TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
How much should AI spending be? Why haven't most VC funds returned capital to investors yet, and who are the true outliers?

How much should AI spending be? Why haven't most VC funds returned capital to investors yet, and who are the true outliers?

2026.10.10
Share
For founders currently fundraising, understanding the capital pressures behind the fund is far more important than focusing on its brand halo.

Author: Venture Curator

Translated by TechFlow

TechFlow Summary: This article uses two sets of hard data to debunk two realities: giant VCs cluster around homogeneous AI projects, whereas emerging managers are actually more willing to bet on outliers; and most VC funds fail to return their initial capital to LPs nine years after inception. For founders currently fundraising, understanding the capital pressures behind these funds matters far more than their brand pedigree.

Emerging VCs are investing in more unconventional startups, and AI spend varies drastically across over 70,000 companies.

Who Exactly Is Investing in VC Outliers? What the Data Shows.

VCs are typically taught that this is a business built on outliers. Since a small number of unexpected companies generate the majority of returns, firms with the largest teams, broadest networks, and deepest pockets should theoretically be best equipped to find them.

But Dan Grey of Odin’s new research, based on Dealroom data, suggests that the largest firms are now the least likely to invest in companies that look fundamentally different from everyone else in the portfolio.

Grey scored every company across the seed-stage portfolios of five mega-funds (a16z, General Catalyst, Lightspeed, NEA, and Accel) between 2023 and 2025, alongside a curated cohort of top-tier emerging managers, based on rarity. Rarity was measured along business (industry, model, customer, geography) and technical dimensions, benchmarked against all companies receiving seed funding from 2018 to 2025. The least typical one-fifth were classified as outliers.

Emerging managers allocated 24.7% of their scoreable seed rounds to outlier profiles (49 out of 198). Mega-funds only invested 11.5% (32 out of 278). That’s roughly a two-fold gap.

This gap isn’t because every small fund is purely contrarian. Many emerging managers are equally heavy on AI like mega-funds: Conviction is 100% AI, South Park Commons is 93.5%. The difference comes from specialized funds with different mandates, such as Lowercarbon, which allocated only 18.8% to AI, and Multicoin, at just 15.4%.

This is why portfolio allocation matters. Aggregating the ledgers of five distinct emerging managers typically reduces concentration by 22.6%. Aggregating five mega-funds only drops it by 2.0%, because all five already sit between 82.7% and 89.5% AI exposure. Out of 462 possible combinations of emerging managers, only two failed to outperform the mega-fund cluster.

The significance extends beyond portfolio statistics. Research on financing risk has long held that whether investors expect follow-on capital determines which experiments get the first check. Odin cautiously notes its data doesn't prove this mechanism. But if the largest pools of capital only invest in projects already widely considered viable, companies outside that consensus are left relying on others to make the first call.

Does this mean emerging managers are simply better at picking winners?

No. "Outlier" here refers to rarity within the funded companies; it does not imply greater innovation or higher likelihood of success. Adding founder background to the scoring narrows the gap to 19.8% versus 14.4%. This group of emerging managers is a curated sample and does not represent the broader market. Plus, smaller funds have their own herd moments: blockchain accounted for 46.7% of their early rounds before dropping to 20.9%.

This is exactly where investors need to exercise caution.

If your thesis is that returns come from outliers, adding another legacy firm mostly just duplicates the same bets. A basket of funds with genuinely distinct mandates is what actually expands scope. But scope alone doesn't justify writing checks. Odin’s own test is this: what gaps does a fund fill in your existing portfolio, at what fee and carry structure, what diluted ownership do you get, and what is the follow-on financing risk?

The real question isn't which fund is the biggest. It's which funds are actually looking where no one else is, because as the data shows, that list is increasingly composed of different players.

Nine Years In, Most VC Funds Still Haven't Returned Money to Investors. What Happens Next?

Founders are often taught to view their VCs as patient capital. Venture capital is a long game; funds are structured to wait for outsized outcomes, and no one on the cap table is rushing for liquidity.

But new data from Carta’s Q2 2026 VC Fund Performance Report shows many funds' patience is wearing thin, as most still haven't returned their investors' money.

Carta tracked net DPI—the ratio of cash actually returned by a fund to LPs relative to each dollar of committed capital—covering roughly 3,000 US venture funds that began investing between 2017 and 2026.

Funds starting investment in 2017 are now about nine years old, nearing the end of a typical decade-long lifecycle. The median fund has returned only 0.37x in cash. Even top-quartile funds sit at just 0.70x. Only the top 10% have returned more than the capital LPs put in, reaching 1.37x.

Younger funds fare worse. The median 2018 vintage returned 0.15x, while the 2019 vintage median sits at 0.04x—meaning after seven years, you're getting back only 4 cents on the dollar.

The value exists on paper. The median 2017 fund holds book values of 1.72x, but only about one-fifth has been distributed as cash. The remainder remains locked up in private companies yet to be sold or taken public.

Another detail explains why this is hard to see from the outside: most of these funds have indeed returned some capital. 86% of 2017 vintages and 70.4% of 2018 vintages have distributed at least partial cash. So almost every fund can tell its LPs that distributions have begun. Rarely can they say they've returned the money.

So why should founders care about their investors' fund books?

Because the clock ticking on your investor's fund doesn't match yours. A fund nearing maturity, facing real cash obligations to LPs, and simultaneously raising its next vehicle has powerful incentives to secure liquidity quickly. This pressure can manifest as pushing for earlier exits, dumping shares on secondary markets, or pulling back on follow-on commitments during your next round.

And this is exactly where founders need to pay attention.

None of this implies your investors will act against your interests. Funds can extend their lives, and many top institutions have strong reasons to continue backing their winners. Carta's data only highlights the cash shortfall; it doesn't dictate how every fund will behave. But the vintage of the fund cutting your check is now a genuine variable, not a footnote. A partner deploying capital from a 2017 or 2018 fund faces an entirely different reality than one doing so from a freshly closed 2025 fund.

The real risk isn't that your investors need liquidity. It's that you don't know which fund your money came from, how much cash it still owes its own backers, and how that dynamic will shape their advice on your next round or exit.

Where the Biggest Money Is in AI: Frontier Models or the Mid-Market?

Most AI coverage fixates on the frontier: each lab's newest, smartest, and priciest model. But Tomasz Tunguz at Theory Ventures argues that the truly significant market is the mid-layer, and spend data backs him up.

The mid-layer now accounts for roughly 40% of AI spend and 30% of token consumption.

By contrast, frontier models capture a thin slice. Anthropic's flagship model Fable 5.1 represented only 3.7% of gateway spend in its first 12 days since launch. Its predecessor peaked at 13.2%, then fell to 4.9% a month after Opus 5 launched at half the price. Among enterprise accounts, frontier model token consumption dropped from 53% in early August to 45% in September.

Price wars are also erupting squarely in this tier. When Anthropic cut prices on its new model in September, OpenAI followed suit roughly 90 minutes later. Prior to that cut, the Opus series maintained steady pricing across four consecutive versions at $5 per million input tokens and $25 per million output tokens. At the entry level, OpenAI slashed Luna's price by 80% in July and another 50% in September.

Three forces are continuously driving mid-layer prices down:

Inter-lab competition. Labs now mirror each other's price cuts within hours rather than quarters.

Open-weight models. On data-released gateways, open models handle the bulk of token volume, priced at 86% less than the blended rate of closed-source models.

Fine-tuning. Cursor Composer 2, fine-tuned on the open-weight model Kimi K2.5, reduced total costs by 86% compared to previous in-house models. Harvey cut per-seat costs by 55% relative to Sonnet 5 while outperforming Fable 5 on quality scores.

This dynamic plays out structurally. The tasks enterprises need AI to perform—summarizing a contract or classifying a support ticket—remain nearly identical year over year. Yet the cost of intelligence capable of meeting this baseline is plummeting. Consequently, products that pass through a fixed requirement threshold become cheaper quarter after quarter, and most actual workloads land in this layer, not at the frontier.

Demand resembles less a pyramid topped by frontiers and more a bell curve with a thick middle. Buyers care about how much intelligence they get per dollar, not peak capability. The unanswered question is whether this mid-market will ultimately commoditize. If it does, the entire economic calculus of the AI market shifts.

For anyone building products on top of these models, it's worth asking: which tier does your product actually require? Paying frontier model prices for mid-tier workloads is now the fastest way to bleed margins.

How Much Should Your Startup Really Spend on AI? The Data Says This.

Ask most founders how much they spend on AI, and you'll get a highly confident figure: seats for ChatGPT and Claude, Cursor invoices, growing API fees, totaling a few thousand dollars a month. It sounds precise, but it isn't. It mashes together three fundamentally different expense categories with only the vendor label in common.

Payment data from Ramp, covering over 70,000 US companies, reveals just 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 650 times the median. When Ramp correlates these expenses with workforce records from Revelio Labs covering 21,559 companies, high-spenders grew headcount by 10.2% over two years, including a 12% increase in entry-level hiring.

Low-intensity users show absolutely no statistically significant changes.

So spend more. But where?

Most founders compare AI invoices to their software budgets, since those line items sit alongside Figma and Notion on the statement. That's the wrong denominator. Increasingly, AI spend substitutes people, not software. A customer service agent is benchmarked against a CS role, not your SaaS stack.

Run the math on a typical seed-stage startup: 12 employees, ~$180k monthly payroll, and $4k monthly AI spend. Against a software budget, $4k looks heavy. Against payroll, it's only 2.2%. Same number, completely opposite conclusions.

The fix is splitting the AI budget into three buckets, each evaluated against its own benchmark:

Pay-line AI (CS reps, SDRs, and coding agents) is benchmarked against the salary of the person you didn't hire. If you can't define the role, it doesn't belong here.

Productivity-line AI (seats, notes, and writing tools) is judged by adoption rates. If usage decays by month two, cancel it.

Cost-line AI (in-product token and API calls) is measured by gross margin at 10x scale, and it must live under COGS on your books.

Mixing all three blurs the signal. Overspending on seats inflates the total, causing founders to ration agents, precisely where the ROI lives. The top 1% winners aren't winning because they spend more; they win because they categorize correctly: leaning hard where AI replaces headcount, staying strict where it merely accelerates humans, and locking in margins where it runs inside the product.

We've broken down the full scoring methodology, including a 30-minute audit workflow you can run on your own bills this week. See "What Benchmarks Should Your AI Spend Actually Use?" for details.

Why Has Venture Due Diligence for Deep Tech Actually Gotten Worse?

You might assume the hardest technologies warrant the strictest scrutiny. A quantum computer, a fusion reactor, or a novel chip architecture either obeys physics and functions or it doesn't, and verifying that demands true domain expertise.

But Marin Ivezic of PostQuantum argues the exact opposite holds true: the larger the deal and the higher the technical bar, the less likely independent experts are brought in to verify scientific feasibility.

This isn't due to investor laziness. Several structural shifts have converged.

The expert layer has vanished.

Historically, investors could tap bank research teams for independent technical assessments staffed by specialist analysts. The 2018 EU MiFID II regulations, which mandated ring-fencing research fees from transaction fees, caused this model to contract. European equity research scales reportedly shrank by ~20%, and headcounts of specialist analysts at major banks also declined.

Expert networks now represent a ~$2.5B industry filling part of the void. But they only work if you know which expert to consult and what questions to ask. In frontier fields, that knowledge is precisely what generalist investors lack.

Capital has flowed toward generalists.

In 2025, private quantum investments hit $4.9B, up 192%. The largest deployers are BlackRock and Nvidia, not specialized quantum funds. Sovereign wealth funds and pension schemes are increasingly participating. A survey of sovereigns found 58% lack the resources to lead a single deal. Capital unable to conduct its own technical reviews ultimately relies on others' judgments.

Lead investor status has replaced due diligence.

A reputable lead investor credit-marks the co-investors. Co-investors assume the lead vetted physical feasibility, but they never see that memo nor can verify it happened. Some rounds skip this entirely: Quantinuum's $600M round had zero named leads.

Deadlines and herding behavior prioritize speed over review.

By late 2023, US funds held a record $3.116T in dry powder. A PM facing deployment pressure gains more from joining a hot deal than delaying it. Furthermore, going wrong alongside peers damages reputation far less than being the lone misstep.

Deep tech lacks mandatory checkpoints.

Biotech serves as a sharp contrast. FDA approval pipelines force stage-gated data disclosure, prompting the industry to build entire due diligence infrastructures around it: scientific advisory boards, specialist reviewers, phased gates.

Quantum has no equivalent gates. The toughest technical assessment of large quantum firms this article could locate came from the Australian government, released via freedom of information requests, not by any investor.

The results are clearly reflected in the numbers.

Private quantum investments totaled $4.9B in 2025, while the entire quantum computing sector generated roughly $1.4B in revenue. Three quantum firms that went public via SPAC in 2021 projected combined 2025 revenue of ~$1.2B, but delivered only ~$138M. And the entities doing the most rigorous verification weren't investors—they were short sellers.

Precedents exist for how this ends. Between 2006 and 2011, VCs poured over $25B into cleantech, losing more than half. Academics studying that cycle warned then that pensions and sovereigns without hardware experience would be the next wave of capital providers.

The proposed solution is straightforward: before co-investing in a deep tech round above a certain size, LPs should demand an independent technical feasibility review that neither the fund commissions nor the founders select. Full disclosure: the author invests in quantum startups and runs a firm that provides exactly this service. But the core argument stands: in deep tech, a name on the cap table has quietly replaced technical validation, and given cheque sizes like these, a few weeks of independent review is trivial.