Summary

Dieffenbach argues venture only compensates LP risk when an allocator can consistently access top-decile managers; below that, even top-quartile returns do not beat the public-market alternative. Mature venture vintages delivered roughly 8% median net IRR; even top-quartile performance was about 15% IRR, 2.5x TVPI and only 1.8x DPI over 15 years. Against QQQ as Carnegie Mellon’s PME, his answer to whether LPs are compensated for venture risk is blunt: “Absolutely not.”
Carnegie Mellon runs its $4 billion endowment as an 85% equity, 15% fixed-income portfolio, with half the total in private assets and just under 25% in venture. That makes CMU roughly 5-10 percentage points overweight venture versus comparable endowments, but its overall private book remained self-funding through the liquidity drought, with buyout distributions contributing most. Venture itself finally became self-funding again this year, for the first time since 2021.
The hidden constraint on mega-funds is ownership: a $7 billion platform averaging 5% entry ownership effectively must create $140 billion of enterprise value before multiplying LP capital. To reach CMU’s 4x-net target after fees, Dieffenbach estimates it needs at least 6x gross, implying nearly $800 billion of exits—close to the entire roughly $850 billion exit value of 2021. Harry Stebbings argues future companies could be much larger; Dieffenbach concedes that possibility but asks, “What is the margin of safety?”
Seed investing remains a picking game where the founder or idea is genuinely non-consensus; consensus seed has become an access war against multi-stage capital. Stebbings argues a $50 million-$100 million fund cannot both buy meaningful ownership in today’s $4 million-$5 million seed rounds and diversify across roughly 30 companies. Dieffenbach’s counter is that Airbnb, Uber, SpaceX and Amazon all initially struggled to raise, preserving a potential moat around investors who can recognize what looks “crazy” before consensus forms.
Scale turns venture franchises into exceptional fee businesses long before it proves they remain exceptional investments. A platform with $15 billion across stacked funds might collect roughly $300 million annually in fees, while a generational $20 billion outcome can barely move a $7 billion vehicle. Dieffenbach would preserve premium economics for true early-stage investing but move scaled growth funds toward long-only fees—roughly “one and 10,” or even budget-based management fees plus 10% carry.
Illiquidity is primarily a pricing problem, not evidence that the IPO market is closed. From 2022 through 2024, IPOs raised fewer dollars than in 2002-2004 despite venture being roughly 10 times larger; meanwhile, private sellers still wanted multiples unsupported by public alternatives such as Microsoft. Public markets are now rewarding selected growth companies again, prompting Dieffenbach’s direct appeal: “Now is the time. Please take your companies public.”
AI can become a foundational technology while still producing a capital-destroying bubble, and OpenAI’s dependence on new funding is the key distinction. Dieffenbach says its unit economics are improving, but a company burning $5 billion-$10 billion annually with a perhaps $70 billion-$80 billion preference stack lacks the control of destiny enjoyed by profitable Google, Meta or self-funding SpaceX. Nvidia is not necessarily overpriced today, yet a cyclical downturn combining lower revenue, compressed earnings and a lower multiple could, in his scenario, produce a roughly 70% drawdown.

Deep dive

1. Adversity made risk—not asset labels—the starting point

At 26, Dieffenbach learned that his lymphoma had progressed enough to require chemotherapy within a week. After “sulking” for roughly 12 hours, he adopted a coach’s maxim—“success in life is 10% what happens to you and 90% how you react”—and decided that “cancer can’t kill me if I don’t stop moving.” Four months later, he was cancer-free.
The lasting change was perspective rather than invulnerability. He calls adversity “beauty in the struggle,” says it builds a stronger person, and would not reverse the experience: “Life is an incredible joy and a blessing,” leaving few professional setbacks capable of taking him down mentally.
Carnegie Mellon applies a similarly explicit risk-first hierarchy to its $4 billion endowment: 85% equity and 15% fixed income, with 50% of the total in private assets and 50% in hedge funds and liquid securities. Within privates—venture, buyout, real estate, natural resources and private credit—the team runs a “best athlete portfolio,” allocating wherever it sees the strongest risk-adjusted return.

2. Venture only pays when access reaches the top decile

CMU’s private portfolio has been self-funding for three years, with buyouts contributing the most distributions and venture detracting the most. Private exposure has therefore stayed near 50% for six or seven years, although venture NAV rose as distributions slowed and was partly offset by markdowns.
Venture represents just under 25% of the entire endowment, nearly half of CMU’s private book and roughly 5-10 points more than comparable institutions. The endowment offsets that overweight by holding less in hedge funds and real assets.
Dieffenbach’s risk comparison is concrete: an industrial property leased to Amazon with rents rising 3% annually has replacement value and stable cash flow; a $100 million venture fund backing “two or three people and an idea” is probably the riskiest asset available. Asked whether LPs receive enough compensation, he answers, “Absolutely not.”
Across mature vintages from roughly 1998 through 2015 or 2016, he cites median net IRR near 8%, top-quartile IRR around 15%, top-quartile TVPI near 2.5x and 15-year top-quartile DPI of only 1.8x. CMU compares venture with QQQ, and only top-decile managers consistently clear that PME. His threshold for a new allocator: “Do you think you’re going to have access to top-decile managers?”

3. Seed economics turn “small and nimble” into a narrow target

Stebbings challenges LP enthusiasm for $50 million-$100 million seed funds. With average rounds of $4 million-$5 million, meaningful ownership may require a $3 million-$3.5 million check; doing that across roughly 30 companies approaches a $90 million portfolio before fees. Smaller funds therefore accept weaker ownership, weaker diversification or awkward $1.5 million checks that are hard to win when elite investors want the allocation.
Dieffenbach’s counter is that consensus founders and ideas are extremely difficult because multi-stage firms can deploy $5 million-$10 million at seed with a cheaper capital base, treating specialist seed funds as “shrapnel.” His remaining opening is non-consensus investing, where rounds are less competitive and price and ownership improve. Stebbings pushes back that even non-AI companies now command rich prices; Dieffenbach concedes, “I hope and pray” capital has not eliminated the picking moat.
The weighting depends on strategy: for scaled multi-stage firms, Dieffenbach estimates 70% access and 30% picking; for small, nimble early-stage funds, he flips that to 70% picking and 30% access. CMU’s practical range begins around an $80 million fund and extends to $400 million-$1 billion, typically with commitments starting near $10 million, but the deciding variables remain people, prior proof and strategy fit.

4. Selling and honest marks now separate managers from storytellers

Dieffenbach defines five venture muscles: sourcing, picking, winning, helping and selling. Selling is the newest institutional capability; he holds up Union Square—where CMU wishes it were an LP—as unusually disciplined about becoming an active seller during years eight through 12.
CMU prefers cash distributions over stock because different LP sale times can create a 1%-2% pricing discrepancy. A manager can sell the entire position immediately and distribute identical cash economics. Most firms failed this test in 2021-2022, although Dieffenbach explains the temptation: median software ARR multiples reached 20x and top-quartile growers reached 40x, making another 2x-3x appear defensible before the market changed abruptly.
Perennial multi-stage firms are among the most conservative markers, often carrying securities at 20%-30% discounts even when secondary indications are higher. Since 2021, CMU has independently underwritten every prospective manager’s top 10 company NAVs using revenue, gross-profit and free-cash-flow trends, classifying each position as overvalued, fair or undervalued.
Hubris ends meetings quickly. Managers who call their strategy easy or describe performance as “shooting fish in a barrel” ignore how rare even a 6x net fund is. One manager underwritten in 2023 still carried OpenSea at $13 billion; challenged by CMU, it promised to revise both the mark and its valuation policy.

5. Off-sheet references reveal the people behind the track record

For established Sand Hill Road and London franchises, Dieffenbach sees little systematic sourcing beyond powerful partners, brand and proximity to S-tier founders; being a mandatory meeting is itself the moat. Esoteric geographies or bootstrapped markets may support automated sourcing, but mainstream venture often contains “a lot of luck”: hustle, introductions and taking enough meetings. Ali Partovi is one of the rare sourcing standouts he names.
Picking requires reconstructing what the investor believed before the outcome became obvious. Dieffenbach recalls thinking Uber and Airbnb sounded absurd—riding with or sleeping beside strangers—while investors such as Mike Maples and Cyan Banister could “see into the future.” CMU asks founders who ignored them, who believed first and why they selected that particular partner.
A new fund normally triggers at least 20 reference calls, only five supplied by the GP. The “golden references” are off-sheet and focus less on strategy opinions than interpersonal risk, partnership dynamics and accurate deal attribution. CMU constructs its own partner-level attribution tables because, after a deal leader retires or leaves, firms may reassign that win to someone still fundraising.
Partnerships most often fracture over incentives and perceptions of who works hardest. Dieffenbach saw more personnel change in the prior two years than in his preceding eight years as an LP: wealthy partners tired of broken cap tables and absent liquidity, while younger partners watched expected carry evaporate and compensation fall perhaps 70%. His preferred “founder friendliness” is hard coaching from a loving, aligned perspective—not avoiding difficult conversations.

6. A venture commitment is a 25-year alignment decision

CMU often waits: about half its new commitments happen in the first observed fund, while the rest follow one or two funds—three to six years of relationship-building. An early-stage fund may require 15 years to wind down, perhaps 18, and CMU intends to back at least three vintages. Dieffenbach therefore describes the decision as a roughly “25-year illiquid relationship,” potentially twice the average US marriage.
GPs should diversify their LP base across endowments, foundations, family offices, founders and perhaps other venture firms. Best case, no investor exceeds 10%; Dieffenbach becomes uncomfortable above 30%, though a strategically aligned $10 million-$30 million anchor can still make sense for a $50 million-$100 million fund.
Selling part of the management company is a “massive red flag” because carried interest—the partnership’s motivational engine—moves to a silent owner who is not grinding beside the team. GP commitment matters too, but CMU evaluates what the amount means to each person rather than comparing nominal dollars; Dieffenbach calls it one of the team’s best forward-looking quantitative indicators.

7. A $7 billion platform can require an entire exit year

Dieffenbach traces mega-fund expansion to SoftBank’s first Vision Fund, after which established venture firms scaled far beyond the roughly $400 million Series A funds, sometimes paired with similarly sized growth vehicles, that had persisted for a decade. Assuming 2010-2017 returns survive that capital expansion “worries us tremendously.”
His worked example is an unnamed $7 billion platform containing a $1 billion early-stage fund, a $2 billion-$3 billion growth fund and a larger opportunity vehicle. Entry ownership fell from about 15% in early stage to 6%-7% in growth and 2.5%-3% in opportunity. Because LPs invest proportionally across the stack, their dollar-weighted ownership was only about 5%.
Dividing $7 billion by 5% implies the manager must invest into companies eventually worth $140 billion merely to deploy the fund’s ownership base. CMU targets 4x net; after 2.5-and-30 early-stage fees and 2-and-20 growth economics, Dieffenbach estimates at least 6x gross is necessary. That means close to $800 billion of exits, versus roughly $850 billion across the record 2021 exit year: “You need an entire year of IPOs and M&A just for this one manager.”
Stebbings’ counterargument is that outcome sizes may compound dramatically—Microsoft could reach $10 trillion, while OpenAI, Anthropic and SpaceX could list near or above $1 trillion. Dieffenbach admits “we could be wrong,” but notes there have been 11 venture-backed $50 billion IPOs, with the two largest being Facebook in 2012 and Alibaba in 2014. He would rather underwrite a fund needing $10 billion-$30 billion of outcomes and retain upside beyond that.

8. Index shows scale can work, but fee math breaks alignment

Index is Dieffenbach’s outstanding scaled exception. He cites its major ownership in Figma, Dream Games and Wiz, plus positions in Scale AI and Revolut, while praising its decision to reduce fund size after 2021 despite unlimited fundraising capacity. “They are the most performance-driven culture that we see.”
Stebbings questions whether a $1 billion-$2 billion platform is trapped between specialist funds and General Catalyst-, Lightspeed- or SoftBank-scale capital. Dieffenbach argues Index still has enough money for unusually large seed through Series B checks, but not so much that individual wins become irrelevant. Its brand also attracts repeat generational founders even when cheaper capital is available.
The arithmetic becomes punishing beyond that point. Ten percent of a $20 billion-$25 billion Figma-like company produces about $2 billion before carry—only around 0.2x for a $7 billion fund, requiring perhaps 15 Figmas. Wiz’s roughly $30 billion-$31 billion outcome returned only about one-third of Insight’s fund, illustrating how “the GDP of a country” can become merely helpful inside an oversized vehicle.
Dieffenbach does not blame GPs: stacked funds totaling $15 billion can generate roughly $300 million in annual fees, among “the best high-margin businesses ever created.” But growth investing in established, fully staffed companies resembles passive long-only public equity. He would reserve 2.5-and-20—or 2.5-and-30 for exceptional franchises—for core early-stage work, while scaled growth moves toward one-and-10 or budget-based fees plus 10% carry.

9. The IPO market is priced, not closed

Deployment speed should match what the GP sold. A declared two-year cycle followed by a fund after two years is acceptable; a promised three-to-four-year period compressed into two demands an explanation because vintage diversification matters. Moving slower is not inherently bad: Dieffenbach credits Mark Suster for recognizing 2021’s excess and selling much of his portfolio, putting DPI into LPs’ pockets.
US venture fundraising was tracking toward its lowest year since 2017, though Dieffenbach qualified that it might reach back to 2016. The principal cause is liquidity: IPOs raised more dollars during 2002-2004 than during 2022-2024 even though the asset class had become roughly 10 times larger. After the dot-com peak, QQQ required 13 years to regain par—yet the subsequent three years still generated more IPO capital than the latest drought.
Dieffenbach rejects the phrase “IPO markets are closed”; price is the clearing mechanism. A $100 million-ARR SaaS company growing 15% at breakeven cannot demand eight-to-10-times ARR when investors can buy Microsoft growing revenue 14% and earnings 17%, with GAAP profit, a dominant moat and annual repurchases of roughly 1% of its shares. LP frustration comes from watching public technology compound while private marks resist that comparison.

10. Circle shows why venture tails punish secondary sellers

Dieffenbach views Harvard’s reported $1 billion sale against a roughly $50 billion endowment as a portfolio refresh, not a capitulation. A reported Yale-CalPERS transaction around a 10% discount looked exceptionally attractive; without seeing its GP and asset mix, he would have guessed nearer 20% because even Yale’s strong portfolio must clear a supply-constrained secondary market.
CMU itself nearly missed the lesson. A 2012 venture fund had one residual asset after 13 years, worth less than CMU’s $1 million internal tracking threshold and carried around a 30% discount to Circle’s last roughly $5 billion round. Dieffenbach discovered it while reading the S-1 and recognizing the GP on the cap table; with Circle later around $50 billion, that single tail position could add approximately three turns to an otherwise realized fund.
In the Yale transaction discussed, CalPERS reportedly bought about $500 million of exposure and got roughly a $100 million write-up from Circle within two months. Dieffenbach admits CMU might also have sold after underwriting the stale position: nobody could confidently predict stablecoins becoming the hottest crypto segment or Circle trading around 100 times EBITDA. Venture’s right tail can arrive in years eight through 13, precisely when sellers assume “there’s not a lot of juice left to squeeze.”
CMU’s venture book became self-funding this year for the first time since 2021, but announced liquidity remains delayed: Wiz awaited regulatory approval, Figma had not yet listed and Dream Games required European clearance. Dieffenbach expects 2026 to deliver more cash and help fundraising, but “one year is not gonna solve the industry’s problem.” His message to managers is immediate: “Please take your companies public.”

11. China and AI-era M&A both carry hidden alignment costs

CMU’s best historical investment was a China fund returning more than 20x net, yet the bar is now extremely high. Dieffenbach cites US restrictions on investing in Chinese AI, semiconductor and defense companies, alongside a structural conflict: managers traditionally raised pari-passu USD and RMB funds, but the two pools can no longer access the same opportunities.
With AI representing roughly 70% of US venture deals in his cited comparison, exclusion from that category can radically change what a dollar-denominated China fund owns. Local-government RMB vehicles may receive the assets USD LPs cannot, creating an alignment problem; meanwhile, many strong Chinese founders have chosen the US, Singapore or London.
Google, Microsoft, Amazon and Meta collectively generate about $600 billion in annual operating cash flow and may prefer strategic acquisitions to marginal buybacks. Yet a 12-month review can make a fast-changing AI target obsolete; Wiz’s 10% breakup fee—the largest cited for an M&A transaction—shows the financial risk. That encourages talent hires and IP licensing that deliver people and technology immediately, though Stebbings notes this workaround cannot replace the revenue and customers that make Wiz worth roughly $31 billion.

12. AI may transform GDP and still destroy today’s capital stack

OpenAI’s unit economics are improving rapidly, but Dieffenbach asks why it raised two of history’s largest venture rounds within 12 months: “It’s ’cause they’re burning $5 to $10 billion a year.” If AI’s financing cycle turns while OpenAI carries what he frames as a $70 billion-$80 billion preference stack, dependence on another enormous equity check leaves it without control of its destiny.
SpaceX is the counterexample: Starlink has reached “escape velocity,” the company is self-funding and secondary tenders do not finance operations. Google and Meta similarly entered public markets with roughly 30%-40% GAAP operating margins. Dieffenbach therefore refuses to call OpenAI or Anthropic slam-dunk trillion-dollar independent companies five years out; the distinction is survivability when capital markets stop cooperating.
He can imagine material GDP impact over 10 years, but not confidently within three to five. OpenAI exhausting its GPUs because users were generating cartoon images illustrates the gap between adoption and productivity. Hyperscalers may deploy roughly $1 trillion of CapEx from 2024 through 2027; Dieffenbach posits a roughly $100 billion US venture run rate, with perhaps 80% directed toward AI. If economic payoff takes a decade, “there will be a lot of pain.”
Dieffenbach does not call Nvidia overpriced for today’s business, but emphasizes “peak earnings and peak multiple.” In a cyclical downturn, revenue might fall 20%-30%, earnings perhaps 40%, and a roughly 38x forward multiple could contract toward a historical trough near 24x—producing an approximately 70% drawdown. He does not predict its timing; he rejects treating the scenario as impossible simply because the long-run AI thesis is right.

Dwarkesh Patel and Noah Smith on AGI and the Economy Zilliz CEO 星爵: Making Open Source Pay, Vector DBs, and AI Idealism