Elon Musk loue ses serveurs à son pire ennemi — et c'est du génie !

The episode of Silicone Carnet opens with a significant development in the AI landscape: Elon Musk, previously critical of OpenAI, has struck a deal with Anthropic, a company he once accused of hating Western civilization. This agreement will see SpaceX lease a massive computing infrastructure to Anthropic, comprising 220,000 GPUs and 300 MW of power, available by the end of the month. This move is seen as a paradox, highlighting how practical needs, like compute power, can override ideological stances, especially when compute bills arrive. The discussion among the guests – Fred Binquet, Fred Lanier, and Fred Montagnon – delves into the strategic implications of this deal. Fred Binquet views it as a move by two somewhat faltering entities: Musk, who might have overextended on his "Colossus" compute project, and Anthropic, which, despite its growth, was facing escalating costs and potential brand damage. He also notes a shared "enemy," implying a competitive dynamic against other AI giants. Fred Lanier sees this as Musk strategically positioning himself on the "analog" axis of competition – infrastructure and energy – to gain leverage over the "digital" software and AI agents axis. He draws a parallel to Rockefeller's strategy, building an empire on infrastructure like railways rather than just oil wells. Lanier highlights a crucial clause in the SpaceX-Anthropic contract: Musk can terminate the deal if he deems Anthropic's actions perilous to humanity, a highly subjective and potentially game-changing condition. He also frames this as a Cournot oligopoly strategy, where Musk makes a significant investment to signal his position and influence market dynamics, creating a difficult situation for regulators. Fred Montagnon offers a more pragmatic view, suggesting that providing infrastructure (like AWS for Amazon) is a sound business strategy, especially given the immense demand for compute power. He contrasts this with the volatile landscape of AI model development, where open-source models are rapidly evolving, making it difficult for companies like OpenAI and Anthropic to maintain a leading edge and predict future profitability for IPOs. He emphasizes that infrastructure providers, like Elon Musk, are better positioned. The conversation then shifts to OpenAI's financial situation and IPO prospects. Fred Lanier expresses skepticism about OpenAI's ability to go public, citing the increasing valuation requirements with each funding round and the competitive pressure from Anthropic. He points to secondary market transactions where Anthropic's valuation is reportedly higher than OpenAI's projected future valuation. Fred Binquet, however, remains more optimistic about OpenAI, pointing to their substantial funding ($120 billion) which gives them time. He also highlights the impressive performance of GPT-55 and Codex, suggesting a strong comeback for OpenAI. He acknowledges that while funding provides time, a future IPO remains a significant challenge, especially if revenue growth slows or competition intensifies. The discussion touches upon the sheer scale of AI growth. Fred Binquet presents a graph showing a strong correlation between compute capacity and revenue for OpenAI, suggesting that compute is the primary bottleneck. He argues that even with smaller open-source models, the inference problem (housing these models) will still require significant compute, which will be dominated by major players. He believes that Musk's deal with Anthropic is a strategic move to secure this essential compute. The debate then moves to the nature of AI companies and their market positioning. Fred Montagnon suggests that AI companies are creating immense value, potentially reaching trillions in revenue, but this is constrained by physical resources like compute and energy. He believes that while economic criteria might seem less important now, they will become crucial when these companies seek to go public. He anticipates potential disappointments in IPO pricing, leading to adjustments in the supply chain, including contract repricing, layoffs, and product shifts. The geopolitical dimension of AI is also explored. The guests discuss Europe's struggle to keep pace with the US in terms of compute infrastructure and AI development. Fred Binquet highlights the immense revenue and profitability of hyperscalers like Google and Amazon in the AI infrastructure space, questioning Europe's equivalent. He suggests that Musk's move into AI infrastructure, including data centers in space, is a strategic calculation to position SpaceX as a hyperscaler rather than just an aerospace company, aiming for higher valuations. The conversation then pivots to regulation. The White House's apparent shift from encouraging a "pragmatic growth" approach to considering pre-market validation of AI models, akin to drug approval, is analyzed. Fred Binquet attributes this shift to a pragmatic, even cynical, assessment of security risks, particularly the potential for AI models to create widespread vulnerabilities in defense and economic systems, rather than a purely moral stance. He contrasts the US approach, focused on preventing monopolistic practices and providing adaptation time, with Europe's focus on consumer welfare and precautionary principles. Fred Montagnon presents an analysis of the sheer volume of legal texts in the US and France, suggesting that regulation is not absent in the US but follows a different logic. He notes that US regulation, particularly in defense, sometimes mandates contracts with smaller companies to foster innovation and new entrants, a stark contrast to Europe's challenges in integrating startups with larger corporations. A third explanation for the US regulatory shift is proposed: a fear of social unrest and a desire to manage the societal impact of AI. The increasing resistance to data centers and the declining enthusiasm for AI among younger generations in the US are presented as evidence of this concern. The guests debate whether this fear stems from job displacement fears or a deeper philosophical question about the compatibility of rapid technological progress with a meaningful human life. Fred Binquet argues that regulation is not the solution to existential questions about the meaning of life in a rapidly changing world. He observes that this confusion seems more pronounced in Europe, where regulatory frameworks are heavily influenced by principles of precaution and fundamental rights, compared to the US's more individualistic approach. The discussion then turns to the potential for AI to address societal challenges like rising costs in healthcare and education, and the impact on employment. While acknowledging job displacement concerns, the guests suggest that AI might also create new opportunities and increase individual productivity, potentially leading to an era of abundance. Examples of call centers still employing humans due to cost-competitiveness and the emergence of new AI-enabled roles are discussed. The episode concludes with a segment on prediction markets, specifically Polymarket. The arrest of a US soldier for insider trading on a secret operation and a French trader allegedly manipulating weather data for bets are presented as examples of the growing scandals in this space. The guests discuss the phenomenon of prediction markets becoming a significant force, even replacing traditional polling methods in the US. Fred Binquet highlights the potential for these markets to provide more robust and less biased estimates when sufficiently deep, but cautions against unrealistic bets. He draws parallels to the history of insider trading, suggesting that these markets, while growing rapidly, are still in their nascent stages and will likely face increased regulation. The concentration of profits among a few "whales" and the comparison to crypto markets are also discussed, with some guests expressing concern about the potential for moral hazard and a "get rich quick" mentality among younger participants. The idea that these markets might reflect a deeper societal desire for individual expression and influence on policy is also debated. The episode ends with a humorous anecdote about a bet on the return of Jesus, illustrating the wide range of topics covered by prediction markets.