Future Inference Will Consume 70% of Compute, Leaving 30% for Training | Silicon Valley Investor Zhang Lu at AIGC2026
At the 2026 China AIGC Industry Summit, Zhang Lu, Founding Partner of Fusion Fund, highlighted that the focus of AI compute demand is shifting from training to inference, with inference expected to account for 70% of compute. Communication in data centers may consume 100 times more electricity than computation, making technologies like optical communication critical. The biggest bottleneck for physical AI is the scarcity of high-quality real-world data. Healthcare, space, and nanorobots are the three most promising application directions.
Article intelligence
Key points
- Inference compute share will rise from 50% to 70%, becoming the core optimization target for AI infrastructure.
- Communication in data centers can consume over 100 times more electricity than computation, driving innovations like optical communication.
- Physical AI faces a data bottleneck: lack of high-quality real-world data, synthetic data cannot replace edge collection.
- Healthcare, space, and nanorobots are the top three AI application areas favored by the investor.
Why it matters
This matters because inference compute share will rise from 50% to 70%, becoming the core optimization target for AI infrastructure.
Technical impact
May affect model selection, inference cost, product capability, and evaluation benchmarks.
The narrative of AI is undergoing a quiet shift, according to Zhang Lu, Founding and Managing Partner of Fusion Fund, a Silicon Valley venture capital firm. Speaking at the 2026 China AIGC Industry Summit, she shared her front-line observations on the new cycle of AI innovation, emphasizing that the real battlefield is moving from models and compute to the "communication layer" of infrastructure and the "data layer" of the physical world.
Zhang Lu pointed out a fundamental change in compute demand: training was once the dominant consumer of compute resources, but inference is rapidly outstripping it. She predicted that inference would account for 70% of total compute demand in the future, up from about 50% today. This shift is driven by the transition from conversational AI to agentic AI, where agents require always-on, continuous inference. Optimizing inference compute is therefore becoming a central challenge for AI infrastructure.
Another surprising insight concerns energy consumption in data centers. While much attention is paid to the electricity used by computation itself, Zhang Lu highlighted that communication—the transfer of data between chips and servers—can consume 100 times more power than computation. Citing a conversation with John Hennessy, Chairman of Alphabet and former President of Stanford, she noted that moving data is far more energy-intensive than processing it. This creates a huge opportunity for new communication technologies, especially optical communication, to drastically reduce power consumption.
For physical AI—which encompasses robotics, autonomous driving, and factory automation—the biggest bottleneck is not architecture or compute, but data. Synthetic data has limitations and blind spots, making the collection of real-world edge data essential. Zhang Lu emphasized the need for new data collection platforms and sensors, such as the high-precision, low-power artificial skin developed by Professor Zhenan Bao's lab at Stanford, which can provide rich tactile data for physical AI training.
Looking at application directions, Zhang Lu identified three key areas: healthcare, space, and nanorobots. Healthcare, she argued, is not just a large market (20% of US GDP) but also a goldmine of high-quality data. AI companies are increasingly partnering with pharma giants like Eli Lilly and Merck to build vertical AI models for specific diseases or treatments. In space, the rise of the space economy—driven by upcoming IPOs like SpaceX—will create native demand for AI and robotics, from infrastructure assembly to space factories. Nanorobots, still in early stages, hold promise for medical applications such as clearing blood clots or targeted drug delivery at the DNA level.
Finally, Zhang Lu stressed that beyond technology innovation, the speed of industry integration is now the true competitive advantage. Fortune 500 companies are ramping up AI budgets to billions of dollars and compressing procurement cycles from months to weeks, providing the real-world feedback loop that fuels further innovation. She encouraged more cross-border collaboration between China and Silicon Valley to accelerate this transformation.