Tech • AI • Robotics • Game
Kimi K3.1 Pro: What Moonshot is hiding behind its strategy
8/10TL;DR
Moonshot is trying to monetize the giant Kimi K2 3.1 Pro by licensing open weights to mostly American cloud hosts because export controls and hardware costs leave it without enough Nvidia capacity to serve the model itself.
KEY POINTS
A model too large for standard deployments
Kimi K2 3.1 Pro currently relies on a 1.56 TB checkpoint, beyond the memory of a standard 8x H200 setup capped at 1,128 GB. Deployment guidance points to at least 8 B300 chips or 16 across two nodes, while production use is said to require roughly 64 accelerators. Each generated token activates about 104 billion parameters, more than triple the 32 billion active in Kimi K2 about fifteen months earlier.
Export controls create a hardware trap
The recommended B300 Blackwell chips are exactly the class of advanced hardware covered by US export restrictions to China. That leaves Moonshot in a bind: it can train frontier-scale systems through leased capacity, but it lacks enough compliant infrastructure to run them broadly for customers. The result is a business model shaped as much by geopolitics as by model quality.
Training capacity existed, inference capacity does not
Reporting has linked Moonshot’s training stack to roughly 20,000 Nvidia GPUs leased via Alibaba Cloud, a major shareholder. The issue is not whether the company could build the model, but whether it can profitably serve it at scale afterward. Kimi K2 3.1 Pro also appears to reuse the same base architecture as Kimi K3, with gains coming mainly from post-training rather than a fresh pre-training run.
Performance is strong, speed is weak
The model is expensive to serve and relatively slow. Public measurements put it near 23 tokens per second with close to 8 seconds of latency, while internal figures cited by the company are around 41 tokens per second. Even the premium High Speed version, sold at double the price, appears to top out around 31 tokens per second in third-party rankings, placing it near the bottom of tested models for throughput.
Moonshot is already rationing access
Limited inference capacity has pushed Moonshot to divert users toward lighter models. Since September 11, the coding tool Kimicode has reportedly defaulted to Kimi K2.8 Preview, with many Kimi K3 requests redirected unless users explicitly enable deeper reasoning. That lighter model was estimated to account for 40% to 45% of September recurring revenue, versus 30% to 35% for Kimi K3.
Open weights double as outsourced infrastructure
The planned weight release on November 4 is central to the company’s strategy. When the earlier Kimi K3 became available, more than 20 hosts, mostly American, began serving it. The logic is straightforward: if Moonshot cannot buy enough chips, it can let foreign cloud providers run the model on their hardware and still collect revenue through licensing.
The license carries a revenue-sharing clause
The “open” approach is constrained by commercial terms. For hosts above $20 million in annual revenue, Moonshot can require custom agreements claiming up to 30% of proceeds. That turns open weights into a form of rent extraction: cloud providers supply the scarce compute, while Moonshot monetizes the right to operate the model.
Lower sticker prices do not mean cheaper use
List pricing fell by about a third to $2 per million input tokens and $10 per million output tokens. But actual task costs appear slightly higher than the prior model because of heavy reasoning overhead. This “overthinking” effect can inflate token consumption without proportional quality gains, meaning customers pay more in practice than list prices suggest.
Margins are being squeezed
Estimates put gross margins for Kimi K3 at roughly 35% to 40%. With service costs still high and list prices cut, margins for Kimi K2 3.1 Pro may fall to about 2.5% to 10%. The gap is currently supported by investor capital, with Moonshot reportedly carrying a valuation near $50 billion and aiming for a Hong Kong listing in early 2027.
Coding gains sharpen cyber risk
Reinforcement learning reportedly pushed coding performance on Terminal Bench 4.0 from 17% to 64%, while legal research and tax benchmarks fell by 9 points each. On cyber evaluations, a safety-limited version still reached 81% exploit effectiveness, versus 90% without filters and 33% for the earlier generation. That leaves the model positioned as highly capable in agentic coding and terminal tasks while trying to remain just below the most politically sensitive threshold.
CONCLUSION
Moonshot is turning hardware scarcity into a licensing business, using open weights to shift inference onto foreign cloud providers while preserving a share of revenue. The success of Kimi K2 3.1 Pro may depend less on its posted API price than on who can host it, at what speed, and under what regulatory pressure.
Ask a question