AI Startups Flashing Record Revenue Numbers Are Not What They Seem

Seed-Stage AI Startups Are Flashing Record Revenue Numbers And Most Of Them Are Not What They Seem

Josipa Majic Predin Apr 08, 2026 at 09:22am EDT Forbes

At Y Combinator’s Winter 2026 Demo Day last month, investors noticed that companies that were only eight weeks old were asking for $5 million at $40 million post-money valuations. Several had already landed six- and seven-figure customer contracts. The default valuation for the batch had, by one tracker’s count, doubled in three years; from $20 million to $40 million post-money. One startup allegedly raised at a $200 million valuation, the highest in six years of Demo Day tracking.

It happened against a backdrop of founders and companies tweeting spectacular revenue milestones almost weekly. Zero to $100 million in annual recurring revenue in twelve months. Zero to $50 million in six. The numbers were so large, and arrived so fast, that they started to feel like the baseline rather than the exception.

Into that environment, Andreessen Horowitz general partner Jennifer Li, who oversees some of a16z’s fastest-growing AI portfolio companies including Cursor, ElevenLabs, and Fal.ai, stated a quiet corrective on TechCrunch's Equity podcast in February. "Not all ARR is created equal," she said, “and not all growth is equal either.”

The distinction matters because seed valuations set the floor for every subsequent round. A company that raises at $40 million post-money at seed needs to justify something close to $150 to $200 million at Series A. If the revenue that anchored the seed price was fragile; pilots that expire, one-time enterprise experiments, or consumer signups that churn in month two, the math collapses over the following eighteen months.

The Mechanics of the Confusion

Annual recurring revenue is, in its original accounting sense, the annualized value of contracted, recurring subscription revenue. It measures money a company is contractually owed on a recurring basis: the kind of revenue you can underwrite because customers are locked into agreements. A company with 500 enterprise customers each paying $10,000 per year has $5 million in ARR in this strict sense, and that number is meaningful because the customers have committed.

What many AI founders are actually reporting is something different: revenue run rate. You take whatever revenue came in during your best recent month, multiply by twelve, and call it ARR. ChartMogul, one of the most widely used SaaS analytics platforms, notes that this conflation has become standard practice, calling the traditional meaning of "annual recurring revenue" almost meaningless for companies built on month-to-month subscriptions. The term has, in effect, been repurposed.

“True” contracted ARR implies a level of revenue certainty. Run-rate ARR implies only that the company had a good month. A founder who closed three large enterprise pilots in January, took that month's revenue, multiplied by twelve, and tweeted the result has told investors something true but incomplete. The pilots may have three-month opt-out clauses. The enterprises may be running experiments with AI budgets that are not yet embedded in core operating spend. The customers may churn when the pilot ends.

The Prompt: Get the week’s biggest AI news on the buzziest companies and boldest breakthroughs, in your inbox.

Li described the missing information on the podcast: "There's a lot of missing nuances of the business quality, retention, and durability that's missing in that conversation." A founder may have had a strong sales month that will not repeat. Or the startup may be full of short-term customers on pilot programs, which means the revenue is not guaranteed to stick around.

The problem is compounded by the specific economics of AI products. Unlike traditional SaaS, which is typically sold on annual contracts with meaningful switching costs, many AI tools are sold month-to-month at low price points. They attract what a16z’s own research calls “AI tourists"; users who pay to experiment with a new product, generate excitement on social media about it, and then cancel when the novelty fades or a cheaper alternative appears. A ChartMogul retention study of 3,500 software companies found that AI-native companies had a gross revenue retention rate of roughly 40 percent, far below the 88 percent typical of B2B SaaS. A company can be growing rapidly at the top line while its cohorts are leaking badly at the bottom.

Who Has Power In This Market

The seed-stage AI market has, over the past two years, bifurcated into two ecosystems that happen to share the same vocabulary.

The first is a small number of genuinely exceptional companies. Cursor hit $100 million in revenue in twelve months. ElevenLabs crossed $330 million ARR in 2025 and is backed by a16z. Lovable, Bolt, and OpenEvidence each posted rapid early traction. These companies have real recurring revenue, enterprise customers on multi-year contracts, and the kind of net revenue retention that suggests customers are expanding their use rather than experimenting and leaving. They are the benchmark everyone else is being compared to.

The second is a much larger group of companies that are benefiting from the halo effect of the first group. Carta data shows the median post-money seed valuation hit a record $24 million in Q4 2025, up from $18 million a year earlier. AI startups specifically carry a valuation premium of roughly 42 percent over non-AI companies at seed stage. AI companies now absorb 42 percent of all seed capital, up from 23 percent before ChatGPT launched in late 2022.

Driving those numbers are several distinct types of investors. Large venture firms - with substantial funds and pressure to deploy capital, are moving earlier in the stack, crowding out smaller funds and driving up prices. Ashley Smith, a general partner at early-stage fund Vermilion, told TechCrunch she can easily find herself priced out of a round when a larger firm moves in. This compression, fewer deals at higher prices, with capital concentrating into fewer companies, is showing up in Carta data as a widening gap between the median valuation and the top decile. The 95th percentile seed valuation in 2025 hit $80.5 million, nearly triple the $28.5 million it was in 2019.

Founders themselves are significant actors in the dynamic. Shanea Leven, founder of enterprise AI platform Empromptu, told TechCrunch that investor expectations have escalated sharply. "The pressure is at an all-time high, not to be a billion-dollar company, but a $50 billion." A friend raising for a non-AI startup took two years to raise half what Leven raised in three weeks.

Then there is Y Combinator itself. The accelerator’s batch composition has shifted dramatically: 88 percent of its Winter 2026 cohort is AI-first, and 56 of 198 companies are building fully autonomous agents. The batch had three times more companies reaching $1 million in annualized revenue than the Winter 2025 cohort. The “YC tax”, the premium investors pay simply because a company went through the accelerator, has not gone away, but it now compounds on top of an AI premium that already inflates starting valuations.

The outliers anchor the entire conversation. Former OpenAI executive Mira Murati raised $2 billion for Thinking Machines Lab at a $12 billion valuation in 2025. That is a seed round. It does not represent the market, but it makes a $40 million post-money valuation for an eight-week-old company feel modest by comparison.

Where Serious People Disagree

Most participants agree that AI is a genuine platform shift and that some of today's early companies will become very large. The disagreement is about what that potential justifies at the seed stage, and whether current metrics are adequate for measuring it.

The bull case, articulated by investors like Marlon Nichols, managing general partner at MaC Ventures, is that current valuations reflect genuinely different company fundamentals. "The best seed-stage companies do not look like traditional seed-stage companies anymore," he told TechCrunch. AI tools have compressed the time from idea to working product from months to weeks. Founders are entering accelerator programs with enterprise customers already paying. The old benchmarks for "too expensive" were calibrated to a world where getting to market took longer and cost more.

Nichols’ last two seed investments were generating more than $2 million in revenue before he wrote checks of $3 to $4 million, valuing them at $25 to $30 million post-money. He points to paid pilots from large enterprises and "a clear line of sight to full commercial agreements" as justification. The background of the founders: relevant experience, track records, reduced early-stage risk enough to justify the price.

The bear case, implicit in Li’s warning and made explicit by retention data, is that the fastest-growing AI companies are outliers being used to price a market where most companies will not replicate their trajectory. ChartMogul’s analysisfound that low-priced AI products (under $50 per month) retain just 23 percent of gross revenue annually, meaning roughly three-quarters of the revenue base turns over each year. Even mid-priced AI products ($50-$249 per month) retain only 45 percent of gross revenue, compared to 88 percent for B2B SaaS.

Cassie Young from Primary Venture Partners has called what is coming a potential "gross retention apocalypse": enterprise buyers who ran AI experiments in 2024 and 2025 will scrutinize whether those tools delivered measurable value. AI tourists will move on to the next product. Model providers will absorb thin-layer AI wrappers as base model capabilities improve. A company with a run-rate ARR of $10 million in January that loses 70 percent of its cohorts by December has not built a $10 million ARR business. What they really have is a conveyor belt that requires constant new customer acquisition just to hold revenue flat with AI tourists.

The question is: is an AI company that is growing fast with high early churn more like a hyper-growth SaaS company in year one, or more like a consumer app that has hit a viral moment and will mean-revert? The answer determines whether high valuations are early-entry prices on durable businesses or speculation on momentum.

Li explains her own portfolio companies hit extraordinary growth rates. She is saying that 5x to 10x year-over-year growth, coupled with strong retention, is both achievable and what serious investors are actually looking for. Growing from $1 million to $5-10 million in year one and then to $25-50 million in year two, with customers who expand their spending over time, will attract investor interest. What will not hold up is growth built on pilots that expire, novelty spend that churns, or run-rate math that assumes the best month will never end.

The Evidence

The retention data is the most important empirical input into this debate, and it is not encouraging for the bull case at scale. The ChartMogul study covering 3,500 software companies found that as of 2025, AI-native companies had a median gross revenue retention (GRR) rate of 40 percent and a net revenue retention (NRR) rate of 48 percent. For context: the median B2B SaaS company retains 82 percent of net revenue annually. Consumer SaaS retains 49 percent. AI-native companies are performing at roughly the same level as consumer apps, which are not typically valued on ARR multiples.

There is a price-point divide within that number that matters. AI-native products selling for more than $250 per month see 70 percent GRR and 85 percent NRR - essentially equivalent to B2B SaaS. Products selling for $50-$249 per month see 45 percent GRR. Products under $50 per month see 23 percent GRR. The retention problem is concentrated in cheap, self-serve AI products. Enterprise-priced, workflow-embedded AI products retain customers at rates that justify SaaS-style valuation multiples. The problem is that most of the companies raising at high seed valuations have not yet proven which category they fall into.

Cursor hit $100 million in revenue in twelve months, a growth rate that has few historical precedents in software. But Li noted in her podcast appearance that even Cursor faced real operational problems: a poorly rolled-out pricing change angered its customer base in late 2024, a warning about what happens when growth outruns infrastructure and judgment. ElevenLabs crossed $330 million ARR in 2025 but started, as its recent a16z documentary showed, as a weekend project. The speed of its growth created hiring and compliance challenges that required significant operational repair.

Hex Security, one of the most sought-after companies at YC's Winter 2026 Demo Day, had reportedly crossed $1 million in run-rate revenue within eight weeks of founding, prompting investors to compete aggressively for allocation. That is genuine traction. But $1 million in run-rate revenue over eight weeks, multiplied out to $6 million in ARR, built on a handful of early enterprise security contracts, is not the same asset as $6 million in contracted annual revenue from customers who have renewed at least once.

A broader structural signal: Carta's 2025 data shows that the bottom 50 percent of U.S. startups combined to raise just 14 percent of all venture capital in 2025. The top 10 percent raised about half. Seed deal count is falling even as valuations rise. The market is concentrating capital into fewer companies at higher prices, which means the companies that do not make it into the top tier face both higher entry valuations from competitors and smaller funding pools if they fail to differentiate.

The crypto boom of 2021 offers interesting potential analogy; projects raised eight-figure seed rounds on whitepapers and founding team pedigree. Many struggled to raise Series A when the market corrected because they had already consumed too much of their potential valuation upside. AI is a more fundamental technology shift than crypto, but the capital allocation pattern has some similarities: large amounts of money chasing a small number of genuinely valuable companies, with a long tail of companies priced as if they are the top tier.

What Could Go Wrong

The most specific failure mode is the pilot cliff. Enterprise AI adoption in 2024 and 2025 was in many cases funded by experimental or innovation budgets that sit outside core IT spending. Companies bought AI tools to learn about AI, not because they had determined they needed a particular tool indefinitely. When those pilots come up for renewal, buyers will apply cost-benefit scrutiny they did not apply at purchase. A startup that reported $5 million in run-rate ARR based on enterprise pilots in its first year may find that fewer than half of those customers renew at full price, if they renew at all.

Crunchbase’s analysis describes this as “experimental recurring revenue”, revenue from pilots, proofs of concept, or innovation budgets that can vanish when corporate priorities shift. Investor Jamin Ball has used the same framing. The companies that labeled this revenue ARR in their fundraising materials will find it very hard to explain to their Series A investors why a metric that looked stable is suddenly shrinking.

The second potential trap involves valuation ladders. A company that raises $5 million at a $40 million post-money seed valuation needs to clear something in the range of $150 to $200 million at Series A to deliver a reasonable return multiple to its seed investors. That requires genuine, durable business metrics, not run-rate math built on a good quarter. If the company's retention data looks like the AI-native median (40-48 percent NRR), investors will discount the ARR figure significantly when pricing the Series A. The likely outcome is a flat round, a down round, or an extended period of stagnation while the company tries to rebuild its metrics.

A third risk is operational. Li flagged this in her podcast appearance: lightning-fast growth creates problems of its own. Hiring the right people for a company growing this fast is genuinely difficult. Cursor's pricing change debacle, ElevenLabs' compliance issues, and the general pattern of early AI companies launching before they have legal, security, and support infrastructure in place are symptoms of the same underlying problem. A company that grows from zero to $20 million in revenue in twelve months may be managing ten times the complexity of what it managed three months earlier. Most early teams are not built for that.

Finally, there is the platform risk. Many seed-stage AI companies are thin layers on top of foundation models; they call OpenAI or Anthropic APIs and add UI, workflow logic, or domain-specific prompting. As base models improve, the value that those thin layers provide gets absorbed by the underlying platform. A company valued on its ability to do X with GPT-4 faces a structural problem when GPT-5 does X natively. Cursor exists because coding assistance required significant additional engineering in 2023. Whether it maintains its advantage as coding assistance becomes a native feature of every development environment is an open question.

Where This Is Heading

The companies that raised on genuine traction, enterprise customers on multi-year contracts, high NRR, workflow integration that creates switching costs, will validate their valuations at Series A. Investors who backed those companies will declare the AI thesis confirmed. The companies that raised on run-rate math and pilot revenue will face harder conversations. Some will have used the time to convert pilots to long-term contracts and build durable businesses. Others will find that their metrics have not improved since the seed round, and that the Series A environment has less patience for the story.

A16z has published frameworks for how it thinks about AI retention that hint at what sophisticated Series A diligence will look like. The firm’s research recommends measuring retention from month three rather than month zero, to exclude AI tourists from the base. M12 divided by M3, how well customers who survived initial churn perform over their first full year, is an early predictor of long-term retention quality. Companies that cannot demonstrate a flattening of their retention curve, and ideally the early signs of expansion, will face skepticism about whether their ARR figure is investable.

The key indicator to watch is enterprise renewal rates in Q3 and Q4 2026. The cohort of companies that signed large enterprise pilots in 2024 and early 2025 will be coming up for their first full-year renewal decision in that window. If the renewal rate is strong, it will validate the current seed pricing. If renewal rates are materially weaker than the initial ARR figures suggested, expect a rapid repricing of the market.

If the Winter 2027 Y Combinator’s batch shows a shift back toward companies with fewer AI companies and more capital-efficient, non-AI businesses, it will suggest that the current premium has begun to normalize. The opposite, continued or increasing AI concentration, would suggest the market believes the retention problem is a temporary early-stage phenomenon rather than a structural flaw in how these businesses are built.

What would need to be true for the optimistic path: retention rates at AI-native companies converge toward B2B SaaS benchmarks as products deepen into enterprise workflows; the companies that raised at high seed valuations use the capital to build genuine workflow lock-in and multi-year contracts; and base model improvements expand the market rather than commoditizing existing players.

What would need to be true for the pessimistic path: enterprise buyers scrutinize their AI experiments at renewal and find that ROI has not materialized; base model improvements absorb the functionality of thin-layer AI products; and a significant portion of seed-stage companies find themselves unable to raise Series A at or above their seed valuations, triggering a down-round cycle.

The Reframe

The most important thing Li's warning surfaces is that ARR has become a single number doing the work of at least four different questions: How much revenue did the company generate recently? Is it contracted or discretionary? Will customers still be there in twelve months? Are customers using the product enough that their spend will grow over time?

In traditional SaaS, annual contracts meant customers had committed. High switching costs meant they were unlikely to leave. Seats-based pricing meant expansion was relatively predictable. The metrics and the business model were aligned.

The market is working through a genuine accounting problem: the tools developed for SaaS do not cleanly map onto AI businesses, and no replacement framework has achieved consensus. Until one does, the safest reading of any headline ARR figure is a starting point for questions, not an answer.