Overview of 20VC: Deepseek Raises $50BN | Wall St's $725BN AI Question | The Rise of Open Source & How it Threatens OpenAI & Anthropic | OpenAI Builds its Own Chip: Jalapeño | The Death of Moats & The New AI Software Winners
This episode is a fast-moving deep dive into the current state of AI markets, centered on talent migration, open source pressure, sovereign AI, infrastructure costs, and the economics of enterprise software in an AI-saturated world. Harry Stebbings, Jason Lemkin, and Rory O’Driscoll argue that AI is entering a new phase where pricing, margins, and workflow automation matter more than raw model hype, and where the winners will be the companies that can ship, retain elite talent, and control cost.
Key AI Industry Stories
DeepMind talent leaves for Anthropic
- The hosts discuss two major departures from Google DeepMind:
- Noam Shazeer (key figure behind the attention paper / Character.AI)
- John Jumper (Nobel Prize winner and AlphaFold co-creator)
- Their move is framed as a sign that top AI researchers increasingly prefer environments where they can:
- work on what they want
- move faster
- ship products
- avoid legacy bureaucracy
- Takeaway: top talent is flowing toward startups with momentum and flexibility, especially OpenAI and Anthropic.
DeepSeek’s massive raise and China’s AI strategy
- DeepSeek is discussed as raising $7.4B at roughly a $50B valuation.
- The round is notable because:
- the Chinese state retains control/voting influence
- the founder is committing substantial capital himself
- the company operates in a strongly sovereignty-driven context
- The hosts argue that Chinese open-source AI is less “open” than it appears because state subsidies are effectively underwriting training and infrastructure.
- Broader point: China is building AI as a strategic national capability, not just a commercial product.
Open source is becoming a real threat
- The conversation repeatedly emphasizes that open source is now a serious competitive force, especially for:
- coding
- inference
- enterprise workflows
- Open source models are described as:
- materially cheaper
- good enough for many workloads
- increasingly competitive with frontier closed-source models
- The hosts believe open source creates a ceiling on pricing and margins for OpenAI and Anthropic, especially in the “middle” of the market.
The $725B AI Question: Who Pays?
AI capex is colliding with real economics
- A major theme is Wall Street’s growing concern over who will ultimately pay for the enormous AI buildout.
- The hosts cite estimates of:
- $700B+ annual hyperscaler capex
- multi-trillion-dollar cumulative AI infrastructure spend over the next several years
- Their core argument:
- current AI revenue is far below the level needed to justify the spend
- to earn a return, AI must either:
- replace a meaningful share of labor, or
- dramatically raise productivity across industries
The revenue math is brutal
- They argue that for the economics to work, AI likely needs to replace or materially augment around 7–8% of labor.
- This makes the “AI replaces everything” story less a slogan and more a revenue requirement.
- If AI spending stays this high, then:
- enterprise customers will demand ROI
- token usage will become tightly managed
- companies will face pressure to lay off labor or become materially more efficient
Token Pricing, ROI, and Enterprise Adoption
Token maxing is giving way to ROI scrutiny
- In 2025, many companies are still spending aggressively on AI just to learn and experiment.
- The hosts think that by 2027, the question will be:
- “Show me the ROI next year.”
- That means AI budgets will likely move from exploratory spend to hard business-case scrutiny.
Intelligence is not free
- A key point is that unlike traditional SaaS, intelligence has:
- theoretically infinite demand
- but meaningful cost
- That means CIOs will need a new discipline:
- deciding where intelligence is worth paying for
- which teams get access
- which workflows deserve premium model usage
OpenAI and Anthropic pricing strategies matter
- They discuss Anthropic’s efforts to encourage prompt caching and reduce inference costs.
- The hosts note that:
- enterprise customers can be highly profitable
- prosumer and max-plan users are often heavily subsidized
- This creates a mixed-margin business where pricing power is under pressure from both:
- open source
- customer efficiency improvements
The Death of Moats
“Your moat can be LLM-lifted away”
- One of the strongest recurring themes is that classic startup moats are becoming weaker.
- The hosts say founders obsessing over “moats” are thinking too traditionally:
- workflows can be lifted
- vendor setups can be migrated
- products can be copied or compressed by AI
- Their view: moats are increasingly temporary unless they are tied to distribution, speed, or cost structure.
LLMs are becoming moat destroyers
- They point to examples where AI can:
- migrate data
- rewrite workflows
- replace consulting labor
- automate software transitions
- This makes it easier for customers to switch vendors and harder for incumbents to defend expensive services.
AI Agents and the Future of Work
Agents are already doing real back-office work
- Harry describes building an AI “VP of Finance” that can:
- create quotes
- build and send contracts
- update Salesforce
- issue invoices
- chase payments
- reconcile QuickBooks
- The point is that current models can already handle multi-step, operational workflows, even if they are not yet perfectly reliable.
“Training agents” as a future job category
- The episode discusses the idea that “training agents” could become a huge category of work.
- The hosts agree that the skillset is changing rapidly:
- prompt engineering is fading
- agent management / orchestration is the new skill
- The job isn’t just writing prompts anymore; it’s:
- designing workflows
- knowing when agents fail
- shaping output quality
- monitoring where automation breaks down
Consulting, Accenture, and AI for SI
Accenture as a case study in disruption
- Accenture is discussed as being under pressure despite also benefiting from helping firms adopt AI.
- The logic:
- AI adoption work is a new opportunity
- but Accenture’s core business is selling labor-heavy consulting seats
- That core model is vulnerable because AI can automate:
- requirements gathering
- documentation
- implementation work
- systems integration
“Selling hours” is becoming harder
- The hosts argue that labor-based services are in trouble if AI can do a large share of the work.
- They contrast:
- seat-based models
- body-based models
- AI-native delivery
- Their view: many consulting firms will be compressed unless they reinvent themselves around AI-first delivery.
OpenAI’s Custom Chip: Jalapeño
OpenAI joins the custom silicon race
- OpenAI’s new custom chip, Jalapeño, is framed as a major strategic move.
- It’s co-developed with Broadcom and is intended to improve:
- performance per watt
- inference efficiency
- long-term cost structure
Mixed reaction
- Jason argues OpenAI should stay focused on winning the model and product layer rather than vertically integrating deeper into hardware.
- Rory counters that:
- cost pressure and open source competition may force frontier labs to control more of the stack
- the “middle” of the market is especially vulnerable
- The broad takeaway: hardware control may become necessary if inference becomes a margin war.
Venture, Fund Strategy, and Startup Economics
Menlo’s $3B fund
- The hosts discuss Menlo’s new fund and suggest its size is rational:
- large enough to participate in major winners
- not so large that it forces bad strategy
- They note that venture firms are increasingly using SPVs and sidecars to access outsized winners without bloating core fund size.
The funding market is favoring efficiency and growth
- They revisit a YC-style critique:
- investors don’t fund revenue
- they fund margin on revenue
- The hosts partially agree, but stress that in AI:
- negative gross margin businesses have still been funded
- many will only work if they scale very fast
- Still, they think the market is starting to punish companies with:
- weak margins
- slow growth
- no path to efficiency
Prediction Markets and Kalshi
Kalshi’s growth is strong
- Kalshi is highlighted as a major beneficiary of regulatory arbitrage in prediction markets.
- It is effectively capturing a lot of sports-betting demand while operating under a different regulatory framework.
- The hosts note the business is growing fast and could potentially IPO.
Meta as a possible competitor
- They speculate that Meta could eventually launch a socially integrated prediction product.
- If so, Facebook/Instagram’s scale and social graph could make prediction markets much more accessible.
- That said, there is real regulatory risk, and the market’s future depends heavily on policy.
Work-from-Home and Company Culture
WFH is increasingly seen as incompatible with high-performance startups
- Ryan Petersen’s “WFH is white collar fraud” comment is discussed as both provocative and outdated.
- The hosts’ real position:
- some remote work is fine
- but the best AI-era startups require intense, in-person, high-output teams
- They argue the new startup playbook is:
- smaller teams
- higher compensation
- more equity
- more urgency
- more office time
Big Takeaways
- AI is shifting from hype to economics.
- The key question is no longer whether models improve, but who pays for them and how margins hold up.
- Open source is now a real strategic threat.
- It compresses pricing, especially in the middle of the market.
- Talent is the scarcest asset.
- Elite researchers want freedom, speed, and strong environments.
- Agents are already changing operations.
- Back-office work, finance workflows, and systems integration are early wins.
- Moats are weaker than founders think.
- If AI can lift and migrate workflows, defensibility must come from speed, distribution, and cost.
- AI may force a labor reset.
- In the next phase, companies will need to prove ROI or cut headcount to fund AI adoption.
Notable Lines and Framing
- “Your moat can be LLM lifted away.”
- “The whole reason OpenAI and Anthropic work is because other idiots spent the $300 billion on their behalf.”
- “Show me the ROI next year.”
- “There’s only one thing worse than a seat-based model, and that’s a model based on bodies.”
- “Winners win and compound.”