Overview of 20VC
Harry Stebbings, Rory O’Driscoll, and Jason Lemkin unpack a fast-moving week in tech focused on the AI model wars, the economics of compute, regulatory pressure between the US and China, and the implications of shrinking exit markets for private equity and venture. The core thesis throughout: we’re moving from a human-led software market to an agent-led one, and that changes which products, vendors, and business models survive.
Biggest Takeaways
-
The AI stack is shifting from “models humans like” to “models agents choose.”
- In a world where autonomous agents make tool/vendor decisions, the winning models are likely the ones embedded in the agentic layer, not just the ones with the best human sentiment.
- OpenAI is seen as regaining momentum, especially with stronger coding performance in GPT-5.5/Codex.
- Anthropic still looks strong, but the discussion suggests its human-loved coding advantage may matter less once agents are the primary users.
-
Compute is no longer just a cost center; it’s a strategic bet with huge capital risk.
- The group argues Sam Altman’s “compute = revenue” framing is directionally true but too simplistic.
- Real success requires both:
- enough compute to serve demand, and
- a model good enough to create demand.
- Anthropic’s compute shortage is framed as the result of overwhelming success, while OpenAI’s issue last year was the opposite: compute strength but weaker models.
-
Private equity’s classic playbook is under pressure.
- Highly levered 2021-era software buyouts are being repriced by AI-driven terminal value concerns.
- Medallia is held up as a warning: not just over-levered, but overpaid for in a category whose relevance is eroding.
- The broader message: if a company has no durable AI story, debt becomes much harder to service.
AI, Anthropic, OpenAI, and the Agentic Future
OpenAI vs. Anthropic
- The panel views OpenAI’s recent coding/product improvements as a meaningful comeback.
- Anthropic is described as having had strong market share gains in coding, but that advantage may be less durable when AI agents themselves decide which model to use.
- Their emerging view:
- Humans may prefer Claude for certain workflows.
- Agents may prefer OpenAI because it is highly capable and easier to operationalize across tasks.
The “agent picks the vendor” thesis
- This was one of the central ideas of the episode:
- Today, humans choose tools.
- Soon, AI agents will choose the tools and models.
- That means many software products become vulnerable if an agent can do the same job directly.
- Products that are mostly workflow wrappers for humans may be bypassed entirely.
Products at risk vs. products with staying power
- At risk / likely to compress:
- Jira
- Confluence
- Outreach
- SalesLoft
- Marketo
- Some project-management and workflow tools that agents can replace or shortcut
- More resilient:
- System-of-record software with durable data gravity
- Infrastructure that agents still need to operate
- Products that become AI-enhanced rather than AI-displaced
Anthropic’s $45B Funding and the Compute Arms Race
- The round is interpreted as a sign of both momentum and constraint:
- Anthropic built a strong model,
- but now needs massive compute to keep up with demand.
- The conversation emphasizes how capital intensive frontier AI has become:
- forecasting demand two years out,
- committing enormous capex,
- and accepting large risk if growth overshoots or undershoots.
- A key insight:
- In software, you could often grow revenue faster than costs.
- In frontier AI, growth itself forces huge forward capex commitments.
Why Google may be the quiet winner
- Google is seen as uniquely advantaged because it has:
- cash flow,
- its own AI stack,
- huge compute capacity,
- and optionality to allocate resources across Gemini, Anthropic, and its core business.
- The panel repeatedly frames Google as a “multiple ways to win” asset.
NVIDIA vs. Google: where to put money
- Risk-adjusted preference: Google
- because it has diversification and existing cash generators.
- Pure AI upside: NVIDIA
- because it is the most direct way to express a bet on explosive AI capex.
- Their takeaway:
- if you want AI exposure and can’t buy OpenAI/Anthropic, NVIDIA is the cleanest pure play.
China Blocks Meta’s Manus Deal
- China blocking Meta’s acquisition of Manus is framed less as a money issue and more as a geopolitical leverage move.
- The panel believes:
- capital may already be out,
- but the real issue is control over technology and talent.
- Broader implication:
- this is another expression of the US–China AI conflict.
- The sanctions/chips debate and deal interference are part of the same strategic pattern.
Private Equity, Medallia, and the Death of Easy Exits
Medallia’s handback to creditors
- Medallia is treated as a landmark event:
- a large equity wipeout,
- with debt now effectively overwhelming the business.
- The important lesson:
- this is not just a leverage problem;
- it’s an overpay problem plus an AI relevance problem.
- In a world where better AI-native competitors exist, legacy customer-experience/survey software looks increasingly expendable.
What this means for PE as an asset class
- The panel believes the old assumption that PE is the “safe” part of the portfolio is breaking down.
- If multiple large buyouts fail, LPs will feel it.
- More importantly:
- one of the traditional exit routes for venture-backed companies is disappearing.
- They argue the exit funnel is narrowing to:
- very large IPOs,
- highly targeted strategic acquisitions,
- and maybe some secondary or internal consolidation.
- The market is moving toward:
- fewer but much bigger winners.
Venture, IPOs, and “Fair Value” vs. Pixie Dust
- The group argues that SaaS no longer gets automatic “pixie dust” multiples.
- Strong companies will still get valued well, but increasingly at fair value, not narrative premium.
- This is bad news for venture funds that depend on a few outsized marks to drive returns.
- Their framing:
- Early stage now needs more diversification and a larger portfolio.
- Late stage is becoming concentrated among only a few companies large enough to go public.
Big IPOs are still possible, but the bar is much higher
- A company may be profitable and large, yet still not be “IPO-ready” in the old sense.
- The panel suggests the new public-market threshold is closer to:
- massive scale,
- strong growth,
- and clear AI relevance.
YC’s Revenue Transparency Push
- Gary Tan’s guidance on “bullshit ARR” is discussed as both:
- a good corrective,
- and a shrewd move to preserve trust in the YC market.
- Main point:
- revenue definitions are often fuzzy in startups,
- and founders should be precise and honest about what counts.
- This is positioned as market infrastructure:
- if buyers and investors can’t trust the numbers, the whole seed market weakens.
Thrive Eternal, Sports Assets, and Durable Non-AI Businesses
- Thrive’s “Eternal” strategy is discussed as distinct from Sequoia’s Evergreen experiment.
- The idea is not just “hold public equities forever,” but rather invest in non-digital, durable assets.
- Sports teams are used as an example of assets that may remain resilient even in an AI-heavy world.
- A side debate notes that:
- AI could change media consumption and digital rights economics,
- but marquee sports properties still have exceptional pricing power.
Robinhood Ventures and AngelList USVC Fund
- These products are viewed as a way for public investors to access private-market exposure.
- The panel’s stance:
- they make sense symbolically and behaviorally,
- especially for getting exposure to the “big three” private assets: OpenAI, Anthropic, SpaceX.
- On fees:
- AngelList’s reported fee structure is debated,
- but the broader point is that private investing already has meaningful fee drag.
- Their conclusion:
- as long as the underlying assets are strong enough, these vehicles can be rational small allocations.
Final Note
The episode’s dominant message is that AI is not just changing product features — it’s changing who chooses software, how compute is allocated, what gets valued, and which business models survive. The winners will be the companies that either:
- become essential to agents,
- control the compute layer,
- or own durable, non-digital assets outside the AI substitution zone.