Ce que Sam Altman vous cache derrière les vrais chiffres d'OpenAI !

This episode of Silicon Carnet discusses the state of OpenAI, the rise of Google Gemini, Elon Musk's lawsuit against Sam Altman, the struggles of French unicorns, and the emerging trend of peptide use in Silicon Valley. OpenAI is facing challenges, as reported by the Wall Street Journal. The company has fallen short of its user and revenue targets, with a significant gap in achieving its goal of one billion users. Sarah Fryer, OpenAI's CFO, is reportedly concerned about the company's ability to honor its infrastructure commitments, which exceed $600 billion. This situation contrasts with Anthropic, which struggles to meet demand due to infrastructure limitations, leading to increased service costs. Pierre Gobille, a guest expert, suggests that the rapid pace of technological change makes long-term predictions difficult. He questions whether companies like OpenAI and Anthropic can generate enough value to justify their massive investments. He identifies three key dangers: the commoditization of large language models (LLMs), the lack of clear general use cases (with Anthropic having an advantage in specialized applications like coding), and the rise of open-source LLMs that threaten to capture value. He likens the situation to early internet infrastructure providers who laid the groundwork but didn't capture the most value, which instead went to companies like Amazon and Google. Arnaud Auger notes that a year ago, OpenAI and ChatGPT were clear leaders in consumer usage, leading to the "winner takes all" perception. However, other models are catching up in performance. While OpenAI continues to grow, the market is becoming more competitive, and Anthropic's revenue losses are accelerating. He emphasizes that the core issue is not the models themselves but computing capacity. Anthropic, despite having powerful models, cannot release them due to insufficient compute resources. Google, with its vast infrastructure, and XAI, with its integrated platform, possess non-negotiable assets in this compute race. Arnaud believes that despite potential bubbles, AI is a fundamental infrastructure that will transform the economy, and OpenAI, if it remains a key player, could still succeed. However, an IPO might be delayed due to ongoing issues. Julien Petit views LLMs as a commodity, albeit a difficult and expensive one to produce. He highlights the complex financial equation, noting that for every euro of revenue, four euros of infrastructure investment are needed, with these investments taking two years to materialize. This creates a challenging scenario where companies must anticipate future demand to invest in compute, risking overcapacity if demand doesn't materialize. OpenAI faces this same problem, while Google Gemini seems better positioned. The discussion then delves into the different strategies for acquiring compute. Sam Altman's strategy for OpenAI involves massive investment in data centers, totaling $600 billion in commitments last year, based on the conviction that compute capacity is the bottleneck for growth. However, this has led to a situation where they have invested heavily but lack the commercial demand to justify it. Anthropic, conversely, has not invested as much in compute but has full demand, leading to frustration among its board members who believe CEO Dario Amodi's cautious approach was a mistake. It's noted that OpenAI's $600 billion commitment is not debt but conditional contracts over 5-10 years. A counter-argument is raised that future models might consume significantly less compute. An MIT study suggests LLMs could reduce inference costs by 90%, implying that Dario Amodi's bet on reduced compute consumption might prove correct. Elon Musk, with his Colossus project, also has colossal compute capabilities, despite model-related issues. Each major player has strengths and weaknesses, making it a "real video game" of strategy. Regarding user numbers, OpenAI's struggle to exceed one billion monthly active users is still a colossal achievement, far surpassing Anthropic's estimated hundred million users. Anthropic's success lies in its B2B focus, selling services built on top of models, rather than just the models themselves. Developers pay Anthropic an average of $2000 per month, compared to OpenAI's B2C users paying $20 per month (and the majority paying nothing). Mistral, a French AI company, adopts a similar B2B strategy, achieving €400 million in revenue and a €12 billion valuation this year, despite lacking a large user base or its own compute infrastructure, relying heavily on American providers. The experts question the valuation of Mistral, pointing out that a service company typically has a valuation of 5 to 10 times its EBITDA, whereas infrastructure companies like OpenAI are valued differently. Capgemini, a service company, has a market cap of €17-42 billion on €25 billion in revenue, while Mistral, with €400 million in revenue, is valued at €12 billion, suggesting a significant premium based on its "sovereign European AI" narrative and perceived future upside. The conversation shifts to the broader AI market, noting that despite global conflicts and economic instability, the Nasdaq and S&P 500 are at record highs, largely driven by the AI narrative. Google, with its integrated infrastructure and data centers, is seen as a strong contender. Gemini has surpassed 750 million users, benefiting from pre-installation in Google products like Gmail and Drive. Google's advertising revenue continues to grow, allowing it to release Gemma 4 as a free, open-source model, commoditizing the market and potentially drawing users away from competitors. However, the exact contribution of Gemini to Google's revenue remains unclear due to opaque financial reporting. Elon Musk's XAI, with its colossal compute capacity (aiming for one million operational GPUs by year-end), is seen as a long-term strategic play, especially for future applications in health, biology, and physics that will require immense computing power. Musk's experience in building industrial infrastructure (Tesla, SpaceX) gives him an edge in this area. While XAI's user base is currently small, its model, Grok, is considered effective by some users. The discussion also touches upon the ongoing lawsuit between Elon Musk and Sam Altman, described as the "process of the century." Musk alleges that OpenAI, initially a non-profit, betrayed its mission by becoming a for-profit entity, enriching Altman personally. The legal battle is complex, with judges aiming to determine conformity rather than justice. Musk's argument benefits from a simple moral narrative of good versus evil. Julien highlights Musk's disdain for OpenAI's initial research-focused approach and lack of security protocols. The judge's decision not to analyze AI risks and to allow public statements from both parties are noted as strategic moves. The potential for this lawsuit to set a precedent for changing the nature of foundations into businesses, with significant financial implications, makes it a highly sensitive issue. Next, the conversation turns to French unicorns, with Julien Petit's study revealing a dramatic downturn. Emmanuel Macron once boasted 38 French unicorns; now only 23 remain. Petit's analysis suggests that €123 billion in value has evaporated across Europe, with €20 billion lost in France alone. His methodology, which applies a framework to assess companies based on factors like revenue, growth, and profitability, has been criticized for not accounting for the "narrative" that often drives startup valuations. However, Petit defends his rigorous approach, stating that 60% of European unicorns have seen their valuations collapse or become inconsistent with market reality. Pierre argues that this is a logical market correction, as the traditional "product-market fit" model no longer applies in a rapidly changing world driven by AI. He proposes new criteria for evaluating companies, such as "velocity fit" (the ability of a team to adapt products quickly), "loop fit" (each usage making the product more valuable), and a dynamic assessment of competitive advantages. He blames VCs for creating overvalued unicorns by investing heavily during the "free money" period of 2020-2022, leading to the current problem of inflated valuations. The role of government in promoting unicorns as "Olympic medals" is also discussed, with a focus on France's comparison to other European countries. The UK, for instance, has lost more value and unicorns than France. The current strategy for startups, it's suggested, should be the opposite of pursuing unicorn status, focusing instead on profitability and avoiding excessive spending. While some French companies like Ledger and Exotec have succeeded by being lean and international, many local or regional European startups lack liquidity and exit opportunities. The FinTech sector, heavily invested in, is highlighted. While companies like Revolut and Trade Republic have found success, others like N26, despite significant funding, are struggling. The distinction between "spend management" (SaaS B2B tools) and true FinTech (like Revolut's comprehensive services) is made. The trend of French startups relocating to the US is examined. Nine out of 38 official French unicorns have moved their social headquarters to the US. While this offers access to a larger market and potential for higher valuations and exits (90% of French startup exits over €200 million are by American companies), it also means facing fierce competition in the "Champions League" of the US market. The example of Sweden is given as a successful European country with a strong track record of creating global champions like Lovable and Legora, thanks to a large fund, historical awareness of a small domestic market, and a global mindset. Pierre suggests that France's best chance for economic growth lies in leveraging its advanced military-industrial complex to foster innovation that can spill over into the civilian sector, rather than trying to create unicorns to compete directly with the US. He acknowledges the moral challenge of attracting engineers to work in defense but points to successful startups like Armaton in this space. Finally, the episode introduces the trend of peptides in Silicon Valley. Arnaud explains that peptides are small chains of amino acids that act as messengers in the body, regulating essential functions. While insulin is a well-known synthetic peptide, the GLP-1 class (e.g., Ozempic, Wegovy, Mounjaro) has gained immense popularity for weight loss, with a market exceeding $71 billion. These peptides regulate satiety, leading to reduced food intake. Carlos admits to using GLP-1 and experiencing increased focus, attributing it to less food intake and stable blood sugar levels. The use of peptides extends beyond weight loss to "biohacking" for optimization, including enhanced focus, improved sleep, and faster recovery (e.g., BPC-157). This trend aligns with the "age of agency," where individuals seek to optimize their biology and environment. Amy Webb's prediction that those who don't embrace human augmentation (via AI or peptides) will be at a disadvantage is cited. While GLP-1s are prescribed in the US, other peptides are less regulated and often sourced from China. The ease of self-injection and the increasing availability of oral GLP-1 formulations are democratizing their use. However, concerns about safety, lack of clinical studies for many peptides, and potential side effects (like sarcopenia with GLP-1s if not combined with strength training) are raised. The libertarian ethos in Silicon Valley, where individuals assert "sovereignty over their own bodies" to test these molecules, is noted. The discussion concludes by emphasizing the need for caution, research, and professional medical advice when considering peptide use, drawing a parallel to the risks and opportunities presented by AI.