Engineering Logic Behind $35B Valuation: Kimi K3's Compute Costs and Production Pressure

$3.5 billion raised, $35 billion valuation, oversubscribed. When Bloomberg released this news, my first reaction was to flip through Kimi K3's public technical report and inference benchmarks. Looking at the parameter sheet, there are a few points worth breaking down.

First, here is a set of data, inferred from existing public information:

Model Scale: Inferred MoE architecture, total parameters ~1.2T, active parameters ~200B
Training Compute: Approx. 2.5e25 FLOPs, assuming cluster size of 100k H100s, training period 3 months
Inference Cost: Single inference (long context 128K) approx. 0.8 RMB, compared to GPT-4's $0.15
API Pricing: Kimi K3 external quote 0.03 RMB / 1k tokens, far below industry average

This pricing strategy is interesting. 0.03 RMB per 1k tokens converts to approx. $0.004, only 1/40th of GPT-4. If inference costs are indeed around 0.8 RMB per call as speculated, the gross margin for a single API call might be negative. This isn't losing money for noise; it's Moonshot AI betting on two things: first, inference optimization can compress costs by another order of magnitude; second, user stickiness in long-context scenarios is high enough to allow later price hikes.

From an engineering perspective, the most noteworthy aspect of K3 is its inference throughput performance. According to third-party evaluations I've seen, on identical hardware specs, K3's tokens-per-second output is 30%-40% higher than models of similar size. This generally comes from two places: one is the expert routing efficiency of MoE, and the other is KV cache compression technology. Moonshot AI published a paper on "Dynamic Sparse Attention" at the end of 2024, with the core idea being to dynamically adjust the number of activated attention heads based on input length, saving about 60% of compute in long contexts. If this step lands, inference costs can indeed be pressed down to below 0.3 RMB per call.

[!note] Key Conclusion: K3's engineering advantage lies not in model parameter scale, but in inference optimization. This allows it to support more users and longer contexts under the same compute budget, which is the core logic Silicon Valley VCs are buying.

Now let's look at the financing amount and usage. $3.5 billion, approx. 23.7 billion RMB at current exchange rates. If this money fully lands, Moonshot AI's plan should be:

• Compute Infrastructure Investment: Approx. $1.5 billion, used for self-building or renting larger-scale training clusters, conservative estimate of H100/B200 purchases exceeding 50k units
• Talent Recruitment: Approx. $500 million, focusing on inference engineering, chip adaptation, and system optimization
• Commercial Operations: Approx. $1 billion, including marketing, enterprise client expansion, and API channel construction
• R&D Reserve: Approx. $500 million, for next-gen model K4 preliminary research and foundational algorithm exploration

There is a hidden concern from a supply chain perspective. I've been tracking the ramp-up data of domestic AI chip production. In Q2 2025, shipments of domestic chips equivalent to H100 were approx. 30k units/month, but yield rates are still around 80%, and the proportion of chips that can stably run training is even lower. If Moonshot AI procures large quantities of domestic chips, they will likely face the dual challenge of model adaptation costs and reduced compute utilization. If they continue relying on imports, geopolitical risks will raise supply chain costs. How much of this $3.5 billion goes to "toll fees" is worth watching.

Comparing peers, Zhipu, MiniMax, and Baichuan all completed new funding rounds in the last six months, but valuations generally ranged from $1-3 billion. Moonshot AI securing $3.5 billion at a $35 billion valuation indicates investors have high consensus on K3's technical route, but it also means Moonshot AI needs faster commercialization landing to support this valuation. Currently, Kimi's C-end user growth is indeed good, with MAU exceeding 50 million in Q2 2025, but paid conversion rate is under 5%. On the B-end, enterprise API call volumes are growing, but major clients are concentrated in long-text scenarios like finance, law, and education, with limited gross margin space.

This image is inside a data center; note the heat dissipation density of the racks. If Moonshot AI wants to self-build a supercomputing center in Beijing or Shanghai, the power density per rack needs to reach over 50kW, which is a test for electrical infrastructure and cooling solutions. Did the engineering team plan liquid cooling solutions in advance? Have they signed long-term agreements with the grid?

Original link: https://www.ithome.com/0/983/170.htm