Author: Venture Curator
Compiled by: Shenchao TechFlow
Shenchao Introduction: This article uses two sets of hard data to expose two things: giant VCs are crowding into homogeneous AI projects, while emerging managers are more willing to bet on outliers; and most VC funds have not even returned principal to LPs nine years after inception. For founders currently fundraising, understanding the capital pressure behind a fund matters far more than looking at brand halo.
Emerging VCs are investing in more unusual startups, while AI spending varies enormously across more than 70,000 enterprises.
Who exactly is investing in venture capital's outliers? What the data shows.
Venture capitalists are usually taught that this is a business about outliers. A handful of companies no one expected contribute most of the returns, so the institutions with the largest teams, broadest networks, and most capital should be best positioned to find them.
But new research by Odin's Dan Grey based on Dealroom data shows that the largest institutions are now the least likely to invest in a company that looks different from every other funded company.
Grey scored every company in the 2023 to 2025 seed portfolios of five mega-funds (a16z, General Catalyst, Lightspeed, NEA, and Accel), as well as a select group of top emerging managers, by rarity. Rarity was measured across two dimensions, business (industry, model, customer, geography) and technology, against all companies that raised seed rounds from 2018 to 2025. The most unusual quintile counted as outliers.
Emerging managers allocated 24.7% of their scoreable seed rounds to outlier profiles (49 out of 198). Mega-funds allocated only 11.5% (32 out of 278). That is roughly a two-times gap.
This gap is not because every small fund is a contrarian. Many emerging managers are just as heavily weighted toward AI as mega-funds: Conviction is 100% AI, South Park Commons is 93.5%. The gap comes from specialist funds with different mandates, such as Lowercarbon with only 18.8% in AI and Multicoin with only 15.4%.
This is why portfolio allocation matters. Combining five different emerging manager books usually reduces concentration by 22.6%. Combining five mega-funds reduces it by only 2.0%, because all five are already between 82.7% and 89.5% AI positioning. Out of 462 possible emerging manager combinations, only 2 failed to outperform the mega-fund group.
The significance of this extends beyond portfolio statistics. Research on financing risk has long held that whether investors expect follow-on capital to be available determines which experiments get the first check. Odin cautiously states that its data cannot prove this mechanism. But if the largest pools of capital only invest in projects already widely considered investable, then companies outside that consensus can only rely on others to make the first bet.
So does this mean emerging managers are better at picking winners?
No. "Outlier" here means rare among funded companies; it does not mean more innovative, nor more likely to succeed. When founder background is added to the scoring, the gap narrows to 19.8% versus 14.4%. The emerging manager group is a curated sample and does not represent the entire market. And small funds have their own herd moments too: blockchain accounted for 46.7% of their early rounds, later falling to 20.9%.
This is exactly where investors need to be careful.
If your thesis is that returns come from outliers, then adding another brand-name fund mostly just adds the same bet. A basket of funds with genuinely different mandates adds range. But range alone is not a reason to write a check. Odin's own test is: what missing thing can a fund add to your existing portfolio, at what fee and carried interest, at what diluted ownership, and with what follow-on financing risk.
The real question is not which fund is largest. The question is which funds are genuinely looking where others are not, because based on the data, that list is increasingly not the same group of people.
Nine years later, most VC funds still have not returned money to investors. What happens next?
Founders are usually taught to view their VCs as patient capital. Venture capital is a long game; funds are set up to wait for big outcomes, and no one on the cap table is in a hurry for cash.
But new data from Carta's Q2 2026 VC fund performance report shows that many funds' patience is running out, because most still have not returned their investors' money.
Carta tracked net DPI, the share of cash a fund has actually returned to LPs per dollar LPs put in, covering roughly 3,000 U.S. venture capital funds that began investing between 2017 and 2026.
Funds that began investing in 2017 are now about nine years old, near the end of a typical ten-year fund life cycle. The median fund has returned only 0.37x cash. Even top-quartile funds have returned only 0.70x. Only the top 10% have returned more than LPs put in, at 1.37x.
Younger funds are worse off. The median 2018 fund has returned 0.15x, and the median 2019 fund is 0.04x, meaning just 4 cents on the dollar after seven years.
Value exists on paper. The median 2017 fund shows a 1.72x paper value, but only about one-fifth of that has been returned in cash. The rest remains locked in private companies that have not yet been sold or listed.
There is a second detail that explains why this is hard to see from the outside. Most of these funds have indeed returned some money: 86% of 2017 funds and 70.4% of 2018 funds have returned at least some cash. So almost every fund can tell its LPs that distributions have begun. But few funds can say they have returned the money.
So why should founders care about their investors' fund books?
Because your investor's fund clock is not the same as your company's clock. A fund nearing maturity, still owing LPs real money, and currently raising its next fund has a strong incentive to get liquidity early. That pressure may show up as pushing you to sell earlier, selling its stake on the secondary market, or becoming more cautious about continuing to invest in your next round.
And this is where founders need to be careful.
None of this means your investor will do something against your interests. Funds can extend their lives, and many top institutions have good reason to keep supporting their winners. Carta's data only shows the cash gap; it does not indicate how every fund will act. But the vintage of the fund that wrote your check is now a real variable, not a footnote. A partner writing from a 2017 or 2018 fund is in a completely different position from a partner writing from a brand-new 2025 fund.
The real risk is not that your investor needs liquidity. It is that you do not know which fund your money came from, how much cash it still owes its own backers, and how that affects the advice they give you on your next round or exit.
Where is the biggest money in AI: frontier models or the mid-market?
Most AI coverage stares at the frontier: each lab's newest, smartest, most expensive model. But Theory Ventures' Tomasz Tunguz makes the argument that the market that truly matters is the middle layer, and spending data supports this.
The middle layer has already taken about 40% of AI spending and 30% of token usage.
By contrast, the frontier model segment is thin. Anthropic's strongest model, Fable 5.1, accounted for only 3.7% of gateway spending in its first 12 days after release. Its predecessor peaked at 13.2%, then fell to 4.9% a month after Opus 5 launched at half the price. In large enterprise accounts, frontier models' share of token consumption fell from 53% in early August to 45% in September.The price war is unfolding at exactly this tier. When Anthropic cut prices on its new model in September, OpenAI followed about 90 minutes later. Before that cut, four consecutive versions of the Opus series held pricing at $5 per million input tokens and $25 per million output tokens. At the low end, OpenAI slashed Luna's price by 80% in July and another 50% in September.
Three forces keep pushing prices at the middle tier down:
Competition among labs. Labs now follow each other's price cuts within hours, not quarters.
Open-weight models. On gateways that publish data, open models run the majority of token volume, at 86% below the blended price of closed models.
Fine-tuning. Cursor Composer 2, fine-tuned on the open-weight model Kimi K2.5, cut total cost by 86% versus its previous in-house model. Harvey cut cost per cell by 55% relative to Sonnet 5, while scoring higher than Fable 5.
That this keeps happening is structural. What enterprises need AI to do—summarize a contract or classify a ticket—barely changes year over year. The cost of intelligence that clears that bar, meanwhile, is falling fast. So the tier that passes a fixed requirement gets cheaper every quarter, and most real work lands in that tier, not at the frontier.
Demand doesn't look like a pyramid with the frontier at the top. It looks more like a bell curve that's fat in the middle. Buyers care about how much intelligence they get per dollar, not peak capability. The open question is whether this middle market eventually becomes a commodity. If it does, the economics of the entire AI market change with it.
For anyone building a product on top of these models, it's worth asking: which tier does your product actually need. Paying frontier-model prices for middle-tier work is now the best kind of margin to lose.
How much should your startup actually spend on AI? Here's what the data says.
Ask most founders how much they spend on AI and you'll get a confident number: seats on ChatGPT and Claude, the Cursor bill, a growing API line, a few thousand dollars a month all in. It sounds precise. It isn't. Because it adds together three completely unrelated kinds of spend whose only common trait is the vendor category.
Payment data from Ramp covering more than 70,000 U.S. businesses shows how fragmented "using AI" has become.
The median company spends $11.38 per employee per month. The top 10% spend $611.
The top 1% spend $7,449, roughly 650x the median. And when Ramp linked that spend to workforce records from Revelio Labs covering 21,559 companies, it found that high-spend companies grew headcount 10.2% over two years, with entry-level hiring up 12%.
The low-intensity users showed no significant change at all.
So spend more. But spend where?
Most founders compare their AI bill to their software budget, because the charges sit on the same statement as Figma and Notion. That's the wrong denominator. A growing share of AI spend replaces people, not software. The comparison for a customer-service agent is a customer-service role, not your SaaS suite.
Run the math on a typical seed-stage company: 12 people, about $180,000 a month in payroll, $4,000 a month in AI spend. Against a software budget, $4,000 looks heavy. Against payroll, it's 2.2%. Same number, opposite conclusion.
The fix is to split the AI budget into three buckets, each judged against its own benchmark:
Payroll-line AI (support, SDR, and coding agents) is measured against the salary of the person you didn't hire. If you can't name that role, it doesn't belong here.
Productivity-line AI (seats, note-takers, writing tools) is measured by usage. If usage decays by month two, cancel it.
Cost-line AI (in-product tokens and API calls) is measured by gross margin at 10x volume, and it belongs in your COGS.
Blend the three and every signal cancels out. Overspend on seats makes the total look bloated, so founders get conservative on agents, which is exactly where the return is. The top 1% aren't winning because they spend more. They're winning because their spend is sorted: heavy where AI replaces labor, disciplined where it only accelerates people, and margin-aware where it sits inside the product.
We broke down the full ranking method, including a 30-minute audit you can run against your own bill this week, in "What Should Your AI Spend Actually Be Benchmarked Against?"
Why has deep-tech VC diligence gotten worse?
You'd think the hardest technology would get the strictest scrutiny. A quantum computer, a fusion reactor, or a new chip architecture either works within the laws of physics or it doesn't, and verifying that takes real expertise.
But Marin Ivezic of PostQuantum argues the opposite is happening: the bigger and more technical the deal, the less likely anyone independent verifies whether the science holds up.
It isn't because investors got lazy. Several structural shifts compounded.
The expert layer disappeared.
Investors used to get independent technical judgment from bank research teams staffed with specialist analysts. That model contracted after Europe's MiFID II rules in 2018 required research to be paid for separately from trading. European equity research reportedly shrank by about 20%, and specialist analyst headcount at major banks fell with it.
Expert networks, now a roughly $2.5 billion business, filled part of the gap. But they only work if you know which expert to ask and what to ask them. At the frontier, that's precisely the knowledge generalist investors don't have.
The money moved to generalists.
Private quantum investment hit $4.9 billion in 2025, up 192%. The largest backers were BlackRock and Nvidia, not dedicated quantum funds. Sovereign wealth funds and pensions are increasingly in these rounds, and a survey of sovereign funds found 58% lack the resources to lead a deal. Capital that can't do its own technical review ends up relying on someone else's judgment.
The lead investor's name replaced the review.
A prestigious lead underwrites everyone who follows. Followers assume the lead verified the physics, but they can't see that review and can't confirm it happened. Some rounds don't even have that: Quantinuum's $600 million raise had no named lead at all.
Deadlines and herd behavior put speed ahead of scrutiny.
U.S. funds ended 2023 sitting on a record $311.6 billion in dry powder. A fund manager facing deployment pressure earns more by joining a hot deal than by slowing it down. And being wrong alongside peers damages reputation far less than being wrong alone.
Deep tech has no forced checkpoint.
Biotech is the useful contrast. The FDA approval process forces data into the open at every stage, so the industry built a whole diligence discipline around it: scientific advisory boards, specialist reviewers, staged checkpoints.
Quantum has no equivalent gate. The most rigorous technical review of a major quantum company this article could find came from the Australian government, disclosed through a freedom-of-information request, not from any investor.
The results show up plainly in the numbers.
Private quantum investment was $4.9 billion in 2025, while total revenue across the quantum computing industry was about $1.4 billion. Three quantum companies that went public via SPAC in 2021 forecast roughly $1.2 billion in combined 2025 revenue and delivered about $138 million. And some of the most solid verification work was done by short sellers, not investors.
There's precedent for how this ends. Between 2006 and 2011, venture capital put more than $25 billion into cleantech and lost more than half of it. Scholars studying that cycle warned at the time that pensions and sovereign funds without hardware experience would be the next to write checks into these companies.The proposed solution is simple: before co-investing in a deep-tech round above a certain size, LPs should require an independent technical feasibility review, one that is neither commissioned by the fund nor selected by the founder. Disclosure: the author invests in quantum startups and runs a firm that provides such reviews. But the core point still stands: in deep tech, a logo on the cap table has quietly replaced technical due diligence, and at those check sizes, the cost of a few weeks of independent review is a rounding error.