Three US providers host Moonshot's Kimi K3 at a tenth the cost of going to the source

Hosting a 2.8-trillion-parameter Mixture-of-Experts model is a memory problem before it is a compute problem. K3 routes each token through 16 of its 896 experts ^[3], under 2 percent of them ^[14], activating roughly 104 billion parameters per forward pass, about 3.7 percent of the total ^[13]. The compute bill is sized by that 104 billion. The memory bill is sized by the whole file, which runs past 1.4 terabytes even after MXFP4 quantization ^[5], and Baseten has said deployments need multi-GPU setups, eight Nvidia GB300 accelerators in concert just to hold the model in memory ^[6]. Split 1.4 terabytes eight ways and each card is carrying at least 175 gigabytes ^[15].

Modal reported 460 tokens per second on its K3 deployment, using a speculative decoder it built for the architecture and calls DFlash ^[8]. At $15 per million output tokens ^[9], one stream at that rate needs about 36 minutes to produce a million tokens and bills roughly $25 for an hour of continuous generation ^[16]. Input is $3 per million and a cached token is 30 cents ^[9], so filling the 1-million-token context window once costs $3 and re-reading it on a cache hit costs $2.70 less ^[18].

The one-tenth figure is the softest number in the account. Crypto Briefing reports the three providers serve K3 at roughly a tenth of the cost of direct access ^[1]; the article does not include Moonshot's own list price. Run the multiple backwards and direct access would sit near $30 per million input tokens and $150 per million output ^[17].

The publisher's reading of the gap is hardware: US providers have Nvidia's GB300 NVL72 systems and AMD's MI350X and MI355X parts, which export restrictions keep from Chinese companies including Moonshot itself ^[7]. The other reading is that the chips explain the speed while compliance explains the sale. All three providers market zero-data retention, dedicated capacity and US-hosted endpoints ^[10], and all three speak OpenAI-compatible APIs, so an existing integration can be repointed with minimal code changes ^[11]. I'd put more weight on the second: a healthcare or government buyer that cannot send prompts to a Chinese endpoint at any price is not comparing $3 with $30. If Moonshot's published price turns out to be within a factor of two of $3 in and $15 out, the export-control story loses most of its force.

Moonshot is not collecting the serving fee on its largest model. It published 2.8 trillion parameters on July 27, 2026 ^[2], the open-weights licence lets anyone download and run them ^[20], and three American infrastructure companies did exactly that ^[1]. The weights travel free. The $15 per million output tokens is booked by whoever owns the eight-GPU box ^[9].