Inference Will Consume 70% of Compute, Leaving 30% for Training: Silicon Valley Investor Zhang Lu at AIGC2026
According to Fusion Fund founding partner Zhang Lu, AI compute demand is shifting from training to inference, which will account for 70% of future compute. Data center communication consumes up to 100x more power than computing itself, making optical communication critical. Physical AI's bottleneck is high-quality real-world data; synthetic data cannot replace edge-case collection. She highlights healthcare, space, and nano-robotics as key application areas. The speed of industry integration is the true competitive edge for AI deployment.
At the 2026 China AIGC Industry Summit, Fusion Fund founding partner Zhang Lu delivered a keynote that reoriented the AI narrative. She argued that the industry has been fixated on models and compute, but the real battlefront is shifting to the "communication layer" of infrastructure and the "data layer" of the physical world.
Zhang projected that inference will soon consume 70% of total AI compute, up from roughly half today, with training dropping to 30%. Unlike training—a one-time investment—inference creates sustained demand, especially as AI agents replace chat interfaces and require always-on responses. This shift makes inference optimization the central challenge for AI infrastructure.
A startling insight from Zhang's talk concerned energy consumption in data centers: communication between chips consumes up to 100 times the power of computation itself. She noted that Alphabet Chairman John Hennessy had previously highlighted this inefficiency. Moving data is far more energy-intensive than processing it locally, which explains the growing importance of optical communication technologies that drastically cut power usage.
Turning to physical AI—spanning autonomous driving, manufacturing, healthcare, and space—Zhang identified the primary bottleneck as data. While synthetic data is advancing rapidly, it cannot replace high-quality real-world data from edge scenarios. She emphasized the need for novel data collection platforms and edge computing solutions. Small models running on devices like Raspberry Pi can now achieve GPT-4-level capability, enabling localized data processing and privacy-preserving AI deployment.
Zhang pointed to three sectors ripe for AI disruption. First, healthcare: with its wealth of high-quality data, it has attracted major partnerships like Eli Lilly–Nvidia and Merck–Google Gemini. Vertical AI models now target specific therapies (e.g., cell therapy, Parkinson’s) and incorporate physical AI for lab automation—such as Medra's robot-run laboratory in San Francisco. Second, space technology: the coming decades will see AI-native and robot-native infrastructure for orbital factories, exploration, and in-space services. Third, micro and nano robots: tiny machines that can navigate blood vessels to clear clots or deliver drugs at the molecular level are entering early commercialization.
In closing, Zhang stressed that technological innovation is only the starting point. What truly drives AI adoption is the accelerating pace of industry integration. Fortune 500 companies are increasing AI budgets from tens of millions to billions of dollars, and procurement cycles have shrunk from months to weeks. This rapid deployment generates the real-world data and feedback loops that fuel iterative improvement, making speed of integration the ultimate competitive advantage.