The AI talent wars have never been crazier.
Within 48 hours, Google lost two of their generational scientists to OpenAI and Anthropic.
I sat down with @jasonlk and @rodriscoll to discuss it, along with the biggest news in tech this week:
- Deepseek Raises $50BN
- Wall St's $725BN AI Question
- The Rise of Open Source & How it Threatens OpenAI & Anthropic
- OpenAI Builds its Own Chip: Jalapeno
My notes below:
1. The Number Three Closed-Source LLM Is Most at Risk
In the closed-source foundation model race, the number three vendor faces severe pressure as multi-model routing spreads across enterprises. Historically, a third-place software or cloud vendor could survive by being cheaper or simpler. But today, that tier is being squeezed by highly capable open-source models, many heavily subsidized by China, making developers less likely to care about a closed-source number three.
2. The Playbook for Building a Startup to Its First 100 Employees Has Changed
Scaling a startup to its first 100 or 200 employees has changed radically from the remote-work era. In today’s hyper-competitive AI landscape, companies built around relaxed schedules and 20-hour remote workweeks will not win. The modern model is lean, elite, highly compensated teams working intensely in person, often six or seven days a week, to survive nonstop product sprints.
3. Why It Makes No Sense for OpenAI and Anthropic to Build Their Own Chips
Frontier AI labs have captured the greatest demand engine in modern technology and should focus entirely on winning enterprise customers. Designing custom hardware is a massive distraction when cloud giants like Oracle, Google, Microsoft, and Amazon are already competing aggressively to provide cheap compute. Those infrastructure vendors absorb the capital risk, letting labs prioritize product growth and customer adoption.
4. Why Vertical Integration Makes No Sense for OpenAI and Anthropic
Closed-source labs scaled rapidly by staying asset-light and outsourcing heavy infrastructure needs. Their model worked because cloud hyperscalers absorbed hundreds of billions of dollars in cumulative CapEx on their behalf. Forcing backward vertical integration down to the chip level would introduce massive capital liabilities and undermine the operational efficiency that made the model work.
5. Show Me the ROI Next Year
Enterprise software is moving from loose AI experimentation to strict financial accountability. Early corporate budgets funded unconstrained “token maxing” to build basic AI fluency, but the 2027 narrative will demand measurable ROI. CIOs will no longer allocate tokens based on compelling pitches alone; they will require verified departmental efficiency gains or clear revenue growth.
(links below)
