The “Lobster” Era Brings A Boon For Large Model Companies

On March 10, MiniMax closed up 22% in Hong Kong trading, rallying 51% over two sessions and surpassing HKD 1,000 per share, with its market capitalization overtaking Baidu for the first time. Since listing in Hong Kong earlier this year, MiniMax’s share price has at one point risen more than 600% above its issue price.

MiniMax reported approximately RMB 540 million in revenue for 2025, a year‑on‑year increase of 158.9%, though the absolute revenue level remained modest. Market enthusiasm has centered on OpenClaw, the open‑source agent framework that has captured global attention. After OpenClaw founder Peter Steinberger reshared the PinchBench rankings, investor focus shifted from the framework itself to the model vendors powering it. PinchBench evaluates model performance within the OpenClaw environment across success rate, response speed and inference cost; MiniMax’s MiniMax‑M2.1 tied Gemini 3 for first place on the success‑rate leaderboard, a key factor behind MiniMax’s sharp two‑day advance.

Prior to the rally, MiniMax introduced MaxClaw, a managed OpenClaw offering, while Zhipu released AutoClaw on March 10; both product launches contributed to intraday share gains exceeding 20%. Subsequent security advisories from the Ministry of Industry and Information Technology prompted price pullbacks, yet institutional support remained evident: Morgan Stanley reiterated an “overweight” rating on MiniMax, and Haitong International initiated coverage of MiniMax and Zhipu with “outperform” ratings.

OpenClaw functions as a standardized agent assembly framework in which developers package discrete capabilities as “skills” that users combine with chosen large models and container environments to create agents capable of executing complex, long‑running tasks. These agents, however, drive substantial token consumption, creating a direct monetization channel for model providers. MiniMax’s AI‑native products accounted for 67.2% of revenue through the first three quarters of 2025, yet that segment’s gross margin was only 4.7%, reflecting persistent monetization challenges. Although cumulative users exceeded 212 million by the third quarter of 2025, paying users numbered only 1.77 million, yielding a conversion rate below 1%, compared with market estimates of 2.7%–3% for ChatGPT.

Enterprise‑oriented Zhipu faces different constraints: localized deployment accounted for 84.4% of its revenue, but such bespoke projects lack recurring economics and require substantial human resources. Zhipu’s cloud deployment gross margin declined from 76.1% in 2022 to –0.4% in the first half of 2025. MiniMax remained loss‑making at the end of 2025, while Zhipu reported net operating cash outflows of RMB 1.327 billion in the first half and held RMB 2.55 billion in cash and equivalents, roughly sufficient for less than one year of operations on a simple estimate.

OpenClaw’s rapid adoption has materially altered the revenue dynamics for model vendors. By early March, OpenClaw’s GitHub star count exceeded 200,000, making it one of the fastest‑growing open‑source projects in the platform’s history. Agents consume tokens at rates far higher than traditional conversational interactions because task execution requires multi‑step planning, tool invocation, visual and contextual processing, and persistent availability; token usage per task can be tens to hundreds of times greater. Reports of rapid token burn—such as 30 million tokens consumed in five hours by a news‑collection skill or 12 million tokens in ten minutes by a developer test—illustrate the scale of consumption. IDC projects that by 2031 Chinese enterprises could operate 350 million active agents, driving exponential token demand growth.

In response, vendors have introduced productized OpenClaw offerings with monetization mechanisms. MiniMax’s MaxClaw adopts a subscription model defaulting to MiniMax’s own models, while Zhipu’s AutoClaw emphasizes one‑click installation and preconfigured skills for domestic office scenarios. Major internet platforms have also mobilized resources to promote agent adoption through compatible products and installation support. For end users, however, token costs have become a tangible concern, with some reporting daily expenditures of RMB 100 or monthly bills around RMB 20,000.

From a market perspective, investor sentiment has shifted: model benchmark performance now serves as a practical selection criterion for users choosing models for agents, and model vendors are prioritizing metrics that drive agent adoption. At MiniMax’s March 10 peak, its price‑to‑sales ratio exceeded 600, implying investors were willing to pay over RMB 600 for each RMB 1 of current revenue. Company disclosures indicate that average daily token usage for the M2 series in February rose more than sixfold versus December, lifting MiniMax’s annual recurring revenue estimate from USD 100 million to USD 150 million within two months. In this environment, the emergence of OpenClaw has opened a new commercial avenue for large model companies and reshaped the competitive and monetization landscape.