Pierre Ferragu

Pierre Ferragu asked for insights into Cisco's commercial momentum on the AI front, specifically how Cisco wins typical use cases with large customers and hyperscalers, highlighting differentiators. He also inquired about the dynamics in new cloud opportunities (NeoCloud, Enterprise), asking if it's a similar use case-focused competitive game and if Cisco is a primary or secondary source for networking for AI clients.

Answer

Chuck Robbins (Chairman and CEO, Cisco Systems) explained that Cisco wins in AI due to its well-received optics portfolio (number one in the world), silicon diversity, programmability of silicon, power efficiency of systems, and strong commercial relationships. For NeoCloud and Enterprise, he noted opportunities across AI Factory with NVIDIA, emphasizing Cisco's role as the network standard when partnering with NVIDIA in the enterprise, and the opportunity to bundle security solutions like AI Defense.

Pierre Ferragu asked for insights into Cisco's commercial momentum on the AI front, specifically how it wins typical use cases against competition, highlighting differentiators. He also inquired about the dynamics in new cloud opportunities, such as neoclouds and broader markets, asking if the competitive landscape is similar and if Cisco is seen as a primary or secondary source for AI networking.

Answer

Chairman and CEO Chuck Robbins explained that Cisco wins through its well-received optics portfolio, silicon diversity desired by customers, programmability of its silicon, power efficiency, and strong commercial relationships. For enterprise AI, he noted opportunities across AI Factory with NVIDIA, where Cisco is often the network standard. He also highlighted the opportunity to bundle security solutions like AI Defense into these AI Factory solutions.

Pierre Ferragu asked why the Model 3 and Model Y have not captured a larger share of their addressable market, similar to how the iPhone dominated smartphones, given their product superiority.

Answer

CEO Elon Musk responded by comparing the current automotive landscape to the pre-smartphone era of flip phones. He argued that the future is not about competing for market share in traditional cars, but about the transition to autonomy, predicting that non-autonomous gasoline cars will become a rarity, much like riding a horse or using a flip phone.

Pierre Ferragu of New Street Research inquired if the upcoming robotaxi service in Austin would be open to public use and asked if Tesla is developing an 'eyes-off' FSD feature for drivers to use for personal tasks like checking email.

Answer

CEO Elon Musk clarified that the initial robotaxi launch will be for Tesla's internal fleet to ensure reliability, with public participation likely next year. He acknowledged the demand for an 'eyes-off' mode and the current 'perverse situation' where drivers disengage FSD to text. He stated that while the capability is not far off, likely a matter of 'low single-digit months,' the company must first prove unequivocal safety to itself and regulators.

Pierre Ferragu asked about the application of Tesla's expanding AI compute power and the operational plan for the ride-hailing network launch in Texas and California.

Answer

CEO Elon Musk explained that real-world AI for vehicles differs from LLMs, requiring heavy training to compensate for limited in-car inference compute. Head of Autopilot Software Ashok Elluswamy added that a key bottleneck is the time required to validate model improvements through real-world mileage. Regarding the ride-hailing launch, an executive noted they will follow all state regulations for safety drivers, but Elon Musk affirmed the goal is to have driverless paid rides sometime next year.

Pierre Ferragu from New Street Research asked for a high-level breakdown of compute allocation between pretraining, reinforcement learning, and inference for large AI models, and which area is seeing the most growth.

Answer

CEO Jensen Huang explained that today, the compute workload is 'vastly in pretraining' because advanced post-training and inference-time scaling techniques are still new. He noted that the ultimate goal is to minimize inference cost by maximizing pre- and post-training. However, he expects all three areas to scale for the foreseeable future, driven by trends like multimodality, which will require more compute and necessitate NVIDIA's continued drive to improve performance and reduce costs.