Episode Summary
Executive Summary: The episode centers on Y Combinator’s smaller, AI-heavy winter batch and the broader venture environment: AI is shifting from wrappers to practical vertical products and agents, but the market is still in a “sampling” phase. The panel also debates rising startup valuations, compressed exit markets, reserve strategy, fund concentration, LP/GP selection, and the survival of only the strongest managers and companies.
Main Topics: YC Winter Batch and the Rise of Practical AI (Priority: 5/5): The hosts discuss YC’s smaller class, sub-1% acceptance rate, and the emergence of AI companies focused on real workflows rather than generic chat interfaces. One breakout mentioned is Leia, an AI assistant for lawyers. Co-pilots vs Agents in AI (Priority: 5/5): Tomash Tangis distinguishes AI co-pilots that boost productivity from agents that perform tasks on behalf of users, arguing that agents can meaningfully replace workflows in support, sales, and security operations. Verticalization, Commoditization, and the AI Stack (Priority: 5/5): The panel argues startups cannot compete head-on with giant model providers and hardware incumbents, so value is likely to accrue in vertical applications, workflow-specific data, and distribution advantages rather than base-model ownership. Startup Valuations and Exit-Market Pressure (Priority: 5/5): Using CARTA Q1 2024 data, the speakers debate whether higher entry prices and depressed exit markets will compress VC returns, especially given closed IPO windows and regulatory constraints on M&A. Portfolio Construction, Reserves, and Fund Math (Priority: 4/5): The discussion covers concentrated vs broad venture strategies, reserve allocation, and how rising entry valuations force larger funds or lower ownership to achieve meaningful fund-level returns. Venture Firm Survivorship and Manager Selection (Priority: 4/5): LPs and GPs describe a thinning market for venture firms, with many first-time managers unlikely to raise sophomore funds. Selection emphasizes teams, durability, and prior hits among top 1% outcomes. Personal Psychology of Venture Investing (Priority: 3/5): Several speakers frame venture as a high-loss, probability-driven game requiring emotional resilience, patience, and comfort with being wrong far more often than right.
Key Arguments: AI value is likely to emerge first in practical, vertical use cases because startups cannot easily outcompete Microsoft, Google, or dominant model vendors in general-purpose software. Co-pilots have already shown large productivity gains, but agents may create deeper labor substitution by executing end-to-end tasks like customer support or security analysis. The AI market is moving quickly enough that today’s model advantage can disappear, so winning startups must adapt to changing architectures and costs. High startup valuations combined with weak exit markets may squeeze venture returns unless firms are disciplined about ownership, reserves, and deployment pacing. Venture firms need strong reserve discipline because follow-on capital can determine whether a fund reaches top-quartile or top-decile performance. The market is overpopulated with VC funds; survival will favor firms with durable teams, clear differentiation, and a track record of backing top 1% companies. At the earliest stages, founder quality, customer intimacy, and product velocity matter more than thematic enthusiasm because most bets will fail. LPs should judge managers by whether they repeatedly accessed and supported breakout companies, not just by a single good or bad fund. There is a strong psychological burden in venture because even correct theses can fail if the wrong company is chosen, making judgment and resilience essential.
Data Points: YC applicants: 27,000 - Gary Tan said the winter YC class came from a pool of 27,000 applicants. YC acceptance rate: <1% - The winter YC batch was described as having an acceptance rate of less than 1%. YC class size: 260 - Participants estimated the class had around 260 companies. YC annual output: ~500 startups/year - The panel noted YC still appears to do roughly 500 companies per year. Productivity gain from co-pilots: 50% to 75% - Tomash cited reported productivity improvements from Microsoft and ServiceNow. GPT-4 cost: about $60 million / ~1M tokens - Tomash used this to illustrate how quickly model economics can shift. Earlier model cost reduction: 120x less expensive - He said the prior model version was roughly 120 times cheaper. Customer acquisition cost increase: 60% over five years - Cited as a broad software market trend reinforcing the value of distribution. Venture AUM growth: $8B to ~$300B - Tomash described venture’s growth during the zero-rate era. Forward public software multiples: 5x to 15x - Used to explain the valuation expansion during the last decade. Rate environment: 5% to 5.5% - Discussed as a drag on exit valuations. Top acquirer share: ~70% of dollars / ~75% of market cap - He argued major strategics are effectively sidelined by regulation in M&A. Hot AI company valuation example: $250K ARR raising $50M at >$250M valuation - Illustrated frothy pricing in AI. Typical VC fund duration to exit: 7-8 years to sell; ~12 years to IPO - Tomash framed current valuation risk in light of long realization timelines. Typical concentrated fund size: 15 shots on goal - Tomash said Theory Ventures will likely make about 15 investments. Fund size: $238M - David Clark described the size of the fund he raised in March 2023. Fundraising timeline: just under 4 to 4.5 months - David Clark said the fund was raised in a relatively short period. Managers in market: 6,500 VC funds - David Clark cited a Preqin-linked statistic on firms seeking capital. First-time managers unlikely to raise fund II: 247+ - PitchBook estimate referenced by David Clark. Common reserve split: 50/50, 60/40, 40/60 - The panel discussed historic reserve allocation patterns. Growth fund performance example: 15x - David Clark said the best-performing fund they invested in was a growth fund with this return profile. Personal investment cadence: 500 first meetings / 1,000 meetings - Jason described his internal progression for researchers to become analysts. Seed-stage check range: $500K to $2M - Tomash described Theory’s typical earlier-stage check sizes.
Pivotal Quotes: "The founders are on the front line. Their talk, you know, it's kind of like if you think of it like a war, you know, there's people on the front." — Jason Calacanis: On why operators have the best information about what’s actually working in AI and why venture is highly dependent on founder feedback. "The thing that really matters, and what's what I think will sustain is distribution." — Tomash Tangis: Explaining why AI startups need technology advantages that translate into customer acquisition advantages. "We know ventures are powerlow industry. We know it's the top 1% of companies that deliver the majority of the exit value." — David Clark: Justifying disciplined LP selection and concentration on managers with repeated breakout hits.
Implications: AI will likely reward vertical, data-rich products and disciplined distribution more than generic model wrappers. Venture returns may tighten as valuations rise and exits stay constrained, forcing better reserves, concentration, and manager selection.
From the Transcript
And just looking at some of the early bets they've made in this space, you know, we're hopeful they'll be able to do that again. Yeah, that resonates with me a lot, which is, you know, it's kind of our jobs as GPs to figure out where the opportunity is and to pick the best teams to pursue those opportunities. But at the end of the day, the founders are on the front line. Their talk, you know, it's kind of like if you think of it like a war, you know, there's people on the front. And when you're on the front and you're seeing what's happening, hand-to-hand combat, how people are using it, they have really good. Good information there, they're going to be able to pivot and find the opportunities. And, you know, we're kind of like generals moving back. And then there's like the LP class, which is enabling the generals to then enable the troops. And there's some metaphor here that works, but we're abstracted a little bit and distanced from the work. And there's advantages to be distance, right? You know, Dave can look at Thomas, myself, and others, and kind of think about how we're thinking about how the front line is engaging.
Transformer, which is a key innovation in the current wave of AI, is no longer really the right thing, should be willing to abandon it for the next version of AI, whether it's like multimodal or any of these other techniques. The thing that really matters, and what's what I think will sustain is distribution. That's what we realized in the software wave: acquiring a customer is actually more expensive each year. Over the last five years, there's been a 60% increase in the cost to acquire a customer across software companies broadly. And so the founder. That we really get excited about are the ones who look at the technology and say, I can build a product. And as a result of building that product in a very particular way, I have a customer acquisition advantage. And so I can acquire those customers at less than my competition can. Ergo, I can grow much faster on less capital. And so that combination of finding a technology advantage that is produced as a go-to-market benefit is that's what we're really after.
That's where psychologically you're comfortable, but where you're still able to perform at a high level is really important. And for us, it just comes down to: you know, we know ventures are powerlow industry. We know it's the top 1% of companies that deliver the majority of the exit value. And so, you know, we have a very simple screen, which is every time we look at a manager, we ask ourselves, you know, how many top 1% companies have they backed? What was their role in doing those deals? You know, were they? Invited in? Were they a bystander or were they actively leading that deal? You know, have they been able to do that multiple times? And the more times that we see them invest in those top 1% companies, the more confidence we have that they'll be able to repeat that. And David, your model allows for a bad vintage from time to time or a sub-average, knowing that it's probabilistic in nature. Yeah, absolutely. You know, you're always going to have good managers who, as you said before, they had the choice over company A, company B, they got the
About This Week in Startups
Jason Calacanis covers startups, tech, markets, media, and all the hottest topics in business and technology. He also interviews the world’s greatest founders, operators, investors, and innovators.