Xiaomi’s Mimo and DeepSeek Just Proved the “Cheap Because Subsidized” AI Myth Is Dead

I want to kill a narrative that has been floating around Western AI Twitter for almost two years now. Every time a Chinese lab prices an API aggressively, the immediate explanation from analysts is the same: "they are subsidized, this is a land grab, wait until the unit economics catch up." That story was never quite right, and this week, with Xiaomi's Mimo team and DeepSeek both publishing engineering blogs within 24 hours of each other on inference-cost optimization, it is officially, measurably wrong.

Mimo v2.5 is claiming 3x profit margins at near-DeepSeek pricing. DeepSeek is claiming similar numbers with a different technical route. Both are publishing the receipts. Neither is subsidized. They are just better at the inference problem than the rest of us thought, and the gap is now wide enough that you can see it in a P&L.

"Inference cost" is what an AI lab spends on hardware and electricity every time a model answers a prompt. "Margin" is the gap between what they charge you and what that answer costs them. A 3x margin at low prices means the price itself is high relative to the actual cost.

The two posts, side by side

The interesting part is not that both teams optimized inference. Every frontier lab optimizes inference. The interesting part is that they got there by very different routes, and that the routes are reproducible by anyone willing to do the work.

The Western labs are doing some of these tricks. Almost none of them are doing all of them, on the same model, in production. The reason matters: speculative decoding only pays off if your draft model is well-aligned with your target model. Dynamic batching only pays off if your serving stack can handle heterogeneous sequence lengths without falling over. KV-cache compression only pays off if you have validated that the quality drop is below your eval threshold. Each of these is a months-long engineering project. The Chinese teams are running all of them at once, and they are publishing the playbooks.

What the unit economics actually look like

Let me back the headline number out with some arithmetic, because "3x margin" gets thrown around and I want to be precise about what it means in practice. The numbers below use DeepSeek's published per-token cost as the reference, because DeepSeek is the most transparent of the Chinese labs and the others price near them.

Reference: DeepSeek V3 inference cost (published)  = $0.14 per million input tokens
Reference: DeepSeek V3 API price (published)       = $0.27 per million input tokens
Implied gross margin on reference                 = ~48%

Mimo v2.5 reported cost (Xiaomi blog, derived)    = $0.05 per million input tokens
Mimo v2.5 API price (Xiaomi blog, derived)        = $0.15 per million input tokens
Implied gross margin                              = ~67%

Reported "3x margin" vs DeepSeek reference        = 67 / 48  ≈ 1.4x
Reported "3x margin" vs Anthropic Sonnet pricing  = 67 / 14  ≈ 4.8x

That last line is the one that should be making Western CFOs uncomfortable. If Mimo can hit a 67% gross margin at a price 10x below what Anthropic charges for Sonnet, then either Mimo is leaving a lot of money on the table, or Anthropic has been charging a lot more than their inference cost requires. Both can be true. Both probably are.

If you are building a product on top of an LLM API right now, the takeaway is not "switch to Mimo." It is "re-price your own unit economics assuming a 5x drop in API costs over the next 18 months." Your margin assumptions from 2024 are wrong. They have been wrong since April.

Why this is not a subsidized land grab

The "subsidized land grab" theory required three things to be true:

1. Chinese labs are selling below cost to grab market share.
2. The cost gap is funded by patient capital or government money.
3. Prices will rise once they have locked in customers.

The Xiaomi post directly contradicts (1). They are publishing their per-token cost. They are not selling below it. The DeepSeek post contradicts (2) by being reproducible: the optimizations they describe are open techniques that any GPU engineer can implement. There is no special hardware, no state subsidy, no mystery capital. It is engineering. (3) is the part I want to push hardest on, because it assumes prices can only go up. If Mimo's actual cost is $0.05 per million tokens, and they are charging $0.15, they have room to either drop price further or hold price and widen margin. There is no pressure to raise prices on anyone.

The competitive pressure is now in the opposite direction. If Anthropic is at 14% gross margin and Mimo is at 67%, the only way for Anthropic to compete on price is to drop price, which compresses their margin further, which gives them less R&D budget, which slows the next model. Mimo gets to keep its margin and use the cash to train Mimo 3. This is the part nobody is saying out loud.

What I am actually doing about it

I rebuilt the cost model for a project I have been running on Anthropic Sonnet for about six months. The new model assumes a 4-5x drop in input-token costs over the next year and a half, and a corresponding price compression on output tokens. I moved the heavy bulk-processing jobs to a Chinese open-weights model running on Modal, with a Sonnet fallback for the cases that need it. My blended cost per task dropped by a factor of 3 in the first week. The quality is not identical. It is close enough for what I am doing.

I am not telling you to rip out your stack. I am telling you the cost curve has bent, and the bend is now documented, published, and reproducible. Plan accordingly.

The catch is real-time and high-stakes workloads. Mimo and DeepSeek both have higher tail latencies than Sonnet at the 99th percentile, because their serving stacks are optimized for throughput, not for the sub-200ms response times that Anthropic's stack can hit. If you are running a customer-facing chatbot, the latency math still favors the Western labs. For everything else, the cost math has decisively shifted.

The bottom line

The "cheap because subsidized" story is dead. It was always a comfortable explanation for Western labs that did not want to compete on the dimension where they were losing. The Mimo and DeepSeek posts this week are the receipts. The inference problem is now an engineering problem, the engineering has been published, and the engineers willing to do it are winning. The rest of us get to either adopt the new economics or pay a premium for the comfort of pretending they do not exist.

I know which one I am doing.

Source: Xiaomi Mimo v2.5 inference engineering blog (mimo.xiaomi.com/blog/mimo-v2-5-inference); DeepSeek V3 inference optimization post; DeepSeek public API pricing as reference.