In 2026, the pace of AI application adoption is exceeding most people’s expectations. Token consumption has grown tenfold in six months, coding agents are replacing integrated development environments (IDEs), and the open-source project OpenClaw completed an entire lifecycle in just four months—from large models to the application layer, every link in the industry chain is being reshuffled. Against this backdrop, the logic by which frontline investors evaluate the sector is also changing.
Recently, at the launch ceremony of the third cohort of the Xingdi AI Application Incubator, Liu Yi’ang, a partner at China Creation Ventures (CCV), delivered an in-depth presentation on “AI Application Entrepreneurship Directions and Investment Opportunities.” He began with the explosive growth in token consumption, traced OpenClaw’s complete cycle from viral success to retreat, analyzed the divergent paths of the Chinese and American AI industries, and concluded with specific advice for AI entrepreneurs.
Token Consumption Scales Up Dramatically, Inference Costs Keep Falling
Liu Yi’ang presented a set of striking figures during his talk. From 2024 to 2026, token consumption grew 150-fold. Using OpenRouter, a US-based token routing and distribution platform, as an observation window, weekly consumption has risen from 2.1 trillion to 24.5 trillion tokens—a more than tenfold increase in six months. He pointed out that no infrastructure in the past—whether power grids, internet traffic, or computing power—has ever achieved this kind of six-month growth rate. Daily token consumption in March has already surpassed 140 trillion.
At the same time, costs are falling rapidly. When ChatGPT 4 was first released, the cost of one trillion tokens was close to 20 yuan (approximately $3); today it has dropped to 0.4 yuan. OpenAI’s new model 5.6 Luna, released at the end of July, cut prices to one-fifth of the previous level. The entire industry’s inference costs are on a steep downward trajectory.
The more critical change is in the consumption mix. Initially, users primarily used AI for chat, which consumed very few tokens; today, a single agent task consumes 50 to 100 times more tokens than a chat scenario. Long-duration tasks—for example, AI continuing to execute work in the background while the user sleeps—can consume tens of thousands of times more tokens overnight than a chat session. Video generation consumes even more, and Chinese companies Kling and Seedance 2.0 have already captured a substantial share of the global market.
Some institutions predict that China’s inference market will grow 370-fold from 2025 to 2030, but Liu Yi’ang believes this forecast is conservative. In terms of volume, China’s overall token consumption far exceeds that of the US, yet the US’s annual recurring revenue (ARR) far surpasses China’s—a difference of orders of magnitude.
Diverging Paths Between China and the US: Big Tech Ecosystems and the Startup Window
Liu Yi’ang noted that the US market is primarily contested by three major players—OpenAI, Anthropic, and Google—whose direction is to build high-capability models with premium pricing. China, by contrast, continues to leverage its traditional strengths: the ability to fight price wars and strong engineering capabilities. Looking at revenue structures, US vendors pursue a high-pricing strategy, while Chinese vendors’ revenue comes mainly from enterprise private deployments. Willingness to pay among C-end consumers and small companies in China is relatively weak, with major customers concentrated among large corporations and state-owned enterprises.
The industry chain is also diverging. US model vendors have already welded models and ecosystems into a unified product. While Chinese tech giants are attempting something similar, they still primarily “sell utilities”—water, electricity, and coal. This means there are still opportunities for entrepreneurs along the industry chain; it has not yet reached a winner-take-all stage.
One of the biggest themes in the first half of 2026 is the workflow shift from chat to agents. Coding agents are the largest global theme and the core battleground for tech giants. Codex, Claude Code, and Cursor are fiercely competing, but the landscape has undergone a subtle shift—the new narrative is “we don’t need IDEs anymore.” New harness tools like Claude Code and Codex can already largely replace traditional IDEs, and traditional IDE paths like Cursor, which lacks proprietary model capabilities, face a massive crisis.
Video generation is another high-growth direction, particularly favored by tech giants because token consumption is enormous. OPC (Outcome-as-a-Service) for customer service and marketing has evolved from concept to industrial scale, with many local governments in China building OPC industrial parks. Liu Yi’ang gave an example: a two-person company serving ten clients can do quite well. If this trend persists, it could disrupt some of the existing dynamics in the SaaS market.
AI for Science has also just emerged this year. The latest generation of AlphaFold no longer permits commercial use, and the team has spun out to form a new company that raised over $2 billion this year. Similar companies in China are also moving quickly.
The Rise and Fall of OpenClaw and an Industry Warning
Liu Yi’ang personally experienced OpenClaw’s complete journey from viral sensation to retreat: he started using it in February and switched to Hermes in May. This individual open-source project surpassed 300,000 GitHub stars within four months, with a peak of 17,000 new stars in a single day—an epic breakthrough in the open-source world. It even sold out the Mac mini—a product line Apple had nearly discontinued, which was revitalized as a result.
More importantly, OpenClaw gave a boost to Chinese-made large models. In the first week after Chinese New Year this year, Kimi’s revenue exceeded its total revenue for all of 2025. The shift from conversation to agents is a topic worth revisiting repeatedly this year and over the next five years.
But OpenClaw quickly retreated. Liu Yi’ang believes security concerns may have been only one factor. As a user, he experienced two core pain points: first, version updates once occurred daily, and after each update all plugins stopped working and required reconfiguration—a painful process; second, token consumption was a black box—users had no idea what the AI was doing, and an infinite loop could blow through five hours’ worth of token quota in five minutes.
He concluded that this is the fate of independent open-source software. OpenClaw’s rise and fall warrants self-reflection, but for the industry as a whole it serves as an alarm—this industry is changing.
After OpenClaw’s retreat, the agent industry as a whole continues to grow. Anthropic is now the international leader, with ARR reaching $45 billion, of which Claude Code—a software product—accounts for over $4 billion in ARR. Note that this is not model revenue but software product revenue, achieved in just 14 months since launch. Codex is also updating frequently and catching up—a compelling two-horse race.
Token Distribution Will Determine Industry Structure; Six Directions Offer Startup Opportunities
Liu Yi’ang made a core judgment: the future AI industry ecosystem will be built on a token distribution business model. Whoever controls the most critical nodes in token distribution will capture the largest rewards in this industry chain. This is the same logic as internet-era giants controlling traffic distribution nodes.
But token distribution differs structurally from traffic distribution. First, the switching cost for large models is extremely low—one command and you’ve switched, essentially seamless and imperceptible. Large models themselves have almost no network effects—unlike WeChat, where users simply cannot switch away. Second, token production has a clear cost; economies of scale exist but are not as strong as in traditional internet industries.
Today’s industry chain architecture is changing: model providers → token routers/factories (Alibaba Cloud’s Bailian, ByteDance’s Volcano Engine, SiliconFlow) → agent layer (Work Buddy, Codex, Claude Code, open-source projects) → end applications. End applications are also distributing tokens; many applications are essentially “wrapping a shell” around token resale. Whether this model can hold remains to be seen.
In the Chinese market, Liu Yi’ang cautioned that while everyone is discussing Zhipu and MiniMax, the real players to watch are Alibaba Cloud and ByteDance—these are the largest ecosystem players in China today. DeepSeek represents a different approach: models only, ultra-high cost-performance, concentrated force for single-point breakthroughs—a strategy inseparable from the capabilities and architecture of Huawei’s Ascend chips in China.
| Startup Opportunity Direction | Core Logic |
|---|---|
| Agent Interaction Layer | MCP, ACP, and other protocols remain immature; the endgame for inter-agent payment protocols is undecided |
| Memory Layer | Large model switching costs are low, but memory switching costs are extremely high; a unified memory layer holds enormous value |
| Ops Layer | The software industry faces earth-shaking change; both MMOPS and LMOPS layers offer strong opportunities |
| Edge Device Layer | Every token produced has a cost; demand for on-device and edge-cloud computing power is exploding |
| Video Generation | Chinese companies hold a monopolistic position globally; a mid-stream, high-certainty track |
| AI for Science | Relatively independent industry chain; requires deep industry penetration to replace low-efficiency links |
Note: The above directions are compiled from Liu Yi’ang’s presentation and represent his personal assessment of AI startup opportunities.
On the edge device layer, Liu Yi’ang specifically noted that the Qwen 35B model can already solve a great many problems. This year’s Mac mini sellout demonstrates that demand for on-device model inference computing power is exploding. Model capability improvements, on-device hardware improvements, and KV cache optimization—advancing on both fronts simultaneously—will create enormous opportunities in this direction.
In video generation, Kling (spun off from Kuaishou) launched in June 2025 and has already surpassed $700 million in ARR. Chinese companies now hold a near-monopolistic position globally. This is a mid-stream, high-certainty track outside of foundation models and big tech competition.
In AI for Science, AI-driven drug discovery has developed rapidly over the past two years. Some Chinese companies, combining AI efficiency with China’s execution efficiency, now have pipelines approaching those of top overseas pharmaceutical companies. However, he cautioned that using AI merely for tools and software is too thin—entrepreneurs must genuinely penetrate industries to replace low-efficiency links.
Finally, Liu Yi’ang issued a clear warning to entrepreneurs: coding has become a must-win battleground for tech giants. If entrepreneurs discover interesting opportunities, he welcomes discussion—but they must be mindful of competition from the giants.