News · 2026-07-07

A startup router is giving away 100 million tokens of Kimi, MiniMax and GLM

Dahl Inference, a third-party API router, is giving away 100 million free tokens across top open-weight Chinese models - including Kimi K2.6, MiniMax M2.7, and GLM 5.2 - as a user-acquisition play. The important correction: this is not a price cut from the frontier labs themselves. It is an aggregator burning margin to pull developers off the big providers, and it is a clean signal of just how oversupplied and cutthroat the inference market has become.

Key facts

• Dahl Inference, an API aggregator/router, is offering 100 million free tokens as a promotion.
• Covered models include Kimi K2.6, MiniMax M2.7, and GLM 5.2 - all strong open-weight models.
• The offer is from the router, not the labs; its pitch reads "Stop depending on big AI. Powerful models. Prices - pennies."
• It reflects compute oversupply and a margin war among inference providers, not a coordinated move by Chinese frontier labs.

First, the attribution, because it is the part most likely to get garbled. When people saw "100 million free tokens for Kimi and MiniMax," the natural read was that the Chinese labs had slashed prices to zero to grab market share. They didn't. Dahl Inference is a middleman - a router that takes your API call and forwards it to whichever backend serves the model - and it is the one eating the cost of those free tokens. The labs make the models; Dahl resells access to them. Getting this right matters, because "frontier lab drops prices to nothing" and "reseller runs a loss-leader promo" are very different stories about the market.

To understand why a router would give away 100 million tokens, you have to understand the business it's in. Inference routers make money on the spread - they buy compute wholesale (or run open-weight models on rented GPUs) and charge a small markup. It is a commodity business with thin margins and near-zero switching costs: a developer can change one line of configuration and move to a competitor. In a commodity market with cheap inputs, the way you win is customer acquisition, and the classic tactic is the loss leader - give the product away, get developers to wire you into their apps, and hope they stay once the free tokens run out. A hundred million tokens is a marketing budget, not a technology breakthrough.

What makes it a signal rather than noise is why the loss leader is affordable right now. Giving away that much inference only makes sense if the underlying compute is cheap and plentiful - and it increasingly is. The rise of strong open-weight models like Kimi, MiniMax, and GLM means routers no longer pay a frontier lab's premium API rate; they can run the weights themselves on commodity GPUs, driving the marginal cost of a token toward the cost of electricity. When the input is that cheap, resellers can afford aggressive giveaways, and the competition among them collapses into a price war. Dahl's own slogan - "Stop depending on big AI. Powerful models. Prices - pennies" - is explicitly aimed at peeling developers away from the expensive incumbents.

This connects to the louder macro story of the week: a growing chorus, from VC newsletters to mainstream finance channels, arguing that AI infrastructure has been overbuilt and that a compute glut is coming. A router able to give away 100 million tokens is a small, concrete data point for that thesis. It is hard to square "compute is scarce and expensive" with "a middleman can afford to hand out nine figures of it for free." The open-weight model ecosystem is the mechanism turning a rumored glut into real, felt price collapse.

The caveat: a giveaway is not a sustainable price, and free tokens are the oldest trick in the developer-tools playbook. What Dahl charges after the promotion, whether its routing is reliable enough to trust in production, and whether the featured models stay competitive are all open questions. The 100 million figure tells you about the state of the market - cheap compute, thin margins, a scramble for users - more than it tells you about Dahl specifically. But as a read on where inference pricing is heading, the direction is unmistakable, and it is down.

Primary source, verified: read the paper →

Ground Truth. (2026, July 7). A startup router is giving away 100 million tokens of Kimi, MiniMax and GLM. Ground Truth. https://groundtruth.day/news/dahl-inference-gives-away-100-million-tokens.html

@misc{groundtruth:dahl-inference-gives-away-100-million-tokens,
  title  = {A startup router is giving away 100 million tokens of Kimi, MiniMax and GLM},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/dahl-inference-gives-away-100-million-tokens.html}
}

Topics: inference · pricing · open-weight-models · china · industry