Observation on the first day of TechCrunch Disrupt 2026: Where are the real bottlenecks in the next stage of AI? Today, I attended TechCrunch Disrupt in San Francisco and listened to several speeches on AI chips, energy, brain computer interfaces, quantum computing, and more. The biggest feeling is that Silicon Valley still has strong long-term confidence in AI, but the focus of discussion is gradually shifting from the growth of model capabilities to computing power, electricity, capital efficiency, and commercialization. Share a few observations that left a deep impression on me. Cerebras: AI computing architecture still has huge room for innovation Speaker Andrew Feldman is the co-founder and CEO of Cerebras, whose previously founded server company SeaMicro was acquired by AMD for approximately $357 million in 2012. Cerebras is currently a worthy challenger for NVIDIA in the field of AI computing architecture. He discussed whether AI can continue to scale and the limitations faced by traditional computing architectures. Cerebras adopts wafer level chip design, attempting to break through the bottleneck of data transmission and inference efficiency. My understanding is that as AI moves from training to large-scale inference, the standards for measuring the value of computing power are changing. The effective tokens, latency, and energy consumption generated per dollar are becoming increasingly important. NVIDIA's ecological advantage remains strong, but the inference market may accommodate more differentiated architectures. Bloom Energy: Electricity becomes the most realistic bottleneck for AI expansion Speaker Bill Thayer is the Senior Vice President and Head of Data Center Solutions at Bloom Energy; Ben Longmier, who was also present, is the founder and CEO of Ambrosia Energy. He participated in the development of SpaceX Starlink Mobile and co founded Swarm Technologies, which was later acquired by SpaceX. This discussion focuses on the energy opportunities brought about by the expansion of AI infrastructure. Bloom emphasizes that fuel cell solutions can reduce reliance on traditional power grid construction cycles, while also having characteristics such as low pollutant emissions, no need for on-site combustion, and low water demand. But I am more concerned about another issue: if Hyperscale and Neocloud face financing pressure, will the delay in data center projects be transmitted to energy equipment orders? The logic of long-term electricity demand is strong, and in the short term, we still need to be vigilant about the capital expenditure cycle. Brain computer interface: commercialization may be closer than imagined Speaker Max Hodak is the founder and CEO of Science Corporation, as well as a co-founder and former president of Neuralink. He has worked with Elon Musk to promote the development of brain computer interface technology. Science Corporation is developing technology to restore visual ability through retinal implantation, and has made clinical progress in recognizing text in some patients. My feeling is that there is already a fairly specific commercialization path for brain computer interfaces in the medical field. There are real needs for visual restoration, neurological disease treatment, and motor function repair. The future of human-computer interaction may undergo significant changes as a result, but clinical validation and regulatory approval remain critical processes. Quantum computing: There is still a gap between technological progress and investment returns Speaker Andrew Dzurak is the founder and CEO of Diraq and a quantum computing researcher at the University of New South Wales in Australia. He has been researching quantum computing based on silicon chips for a long time, and Diraq is committed to using mature semiconductor manufacturing processes to achieve quantum chip scaling. He discussed the possibility of commercial applications of quantum computing in the next five years and which technological milestones investors should pay attention to. I believe that quantum computing and AI have great potential for long-term complementarity, especially in the fields of materials science, drug development, and complex computing. However, the scale of logical qubits, error correction capability, and practical computational advantages still need to be continuously verified. Technological breakthroughs may still have a long way to go before listed companies can realize their revenue. The startup company took a quick look at the consumer section and found a few novel and fun options, but mainly focused on listening to speakers and visiting exhibitions tomorrow.