Elon's $60B Cursor Bet, Claude kills SaaS, and OpenAI's Mass Departures | EP #249

Dave reports the tool is incredible but 'dog slow,' likely because Anthropic is compute-constrained. Alex and Salem both emphasize that Anthropic desperately needs an ISV strategy and roadmap to avoid destroying the software ecosystem. Peter says the company is 'very PhDish' and needs to think about its power. Yet the capability is real: Claude already generates most of Anthropic's own code, and rumors suggest Google DeepMind engineers are using Claude over Gemini. The hosts advise any startup to make its tech stack model-agnostic so you can swap easily when better or cheaper options emerge. Dave demonstrates that now you can literally tell Claude to switch half of your AI agents to another vendor and it will do it.

This advice emerges from the Claude Design and XAI speech discussions. Alex notes that many SaaS companies are 'one unhobbling away' from replacement, and the surviving ones will be those that have data moats, regulatory expertise, or vertical industry knowledge. Salem gives examples like Vocara, which operates in regulated insurance and mortgage space where specific compliance knowledge is the moat. Dave describes launching a 10,000 concurrent agent swarm and being able to tell Claude to swap half the agents to a different vendor with a single line. The key is that the user experience and customer relationship are what matter, not the underlying model.

Responding to a listener critique about water use, Dave explains that data center water use is microscopic compared to agriculture. However, because evaporative cooling is concentrated in water-scarce sunny regions, it causes friction. He argues simply adding more solar panels and switching to mechanical cooling eliminates the controversy. This is not a fundamental problem, just an engineering choice.

This emerges from the hosts' response to criticism about ivory tower mentality. Peter recounts his 11-year struggle to get Zero-G approved by the FDA, emphasizing persistence. Salem shares arriving in Silicon Valley at 40 with $2,000. Dave talks about living below poverty line, duct-taping his muffler. They all stress that AI amplifies capability, but without a deep why, you won't stick it out. The MTP concept is central to Peter's philosophy: you find a problem you'd work on even without pay.

The hosts earlier listed three roles: founder, joiner, investor. Peter encourages those who feel they can't be founders to find an entrepreneurial company and bring their experience. Dave explicitly states that 'you either need to be a founder or a joiner or an investor because all other career paths are going to disappear.' This reframes the 'entrepreneur or bust' narrative to a spectrum of agentic paths.

Peter shares the story of a young man in Morocco who described his skills and location to ChatGPT and designed a new career. Alex notes that AI is good at identifying problems, referencing a company (Henry) that uses AI operators to make everyone an investor/entrepreneur. Salem recounts a fishing village in Vietnam using a solar panel and internet research to build a solar-powered boat—permissionless disruptive innovation. The cost to start a business has dropped 99.7% since 2005, and AI amplifies this further.

Alex argues that every time a SaaS stock faces a mini-apocalypse from an unhobbling, it underscores that any company merely scaffolding around models is at risk. He advises companies to build vertically integrated native capabilities or physical world integration. Salem stresses Anthropic's product experience is poor—'dog slow'—and they need a Steve Jobs/Jony Ive type to deliver out-of-box quality. He also demands an ISV strategy so software developers can work with rather than against them. Dave notes Anthropic doesn't have enough compute, and the launch seems designed to shock markets without a clear plan. The hosts agree that AI is now the control layer over design, coding, and more, making it easier to list what isn't at risk. They highlight upcoming victims: legal research (LexisNexis), business intelligence (Tableau), Bloomberg terminals, Workday, etc. Salem says legacy media lost its monopoly on reality construction, and trust will be rebuilt through direct voice channels like podcasts.

Alex highlights human arrogance: the San Francisco engineers believe entry-level jobs will be replaced but not their own. He sees this as evidence of recursive self-improvement: most code at Anthropic is now generated by Claude, and rumors suggest Google DeepMind researchers are also using Claude. Dave notes that the bias in such surveys has shifted from exaggeration to downplaying, because speaking openly about capabilities is now risky. Peter asks how large corporates will adopt. Salem predicts a horde of startups building AI-native will emerge while big companies struggle because their workflows are human-to-human; they'll end up acquiring startups to compete. The conversation also touches on global adoption speed: San Francisco is at warp speed; Europe is sublinear. Alex mentions Meta's upskilling program for fiber laying as a temporary patch. Peter says the senior execs at OpenAI were often founders, not ladder climbers, so the entrepreneurial path is the answer.

Alex argues that Moore's Law and Dennard scaling ended long ago; most weights are likely wasted on world knowledge that could be externalized. He would prefer a race to compress capabilities into fewer parameters, i.e., intelligence density. Dave appreciates the transparency but also notes Elon is 'the god emperor of brute forcing' and XAI leads in total compute power (2 GW by year-end). The hosts debate whether frontier labs would release AGI once achieved or hoard it for internal science breakthroughs, and they note that 'AGI' definitions are malleable (OpenAI's AGI = $100B revenue). Alex mentions ternary quantization and binarization as an alternative path. Elon is aware of the density argument but may be constrained by his team's expertise.

Alex recalls the great Anthropic schism triggered by a shift away from broad research. He predicts a similar event, potentially resulting in a new, well-funded frontier lab with a novel AGI approach. Dave reveals that the compensation numbers are 'wacky': Mark Chen was offered $1B by Meta and refused. Senior people like Kevin may have ~0.1-0.2% equity vesting, worth $500M+ if they stayed; departures free up massive comp for hiring. Peter notes Kevin is a friend who texted him. The restructuring is driven by capital efficiency for IPO, focusing on near-term revenue, and AI for science and Sora may be deprioritized.

Alex presents three layers of analysis: (1) Cursor has observed user behavior across all frontier models and may have fine-tuned a Chinese open-weight model on that data, making it one of the best paths to acquire Anthropic-level codegen interaction; (2) The acquisition opens XAI's hyperscaler infrastructure to third parties, with Cursor as first major tenant, marking the birth of an orbital cloud; (3) The valuation, while high, is a drop in the bucket vs. SpaceX's overall cap. Salem says the Cursor team lacks a natural public-company CEO, so this deal provides liquidity and safety. Peter notes Elon's admission that they need to catch up in coding, and Dave adds that the coding tool layer is now elevated to first-class status in the AI stack.

Alex sees the Dyson swarm as becoming the de facto economy, with the majority of solar-system intelligence hosted on data centers within decades. He argues it's fundamentally different from railroad buildout. Salem stresses that this private-led speed will move civilization 100x faster. Dave notes the F-35 program wastefulness vs. private efficiency, but believes the buildout must be even bigger and faster—'the most important thing humanity has ever done.' Peter wants more, and says whatever country does it first will have insane growth. The hosts discuss whether government should fund sovereign AI infrastructure, with Salem warning that nations not investing will be left behind. Alex predicts peak AI CapEx at 25-33% of GDP, then declining due to automation and post-scarcity, rendering GDP irrelevant. The group agrees the definition of economy must change.

Alex sees helium shortages hitting memory chip facilities in South Korea within weeks. He argues this is completely unacceptable and that it's 'high time to start helium startups'—including helium-3 mined from the moon. Peter notes that Taiwan will sacrifice every comfort to keep chip fabs running, but it's uncertain. Salem frames this as a system shock with multiple cascading failures, accelerating deglobalization as countries realize they can't depend on global supply chains. Dave notes the US is largely insulated on oil, but East Asia (Korea, Japan) will be hit hardest. Alex makes a broader point that when the war ends, the response must be redomestication of critical materials.

Alex explores three hypotheses: (1) algorithmic advances have made image generation so fast and cheap it no longer competes with codegen compute; (2) images might be an intermediate representation for code UI generation, directly aiding the codegen race against Anthropic (which doesn't allow image gen); (3) it's a buzz move for retail investors ahead of IPO. Dave notes that Claude is terrible at diagrams, so OpenAI could fill that gap for white-collar workers, becoming relevant in architecture and design documents. Peter adds that he uses Claude for code, Gemini for planning, and currently no need for OpenAI, so this brings OpenAI back into the stack.