The barbellization of the software world: Is AI about to kill the middle ground?

Original author: Venture Curator

Original compilation: TechFlow

TechFlow Introduction: 537 U.S. unicorns are still carrying old valuations from the 2021 bubble era, and the "AI-native" label is shifting from a bonus point to a risk factor waiting to be exposed. For practitioners who are fundraising or holding options, this article helps you distinguish which valuations are real and which are merely paper numbers that no one has punctured.

537 U.S. unicorns are still carrying valuations from the 2021-22 bubble era. Are they still real?

Founders and early employees are often taught to treat their last round valuation as a scoreboard. It is the number on the cap table, the number in the press release, and the number everyone uses to estimate the value of their equity.

But new data from PitchBook's Q3 2026 quantitative outlook report shows that for most U.S. unicorns, that number has not been truly tested by the market for several years.

PitchBook studied all active U.S. unicorns, categorizing them by the year of their most recent priced financing round.

Of the 964 active U.S. unicorns, 312 last raised in 2021, and another 225 in 2022. This means 537 companies—about 56% of the entire unicorn population—are still carrying valuations set during the most expensive fundraising period in venture capital history.

Cash is not coming back to match those numbers. VC fund distributions account for only 7.9% of net asset value, nearly half the long-term average of 14.5%. PitchBook notes that fund returns are being directly dragged down by the inflated valuations of 2020 to 2022.

When these companies actually test their prices, they often cannot hold up. Public market investors refuse to pay private market prices. Chime's IPO valuation was significantly lower than its private peak, and secondary market buyers offer far larger discounts to companies whose last round was years ago compared with those that recently completed financing.

There is a second trend hidden behind this. Many of these companies are not stalled because they cannot continue operating, but because they actively choose not to raise or go public, since any new pricing event could force them to face a lower number. Staying private and staying quiet keeps the 2021 valuation on paper. This is why the market is simultaneously seeing record paper value and record trapped value.

So is the unicorn boom real, or is it just a number no one has verified?

For many companies, both statements can be true. Some companies have genuinely grown into their 2021 valuations and could easily sustain them today. But the market has no way to confirm which companies fall into this category, because they have not raised or sold at a real price since then. A valuation from four years ago is not a lie, just unverified.

This is exactly the part founders and employees need to be careful about.

If your company's last financing was in 2021 or 2022, the valuation on the cap table is the most optimistic number your equity has ever had, and it does not necessarily equal today's actual value. This affects how you discuss your next round, how you think about exercising options, and how you explain to your team what their options mean. The same goes for investors: if a fund shows strong unrealized gains on these companies, it is only reporting book numbers, not real cash.

The real risk is not that the valuation may be lower than it appears, but that you make financing, hiring, or option exercise decisions while treating the number set in 2021 as the price the market is still willing to pay you today.

Calling your startup "AI-native" might actually be hurting your valuation?

Calling yourself an AI company can still command a valuation premium. But investors are increasingly asking a second question: how much of this company is truly defensible because of AI?

The data in this analysis shows that the market is beginning to sort companies with the same "AI-native" label into very different buckets.

The AI premium is still real: Carta's H1 2026 data shows that AI startups raise roughly the same amount of money at seed stage as non-AI companies, but at valuations about 50% higher.

But not every AI company gets the same premium: the analysis values commoditized AI wrapper companies at 3 to 8 times revenue, vertical AI products with sticky proprietary data at 10 to 20 times, and companies that truly own intellectual property plus proprietary data at 25 to 40 times.

The more mature the company, the stricter the scrutiny: at seed stage, companies without protectable intellectual property are discounted by 20% to 30% relative to more defensible peers. By Series A, that gap widens to 30% to 40%.

This explains why simply adding AI to a product is becoming increasingly meaningless.

Investors increasingly want to know what is left when you take away the AI label.

Could another startup build the product by calling the same foundation model API? Does each new customer generate proprietary data that improves the product? Is the software embedded deeply enough in workflows that replacing it would be painful? Does the unit economics still hold when inference volume grows 10x?

These questions point to four areas founders should examine before their next round: technical differentiation, proprietary data, workflow lock-in, and unit economics at scale.

The practical test for technical differentiation is surprisingly simple: swap out the underlying model. If replacing your primary model provider with another comparable model barely changes the product, investors may not see where the technical moat is.

For proprietary data, user numbers alone are not enough. A stronger signal is whether increased usage generates data that makes the product measurably better over time, such as improvements in accuracy, resolution time, or match quality.

Workflow depth is another clue. If customers who integrate the product more deeply consistently have better retention and expansion, founders have actual evidence of switching costs rather than just claiming the product is sticky.

Then there is the economic model. An AI product may have healthy margins at current usage, but that may be a completely different story when inference volume scales 10x. Working through those numbers before investors do can reveal whether the business truly gets stronger as it scales.

This makes the evolution of the "AI-native" premium very interesting.

Eighteen months ago, merely being associated with AI could materially strengthen a fundraising story. As more startups use the same language, investors have more reason to look beneath the surface.

The AI label gets you into the conversation. The increasingly valuable part is proving why what you are building becomes harder to replicate as it grows.

The barbellization of software: AI is about to kill the middle layer of the software market.

AI is making software dramatically cheaper and faster to build. That sounds like good news for startups, but it may also erode some of the moats that have protected software companies for decades.

Conviction's Mike Vernal made an interesting comparison: software may be heading toward the same "barbell" structure that formed after the internet changed the newspaper industry.

Before the internet, regional newspapers benefited from expensive distribution costs. Once distribution became nearly free, the middle layer largely disappeared. A few global publications became much larger, while thousands of independent newsletters and niche publications emerged.

Vernal believes AI may create a similar structure in software.

Traditional software moats usually come from three things:

Building the product is expensive

Switching systems is painful

Integrations create powerful ecosystems.

AI weakens all three. Competitors can replicate features faster, migration can become more automated, and AI can make integration significantly easier.

The likely result is a market that is extremely strong at both ends:

Massive platforms: A few companies may own entire procurement categories, such as sales, marketing, finance, HR, legal, or healthcare. Instead of selling just one narrow tool, they keep expanding until they become the system customers use for almost everything.

Micro software businesses: At the other end, AI coding tools may allow millions of people to build highly specific software for themselves, their own companies, or small groups of customers. Some of these may grow into profitable niche businesses.

The uncomfortable middle: Mid-sized point solutions may face the greatest pressure. If large platforms can keep adding their features and small teams can cheaply rebuild narrow products, then defending as a standalone tool becomes harder.

The Amazon analogy explains what the new moat may look like. Amazon started with something relatively easy to copy: selling books online. Its defensibility came from decades of relentless reinvestment in logistics, infrastructure, technology, and distribution.

AI may push software companies toward the same strategy: build faster, keep expanding, reinvest, and make the overall product increasingly difficult to replicate.

For venture-backed software startups, this creates an interesting strategic question: if building software is nearly free, is your moat still a good product, or is it everything you can build around it?

How is your seed round valuation actually determined? (It has almost nothing to do with your company.)

Founders often assume valuation is mainly about traction, market size, or how well they pitch.

But there is another constraint most founders can barely see: the economic model of the VC fund sitting across the table.

Imagine you are pitching a $28 million seed fund. It typically writes a check of around $600,000 and wants about 6% to 8% ownership.

At 7% ownership, a $600,000 investment implies a post-money valuation of about $8.6 million. At 6%, it implies $10 million. So if you raise at an $18 million valuation, that fund may simply be unable to make the investment within its portfolio model. Pasted markdown

This creates a simple chain:

Fund size → check size → target ownership → feasible valuation ceiling

This matters more today because seed valuations have risen. The median Carta seed post-money valuation cited in the analysis is $24 million. A $50 million fund investing $1.5 million at that valuation would only get 6.2% ownership, potentially below its target. Pasted markdown

This also changes how founders think about round size.

If you tell a fund you are raising $3 million, and the lead investor wants 15%, you have effectively already implied a $20 million post-money valuation before anyone explicitly discusses valuation.

A more useful sequence is:

Calculate how much capital you need to reach the next milestone with 18 to 24 months of runway.

Add a buffer.

That is your round size.

Divide it by the dilution you are willing to accept.

Then go find funds whose check size and ownership model can support that valuation. Pasted markdown

There is another number founders often overlook: the option pool. The lead investor may require a 10% to 15% option pool to be established before the investment, which dilutes existing shareholders before the investment lands. The analysis argues that negotiating the option pool based on actual hiring plans sometimes has a bigger impact on founder ownership than fighting for an extra $1 million or $2 million in headline valuation. Pasted markdown

The simplest way to avoid wasting weeks with the wrong fund is to ask this question on the first call:

"What is this fund's typical first check size, and how much equity do you usually aim to own?"

Those two numbers let you quickly estimate whether the fund's economic model supports your round.

So do not treat every valuation pushback as a rejection of your startup. Sometimes your company is fine. It is just that your round does not fit the fund's math.

If ChatGPT can do what your app does for free, why would anyone pay?

This is becoming one of the biggest questions facing consumer app founders.

If users can open ChatGPT, Claude, or Gemini and get roughly the same result in seconds, why would they pay $10 or $20 a month for a standalone app?

RevenueCat's Daphne Tideman argues that the answer is not simply "add more AI." AI-powered apps already monetize well, generating 41% more revenue per paying user than non-AI apps, with 52% higher trial conversion. But their 12-month retention rate is only 21.1%, compared with 30.7% for non-AI apps.

In other words: AI can bring people in the door. But it does not necessarily give them a reason to stay.

Look at the comparison between Chegg and Duolingo.

The difference lies in what exists beyond the answer.

RevenueCat suggests apps can build six things that are harder for a blank LLM chat to replicate:

Structure: turning an answer into a complete workflow or experience.

Memory: understanding users over time, making the app more useful.

Habits: using streaks, reminders, accountability, and gamification to bring users back.

Precision: mastering specialized data or expertise to produce better results.

Connection: building community, identity, personality, or emotional attachment.

Physical and digital integration: combining software with sensors, devices, or real-world data.

Duolingo is a great example of stacking several of these together. ChatGPT can teach languages, but Duolingo spent years building streaks, leaderboards, progress systems, reminders, and gamification. Duolingo reportedly saw a 17% increase in learning time and tripled its number of highly engaged learners after introducing leaderboards.

There is also a simple test founders can do: the blank chat box test.

Open ChatGPT or another LLM, describe exactly the problem your users come to your app to solve, and then compare its answer with what your product delivers. If most of your product's value can be reproduced with one prompt and one response, you have a differentiation problem. If your app adds structure, accumulated data, habits, precision, connection, or real-world components, that may be where your moat lies.