The real advantage of Chinese AI companies in this round is that many teams haven’t yet caught the “research above engineering” disease. In many Silicon Valley companies, researchers are seen as royalty, while infrastructure work is considered grunt work—everyone wants to move toward research. But by now, with large models at this stage, research no longer happens solely in algorithms and papers. What sets teams like DeepSeek and Kimi apart is that algorithms, data, and infrastructure haven’t started blaming each other. Everyone works toward the same goal, and whoever encounters a problem steps in to solve it. At this frontier of model development, infrastructure is research.

Yang positions massive long-context scaling, from 128K to gigabyte windows in just two years, as AI's equivalent of RAM, making organizational execution the primary moat over raw model capabilities.

AI has split into two camps: closed-source US models and open-weight Chinese models. Restricting access to lower-cost alternatives could weaken competitiveness, raise costs,and ultimately hurt innovation. In AI, economics matters as much as intelligence. #AI #GenAI #OpenSource

exactly. if the best possible reason they can come up with is that increased capex cant be justified due to open source model undercutting their margins with no clear violation of IP law, then that sounds like a bad use of capex! get smarter about monetizing or stop spending all that capex you idiots!

- DeepSeek V4 Pro was launched on April 24, 2026 - Minimax M3 was launched on June 1, 2026 - GLM 5.2 was launched on June 13, 2026 Why do I have a feeling that they’ve been out since last century?

Enterprise AI ROI will probably look late in the macro data. A three-person company can quietly ship like twenty while a 30,000-person company spends the year negotiating access controls.

A Chinese lab dropped the biggest open-weight AI model ever, 2.8 trillion parameters, for free. It matches models OpenAI and Anthropic charge for. Chip stocks fell into a bear market within hours. Nothing kills a moat faster than a competitor who gives your product away.

Here's a question worth sitting with: Are AI models already becoming a commodity? Palantir's CEO said lab tokens are "providing no value" while actually creating risk for big companies. Now OpenAI and Anthropic are giving startups compute essentially for free in exchange for a future revenue share. If their customers aren't generating revenue, "can they actually pay the money to rent out the entire compute and how long can they do it?" That's the real risk hiding in the AI value chain right now.