Cheap AI Models are a Game-Changer
Welcome to MktContext! I am a professional money manager, trader, and investor who has been timing and beating the market for over a decade. We specialize in predicting market direction by studying the economy and market signals. Join 13,000 subscribers at MktContext.com for our weekly deep dives and analysis!
Last week, Chinese startup Moonshot AI released one of the largest AI models ever built. Within 24 hours, the model “Kimi-K3” climbed to the top of the leaderboard for coding, beating Anthropic and OpenAI:
The announcement reminds us of DeepSeek’s entrance back in Feb 2025. The combination of cutting-edge capabilities, aggressive pricing, and open distribution is threatening the US players. Moonshot doesn’t need to be the best; only close enough, cheap enough, and accessible enough to make customers question why they are paying several times more for American systems.
K3 itself is a marvel of AI engineering. It is efficient in routing queries to specific parts of the model, allowing enormous scale without the full computation cost. It supports a context window of 1M tokens for input, enough to analyze a large codebase or document collection. This underscores just how rapidly China’s AI developers are closing the gap with US rivals.
What do we make of this announcement? Investors will be more concerned about this than DeepSeek. Given the industry’s unprecedented spending spree, they need every dollar of revenue to justify the capital spending. Competition and substitutes erode the ROI of that spending.
It is only a matter of time before Trump bans the use of Chinese AI models on the basis of national security concerns (feeding inputs to state actors and such). This will further strain an already-frail trade relationship between the US and China, potentially sparking Trade War 3.0.
Cheap Models are a Game Changer
It’s worth asking why cheap models like Kimi-K3 are so impactful to the AI discussion. Companies are racking up massive expenses from their employees’ AI usage. Rather than capping usage (which hampers productivity), CEOs are trying to be more efficient with tokens.
In practice, this means routing routine tasks to cheaper, open-source models like DeepSeek, Qwen, Kimi, and Llama. Only the genuinely hard queries get escalated to expensive frontier models. This chart shows just how much open-source models have grown:
As a direct result, token usage continues to skyrocket (black line) but the cost of those tokens are collapsing (colored bars):
The vast majority of the world’s AI workloads can be done by cheaper, open-sourced models which don’t rely on expensive compute from Nvidia. The essence of commoditization.
This is bad news for the frontier model makers (OpenAI, Anthropic, Google, Meta) but great news for hyperscalers because even lagging models need to run on someone’s data centers. The cloud providers (Microsoft, Google, Amazon) provide that infrastructure.
Remember, cheap tokens don’t shrink usage. It proliferates it. That means more revenues for the cloud providers that help enterprises deploy AI widely. This realization is why Mag 7 stocks have moved opposite of the broader AI sector.
Get the rest of this article, including our portfolios and trades, at MktContext.com!
Join 13,000+ investors who are timing the market!