Kimi K3 Tokenomics: I Spent $300 So You Don’t Have To

NerdingOut-on-AI

SubscribeSign in

Kimi K3 Tokenomics: I Spent $300 So You Don’t Have To<br>A unit-economics teardown of self-hosting a frontier open model. The GPU bill was mine. The numbers are yours, for free.

Ram<br>Jul 29, 2026

Share

The hosted API charges $15 per million output tokens. So my idea was to rent the GPUs, serve the model yourself, charge a little less than $15, and collect the spread.<br>Thanks for reading NerdingOut-on-AI! Subscribe for free to receive new posts and support my work.

Subscribe

Failed to render LaTeX expression — no expression found<br>It is not free money. I know because I rented the GPUs and measured it.<br>To find: what it actually costs to serve K3, what you’d have to charge, and why the market has already closed the gap you think you see. You don’t have to spend the $300. I already did.

TLDR: For an open model, the output price is frozen at the anchor →$15/1M and the only two moats are speed and utilization. Measured serving cost is $12–16/1M: parity, not margin. A single fully-booked 8×B300 box grosses around $8k/month at absolute best (~$50k revenue − $41k of GPU) and loses money below ~75% utilization. It can push ~3.3B output tokens/month a few thousand real users but only after you fix a latency bug that has nothing to do with the GPU. This is a scale-and-distribution business wearing the costume of a margin business. Own the GPU and you will not win. Keep it busy and make it fast, and maybe you will.

The thesis everyone has, and why it’s wrong

An open model at the frontier, weights you can just download, and a hosted API charging $15/1M output. Surely you rent the box, serve it for $5, charge $12, and undercut the field.<br>The whole thesis rests on a gap between your cost and the market price. This post is that gap collapsing to zero and the discovery that the real fight was never about price. So rather than argue, I booked the hardware. First lesson, before a single token served: finding out is not free.<br>Part 1 — The cost of measuring cost

Kimi K3 is a 2.8-trillion-parameter MoE. In MXFP4 the weights are 1.68 TB, and that single number dictates the entire hardware story:<br>It does not fit on 8×B200 (1,440 GB). The smallest single node that works is 8×B300 — 2,144 GB — at $56.79/hr .

Every cold start streams 1.68 TB onto the box. Every config change is another cold start, and another 1.68 TB.

Half the B300 nodes I was allocated came up with the NVSwitch fabric stuck initializing CUDA error 802, GPUs visible but unusable. Seems like modal’s issue.

Part 2 — Serving economics is one equation

Strip away everything and the economics reduce to a single line. Cost is fixed GPU-time; revenue is only the tokens you actually bill:<br>cost_per_1M_output = cluster_$/hr × 1e6 / (billable_tok/s × utilization × 3600)<br>I ran the saturation sweep of warm cache, concurrency cranked until throughput flat-lined. It plateaued at 1,406 output tok/s . That is the ceiling for this configuration TP8, vanilla vLLM, this workload not a law of physics, and Part 6 is entirely about beating it.

Just a plain vllm recipe, no extra optimizations : https://recipes.vllm.ai/moonshotai/Kimi-K3<br>Feed that into the equation and unit cost is a hyperbola in utilization. On 8×B300 at 1,406 tok/s, I break even against the $15 market price at roughly 75% sustained utilization . So the real cost is $12–16 per 1M output , depending entirely on how full I keep the box.<br>Part 3 — The plot twist: everyone charges exactly $15

I assumed $3-in / $15-out would look greedy next to the specialists, so I pulled the Artificial Analysis provider board for K3. Fireworks, Modal, Together, Moonshot — all $3/$15, to the dollar. For an open model, price has commoditized to the anchor. Nobody is competing on the sticker.<br>The input side is nearly free industry-wide: $3/1M raw, $0.30/1M cached, so agent/RAG traffic at ~90% cache hit lands around $0.57/1M effective. That leaves the entire competitive surface as three things — output price, output speed, and time-to-first-token — and on two of the three I get destroyed.<br>The provider board, ranked by output speed (all $15/1M output unless noted; blended = 3:1 input:output per Artificial Analysis):<br>Fireworks — 163.6 t/s · 1.09s TTFT · $2.31 blended

Modal — 142.7 t/s · 1.64s · $2.31

Nebius — 127.6 t/s · 1.66s · $4.20 blended

Kimi (official) — 32 t/s · 4.87s · $2.31

Baseten / Cloudflare — ~460 t/s (Baseten’s DFlash figure) · quote pricing

Me, naive self-host — decode fast, 3–6s TTFT · cost $12–16

Two data points deserve a hard stare. Nebius charges a premium and still sells, because it ships 128 t/s alongside it: you can price above the anchor, but only if you’re fast. And the punchline that should end the “just self-host” argument on its own, the first-party Moonshot API is the slowest provider on the board (32 t/s, 4.87s TTFT). The specialists beat the people who made the model by 4–5× on speed, at the identical price. This is...