Venture Curator’s article, translated by TechFlow, brings several strands of the software market into one frame: stale unicorn valuations, the changing AI premium, the possible hollowing out of mid-market software, seed-stage valuation mechanics, and the retention problem facing consumer AI apps. The throughline is that AI is changing what the market is willing to value, and why.
More than half of active U.S. unicorns still carry 2021–2022 pricing
According to PitchBook’s Q3 2026 Quantitative Perspectives report, cited in the article, 312 of 964 active U.S. unicorns last completed a priced round in 2021, and another 225 did so in 2022. That leaves 537 companies, or about 56% of the total, still marked at valuations set during the most expensive stretch in venture-capital history.
For founders and early employees, the last round’s valuation often becomes the company’s scoreboard. It sits on the cap table, appears in press releases, and is frequently used as a proxy for what their equity is worth. The article argues that for most U.S. unicorns, those numbers have not been meaningfully tested by the market in years.
PitchBook also reported that venture fund distributions were only 7.9% of net asset value, nearly half the long-run average of 14.5%. In the report’s framing, inflated valuations from 2020 to 2022 are directly weighing on fund returns.
When companies do face a real pricing event, the article says many fail to hold up. Public-market investors have not been willing to pay private-market prices. Chime, for example, went public at a valuation far below its private-market peak, and secondary buyers have applied steeper discounts to companies whose last financing happened years ago than to businesses that raised more recently.
A second pattern sits underneath that. Many of these companies are not frozen because they cannot operate. They are choosing not to raise and not to list, because any new pricing event could force them to accept a lower number. Staying private and staying quiet lets a 2021 valuation remain on paper. That, the piece argues, is how the market can produce both record paper value and record trapped value at the same time.
So the unicorn boom may be both real and unverifiable. Some companies may genuinely have grown into their 2021 prices. The problem is that the market has no way to confirm which ones have, because they have not raised or sold at real prices since then. A four-year-old valuation is not necessarily a lie. It is unverified.
The article says this is where founders and employees need to be careful. If a company last raised in 2021 or 2022, the cap-table valuation may simply be the most optimistic number their equity has ever had, not its current value. That matters when discussing the next round, deciding whether to exercise options, or explaining to employees what their options are really worth.
The same logic applies to investors. If a fund is showing strong unrealized gains in these companies, it is still reporting book value, not returned cash. The deeper risk is not only that valuations may be lower than they look. It is that financing, hiring, and exercise decisions may still be anchored to a price the market no longer accepts.
The AI premium still exists, but investors are splitting companies into very different buckets
The article then turns to AI. Calling a startup an AI company can still command a premium, but investors are asking a more pointed second question: how much of the business is actually defensible because of AI?
Carta’s data for the first half of 2026, as cited in the piece, shows that AI startups raised roughly the same amount as non-AI startups at seed, but at valuations about 50% higher. So the premium is still there.
What has changed is how selectively it gets applied. The analysis places AI companies into distinct valuation bands:
- Commoditized AI wrappers at roughly 3x to 8x revenue.
- Vertical AI products with sticky proprietary data at roughly 10x to 20x revenue.
- Companies with both intellectual property and proprietary data at roughly 25x to 40x revenue.
Scrutiny rises as companies mature. At seed, businesses without protectable IP may take a 20% to 30% discount relative to stronger peers. By Series A, the gap widens to 30% to 40%, according to the article.
That is why simply adding AI to a product is becoming less meaningful. Investors want to know what remains once the label is stripped away.
The questions are basic but hard to dodge. Could another startup recreate the product by calling the same foundation-model API? Does each new customer generate proprietary data that improves the product over time? Is the software embedded deeply enough in workflow that replacing it would be painful? If inference volume rises 10x, do the unit economics still hold?
The article reduces that checklist to four areas founders should review before their next round:
- Technical differentiation.
- Proprietary data.
- Workflow lock-in.
- Unit economics at scale.
On technical differentiation, the proposed test is simple: swap out the underlying model. If changing the primary model provider to another comparable model barely changes the product, investors may conclude there is no real technical moat.
On proprietary data, user count alone is not enough. A stronger signal is whether higher usage creates data that measurably improves the product over time, whether in accuracy, resolution time, or match quality.
Workflow depth shows up in retention and expansion. If customers that integrate the product more deeply consistently retain better and expand more, founders have evidence of switching costs rather than a generic claim of stickiness.
Then there is the business model. An AI product may look healthy at current usage, but a 10x jump in inference can produce a very different cost structure. The article suggests running those numbers before investors do.
Its conclusion on the “AI-native” premium is narrow and practical. Eighteen months ago, being associated with AI could materially strengthen a funding story on its own. As more startups adopted the same narrative, investors found more reasons to look beneath the surface. The label may get a company into the conversation. The valuable part now is proving why the product becomes harder to copy as it grows.
AI may push software into a barbell structure
The article’s next argument is about market structure. AI is making software cheaper and faster to build. That sounds positive for startups, but it may also weaken the classic moats software companies relied on for decades.
Conviction’s Mike Vernal compares the shift to what the internet did to newspapers. Before the internet, regional newspapers benefited from expensive distribution. Once distribution became nearly free, much of the middle disappeared. A few global publications grew much larger, while thousands of independent newsletters and niche outlets also emerged.
Vernal argues that AI could create something similar in software. Traditional software moats usually came from three things:
- Products were expensive to build.
- Switching systems was painful.
- Integrations created powerful ecosystems.
AI weakens all three, the article says. Competitors can copy features more quickly. Migration can become more automated. AI can also make integration much easier.
The likely result is a market with strength at both ends. On one side, giant platforms may come to own broad purchasing categories such as sales, marketing, finance, HR, legal, or healthcare. They would not just sell a narrow tool. They would keep expanding until they became the system customers use for almost everything in that domain.
On the other side, AI coding tools may allow millions of people to build highly specific software for themselves, their companies, or small customer groups. Some of those products could grow into profitable niche businesses.
The uncomfortable place is the middle. Mid-sized point solutions may face the most pressure. If large platforms can continuously absorb their features while small teams can cheaply rebuild narrow products, defending a stand-alone tool becomes harder.
The article uses Amazon as a metaphor for what a new moat might look like. Amazon began with something relatively easy to copy: selling books online. Its defensibility came later, through decades of reinvestment in logistics, infrastructure, technology, and distribution.
Software companies, in this view, may need to follow the same pattern under AI: build faster, expand continuously, reinvest, and make the total system harder and harder to replicate. For venture-backed software startups, that creates a strategic question. If building software is close to free, is the moat still a good product, or everything built around it?
Seed valuations are often constrained by fund math before founders realize it
Founders often assume valuation is mostly a function of traction, market size, or how well they pitch. The article says another force often matters just as much: the economic model of the venture fund sitting across the table.
It gives a simple example. Imagine pitching a $28 million seed fund. That fund typically writes checks of about $600,000 and wants roughly 6% to 8% ownership. At 7%, a $600,000 investment implies about an $8.6 million post-money valuation. At 6%, it implies $10 million. If a company is raising at an $18 million valuation, that fund may not be able to make the investment work inside its portfolio model at all.
The logic forms a straightforward chain:
- Fund size.
- Check size.
- Target ownership.
- Feasible valuation ceiling.
This matters more now because seed valuations have risen. The article cites Carta data putting the median seed post-money valuation at $24 million. A $50 million fund investing $1.5 million at that valuation would end up with 6.2% ownership, which may fall short of its target.
That changes how founders should think about round size. If a founder says the company is raising $3 million and the lead wants 15%, the implied post-money valuation is already $20 million, whether anyone has stated the number or not.
The recommended sequence is different. First calculate how much capital is needed to reach the next milestone with 18 to 24 months of runway. Then add a buffer. That gives the round size. Divide that by the dilution you are willing to accept. Then go find funds whose check sizes and ownership targets can support that valuation.
The article also highlights another number founders often overlook: the option pool. A lead investor may ask for a 10% to 15% option pool to be created before the financing closes, which dilutes existing holders before the new money comes in. The analysis argues that negotiating the pool based on real hiring plans can matter more to founder ownership than gaining an extra $1 million or $2 million on the headline valuation.
To avoid spending weeks with the wrong investors, the article offers a blunt question founders can ask on the first call: “What is the fund’s typical first check, and how much ownership do you usually want?” Those two numbers can reveal quickly whether the fund’s economics fit the round.
The broader point is that valuation resistance is not always a verdict on the company. Sometimes the business is not the issue. The round just does not fit the fund’s math.
In consumer AI, monetization can be strong while retention stays weak
The article’s final section looks at consumer AI apps. The challenge is easy to state: if a user can open ChatGPT, Claude, or Gemini and get roughly the same result in seconds, why pay $10 or $20 a month for a standalone app?
RevenueCat’s Daphne Tideman argues that the answer is not simply “add more AI.” RevenueCat’s data shows AI-powered apps are monetizing well. Revenue per paying user is 41% higher than for non-AI apps, and trial conversion is 52% higher. But 12-month retention is only 21.1%, versus 30.7% for non-AI apps.
That leaves a clear picture. AI can get people through the door. It does not automatically give them a reason to stay.
The article points to the contrast between Chegg and Duolingo. The difference is not just the answer itself. It is everything around the answer.
RevenueCat suggests six categories of value that are harder for a generic LLM chat interface to copy:
- Structure: turning an answer into a full workflow or experience.
- Memory: learning about the user over time so the product becomes more useful.
- Habit: bringing users back through streaks, reminders, accountability, and gamification.
- Precision: using domain expertise or specialized data to deliver better results.
- Connection: creating community, identity, personality, or emotional attachment.
- Physical-digital combination: tying software to sensors, devices, or real-world data.
Duolingo is presented as a strong example because it stacks several of these features together. ChatGPT can help teach a language, but Duolingo spent years building streaks, leaderboards, progress systems, reminders, and game mechanics. The article notes that, according to reports, study time rose 17% after Duolingo introduced leaderboards, and the number of highly engaged learners tripled.
For founders, the piece offers a practical test: the blank-chatbox test. Open ChatGPT or another large language model, describe as accurately as possible the job a user comes to your app to get done, and compare the response with what your product actually delivers.
If most of the product’s value can be recreated with one prompt and one response, there is a differentiation problem. If the app adds structure, accumulated data, habits, precision, connection, or real-world components, those features may be the actual moat.
The article ends on a simple point. AI itself may not be the moat. The better question for founders is what their product gives users that a blank chat box cannot.
In RevenueCat’s view, the winners will use AI underneath the product and compete on the experience built around it.