Which AI startups have the most durable revenue?
In our AI infrastructure market deck, you will find everything you need to understand the market
SUMMARY
Databricks has the most durable AI revenue overall today, while Glean and Harvey have the strongest revenue quality among the younger applied-AI startups.
The biggest AI revenue numbers are becoming less informative on their own. Companies can annualize a few months of explosive consumption long before enough renewal cycles have passed to show whether customers will still be spending at the same level two years later.
The strongest revenue increasingly sits with companies that own something customers still need regardless of which model is best: data infrastructure, company context, legal workflows or customer-service operations. Model independence is starting to look like a revenue-quality advantage.
Harvey has the best explicit retention evidence in the group. Its previously disclosed 98% gross revenue retention and 167% net dollar retention suggest that growth is coming from customers staying and expanding rather than from constantly replacing churned revenue.
Glean's advantage looks more structural. More than 85% of customers reportedly use it across at least five departments, meaning the product is spreading through organizations instead of remaining trapped inside an AI innovation team.
Anthropic and OpenAI are much larger, but a model dollar is still easier to contest than a deeply embedded workflow dollar. Enterprises can reroute API or coding workloads faster than they can rebuild years of governed data pipelines or specialized legal processes.
OpenAI has one important protection that most AI startups do not: several different revenue engines. Consumer subscriptions, APIs, enterprise products, coding and advertising make the business less dependent on any single customer type or monetization model.
Sierra may have the most interesting pricing structure. Charging for completed outcomes rather than raw tokens means falling inference costs can improve its economics without automatically forcing the price of the customer outcome down with them.
ElevenLabs becomes more durable as it moves beyond voice generation itself. The closer its agents get to contact-center software, customer records, routing and recurring communication workflows, the less the business depends on simply having the best speech model.
Scale AI provides the clearest warning about customer concentration. The Meta transaction and subsequent customer reactions showed that a large revenue base can still be fragile when strategically important buyers have the ability to move a meaningful share of spending quickly.
Cheaper inference should widen the gap between companies selling access to intelligence and companies owning the workflow around it. The latter can swap models, route requests more efficiently and keep charging for the business process customers actually care about.
This market map, featured in our AI infrastructure market deck, highlights top companies and startups in the AI infrastructure market
Why is everyone suddenly questioning the quality of AI startup revenue?
AI startup revenue is growing so absurdly fast right now that the headline ARR number tells us less than it used to.
Anthropic recently passed a reported $65 billion annualized revenue run rate. OpenAI is above $40 billion. Databricks says it has reached $7 billion. ElevenLabs crossed $500 million ARR after ending 2025 around $350 million. Glean passed $300 million, while Harvey is now around $350 million.
Those numbers would have been almost unimaginable for private software companies a few years ago. The problem is that AI companies can now add billions of dollars of annualized usage before we have seen enough renewals to know whether the revenue will stay.
The distinction shows up especially clearly in consumer AI. RevenueCat examined more than 75,000 subscription-app developers this year and found that AI-powered subscription apps were losing annual subscribers roughly 30% faster at the median than non-AI apps. Huge initial demand clearly does not guarantee durable demand.
Enterprise revenue can hide the same problem in a different way. A company might sign a giant AI commitment, use the product heavily for six months, then shift the workload once another model becomes cheaper or better. Consumption revenue can also surge because customers temporarily run enormous amounts of inference.
So these days we care less about which AI company crossed the biggest ARR milestone first. We want to know what customers would actually have to change inside their businesses before that revenue disappeared.
What does “durable AI revenue” actually mean?
Durable AI revenue is revenue that should still be there after the novelty fades, model prices fall and customers become much pickier about where they spend their AI budgets.
We look for several things at the same time. Customers need to renew. Ideally, existing customers also spend more over time. The product should become tied to data, workflows, permissions or business processes that cannot be moved overnight. The customer base should be broad enough that losing one buyer does not wreck the company. And the economics should still look attractive after paying for the computing power required to deliver the product.
A frontier model can have enormous revenue and still face fairly low switching costs on certain workloads. An enterprise-search company with much less revenue can become harder to remove if thousands of employees rely on it every day and it has already connected to dozens of internal systems. A legal AI company can become unusually sticky once law firms build internal workflows and agents around it.
That is why we rank revenue durability differently from revenue size.
| What we look at | Stronger revenue | Weaker revenue |
|---|---|---|
| Renewals | Customers stay and expand | Constant replacement of churned customers |
| Product use | Daily production workflow | Occasional experimentation |
| Switching effort | Data, integrations and processes must move | Another tool can be adopted quickly |
| Economics | Software-like margins | Heavy compute or service costs |
| Customer mix | Many meaningful customers | Revenue concentrated in a few buyers |
As this chart shows, and as featured in our AI infrastructure market deck, search interest in AI infrastructure has risen sharply
Does Databricks have the most durable AI revenue?
Databricks currently has the strongest overall revenue base among the major private AI companies we reviewed.
The company says annualized revenue has reached about $7 billion and is still growing around 80% year over year. Databricks is also cash-flow positive on an adjusted basis, according to CEO Ali Ghodsi.
More interesting than the total is where the money comes from.
Its core data-warehouse business alone is running at roughly $1.5 billion and is still growing around 100%. Lakebase, the database Databricks launched for AI-agent workloads, has already crossed a $100 million revenue run rate. AI is being layered onto a large enterprise data business rather than carrying the whole company by itself.
That lowers the risk quite a bit.
A company might decide next year that Claude is better than GPT, or that an open model is good enough for a particular task. It still needs somewhere to store, organize, govern and process its proprietary data. Databricks can support those changing model choices instead of betting the company on one model remaining the best.
The switching cost is real too. Large enterprises build pipelines, governance systems, analytics processes and production applications on Databricks. Moving that infrastructure is much more painful than changing the chatbot employees use.
Databricks does have one awkward classification issue: it was founded in 2013, long before the generative-AI boom. If we restrict “AI startup” to younger companies built primarily around the current AI wave, Glean and Harvey become the more interesting answers.
If you want more recent data on this point, please see our latest AI infrastructure market report.
Why does Glean’s $300 million revenue look better than most AI ARR?
Glean has one of the strongest revenue profiles in enterprise AI today because customers are spreading it through their companies instead of keeping it as a small experiment.
Glean passed $300 million in annual recurring revenue after reaching $100 million only 15 months earlier. Its Fortune 500 customer count nearly doubled year over year, and the company says more than 85% of customers now use Glean across at least five departments.
That last figure is especially useful.
A tool being used by one innovation team can disappear during the next software-budget review. A product running across engineering, sales, finance, support and operations has already become a different kind of purchase.
Usage also looks unusually habitual. Glean reported a weekly-active-to-monthly-active ratio around 45%, which it says is more than twice the typical SaaS benchmark. Earlier data showed an average employee performing about five Glean searches per day.
The product architecture helps too. Glean connects company documents, applications, user permissions and internal knowledge, while customers can choose among more than 40 underlying AI models. Its latest product updates push further in that direction with model routing and controls designed to choose cheaper models when a frontier model is unnecessary.
That puts Glean in an attractive position as AI gets cheaper. Better models can improve the product without automatically replacing Glean.
There is one caveat worth keeping. TechCrunch pointed out this year that some of Glean's $300 million figure comes from consumption-based pricing, so the number should not be treated as perfectly equivalent to classic fixed-subscription ARR.
Even after making that adjustment, the pattern looks strong. Glean went from roughly $100 million to $200 million in nine months, then to $300 million about six months later, while Microsoft, Google and other large vendors were moving aggressively into enterprise AI search. Competition intensified and Glean still accelerated.
This chart, included in our AI infrastructure market deck, shows annual VC investment in AI infrastructure startups
Does Harvey have the best revenue quality among young AI startups?
Harvey currently has the cleanest retention evidence we have found among the large generative-AI application startups.
The legal AI company reported 98% gross revenue retention and 167% net dollar retention in its 2025 Q3 update. Those two figures answer different questions. Harvey was keeping almost all of its existing revenue, and the customers who stayed were spending much more than before.
Usage supported the same story. Queries per user rose 60% year over year, seat utilization reached 77%, and Harvey said some firms were seeing adoption above 90%.
The company has continued growing since then. Recent reporting in The Times puts Harvey at about $350 million ARR and more than 200,000 lawyers across over 2,400 organizations. Harvey was around $50 million ARR at the end of 2024 and roughly $195 million at the end of 2025.
That trajectory becomes much more convincing when combined with the retention numbers. Harvey has not had to replace an enormous pile of lost customers every year to produce its growth.
Legal workflows also create natural friction against switching. Firms connect internal information, train lawyers, establish security procedures and increasingly build custom agents around the product. Harvey said customers had already created more than 25,000 custom agents in an earlier update.
The threat is getting more serious, though. The Financial Times recently reported that firms including Kirkland & Ellis, Freshfields and Goodwin Procter are building or customizing their own AI systems. Roughly one in five large law firms is now reportedly doing some form of bespoke AI development. OpenAI is also moving deeper into legal workflows.
Harvey therefore has a very strong business today, but it still has to prove that law firms want Harvey to remain the layer through which those customized systems operate.
If you want more recent data on this point, please see our latest AI infrastructure market report.
Is Anthropic’s $65 billion revenue actually durable?
Anthropic has built the strongest enterprise revenue machine among frontier-model startups, although its revenue is easier to contest than Databricks, Glean or Harvey revenue.
The pace is extraordinary. Anthropic's annualized revenue run rate went from roughly $9 billion at the end of 2025 to $47 billion in May and more than $65 billion by the end of July, according to figures reported by Bloomberg and Axios. Preliminary second-quarter revenue reportedly exceeded $11.5 billion, more than 14 times the same quarter a year earlier.
That growth is increasingly enterprise-driven.
Anthropic said earlier this year that more than 500 customers were spending at least $1 million annualized, compared with only about a dozen two years earlier. The count has since climbed further. Eight of the Fortune 10 were already Claude customers.
Claude Code is probably the clearest example of Anthropic creating something stickier than generic API usage. The product had already passed a $2.5 billion run-rate revenue level earlier this year, and enterprise customers represented more than half of that revenue. Business subscriptions had quadrupled since the beginning of the year.
Still, coding shows why Anthropic cannot take this revenue for granted.
Developers can use Claude Code, OpenAI Codex, Cursor and other tools that increasingly support several underlying models. Companies routinely test models against one another. Switching a coding workload is much easier than moving years of governed enterprise data.
Anthropic also has enormous computing obligations. Recent agreements with Lambda and several other infrastructure providers have pushed its reported future infrastructure commitments well above $100 billion. Massive demand can support those commitments, but the economics of frontier-model revenue depend heavily on inference efficiency.
There is another reason to treat the $65 billion figure carefully. Axios recently reported that Anthropic records some cloud-channel sales gross and then books partner payments as expenses. OpenAI uses a different treatment for comparable business. The two headline revenue figures therefore cannot be compared dollar for dollar.
Anthropic still belongs near the top of this ranking. We simply trust a deeply embedded workflow dollar more than a model dollar that may be rerouted next quarter.
This chart, included in our AI infrastructure market deck, shows why CoreWeave is winning in AI infrastructure
Is OpenAI’s revenue safer because it comes from so many places?
OpenAI has the most diversified revenue engine in private AI right now, which gives it real protection if one part of the business slows.
Bloomberg recently reported that OpenAI had passed a $40 billion annualized revenue run rate, roughly double its pace at the end of 2025. Growth is coming from consumer subscriptions, enterprise products, APIs, coding and a surprisingly fast-growing advertising business.
OpenAI now says its products reach more than two million businesses, twice the number from a year earlier. Enterprise adoption gives the company much more depth than the consumer ChatGPT business alone would suggest.
Advertising adds another layer. ChatGPT Ads recently passed a $1 billion annualized revenue run rate less than 200 days after launch, with tens of thousands of advertisers already using the system. OpenAI can now make money from users who never buy a subscription.
Very few startups have that flexibility.
If consumer subscriptions weaken, enterprise agents can grow. If API prices keep falling, advertising can expand. Coding has also become a meaningful driver of recent growth. More than one billion weekly ChatGPT users give OpenAI a distribution channel that no other startup currently matches.
The weakness remains model competition. Enterprises increasingly expect to use several models, and OpenAI must keep its products good enough that customers continue sending workloads its way.
OpenAI's diversification makes the revenue safer than it looked two years ago. We still rank Databricks, Glean and Harvey slightly higher on pure durability because more of their value survives even when the identity of the best model changes.
Could Sierra’s pay-for-results model create unusually durable revenue?
Sierra has one of the smartest revenue models in AI today because customers usually pay for completed business outcomes rather than raw tokens.
The company entered its third year with more than $150 million ARR. More recent estimates place the business around the $200 million range, although Sierra has not publicly confirmed that newer figure with the same precision.
The pricing model deserves more attention than the ARR milestone.
Sierra can charge customers when its agents successfully resolve a support request, save a cancellation, complete a sale or produce another agreed result. If an interaction fails and needs human escalation, the company may not get paid.
That lines Sierra's incentives up unusually closely with the customer.
It also changes what happens when inference gets cheaper. Sierra's cost of delivering a successful result can fall while the economic value of solving the customer's problem stays similar. A company simply reselling tokens has much less room to capture that improvement.
Large companies are already using Sierra at scale. The company says more than 40% of the Fortune 50 use its platform, and its agents have handled billions of customer interactions.
The short history keeps Sierra below the leaders for now. We have not yet seen several years of renewal cohorts through changing AI budgets and stronger competition. But the basic revenue structure is one of the best we found.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows annual funding in AI infrastructure startups
Is ElevenLabs building durable enterprise revenue or riding a voice-AI boom?
ElevenLabs increasingly looks like a real enterprise communications platform rather than a temporary voice-generation craze.
The company ended 2025 around $350 million ARR and crossed $500 million during the first four months of this year. That means ElevenLabs added roughly $150 million of annualized revenue in about four months after already reaching substantial scale.
The customer mix has also changed.
Deutsche Telekom, Salesforce, Santander, NVIDIA, KPN and other large companies are using ElevenLabs across customer support, sales, advertising and other communication workflows. Several enterprise customers have also invested in the company.
Those deployments are more interesting than consumers generating celebrity-style voices. A voice agent connected to contact-center software, customer records, routing systems, brand rules and multilingual support becomes part of an operating process.
ElevenLabs is deliberately moving in that direction through ElevenAgents. The company wants to own more of the conversational workflow surrounding the voice model rather than depend entirely on having the most realistic speech synthesis.
That move is essential because high-quality voice generation itself is becoming cheaper and more common. OpenAI, Google and other model providers can increasingly generate strong speech directly.
ElevenLabs has the revenue scale and enterprise adoption to make our shortlist. What it lacks is the kind of public retention history Harvey has already disclosed.
Why did Scale AI fall down the durability ranking?
Scale AI dropped in our ranking because the Meta transaction showed how quickly strategically concentrated AI revenue can become unstable.
Scale built a huge data-labeling business supplying frontier labs. Then Meta bought a 49% stake and hired founder Alexandr Wang to lead its superintelligence effort.
Several major customers immediately questioned whether Scale could still be treated as a neutral supplier. OpenAI stopped using Scale. Google also pulled back before later resuming some work.
That episode gave us something more useful than another ARR announcement: a real stress test.
The company survived it. Under CEO Jason Droege, Scale says sales have rebounded and expects revenue to exceed $1 billion this year. Its data business has returned to profitability.
Scale is also trying to reduce its dependence on frontier labs by moving into enterprise and government applications. Forbes reported that this newer application business was already running around $200 million annualized by the end of 2025, with customers including BP, Mayo Clinic and Allianz.
The recovery is impressive, but the episode exposed a weakness that companies such as Glean or Harvey have not faced to the same extent. A handful of strategically important buyers could materially change Scale's revenue trajectory.
Scale may eventually rebuild a much more diversified business. For now, its revenue deserves a larger risk discount.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, compares the main business model options for AI cloud infrastructure providers
Are enterprise AI startups always safer than consumer AI companies?
Enterprise AI revenue tends to be more durable today, although an enterprise contract means very little if employees barely use the product.
We can see the difference when companies publish real usage data.
Harvey reported 77% seat utilization alongside extremely high retention. Glean says more than 85% of customers deploy the platform across five or more departments. Anthropic has rapidly increased the number of customers spending more than $1 million per year. Sierra says its agents are already running inside more than 40% of the Fortune 50.
These are much more useful facts than the number of enterprise logos on a website.
Consumer AI has the opposite advantage: gigantic distribution. ChatGPT shows how powerful that can become when hundreds of millions of people develop a habit around one product. OpenAI can now monetize that audience through subscriptions, commerce, advertising and business products.
But the broader consumer data remains weaker. As mentioned earlier, RevenueCat found materially higher annual-subscription churn among AI apps than among non-AI apps.
So we would generally choose an AI product that has become part of a company's daily operations over one people open occasionally when they feel like experimenting.
What happens to AI startup revenue when inference gets much cheaper?
Cheaper inference should make Glean, Harvey and Sierra stronger while putting more pressure on companies whose advantage mainly comes from selling access to an expensive model.
This is one of the best tests of AI revenue quality.
Glean can send a routine request to a cheaper model while keeping the company's context, permissions and integrations intact. Its newest platform updates are explicitly designed around model choice and automatic routing.
Harvey can also change the underlying model mix while continuing to sell the legal workflow surrounding it. The lawyers are paying to complete legal work faster, not because they care which GPU produced every token.
Sierra has perhaps the cleanest setup. If the company charges for a resolved support case and inference becomes 70% cheaper, the customer's willingness to pay for the resolved case does not automatically fall by 70%.
Frontier labs face a tougher equation. Lower inference prices can create much more usage, but they also push down the revenue collected per unit of intelligence. OpenAI and Anthropic need demand and capability to grow faster than prices fall.
That tension will probably become much more visible over the next few years.
This chart, featured in our AI infrastructure market deck, shows the share of revenue generated by each customer segment in the AI infrastructure market
Which AI revenue numbers should we be most skeptical of?
We should be most skeptical of enormous annualized revenue figures built from a short burst of consumption when the company gives us little information about renewals, margins or customer concentration.
Anthropic provides a good example of why definitions matter. Its reported $65 billion run rate is genuinely enormous, but recent reporting shows that Anthropic and OpenAI account for some cloud revenue differently. Reading $65 billion versus $40 billion as a clean head-to-head comparison would therefore be misleading.
Glean offers another smaller example. The company calls its $300 million figure ARR, yet TechCrunch reported that some customers pay partly through consumption pricing. A portion of that number behaves more like annualized usage than a conventional fixed SaaS subscription.
Customer concentration can be even more important.
CoreWeave, although already public and therefore outside our startup ranking, shows the extreme case. Its regulatory filings said Microsoft represented about 67% of 2025 revenue. Long contracts can make the revenue predictable on paper while leaving the business heavily exposed to one customer's decisions.
Retention data is much harder to hide behind explosive new-logo growth. That is why Harvey's disclosed gross and net retention numbers carry so much weight in our ranking.
| Revenue evidence | How much weight we give it |
|---|---|
| Gross revenue retention | Very high |
| Net revenue retention | Very high |
| Expansion across real workflows | High |
| Multi-year production contracts | High |
| Gross margin | High |
| ARR by itself | Medium |
| Short-term usage run rate | Low |
| Funding or valuation | Almost none |
So which AI startups have the most durable revenue today?
Databricks has the most durable AI revenue overall today, while Glean and Harvey have the strongest revenue quality among the younger applied-AI companies.
Databricks wins because its roughly $7 billion revenue base sits under customer data, governance and production workloads that remain necessary even when the preferred AI model changes. The company is still growing around 80% while producing positive adjusted cash flow.
Glean comes next. Its $300 million revenue base is much smaller, but the underlying behavior looks excellent: more than 85% of customers use the product across five or more departments, Fortune 500 adoption has nearly doubled, daily usage is high and the platform can work across dozens of models.
Harvey probably has the most convincing pure retention data. The company is now around $350 million ARR, and its previously disclosed 98% gross revenue retention and 167% net dollar retention showed that customers were staying and expanding very aggressively.
Anthropic ranks highest among the frontier labs. Its current revenue scale is astonishing, Claude Code has become a large enterprise business of its own, and million-dollar customers have multiplied quickly. The reason Anthropic sits below the first three is straightforward: companies can move model workloads faster than they can move enterprise data systems or rebuild specialized operating workflows.
OpenAI follows closely. Its revenue is now spread across consumer subscriptions, APIs, enterprise products, coding and advertising, giving the company much more resilience than a single-product startup. Its distribution is also unmatched.
Sierra and ElevenLabs deserve to be in the next group. Both are building revenue around real enterprise operations, and Sierra's outcome-based pricing is especially attractive. We simply have less renewal history for these companies.
Scale AI sits lower because the Meta deal exposed how damaging strategic customer concentration can become, even for a company with enormous demand.
The broader pattern is getting clearer. The safest AI revenue is clustering around companies that own something customers still need even when the model underneath changes: their data infrastructure, company context, legal workflows, customer-service operations or other deeply embedded processes. Companies selling intelligence itself can become vastly larger, but they have to keep winning the technology race to protect that revenue.
| Rank | Company | Revenue durability today | Main reason |
|---|---|---|---|
| 1 | Databricks | Very high | Deep data infrastructure and model independence |
| 2 | Glean | Very high | Enterprise context, integrations and broad internal adoption |
| 3 | Harvey | Very high | Exceptional retention and embedded legal workflows |
| 4 | Anthropic | High | Huge enterprise adoption and Claude Code |
| 5 | OpenAI | High | Unmatched diversification and distribution |
| 6 | Sierra | High but less proven | Outcome pricing tied to completed work |
| 7 | ElevenLabs | High but less proven | Voice moving into recurring enterprise workflows |
| 8 | Scale AI | Mixed | Large revenue with a recent concentration stress test |
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows how GPU cloud infrastructure technology has evolved over time
OUR METHODOLOGY
This analysis ranks AI startups by revenue durability rather than revenue size. We broke durability into the factors most likely to determine whether today's revenue survives over time: customer retention, depth of production use, switching friction, customer concentration, underlying economics and resilience as models become cheaper and more interchangeable.
We gave the most weight to evidence showing what existing customers actually do. Disclosed gross revenue retention and net dollar retention carry more weight than headline ARR growth, followed by seat utilization, expansion across departments and workflows, large-customer penetration, integrations, custom processes and other evidence that a product has become part of day-to-day operations.
We did not treat every reported ARR or revenue-run-rate figure as equivalent. Fixed subscriptions, consumption-based pricing and annualized short-term usage can behave very differently, while cloud-channel accounting can also change how two companies present economically similar revenue. Where the revenue definition required more caution, we reflected that in the ranking rather than assuming every headline number was directly comparable.
We also looked at what would happen if the underlying model changed or inference became dramatically cheaper. Companies that own customer data, permissions, workflow context or completed business outcomes can often switch models without losing the customer relationship. Companies whose value depends more directly on access to a particular model face a tougher durability test.
No single metric determined the final order. Harvey's retention figures, Glean's deployment breadth, Databricks' infrastructure position, Anthropic's enterprise adoption, OpenAI's revenue diversification, Sierra's outcome-based pricing and Scale AI's customer-concentration stress test all measure different parts of the same question. We combined those signals and reserved the strongest judgments for companies where several recent indicators pointed in the same direction.
Key sources included RevenueCat's State of Subscription Apps 2026 for AI-app retention, Databricks' latest operating figures, Glean's $300 million ARR update, TechCrunch on Glean's pricing mix, Harvey's retention and usage update, Harvey's custom-workflow data, The Times on Harvey's more recent revenue and adoption, and the Financial Times on bespoke legal-AI development.
For frontier models and the next group of application companies, key sources included Anthropic's enterprise and Claude Code disclosures, Bloomberg on Anthropic's revenue run rate, Axios on Anthropic and OpenAI's cloud-revenue accounting, OpenAI's business-adoption update, OpenAI's advertising update, Sierra's operating update, Sierra's outcome-pricing explanation, ElevenLabs' $500 million ARR announcement, Scale AI's business update, Bloomberg on OpenAI's response to the Meta transaction, Forbes on Scale AI's recovery and business mix, and CoreWeave's 2025 Form 10-K for the customer-concentration comparison.
In our AI infrastructure market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Which AI startups are actually profitable?
- Which AI infrastructure startups generate the most revenue today?
- Which AI startups are growing without burning billions?
