The Twenty Minute VCDeepSeek Raises at $50B | The Rise of Open Source vs OpenAI & Anthropic | OpenAI Builds Own Chip
CHAPTERS
- 0:00 – 5:58
DeepMind talent exodus: why top researchers are leaving Google for Anthropic
The discussion opens on two high-profile departures from DeepMind—Noam Shazeer and Nobel laureate John Jumper—and what it signals about Google’s ability to retain elite talent. The group contrasts Google’s historical research culture with today’s need for speed, autonomy, and massive compensation packages at frontier labs.
- •Frontier researchers optimize for autonomy, focus, and environment more than money
- •Google’s bureaucracy and ‘failure to ship’ push product-minded leaders to newer labs
- •Anthropic/OpenAI can outbid and offer fewer constraints as ‘unconstrained’ new entrants
- •Shazeer vs. Jumper illustrate different motivations: shipping vs. pure science initiatives
- 5:58 – 13:21
Momentum compounds: incumbents vs. new labs and the “winner flywheel”
Rory and Jason widen the lens: talent flows toward perceived winners, and momentum becomes self-reinforcing. They argue Anthropic is unusually strong at both research culture and tactical execution, while Google is relevant but not the default choice for developers day-to-day.
- •New labs can promise ‘everything to everyone’ due to stock/currency and lack of legacy constraints
- •Google is financially strong but not the product default for many developers
- •Research isolation (e.g., keeping sales away) reinforces elite-lab attraction
- •Market perception and internal velocity matter as much as raw resources
- 13:21 – 18:31
Why being #3 in AI is dangerous: routing, price pressure, and open-source swamping
Jason lays out why the third-place closed model is uniquely vulnerable as multi-model routing becomes standard. Open-source alternatives—often subsidized—create a cost/performance squeeze that can erode developer mindshare and economics for the #3 provider.
- •Enterprises increasingly route workloads across multiple models to manage cost/performance
- •Open source isn’t ‘free like Linux’—but can still be materially cheaper at scale
- •The #3 closed model struggles: not best, not cheapest, and less exciting for researchers
- •Anthropic’s prompt-caching push is framed as a direct competitive move against open-source economics
- 18:31 – 20:23
Sovereign models and Europe’s tradeoff: efficiency vs. national security
The conversation turns to “sovereignty” as a rationale for regional champions like Mistral. Rory argues sovereignty reframes the market: a firm can be ‘#1 in Europe’ rather than #4 globally, at the cost of higher prices and potentially weaker models—unless security considerations justify the tax.
- •Sovereign model claims translate into protected regional market leadership
- •Economic efficiency declines when regions choose less competitive local alternatives
- •Government priorities (security, control) can outweigh pure market logic
- •Sovereignty debate intensifies as open-source leadership is often China-based
- 20:23 – 26:09
China’s AI parallel universe: DeepSeek, subsidies, and governance control
Jason reports firsthand observations from China: OpenAI/Anthropic access restrictions, reliance on DeepSeek and Gemini, and signs of state influence. DeepSeek’s funding structure and state voting control are interpreted as evidence that AI is treated as national infrastructure.
- •OpenAI/Anthropic are blocked (or hard to access) in China; Gemini remains available
- •DeepSeek appears ‘crippled’ domestically (limited web search, different data behavior)
- •Massive subsidies are framed as small relative to sovereign stakes (e.g., compared to defense spend)
- •DeepSeek’s round structure signals state governance priority over conventional investor rights
- 26:09 – 27:30
Open-source benchmark leaps and the distillation question: ceiling on frontier pricing
Rory reacts to reports of Chinese open-source models beating frontier benchmarks in specific domains like coding. The main takeaway is competitive intensity: multiple Chinese labs iterating quickly create a meaningful cap on what closed providers can charge, regardless of exact training methods.
- •Several Chinese open-source families now sit near frontier performance in key tasks
- •Competitive abundance in open source creates a pricing ‘drag’ on closed vendors
- •US companies increasingly build on open-source bases for custom models
- •Distillation/learning-from-frontier remains an unresolved part of the narrative
- 27:30 – 29:54
The infrastructure bill arrives: DRAM shocks and AI’s resource reallocation
The panel discusses surging memory prices and how AI CapEx demand ripples through the economy. Rory frames it as classic resource allocation: data centers bid away inputs, pushing prices up for consumer products, electricity, and even regional housing.
- •DRAM and memory costs spike due to AI infrastructure demand (Apple cited)
- •Price mechanism reallocates scarce inputs from consumer goods to data centers
- •Second-order effects: higher device prices, higher electricity costs, local cost-of-living pressure
- •AI buildout is large enough to affect multiple unrelated markets simultaneously
- 29:54 – 37:34
Wall Street’s $725B question: who pays for the trillion-dollar AI buildout?
The core macro debate: hyperscalers are spending hundreds of billions annually on AI CapEx, implying massive future revenue must materialize. Rory and Jason argue demand may be ‘infinite’ at the right price, but the system still needs provable enterprise ROI to sustain spend.
- •Goldman-style forecasts imply $700B+/year AI CapEx and multi-trillion cumulative spend
- •Current AI revenue levels lag far behind implied required revenue to justify investment
- •AI demand is enormous but price-sensitive; token costs force allocation decisions
- •The key uncertainty is when (not how) the narrative meets financial constraints
- 37:34 – 42:55
From token-maxing to ROI: the enterprise reckoning (and labor displacement math)
They forecast a shift from ‘spend to learn’ budgets toward ROI-gated allocation of tokens by 2027. Rory quantifies the challenge: to justify ~$1T of AI revenue, productivity gains must be huge—implying meaningful labor displacement even if profits don’t rise due to parity adoption.
- •2025–26: experimentation and AI fluency; 2027: budget holders demand measurable ROI
- •Tokens become a managed resource with internal chargebacks and performance gating
- •Industry-wide adoption may eliminate profit gains even if productivity rises (parity tax)
- •Meeting revenue hurdles likely requires major productivity improvements and job disruption
- 42:55 – 48:37
Agents in practice: building an “AI VP of Finance” and why ‘agent mastery’ matters
Jason shares a concrete case study: building an agent that handles billing, contracts, Salesforce updates, invoicing, and bookkeeping workflows. They argue the emerging skill isn’t prompt engineering, but operational competence in supervising agents—understanding failure modes, shortcuts, and reliability.
- •Real workflow automation: contracts → signatures → invoicing → collections → bookkeeping
- •Agents reduce variable contractor hours even without explicit headcount cuts
- •Prompt engineering is fading; the durable skill is supervising/structuring agents and loops
- •Agents exhibit ‘goal-seeking laziness’ (e.g., admitting it didn’t read a full contract)
- 48:37 – 52:43
Gross margin becomes the new growth: why fast revenue isn’t enough
A founder/investor debate on whether weak gross margins are now a funding killer. Rory notes many AI winners scaled despite poor early margins, but Jason stresses that as the market matures, businesses with bad unit economics must achieve extreme growth—or they fail outright.
- •Debate: investors historically funded hypergrowth even with bad gross margins
- •AI-native companies have sometimes ‘grown out’ of negative margins—until growth slows
- •As competition and bundling intensify, margin-improvement may become harder
- •Rule of thumb: if you’re margin-negative without massive growth, you’re at high risk
- 52:43 – 58:15
Menlo’s $3B fund: disciplined sizing, SPVs, and the new venture capital structure
They unpack why Menlo raised $3B despite major AI wins like Anthropic, arguing smaller core funds can preserve strategy while using SPVs for outsized opportunities. Rory emphasizes the constraint: there are only so many places to deploy multi-billion checks without distorting returns.
- •Why ‘not more’ can be rational: fund size dictates strategy and feasible deployment set
- •SPVs/sidecars enable concentration in rare trillion-dollar outcomes without bloating the main fund
- •Smaller funds can improve risk-adjusted outcomes if access to follow-on capital remains
- •Venture dollars concentrating into a few mega-deals changes optimal fund construction
- 58:15 – 1:04:22
Kalshi’s surge: regulatory arbitrage, prediction markets vs. sports betting, IPO risk
Kalshi’s run-rate and IPO chatter leads into an explanation of its business: functionally sports betting framed as prediction markets under federal jurisdiction. They discuss valuation heuristics, competitive threats from Meta, and the persistent tail risk of regulatory reversal.
- •Kalshi benefits from CFTC-style positioning and avoids state-by-state licensing burdens
- •The product experience resembles betting, even if structured as a matching/clearing market
- •Meta could amplify social betting, but demographic fit and regulatory appetite are unclear
- •Regulatory shifts (future administrations, state lawsuits) are key business risks
- 1:04:22 – 1:12:32
Accenture down 40%: AI compresses systems-integration economics and ‘bodies-based’ models
They argue consulting is uniquely exposed to AI because much of the work is repeatable, document-heavy, and code-adjacent—ideal for LLM automation. Accenture’s AI-adoption advisory may grow, but its core SI revenue faces pricing compression as AI-first competitors bid dramatically lower.
- •GenAI advisory demand grows, but legacy SI work faces structural margin pressure
- •AI-for-SI startups can underbid large consultancies by replacing ‘body-hours’ with tooling
- •Bodies-based billing is brittle: fewer people needed breaks the business model and pricing
- •LLM-enabled ‘data/app lifts’ shrink multi-year migrations into weeks, destroying incumbents’ moats
- 1:12:32 – 1:17:42
Work-from-home backlash and the ‘new startup operating system’: small teams, high intensity
A heated segment argues that top-performing AI startups require intense, in-person execution to keep up with relentless competitive sprints. Jason frames it as a bifurcation: stable jobs and lifestyle vs. high-upside equity outcomes that demand extreme pace and tighter team structures.
- •Claim: AI competition is ‘endless sprints’—companies can’t afford low-intensity execution
- •Preference for smaller, higher-paid, equity-heavy teams with more in-office collaboration
- •WFH critique blends two issues: effort levels and effectiveness of coordination
- •Cultural shift: what was once considered toxic hustle language is increasingly normalized in AI
- 1:17:42 – 1:28:01
OpenAI’s “Jalapeño” chip and the ‘hollowing out’ of the AI market’s middle
They close on OpenAI’s custom inference chip announcement and debate whether deep vertical integration is distraction or defensible strategy. Jason ties it to a bigger competitive threat: open source eroding the ‘middle tier’ of model pricing, forcing frontier labs to lower costs or lose share.
- •Custom chips could improve performance-per-watt and cut inference costs materially
- •Rory questions vertical integration: clouds and chip vendors already compete to serve OpenAI
- •Jason argues cost reductions defend against open-source disruption in the ‘middle market’
- •Thesis: the market polarizes—frontier and tiny models exist, but the profitable mid-tier is at risk
