primarily through people using the product to make presentations, websites and documents, then sharing the output with others.

That product-led loop helped Gamma reach 70 million users by the time it announced $100 million ARR. By early 2026, it was approaching 100 million users.

Gamma is now expanding into image generation and competing more directly with Canva and Adobe. That will probably push spending higher. Still, the first $100 million of ARR was built with a cost structure that would have looked almost impossible for a traditional software startup a decade ago.

Is Lovable actually efficient, or are we being fooled by its huge revenue number?

Lovable looks extremely efficient today, but we still do not have enough public cash-flow data to call it another Gamma or HeyGen.

Lovable passed $400 million in annualized revenue in February with only 146 employees. By June, the company said the figure had moved above $500 million. That puts Lovable well above $3 million of annualized revenue per employee if headcount has remained anywhere close to its earlier level.

Usage is equally large. Lovable recently said it hosts around 60 million projects attracting roughly 900 million monthly visitors.

But the company's spending needs are moving up quickly as well. Lovable has now raised roughly $950 million, including a $400 million Series C at a $13.3 billion valuation. It signed a multiyear Google Cloud agreement that is expected to increase usage fivefold, and it has started training its own internal AI models alongside using third-party frontier models.

None of those decisions proves Lovable is burning heavily. They simply make the economics harder to judge from the outside.

The sharp conclusion we can defend is that Lovable has extraordinary organizational efficiency. Profitability remains unproven publicly.

Why doesn't Cursor count as a low-burn AI winner yet?

Cursor does not qualify as one of the cleanest low-burn AI startups today because its $2 billion-plus annualized revenue has come with much weaker gross margins than the headline growth suggests.

Cursor's expansion has been almost absurdly fast. It crossed roughly $100 million ARR in early 2025, reached $1 billion later that year and doubled its annualized revenue again to more than $2 billion by February 2026. Enterprise customers now produce around 60% of revenue.

But TechCrunch reported in April that Cursor had operated with negative gross margins until recently. In other words, serving customers could cost more in model usage than Cursor collected from them.

That problem was especially severe among individual developers. Cursor improved the economics by introducing its own Composer model and routing more workloads toward cheaper external models. Enterprise sales have since become gross-margin positive, while some individual accounts reportedly continue losing money.

The concern grew further when TechCrunch later reported that even the more than $2 billion financing Cursor had been preparing might not have been enough to carry the company to break-even.

Cursor absolutely belongs among the fastest-growing AI startups. Putting it beside HeyGen or Gamma on capital efficiency would hide the most important difference between them.

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

Is Glean becoming another efficient AI giant?

Glean is now big enough to deserve serious attention, but we cannot yet prove that its $300 million revenue scale comes with the same efficiency as Gamma, HeyGen or ElevenLabs.

Glean said in May 2026 that it had passed $300 million ARR, up from $100 million only 15 months earlier. It needed nine months to climb from $100 million to $200 million, then only six months to add the next $100 million.

The customer base is also getting stronger. Glean says its Fortune 500 customer count almost doubled year over year, while more than 85% of customers use the product across at least five departments.

There is an interesting economic angle here too. Glean is increasingly selling itself as a way for enterprises to lower their AI bills. CEO Arvind Jain told TechCrunch that reducing unnecessary token consumption has become a major selling point as companies discover how expensive widespread AI deployment can become.

We should still be careful with the “$300 million ARR” label. TechCrunch noted that Glean offers consumption-based pricing alongside subscriptions, so part of the figure behaves more like an annualized revenue run rate than fixed recurring revenue.

Glean has also raised far more capital than Gamma or HeyGen and has not published comparable profitability figures. The growth is clearly real. The low-burn claim remains less certain.

Why are these AI startups so much cheaper to run than OpenAI or xAI?

The most efficient AI startups usually solve one expensive customer problem instead of trying to build the world's best general-purpose intelligence.

The difference in spending is enormous. OpenAI's shareholder financials reported by The Information showed $8.6 billion of research and development expense in a single quarter of 2026. xAI spent billions building compute capacity and training frontier models during 2025 and 2026.

Gamma can buy much of the intelligence it needs from model providers. HeyGen concentrates its research on AI video and avatars. Midjourney focuses on visual generation. ElevenLabs develops its own models but stays heavily concentrated on voice, audio and conversational agents.

These companies also sell outcomes people already understand. A customer pays ElevenLabs to handle calls, HeyGen to produce videos and Gamma to create presentations. The purchase can often replace an existing workflow with visible labor or production costs.

Frontier research has no equally tidy boundary. Making the next model better can require more GPUs, more researchers, more data and another enormous training run. Once a company decides it must remain near the absolute frontier across reasoning, coding, multimodality, agents and consumer products, spending can rise almost without limit.

The companies in this article have largely avoided that race.

Should AI startups build their own models or keep paying OpenAI and Anthropic?

For a fast-growing AI startup, the most economical answer increasingly seems to be a mix: rent frontier intelligence where it helps, then build specialized models when third-party API bills become too painful.

Using OpenAI, Anthropic or Google is incredibly efficient at the beginning. A startup can launch a sophisticated AI product without spending hundreds of millions training a model.

Scale changes the calculation. Cursor showed what can happen when customers use large amounts of expensive third-party inference under fixed-price subscriptions. Model bills pushed the company into negative gross margins.

Cursor's own Composer model helped improve those margins. Lovable has now started training internal models too. ElevenLabs followed this route much earlier, building specialized voice technology while avoiding the cost of creating a general-purpose language model.

This gives us a useful dividing line. Specialized model ownership can make a large AI product cheaper to run. Trying to compete simultaneously with OpenAI, Anthropic and Google across general intelligence puts a company back into the capital-intensive race.

The cheapest architecture may therefore change as an AI startup grows. Renting models is hard to beat when usage is small. At hundreds of millions of dollars in revenue, selective vertical integration starts making much more sense.

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

Are tiny AI teams really producing millions of dollars of revenue per employee?

Yes. Several of today's fastest-growing AI startups are already producing $1 million to $3 million or more in annualized revenue per employee, which would have looked extraordinary in traditional private software.

Gamma crossed $100 million ARR with approximately 50 employees, or roughly $2 million per employee. Lovable reached $400 million with 146 employees, around $2.7 million. HeyGen had roughly 136 people around its $200 million milestone, which comes to about $1.5 million.

Midjourney looks even more extreme. Forbes currently lists about 60 employees and $300 million of 2024 revenue, which gives us approximately $5 million per employee using the older revenue number.

For comparison, SaaS Capital's 2026 private SaaS benchmark puts median revenue per employee around $141,000. These AI companies can therefore produce ten, fifteen or even twenty times as much revenue per worker as a normal private software business.

We should not turn that ratio into a universal ranking. Cursor also produces enormous revenue per employee while historically struggling with gross margins. A dollar of revenue that costs $1.20 to deliver is less attractive than a dollar that costs $0.20.

Still, the staffing pattern is too consistent to ignore. AI is allowing some software businesses to become very large before they need the workforce that previous generations of SaaS companies would have built.

Company	Revenue or ARR used	Approx. employees around milestone	Revenue per employee
Midjourney	$300M revenue	~60	~$5.0M
Lovable	$400M ARR	146	~$2.7M
Gamma	$100M ARR	~50	~$2.0M
HeyGen	$200M ARR	~136	~$1.5M
Private SaaS median	—	—	~$141K

Will cheaper AI models make these startups even more profitable?

Cheaper inference should help application-layer AI startups significantly, although competition will eventually force them to share some of the savings with customers.

Model prices have fallen rapidly while capabilities have improved. An application company can benefit from those reductions without financing the research that created them. It can also move workloads between OpenAI, Anthropic, Google, open-source models and internal systems as the relative economics change.

Cursor is already doing this. Switching more workloads toward Composer and cheaper models improved gross margins. Glean now actively routes AI requests across models partly to control enterprise token costs.

The effect becomes powerful at scale. If an AI company collects $200 million from customers and spends $80 million on model inference, cutting that inference bill by 40% creates $32 million of additional gross profit before changing anything else.

Customers will eventually demand lower prices as AI becomes cheaper, so those savings will not all stay with vendors. But application companies still have much more freedom to shop for cheaper intelligence than frontier labs have to avoid the cost of inventing the next model.

That is one reason we expect the economics of strong AI applications to improve faster than those of frontier research labs.

What could ruin the economics of Midjourney, ElevenLabs, HeyGen and Gamma?

Commoditization is the biggest threat because today's profitable AI startups are often charging for capabilities that are becoming cheaper and easier to reproduce.

ElevenLabs already appears aware of this. Simple synthetic speech is rapidly becoming commonplace, so the company has moved toward voice agents, enterprise workflows, music and a broader audio platform.

Midjourney faces relentless image and video competition from Google, OpenAI, Adobe and open-source systems. HeyGen's avatar advantage could narrow as general video generation improves. Gamma has Microsoft, Google, Canva and Adobe all pushing AI deeper into products their customers already use.

The danger is particularly obvious for applications whose only advantage is a nicer interface around somebody else's model. Falling model prices help their gross margins, but they also make it easier for another startup to build something similar.

The stronger companies are already moving deeper into workflows. ElevenLabs wants to operate customer conversations rather than merely generate voices. HeyGen is embedded in corporate video production and localization. Gamma is expanding from presentation generation toward a broader visual communication product.

Low burn gives these companies room to adapt, which is one of its biggest strategic advantages. A profitable startup can survive several wrong product bets. A company consuming hundreds of millions every quarter has much less time.

So which AI startups are growing without burning billions?

Midjourney, ElevenLabs, HeyGen and Gamma are the clearest answers today, with Lovable close behind but still lacking the financial disclosure needed to put it in the same group.

Midjourney gets our strongest overall capital-efficiency judgment. It built a profitable business producing hundreds of millions of dollars of revenue without outside funding.

ElevenLabs is the most impressive example among startups doing serious proprietary-model research. Passing $500 million ARR would already be remarkable; combining that growth with an estimated $116 million of profit in 2025 makes the company stand apart.

HeyGen has perhaps the cleanest venture-backed record. More than $200 million ARR against roughly $74 million raised and only around $25 million of reported lifetime burn is an exceptional ratio.

Gamma gives us the purest application-layer example. It crossed $100 million ARR with around 50 employees, says it has been profitable since 2023 and needed only $23 million of initial funding to get there.

Lovable may eventually join that top group. More than $500 million in annualized revenue before its third birthday is extraordinary, and its revenue per employee is among the highest we found. For now, the missing cash-flow numbers keep us from making the stronger claim.

Glean is another company worth watching closely. Its top line has tripled to $300 million in 15 months, although its mixed consumption model and lack of public profit data make the economics harder to compare directly.

Cursor gives us the useful counterexample. Its $2 billion-plus annualized revenue is larger than everyone else on this list, yet negative gross margins persisted until recently. Revenue growth and financial efficiency can diverge dramatically in AI.

The broader answer is pretty clear. Building a major AI company does not inherently require burning billions of dollars. The companies that spend at that level are generally financing frontier models, huge compute infrastructure or both.

The more capital-efficient winners are following a narrower path: solve one valuable problem, keep the team unusually small, avoid unnecessary frontier-model research and control inference costs as usage scales.

That model is already producing some of the fastest-growing and most profitable businesses in AI. The surprising part is no longer that profitable AI startups exist. It is how large several of them have become without ever developing the appetite for cash that defines the frontier labs.

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

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

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