Fireside Chat, Interview, Podcast
DeepSeek Raises at $50B | The Rise of Open Source vs OpenAI & Anthropic | OpenAI Builds Own Chip
- DeepMind Talent Exodus: Two generational scientists left DeepMind within 48 hours:
- Noam Shazeer (co-creator of the Transformer "attention" paper) moved to Anthropic.
- John Jumper (Nobel Prize winner, AlphaFold co-creator) moved to Anthropic.
- Talent Market Dynamics:
- Newer companies (Anthropic, OpenAI) are attracting top researchers by offering autonomy to work on specific research goals without the bureaucratic constraints of established incumbents.
- Turnover is driven by a "vibe" of unconstrained research environments and the ability to ship products, which incumbents like Google struggle to offer to elite researchers.
- Researchers prioritize specific environments where they can pursue what they want over large cash payouts, with compensation packages reaching hundreds of millions of dollars for acqui-hires.
- DeepSeek Series A Financing:
- DeepSeek closed a $7.4 billion Series A round at a valuation approaching $50 billion.
- Voting Rights: Only the Chinese government holds voting rights; founders and investors do not, ensuring state control over governance.
- Founder Commitment: The founder personally committed 20 billion yuan (approx. $3 billion), representing 40% of the round.
- Strategic Context: The round highlights China's strategy for AI sovereignty, using massive state subsidies to create open-source alternatives that compete directly with US firms, effectively decoupling their economy from US AI infrastructure.
- Open Source Market Realities:
- China's Subsidization: The "open source" label is increasingly a facade where China funds the training costs, allowing Chinese models to compete on price and performance with US closed-source models.
- Performance Parity: Chinese open-source models (e.g., Zipu AI's GLM 5.2) are now beating GPT-5.5 on coding benchmarks, creating a competitive ceiling for US vendor pricing.
- The "Flabby Middle" Threat: The middle tier of the AI market is at risk of being hollowed out by cheaper, capable open-source models, forcing US providers to find ways to lower inference costs to compete.
- Wall Street Investment Thesis (The $725 Billion Question):
- CapEx vs. Revenue: Hyperscalers are projected to spend $725 billion (potentially reaching $1 trillion including electricity) on AI CapEx annually from 2026–2031, implying a need for trillions in annual revenue to achieve returns.
- Current Gap: AI currently generates roughly $100 billion in revenue against $700 billion+ in spend, creating a massive profitability gap.
- ROI Shift: Enterprise spending is expected to shift from "token maxing" in 2025 to strict ROI verification in 2027, where budget allocation will be tied to proven productivity gains or labor displacement.
- Labor Displacement: To justify trillion-dollar spends, AI must drive significant labor force disruption, potentially reducing headcounts by 10–15% in knowledge workers to achieve necessary margin improvements.
- Infrastructure and Pricing Trends:
- Memory Costs: DRAM contract prices rose 90–95% in Q1, with some costs increasing 4–5x due to AI infrastructure demand.
- Price Transmission: These infrastructure costs are being passed to consumers via higher prices for hardware (e.g., iPhones) and cloud services, as companies like Apple opt to reduce margins or volumes rather than absorb costs.
- Enterprise Rationing: Companies are moving away from unlimited token consumption to rationing based on department ROI, potentially favoring "power users" while cutting budget for low-productivity teams.
- Kalshi IPO and Betting Markets:
- Revenue Run Rate: Kalshi has reached a $2 billion revenue run rate and is preparing for an IPO, potentially at a 10x multiple.
- Regulatory Arbitrage: Kalshi operates as a prediction market (regulated by CFTC) rather than a sportsbook, allowing it to avoid state-by-state gambling licenses required by competitors like DraftKings.
- Competitive Risk: Meta (Facebook) could launch an integrated betting product that leverages its massive user base, potentially threatening Kalshi's dominance if they can navigate regulatory hurdles similarly.
- Consulting and SI Disruption:
- Accenture Decline: Accenture stock has dropped ~40% year-over-year, driven by the disruption of its core business model (selling hours/seats) by AI.
- Model Vulnerability: The "body-based" billing model is structurally vulnerable to AI automation, which can perform tasks like ERP implementation or data migration at a fraction of the cost and time (e.g., 30 days vs. 5 years).
- Margin Compression: As AI reduces the cost of service delivery, traditional consulting firms struggle to pass on price hikes, while AI-native competitors can bid significantly lower (e.g., $15M vs. $80M).
- OpenAI Custom Chip Announcement:
- Jalapeno Chip: OpenAI announced a custom inference chip co-developed with Broadcom, claiming a 50% reduction in inference costs compared to typical GPUs.
- Strategic Debate:
- Support: Building custom chips is necessary to protect the "flabby middle" of the market by undercutting open-source pricing on inference costs.
- Critique: Vertical integration into chips is a distraction given the abundance of hyperscaler competitors (NVIDIA, AMD, Google TPU) willing to provide cheap compute, and the decision was likely made before current cost pressures intensified.
- Market Reaction: Cerebras stock dropped 16% following the announcement, signaling investor concern over the loss of a $20 billion chip order.
- Work Culture and Hiring Standards:
- Return to Office: There is a growing trend in high-growth startups and VC-backed firms to require rigorous, 6+ day work weeks in-office, rejecting "work from home" as a fraud for low-output teams.
- Binary Choice: The tech market is bifurcating between high-equity, high-risk roles (requiring intense effort and office presence) and lower-paying, stable corporate roles; "middle-ground" flexible arrangements are increasingly rejected for new investments.
- Agent-Centric Productivity: AI agents are now capable of performing complex white-collar tasks (e.g., an AI VP of Finance) with higher accuracy than humans, shifting the required human skillset from "doing the work" to "managing and auditing agents."