Chen Xun on Technology Cycles, Large Models and Founding

Zhang Xiaojun talks with founder and investor Chen Xun about technology cycles, large models and career choices. His market judgments and predictions reflect the interview period and are not investment advice.

Chen introduces his studies in physics and electrical engineering at Peking University and Stanford, and his experience moving between technology founding and investing.

He recalls his first startup and its acquisition by ASML, then his work on Silver Lake’s Asian business and SoftBank Vision Fund investments.

He argues machine intelligence will not stop with language models and may learn from more modalities or environmental interaction. This is a research outlook, not an achieved capability.

He uses rumination as a metaphor: language models learn from expressed human experience, while direct observation may provide different information. The analogy does not fully describe model training.

Reviewing investments in the deep-learning era, he says many applications previously required separately collecting, cleaning and labeling data and training specialized models, a substantial implementation effort.

He says foundation models let multiple applications reuse some capabilities, changing development cost structures; adaptation, inference, maintenance and safety still cost money.

For an enterprise, he says the first step is identifying customers and needs, then choosing a large model, small model or combination rather than pursuing size alone.

He uses the “society of minds” idea to discuss multiple models or modules competing and coordinating. His brain analogy is a possible architecture, not a settled neuroscience claim.

Combining models also requires selection, coordination and evaluation. Whether it works better than one model depends on the task, data and system design.

He discusses an ecosystem of foundation and derivative models, imagining different participants supplying general capability, domain adaptation and applications.

He describes a platform he is involved with, intended to support adaptation and collaboration between model providers and users. This is company positioning, not an independently validated performance conclusion.

He emphasizes model intellectual property, business data security, personal privacy and commercial allocation as collaboration issues; technical mechanisms and actual contracts require specific verification.

Asked whether OpenAI will lead indefinitely, he says it is uncertain whether technical leadership becomes a durable commercial barrier and competitors will also learn and develop.

He acknowledges that no company’s continued advantage or loss can be asserted with certainty; outcomes depend on future actions and market changes.

He says industries have different needs, data and workflows, leaving work between general capabilities and specific applications.

On cooling global enthusiasm, he distinguishes short-term sentiment from long-term technical development: a quieter period may give teams time to solve practical problems, but does not ensure success.

He says technological, economic and social cycles interact. The framework is a lens on history, not a precise market-prediction tool.

He thinks teams that keep validating needs and products after a boom may gain learning opportunities. Persistence alone cannot replace funding, capability and market feedback.

He discusses moving from model training and fine-tuning toward inference and use at scale, highlighting cost, performance, deployment and user experience.

His expectations for application growth over the following years are forecasts; actual timing depends on technology, demand, policy and economic conditions.

On longer-term technological change, he says it is hard to predict the year of the next major leap and calls for humility about timetables.

He discusses compute supply and demand and resource scheduling, saying better use of existing resources is one path. Hardware, energy and costs differ across regions.

He describes his company’s plan to connect compute in different locations. These are company claims and plans from that period, to be assessed against actual service and evidence.

The conversation returns to his SoftBank experience. He says investors’ judgments are shaped by cycles and organizations, while personal recollection cannot represent every decision.

He compares the 2002 startup environment with the market at the time, suggesting long-term opportunities can arise in downturns; this is not a guarantee for any project.

He links career decisions to technology trends, saying individuals should distinguish what they can control from broader environmental changes.

Chen describes investing as bottom-up: besides macro trends, investors need to study individual companies, teams and business models.

On later-stage investing, he emphasizes cash flow and mentions EBITDA as a proxy. EBITDA is not cash flow; capital expenditures, working capital and other items still require examination.

His “aesthetics” of business models means a preference for sustained value and cash generation. Real evaluation needs financial data, not narrative alone.

He compares early- and later-stage investing: the former often has less historical data, while the latter can analyze more operating performance, but both require judgment and risk management.

He discusses why he moved between investing and founding, saying different roles change one’s experience of risk, execution and organizations.

He revisits his product vision for secure use of data, compute and models, emphasizing incremental development. Its safety and effectiveness still require independent testing.

For AI founders, he advises focusing on specific problems, resource constraints and sustained operations rather than hype alone. The right path varies by team.

He discusses serving on a U.S. company board: professional work can cross cultures, while differences in shared memories and everyday conversation can still affect personal experience.

He closes by citing founders’ emphasis on saving, earning and raising money, stressing cash and financing planning. It is not a complete operating plan for every company.

The interview ends with long-term reflections on technology and entrepreneurship. His views about future models, industry and markets should be read in the context of the recording.