Since I’ve been challenged with uploading videos am posting this note here. A tad long but highly relevant to the current AI discourse.
Also
@gladstein tagging you as I believe Chinese open source models is a focus of yours. Hit me up if you want to have a conversation.
China Morning Missive
Another DeepSeek “Moment”
Well, here we are. Last week ends with Anthropic pulling Fable 5. Now, just hours ago, the Trump administration confirms that DeepSeek and dozens of other Chinese firms will not be added to the Entity List.
Much will be obviously made of that decision and yet this decision wasn’t even the most consequential announcement made today when it comes to the ongoing AI rivalry between American and China.
What could prove to be a major breach is the Microsoft announcement that it will be shifting its AI enabled assistant, Copilot, from flat-rate subscriptions to usage-based pricing. An obviously necessary move for a platform whose users were burning through compute faster than any subscription fee model could absorb. Expected. What was not expected came next: Microsoft stated that it was exploring the adoption of DeepSeek V4 to support a lower-cost option for users.
There was no commitment mind you, but the mere admission that a lower cost option was required and that a Chinese open source LLM was being considered defines where the AI contest between American and China now stands. Benchmark rankings among the American frontier models are no longer the primary variable in driving global adoption.
Price for performance is now the priority and on that measure the hyperscalers, not only their users, are adjusting behavior in real time. If Microsoft is having to reach for a Chinese open-source model to maintain adoption for its flagship assistant, the benchmark wars may have already been settled.
None of this should surprise anyone who watched the two countries diverge in how AI ambitions were defined. America set out to build God in a box, indifferent to the cost to meet the buildout of infrastructure or the demand for compute, so long as the markets keep funding the activity. No one asked whether a scalable product-market fit existed. No one, at least so far, has cared.
China set a different objective: not scalability alone but applicability and – critically – how AI could be deployed across industry to widen the manufacturing moat. What isn’t discussed enough is the degree to which China’s models have already been integrated throughout the economy. Even the decision to remain committed to an open-source strategy has delivered the competitive advantage of iterative speed. Each of the various players in China are leveraging the gains achieved by one model to advance their own model. It is for this very reason why there seems to be a new and improved Chinese LLM coming to market every single week.
The American obsession continues to produce models that myopically measure success through benchmarks even if their models are cash infernos with no end in sight. The Chinese discipline produces models cheap enough to support global adoption that is so attractive as to pull the attention of a leading American hyperscaler.
Here, however, is the message which needs to be conveyed.
You cannot be a little bit pregnant, and China is winning the race for global AI adoption. American policymakers and corporate executives either take on China with serious intention or admit there is no political courage (or financial bandwidth) to do what is necessary and be done with the façade. Make no mistake, the capability to address the challenge is well within American reach. The question is whether anyone has the stomach to do what is necessary and make the hard choices.