The AI Smile Curve: Profit Distribution and Future Outlook

Recent reports suggest that 80% of AI startups will fail within five years, aligning with a roughly 20% three‑year survival rate for Chinese startups.

The author surveyed dozens of AI companies and observed that the industry’s high cash burn often leads to the belief that AI cannot be profitable.

Applying the classic "smile curve"—which shows high profits at the design and marketing ends and low profits in manufacturing—to AI reveals a similar pattern.

In the AI value chain, the horizontal axis consists of GPU manufacturing/cloud computing, large‑model development, and AI applications. Profit concentrates at the GPU and application ends, while the middle segment (large models) yields relatively low returns.

GPU manufacturers (e.g., Nvidia) and cloud providers (e.g., AWS) dominate the upstream, controlling essential compute resources, whereas AI application developers capture downstream market value through products such as autonomous driving, medical diagnostics, and smart home solutions.

Large‑model R&D faces high infrastructure costs, steep technical barriers, and limited monetization, resulting in a loss‑making business model.

Examples: OpenAI’s annual operating cost (~$8.5 B) far exceeds its revenue ($3.5‑4.5 B); Anthropic’s model training can cost up to $1 B per model, with future estimates reaching $10‑100 B. Chinese AI “unicorns” also burn billions without achieving profitability.

Three forward‑looking hypotheses are proposed:

GPU manufacturing valuations will likely correct as competition from AMD, Intel, and Huawei intensifies, challenging Nvidia’s dominance.

Large‑model firms will pursue consumer‑facing killer applications, leveraging their generality, technical expertise, data, and compute resources to find profitable use cases.

Approximately 80% of current AI applications will be eliminated due to "creative destruction" triggered by successive model upgrades, which raise performance expectations and entry barriers.

The AI smile curve may eventually invert, and the industry should watch for shifts in profit distribution across the value chain.