In the telecoms business back in the early 2000s we called them “who are we”, because that was roughly the reaction around the table whenever Huawei, then a little-known Chinese manufacturer, surfaced in a client discussion. Soon enough, the joke stopped being funny. By the summer of 2006, 20 years ago almost to the month, I was in an epic battle against them at one of our oldest clients in the Middle East. We won it, eventually… but I had to throw everything I could muster at them.
Their approach then was cheap in the most literal sense of the word. Huawei went after island nations, because giving away the best part of a 3G network for a small territory costs a fraction of attempting the same thing in a market like the United Kingdom, and a pin on a reference map looks identical either way. The pins multiplied. The technology caught up, surpassing everyone else in certain areas, and the prices rose accordingly. Huawei stopped being the cheap option a long time ago.
The same thing is happening to artificial intelligence, only faster. This month free stopped being cheap... read on.
The wrong way round
Moonshot AI, the Beijing startup backed by Alibaba, released Kimi K3 on July 16, 2026 and charged $3 per million input tokens and $15 per million output tokens. Its own previous flagship, Kimi K2.6, costs $0.95 and $4.00 respectively. In a single generation Moonshot raised the price of its best model by more than three times on input and nearly four times on output.
Eleven days later, on July 27, 2026, Moonshot published the full weights of Kimi K3, the downloadable model files, that anyone can grab and run for nothing at all. The price went up and yet the product became free in the same fortnight, from the same company, for the same model.
The rule of thumb until now was simple: Chinese models are cheap, American models are expensive, and the price premium is justified by performance. Independent evaluation now places Kimi K3 in fourth position among 189 models on the Artificial Analysis Intelligence Index. The performance gap narrowed from one direction, the price gap from the other.
Huawei’s discounts worked because the buyers had no money: telcos were still reeling from the dotcom collapse and the licence fees that came with 3G. There is a second aspect that matters more: a network is bought once, as a capital expenditure, so a good discount wins a client contract where the discount can be offset over the next decade in service fees and expansions. An AI model is charged by the token, so a price cut here is not a foot in the door; it is giving away a chunk of the bottom line, every month, year after year, for as long as the customer keeps using it.
Copy that
Distillation is the practice of training a new model on the outputs of an AI frontier lab’s model. It cuts the cost and the time of a launch enormously, because the steep investment is the first training run and the copy does not pay for it. Anthropic argues this is how Chinese labs stay within a few months of the American frontier without the volume of advanced chips that training from scratch would demand, and it has accused Alibaba, Moonshot’s backer, of running the largest known distillation attack against its models. Alibaba denies it.
So if copying is cheap, how come the copy is not? Kimi K3 costs more than three times what K2.6 did.
The reason is simpler than it looks. A model’s price has almost nothing to do with what it costs to build. The investment in training is a fixed cost paid well before the commercial launch, and it does not change whether one company uses the model or 10,000 do. Simply put, AI labs price on what they can get away with... whatever they believe the market is willing to pay!
Which is where the fine print comes in; while the open weights are free, the terms attached to them are not. Any company running Kimi K3 as a service has to go back to Moonshot and agree separate terms once its revenue passes $20m in total across any 12 consecutive months. A separate clause is triggered as a client scales: any product with more than 100m monthly users, or more than $20m in revenue in a single month, has to display the Kimi K3 name in its interface. Download it, run it, build on it... but once you succeed with it, Moonshot shows up as a partner.
The price and the bill
On July 24, 2026, Anthropic released Claude Opus 5 at $5 per million input tokens and $25 per million output, maintaining the same pricing as its predecessor Opus 4.8. It outperforms the model it replaces, and it arrived eight days after Moonshot raised its own pricing with Kimi K3… so Anthropic could reasonably have charged more for Opus 5, but it did not.
The price gap between the two models is closing. On input, Opus used to cost 5.3 times Kimi and now costs 1.67 times. On output, the difference has gone from 6.25 times to 1.67 times. However the published price list is just one element of the final bill: models differ in how many tokens they burn to finish the same job, providers extend discounts for cached prompts and batch requests, while hallucinations cost time and effort to catch.
Downloading Kimi K3’s open weights is free of charge, but running them implies a serious investment. The model needs a cluster of specialist chips to run. Moonshot’s own guidance for production is 64 accelerators or more, which in Nvidia H100 terms means eight servers: an investment between $2m and $2.6m to buy, or around $1.3m a year to rent the same capacity. None of that money goes to Moonshot, and none of it can be spent fast enough: delivery lead times on AI hardware of that class currently run six to 12 months, and the data centre space and power have a waiting list of their own.
Access denied
Which brings us to control, an argument that trumps all. A model already downloaded cannot be taken away… at least not easily. Organisations running live work on Anthropic’s Fable 5 and Mythos 5 found out the hard way back in June, when access was suspended for roughly three weeks to comply with United States export controls, and restored on July 1. A model running on a company’s own hardware is immune to that. For companies that have built processes and products around certain models, continuity is a separate criterion from price or performance, and probably the more critical one in certain cases.
That argument has powerful friends. A coalition of 25 technology companies, among them Nvidia, Microsoft, Meta, IBM, Dell and Hugging Face, published an open letter on July 24 defending distillation as legitimate practice and urging Washington not to restrict downloadable models. Nvidia’s chief executive Jensen Huang used the first post he has ever made on X to promote it, and listed sovereignty among his reasons. Within a day the list doubled to 50, with OpenAI and Google adding their names (Anthropic never did). What governments should do about any of it deserves its own issue of The Compute, together with the policy and regulatory analysis required, and it will get one.
Huawei stayed cheap for exactly as long as it needed to establish a footprint. What it charged in 2006 said more about where it stood in its own growth than about what an actual 3G network cost to build, and the same goes for every model price quoted nowadays. The token rate is the number everyone reads on the price list, but it is the one element least likely to decide what a company actually spends on it.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
Last weekend we spent a few days in the French Pyrenees, somewhere above 1,700 metres, where the air is thin and the breeze is cool. This issue was written at that altitude with the Andorran jazz pianist Kic Barroc playing in the background. His second album, Entor, was written about these same mountains, which made it the obvious thing to have on.
There was a re-watch of Project Hail Mary with the kids, as thrilling as the first time in the cinema. There were also crêpes, made à la Jacques Pépin and frankly outstanding. But it all started with an Instagram rabbit hole through Veyrix’s mashup of “Teach me How to Dougie x Only Time”, Cali Swag District from 2010 welded to Enya from 2000, which went considerably further than planned.
References
Kimi K3: Open Frontier Intelligence. Moonshot AI. https://www.kimi.com/blog/kimi-k3
Kimi K3 model weights. Moonshot AI, via Hugging Face. https://huggingface.co/moonshotai/Kimi-K3
Kimi’s open model K3 nears GPT-5.6 Sol and Fable 5 while signalling the end of super cheap Chinese AI. The Decoder. https://the-decoder.com/kimis-open-model-k3-nears-gpt-5-6-sol-and-fable-5-while-signaling-the-end-of-super-cheap-chinese-ai/
Kimi K3 Is Live: Pricing, Benchmarks, and the Wait for Public Weights. Trilogy AI. https://trilogyai.substack.com/p/kimi-k3-is-live-pricing-benchmarks
Moonshot AI releases weights for Kimi-K3. Tom’s Hardware. https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-ai-releases-weights-for-kimi-k3-firing-a-shot-across-the-bow-of-openai-and-anthropic-open-weight-model-performs-almost-as-well-as-frontier-models-while-being-2-3x-easier-to-run
Introducing Claude Opus 5. Anthropic. https://www.anthropic.com/news/claude-opus-5
Anthropic’s Claude Opus 5 AI model rivals Fable 5 and is cheaper. CNBC. https://www.cnbc.com/2026/07/24/anthropic-claude-opus-5-ai-fable-5-cost.html
Our position on open-weights models. Anthropic. https://www.anthropic.com/news/position-open-weights-models
Nvidia and 24 other companies sign open-weights letter as Washington weighs Chinese AI model ban. Tom’s Hardware. https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-and-24-other-companies-sign-open-weights-letter-as-washington-weighs-chinese-ai-model-ban
Huang’s open-weights letter doubled to 50 without Amazon and Anthropic. Forbes. https://www.forbes.com/sites/sandycarter/2026/07/25/huangs-open-weights-letter-doubled-to-50-without-amazon-and-anthropic/
NVIDIA H100 Server Price in 2026: What an 8-GPU HGX System Costs. Mercatus. https://www.mercatus-ai.com/blog/h100-server-price
Kimi K3: benchmarks, pricing, hardware requirements, and self-hosting. Northflank. https://northflank.com/blog/what-is-kimi-k3-self-hosting
Update on Fable 5 and Mythos 5 access. Anthropic. https://www.anthropic.com/news/fable-mythos-access