Silicon Valley's token giants are being hammered by their own weapon: Price Dumping.
Chinese token vendors are weaponising Silicon Valley’s favourite playbook against it 1, and there’s no easy escape:
Silicon Valley’s Tech Dominance Scheme is simple and brutal:
Pick a winner, subsidise their to build scale, dump the competition to the bottom of their investor’s wallets - don’t worry, nobody has more money than Silicon Valley Techbro investors and economies of scale and data acquisition work in your favour and, voila, you get a tech duopoly that carves up the market and makes all the money back tenfold.
Worked like a charm in the last decade, is completely failing now because anything short of disconnecting international markets, markets required to support the massive valuation of these companies whose investments have reached “Must capture single digits of global knowledge work GDP to be profitable”, will be needed to make it work.
Why? Because there’s a massive economy that converts the investor subsidised token plans into API access for blocked countries while harvesting the kind of data at scale that trains better models.
Much of Anthropic's success with Claude can be attributed to the virtuous flywheel of “more real world training data from more users create better models creates more training data”, but of course, that relies on being able to keep the training data to yourself.
Which is impossible when the product you’re selling produces training data as an unavoidable side effect, as Nvidia pointed out so well in their SLM paper last year 2:
Every proprietary model is producing the information needed to train it’s replacement with every interaction.
It's the incentives...
Leaving the irony aside of the scrapers now trying to gain protection for their ill gotten gains 3 before it is too late (and the recent GLM release shows it’s probably too late already), much of Anthropic’s shrill FUD around Mythos seems to be motivated around just that ... lobby for government intervention because they can’t escape the incentive system they have created:
- By blocking access in countries, legal demands shifts to distributed grey markets, a cat and mouse game no industry has ever solved (see: Illegal streaming, torrents, etc.) and a vibe coded company running on fumes in the frontier has no hope of defeating.
- By price dumping actual costs with massively subsidised plans, the financial incentives for resellers and the potential payoff are sky high.
- By aggressively scraping for data and hyping data as a moat initially, the overall value of usage data itself creates massive incentives for middle-man operations aimed at producing eventual SLMs or fine tuned purpose models (see Cursor, etc)
The token black-markets are large enough to register
As the ChinaTalk essay points out, this shadow token economy has massive downsides, from credit card fraud to simple rug pulls - sell access to one model, serve another and the problem is pervasive enough for Chinese vendors, such as Kimi AI, to take some action: Their Vendor Verifier (linked below in further reading) was launched to enable customers to validate that they are being served the authentic goods.
A game of cat and mouse
Recent aggressive pushes for identity validation are just the last desperate attempt to defend against these dynamics, each successive layer adding more and more friction to user acquisition efforts, degrading the expensive capital subsidising the model.
The excellent ChinaTalk essay, linked below, also makes the case that the available of cheap, Silicon Valley investor financed models may be materially responsible for the low token cost offered by Chinese vendors in country.
We don't have enough information to support that conclusion at this point, at least DeepSeek has shown their economics can work, but it's definitely conceivable and would indicate global token price shocks when the era of cheap token plans ends.
The Mythos ban doesn't help, at all
Trump's shock and awe Mythos ban, whatever the real intent was, is a major own goal for US AI supremacy in this context:
Where previously companies could still pretend that just going "with the best model" was a viable strategy and locking into the corresponding harness a natural byproduct, that illusion has been replaced by clear and present business continuity / sovereignty concerns, driving a wave of interest in open source harnesses, routers (OpenRouter, Sakana Fugu), and of course open weights models. ... each one of them breaking Anthropic's iron hold on the highly valuable interaction data needed to drive continued model performance.
In the end, every player moving between user and model is capable of mounting distillation attacks, driving down model creator margins and there is no real defense because there is not real moat. Cursor, a Visual Studio Fork selling for billions to SpaceX shows just how valuable that data is for model creators.
The next act will be US threats and coercion on Europe and other countries to come to OpenAI and Anthropic’s aid or lose access to top end models. Pax Silicia, expanded to the model layer.
Chinese token vendors are weaponising Silicon Valley's favourite playbook [1] against it, and there's no easy escape: Silicon Valley's Tech Dominance Scheme is simple and brutal: Pick a winner… | Georg Zoeller
Chinese token vendors are weaponising Silicon Valley’s favourite playbook against it [^1], and there’s no easy escape: Silicon Valley’s Tech Dominance Scheme is simple and brutal: Pick a winner, subsidise their to build scale, dump the competition to the bottom of their investor’s wallets - don’t worry, nobody has more money than Silicon Valley Techbro investors and economies of scale and data acquisition work in your favour and, voila, you get a tech duopoly that carves up the market and makes all the money back tenfold. Worked like a charm in the last decade, is completely failing now because anything short of disconnecting international markets, markets required to support the massive valuation of these companies whose investments have reached “Must capture single digits of global knowledge work GDP to be profitable”, will be needed to make it work. Why? Because there’s a massive economy that converts the investor subsidised token plans into API access for blocked countries while harvesting the kind of data at scale that trains better models. Much of Anthropic's success with Claude can be attributed to the virtuous flywheel of “more real world training data from more users create better models creates more training data”, but of course, that relies on being able to keep the training data to yourself. Which is impossible when the product you’re selling produces training data as an unavoidable side effect, as Nvidia pointed out so well in their SLM paper last year [^2]: ==Every proprietary model is producing the information needed to train it’s replacement with every interaction==. ### It's the incentives... Leaving the irony aside of the scrapers now trying to gain protection for their ill gotten gains [^3] before it is too late (and the recent GLM release shows it’s probably too late already), much of Anthropic’s shrill FUD around Mythos seems to be motivated around just that ... lobby for government intervention because they can’t escape the incentive system they have created: - By blocking access in countries, legal demands shifts to distributed grey markets, a cat and mouse game no industry has ever solved (see: Illegal streaming, torrents, etc.) and a vibe coded company running on fumes in the frontier has no hope of defeating. - By price dumping actual costs with massively subsidised plans, the financial incentives for resellers and the potential payoff are sky high. - By aggressively scraping for data and hyping data as a moat initially, the overall value of usage data itself creates massive incentives for middle-man operations aimed at producing eventual SLMs or fine tuned purpose models (see Cursor, etc) ### The token black-markets are large enough to register As the ChinaTalk essay points out, this shadow token economy has massive downsides, from credit card fraud to simple rug pulls - sell access to one model, serve another and the problem is pervasive enough for Chinese vendors, such as Kimi AI, to take some action: Their Vendor Verifier (linked below in further reading) was launched to enable customers to validate that they are being served the authentic goods. ### A game of cat and mouse Recent aggressive pushes for identity validation are just the last desperate attempt to defend against these dynamics, ==each successive layer adding more and more friction to user acquisition efforts, degrading the expensive capital subsidising the model==. The excellent ChinaTalk essay, linked below, also makes the case that the available of cheap, Silicon Valley investor financed models may be materially responsible for the low token cost offered by Chinese vendors in country. We don't have enough information to support that conclusion at this point, at least DeepSeek has shown their economics can work, but it's definitely conceivable and would indicate global token price shocks when the era of cheap token plans ends. ### The Mythos ban doesn't help, at all Trump's shock and awe Mythos ban, whatever the real intent was, is a major own goal for US AI supremacy in this context: Where previously companies could still pretend that just going "with the best model" was a viable strategy and locking into the corresponding harness a natural byproduct, that illusion has been replaced by clear and present business continuity / sovereignty concerns, driving a wave of interest in open source harnesses, routers (OpenRouter, Sakana Fugu), and of course open weights models. ... each one of them breaking Anthropic's iron hold on the highly valuable interaction data needed to drive continued model performance. ==In the end, every player moving between user and model is capable of mounting distillation attacks, driving down model creator margins== and there is no real defense because there is not real moat. Cursor, a Visual Studio Fork selling for billions to SpaceX shows just how valuable that data is for model creators. The next act will be US threats and coercion on Europe and other countries to come to OpenAI and Anthropic’s aid or lose access to top end models. Pax Silicia, expanded to the model layer. [^1]: [https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens-in](https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens-in) [^2]: [https://research.nvidia.com/labs/lpr/slm-agents/](https://research.nvidia.com/labs/lpr/slm-agents/)
linkedin.com 
