💥Why LPs Pass, Where to Make Money in AI, New Fund Performance Data, VC Tool Budgets & More

1. Key Themes

Theme: AI demand and pricing are consolidating in the cheap "middle" of the model market

Mid-tier models hold the budget while frontier demand thins

Tunguz argues that price competition concentrates where most business workloads actually run, and frontier adoption is weaker than the hype implies.

• "Anthropic's Fable 5.1, its most capable model, took only 3.7% of gateway spending in its first twelve days, and frontier models fell from 53% to 45% of large corporate token consumption between early August and September."
• "A chart in the piece shows the mid tier claiming 40% of spend and 30% of tokens, which Tunguz describes as demand shaped like a bell curve with a fat middle."

Open and fine-tuned models are collapsing prices underneath

Customers are actively substituting cheaper open-weight models, sometimes beating frontier quality.

• "Open models now run most gateway token volume at an 86% discount to closed models, and large customers fine-tune open-weight models to cut costs further, e.g. Harvey cut cost per cell 55% against Sonnet 5 while scoring above Fable 5."

Theme: VC fund performance is deteriorating, and newer vintages are starting from a weaker base

Late-2010s vintages are giving back their 2021 gains

• "Median net IRR for 2017 funds stood at 22.0% around 18 quarters in and has since fallen to 9.3%, while 2018 funds dropped from 22.3% to 6.3%."
• "The 2017, 2018 and 2019 vintages reached their highest median IRRs roughly 18, 14 and 10 quarters in, which for each lands around 2021, the height of the valuation boom."

2021-2022 vintages never got the boost

• "At the same age, 2021 funds sit at a 1.6% median IRR after 18 quarters and 2022 funds at 4.4% after 14, against 22.0% and 22.3% for 2017 and 2018 funds."
• Caveat from the author: "The slide is not a final verdict: a fund's IRR keeps moving with every new valuation and exit, so these vintages can still recover before they wind down."

Theme: Emerging managers outperform but struggle to raise

Outperformance exists, yet institutional capital waits for proof

• "Across 2000-2022 vintages, Preqin data shows first-time VC funds beating established managers in many cohorts, with a median net IRR premium of up to 15% in certain vintages."
• "Diligence and monitoring cost an LP about the same whatever the check size, and proof now means DPI, disciplined portfolio construction and institutional-grade reporting, which a first-time fund with unrealized holdings rarely shows."

Operator backgrounds predict top investor performance

• "Of nearly 50 investors who appeared on the Midas List seven or more times between 2016 and 2025, almost 70% had founder or operator experience, and only about 20% rose purely through career VC."

Theme: AI-enabled services roll-ups as a margin-arbitrage play

Buy low-multiple services firms and rebuild delivery with agents

• "Traditional services firms run roughly 5-10% EBITDA margins, and the roll-up thesis holds that agent-rebuilt delivery can reach 30-40% on the same clients and invoices, lifting profit 3-4x."
• "Buyers pay low multiples for service businesses because they assume the margins can't improve, Isenberg argues, so raising margins makes the business worth more. General Catalyst has put more than $750M of a $1.5B allocation behind this strategy."

Theme: VC firms' data/AI tooling budgets are catching up to engineering payroll

Tooling spend reaching parity with engineering headcount

• "The ratio has moved from roughly 2:1 in 2025 to near parity in 2026."
• "Funds under $100M report an estimated average of $45K on data, tools & tokens against $40K on engineering staff, making them the only cohort where tooling outspends people."
• "$1B+ funds report $341K on engineering HR against $247K on tooling, roughly 1.4:1."

2. Contrarian Perspectives

Frontier models are not where the money is. The consensus assumes the most capable model captures the value, but the data shows demand fattening in the middle and moving away from the top.

• "Frontier models fell from 53% to 45% of large corporate token consumption between early August and September."
• "If demand keeps shifting toward cheaper models, as Tunguz suggests it may, the same work keeps getting cheaper to run."

Founders should rarely ask for supervoting shares. Control structures are often treated as a founder-friendly must-have, but Strebulaev argues they carry signaling and valuation costs and deliver less than assumed.

• "Strebulaev advises founders to rarely ask for supervoting, since it signals an expected shareholder conflict and can cost valuation."
• "Supervoting delivers no board seats, cannot override protective provisions, and is irrelevant in any vote requiring a majority of preferred, so it mainly shields founders from being outvoted by common holders."

The "safe" LP approach to emerging managers is actually costly. Waiting for proof looks prudent but forfeits access.

• "Team8 argues that waiting has a price: by the time a new manager looks safe to back, its fund is often already full."
• "First-time VC funds beating established managers in many cohorts, with a median net IRR premium of up to 15% in certain vintages."

3. Companies Identified

Anthropic

• Description: AI lab; maker of the Fable and Sonnet model families referenced in the piece.
• Why mentioned: Its newest frontier model illustrates weak frontier uptake.
• Quote: "Anthropic's Fable 5.1, its most capable model, took only 3.7% of gateway spending in its first twelve days."

Harvey

• Description: Legal AI company.
• Why mentioned: Case study in fine-tuning open-weight models to cut cost while beating a frontier model.
• Quote: "Harvey cut cost per cell 55% against Sonnet 5 while scoring above Fable 5."

Figma

• Description: Design software company.
• Why mentioned: Example of supervoting separating control from ownership before IPO.
• Quote: "Before Figma's IPO, Dylan Field owned roughly 9% of the equity yet held just over half the voting power, through 15-vote Class B shares and a proxy from his departed co-founder."

General Catalyst

• Description: Venture capital firm.
• Why mentioned: The largest named backer of the AI roll-up strategy, with a deal-structure tactic worth copying.
• Quotes: "General Catalyst has put more than $750M of a $1.5B allocation behind this strategy." and "Paying part of the price in shares, as General Catalyst does with about 30%, gives the seller a reason to stay and hand over their client relationships."

Carta

• Description: Cap table and fund administration platform.
• Why mentioned: Source of the median net IRR fund performance data.
• Quote: "Carta's Hamza Shad shared the latest fund performance data, which tracks median net IRR by fund age for every vintage since 2017."

Team8

• Description: Venture/company-building firm.
• Why mentioned: Published the emerging manager report.
• Quote: "Team8 partner Aaron Dubin published The Emerging Manager Paradox, a report asking why first-time funds attract so little institutional capital despite a long record of outperformance."

Preqin

• Description: Alternative assets data provider.
• Why mentioned: Data source showing first-time fund outperformance.
• Quote: "Preqin data shows first-time VC funds beating established managers in many cohorts."

Standard Metrics

• Description: AI-driven portfolio management/data platform (newsletter sponsor).
• Why mentioned: Sponsor pitching portfolio data infrastructure for AI tools.
• Quote: "Standard Metrics builds that foundation for 150+ investment firms. We collect, extract, and verify portfolio data, then connect it to the AI tools your team already uses through MCP."

HSBC

• Description: Global bank.
• Why mentioned: Partner for the DDVC Builders Workshop hackathon.
• Quote: "DDVC Builders Workshop in London Oct 7 - join our first hands-on builders hackathon with HSBC."

4. People Identified

Tomasz Tunguz

• Description: Investor and writer (Theory Ventures).
• Why mentioned: Author of the analysis on AI price competition in the middle tier.
• Quote: "Tomasz Tunguz argues in The Most Important Market in AI is the Middle that AI price competition is concentrated in the mid tier, where most multi-step business workloads run."

Ilya Strebulaev

• Description: Stanford professor.
• Why mentioned: Explains the mechanics and limits of supervoting.
• Quote: "Stanford's Ilya Strebulaev explains in Ten Votes to Your One how supervoting lets founders control companies they barely own, and where that control stops."

Dylan Field

• Description: Figma founder.
• Why mentioned: Case study of control with minimal ownership.
• Quote: "Dylan Field owned roughly 9% of the equity yet held just over half the voting power."

Hamza Shad

• Description: Carta analyst.
• Why mentioned: Shared the fund performance data.
• Quote: "Carta's Hamza Shad shared the latest fund performance data."

Aaron Dubin

• Description: Partner at Team8.
• Why mentioned: Author of The Emerging Manager Paradox.
• Quote: "Team8 partner Aaron Dubin published The Emerging Manager Paradox."

Greg Isenberg

• Description: Entrepreneur and content creator.
• Why mentioned: Author of the AI roll-up playbook.
• Quote: "Greg Isenberg published a playbook on AI roll-ups, arguing that buying services firms from retiring owners and rebuilding delivery with AI agents can roughly triple a firm's profits."

Andre Retterath

• Description: Author of Data Driven VC.
• Why mentioned: Publisher of the newsletter and its landscape report.
• Quote: "Our 2026 DDVC Landscape Report covered how firms split budget between engineering staff and data, tools & tokens."

5. Operating Insights

Default to mid-tier or open models for high-volume work. Reserve frontier models for the steps where extra quality clearly pays off.

• "For VC firms building their own AI workflows, that means defaulting to mid-tier or open models for high-volume tasks like deal screening and research summaries, and reserving frontier models for the few steps where the extra quality clearly pays off."

Treat data, tools and tokens as their own budget line with ROI tracking.

• "Spend on data, tools & tokens is now close to what they pay the engineers who use them, so it needs its own budget line and ROI tracking."

Structure roll-ups so sellers stay invested, and integrate before you scale.

• "The biggest risk Isenberg flags is buying firms faster than you can integrate them. Owners often prefer selling to a person over a fund, and paying part of the price in shares, as General Catalyst does with about 30%, gives the seller a reason to stay and hand over their client relationships."

6. Overlooked Insights

Proxies and sunset clauses are the practical alternatives to supervoting. Founders have cheaper control tools, and investors are increasingly adding limits.

• "Voting proxies from departing co-founders are the cheaper route, and investors increasingly add sunset clauses that remove extra votes after a set period or the founder's exit."

LP diligence costs are fixed regardless of check size, which structurally disadvantages small first-time funds. Team8's suggested workarounds are notable.

• "It suggests standardized fund admin, spreading bets across several new managers, and investing before a fund formally launches so performance signals arrive sooner."