Why AI is Collapsing: How China is Winning.

The AI Spending Avalanche and Early Warnings [00:00]

• Companies are experiencing a rapid and unmanageable surge in AI expenses.
• Tesla has implemented a strict $200 per week cap on employee AI usage, a move that signals significant cost concerns for a company heavily invested in AI.
• Uber reportedly spent its entire $3.4 billion AI budget within a mere 4 months, highlighting an extreme overshoot in financial projections.
• Microsoft has ceased offering its engineers access to certain AI coding tools due to prohibitive costs.
• Meta has also begun monitoring and restricting its staff's AI consumption.
• Meanwhile, a Chinese food delivery app has successfully trained a 1.6 trillion parameter model using chips that are subject to US export bans, and then released it for free.

This is what the top of a bubble looks like.

The AI Cost Cycle Explained [01:22]

• The emerging AI cost cycle involves companies rushing to adopt AI, leading to exploding bills, followed by spending caps and a default to cheaper Chinese models.
• This trend is increasingly seeing AI models trained on Chinese chips and GPUs, progressively excluding US AI companies from the market due to China's significantly lower operational costs, estimated at 10 times cheaper.
• While American labs may still produce the best "frontier" AI models, their high cost is becoming a significant barrier when a "good enough" alternative is available at a tenth of the price.

The question is whether the best matters when good enough costs a tenth as much and your CFO just found the invoice.

Corporate AI Cost Controls in Action [02:02]

• The panic over AI costs is already manifesting at major global corporations.
• Tesla's decision to cap employee AI spend at $200 per week starting July 6th is a direct response to unsustainable expenses.
• Uber's CTO acknowledged that their projected AI budget was "blown away already" after 5,000 employees adopted AI coding tools, leading to an 84% usage rate and depleting their 2026 AI budget in four months, with individual engineers consuming $500 to $2,000 monthly.
• Microsoft, despite its significant investment in OpenAI, canceled many internal cloud code licenses, pushing engineers back to more economical in-house tools.
• Meta's internal tracking system, dubbed "Cloudonomics," revealed employees consumed 70 trillion tokens in a single month, prompting management to steer users towards Meta's own less expensive models.

The tools work too well, people use them constantly over relying on them and under per token pricing, the constant use does not get cheaper with scale, it gets catastrophic.

The Shift to Chinese AI Models and Hardware [03:20]

• Companies are actively switching to Chinese AI models to cut costs, seeking the cheapest capable option for routine work.
• AI startup Lindy migrated 100% of its production traffic from Anthropic to DeepSeek, saving millions and improving performance.
• Coinbase now directs its engineers to two Chinese open-weight models, GLM and Kimi, effectively halving its AI expenditure.
• Airbnb utilizes Alibaba's Kwen for customer service, drastically reducing resolution times from 3 hours to 6 seconds due to its speed and cost-effectiveness.
• Cursor, a $3 billion coding tool, was found to be built on a Chinese model, indicating a widespread adoption trend.
• Chinese open models are significantly cheaper, running 5 to 30 times less expensively than their US counterparts; for example, GLM costs $544 for the same work that costs Claude Opus $4,800.

And the answer is they're already switching to China.

The IP Imperative and the Rise of Self-Hosting [05:22]

• Beyond cost, strategic IP concerns are driving companies to reconsider their reliance on third-party AI labs.
• Palantir CEO Alex Karp articulated that companies providing data to frontier labs risk having their "alpha" (competitive advantage) built by others, emphasizing the need to control the "means of production."
• The model of paying per token is seen as a way for AI labs to quietly absorb user expertise, data, and IP.
• Microsoft's "Frontier" initiative aims to embed engineers with customers to build AI systems that the customer owns and controls, aligning with the idea of a "learning loop" that allows model swapping without losing institutional knowledge.
• Both Palantir and Microsoft are publicly advising enterprises that frontier AI labs extract IP and that a full-stack ownership is crucial.
• The most suitable models for self-hosting are predominantly Chinese open-weight, self-hostable, and free to fine-tune on private data.
• Corporations are increasingly hesitant to send proprietary data to labs like OpenAI or Claude, fearing IP extraction and the development of competing versions.

They want to know they own the means of production. It's not being transferred to someone else.

China's Hardware Independence and the Full Stack [07:24]

• US export controls aimed to restrict China's access to advanced AI hardware, but this has inadvertently spurred domestic innovation.
• Meituan, a Chinese food delivery app, open-sourced LongCat, a 1.6 trillion parameter model trained entirely on domestic Chinese chips, without relying on Nvidia or AMD hardware.
• This achievement marks the first model of its size to undergo full training and inference on Chinese silicon, overcoming a significant technical hurdle.
• On real-world coding tasks, LongCat outperformed ChatGPT 5.5.
• This development signifies that the dependence on US hardware can be severed, with Chinese models and chips forming a complete, cost-effective AI ecosystem.
• The export controls, rather than halting China's AI progress, forced the nation to develop its entire technological stack independently and offer it freely.

Meituan just proved that link can be cut too. That you can go from Chinese models to Chinese chips top to bottom and still hit frontier level coding.

The Confluence of Cost, IP, and Hardware: A Collapsing Bubble [08:31]

• US companies are capping AI spend due to unsustainable costs, rerouting workloads to cheaper Chinese models.
• These Chinese models are now being trained on Chinese chips, creating a fully domestic and cost-competitive AI ecosystem.
• IP protection concerns are driving a move towards self-hosting, which further directs companies to Chinese open models.
• Each element in this chain—the AI labs, the cloud providers, and the chip manufacturers—sees American companies being excluded.
• While US labs may still produce superior models, the market is being redefined by the cost-effectiveness of open models that can be run on private servers.
• This scenario, characterized by rising AI capital expenditure, capped usage, and departing customers, suggests the formation of an AI bubble. The fact that a food delivery company achieved a complete, frontier-level AI stack without US involvement is a critical indicator.

The question was never whether this hits the US AI industry, it's how long the applause lasts before someone looks at the invoice.