Moonshot AI, a Chinese company, just released AI model Kimi K3. Kimi K3 is a 2.8 trillion-parameter open-source AI model that the company claims is the largest ever built. With Kimi K3, Moonshot AI believes it can compete with Claude Opus 4.8 and GPT-5.5 across several benchmarks, trailing the leading systems only marginally. The news is not about Kimi-3’s capabilities but what open-source models may do to pricing power for the AI industry.
Open source refers to AI models whose underlying weights, architecture, or training code are made publicly available. Thus, users can download, run, modify, or build on them without paying licensing fees to the developer. This contrasts with closed or proprietary models like ChatGPT or Claude, which are only accessible through a paid API or app, with the underlying weights and training methods kept private.
Research from SemiAnalysis warns that “the rising share” of open-source capability “would fundamentally erode” any moat if the gap continues to close. While pricing may be a concern to the AI model industry, MoonShot AI security issues may be a problem for users. For instance, OpenAI disclosed that a supply chain attack linked to North Korea compromised a developer tool used by MoonShot AI. The Atlantic Council has warned that self-hosted open-weight models “can’t be fully tested or inspected,” leaving enterprises exposed.
This is precisely why OpenAI’s $10 billion custom chip partnership with Broadcom () is important. Purpose-built, dedicated hardware lets an AI lab control the full stack, model, silicon, and data pipeline, rather than exposing its customers to whatever an open-weight file contains. For chipmakers and data centers, who wins the model war has little impact, as training the MoonShot AI model consumes enormous compute power.
The graphic below is courtesy of Arena.AI via ZeroHedge.
Momentum Mash
The key momentum ETF (MTUM) has underperformed the S&P 500 by 7% over the last 20 days. Other than gold miners, which have given up over 18% to the market, MTUM is the worst short-term performer. The second graphic shows the top ten holdings of the MTUM ETF.
As shown, chip companies such as Micron (), AMD (), Intel (), and Broadcom are the most oversold. However, their scores are not very oversold, indicating that they have more room to fall.
The spectacular gains these stocks experienced help explain why the scores remain tame despite the sector’s rout. It’s worth noting that the high-beta ETF (SPHB) holds some of the same stocks as MTUM; thus, it is underperforming, as is MTUM.
Overall market breadth is good, with most sectors clustered within ±25 points of fair value. As the first graphic shows, the rotation is not necessarily value vs growth, as we have typically seen over the last few years; instead, prior underperformers seem to be taking charge.
It’s worth noting that emerging markets are underperforming for the same reason as momentum: chip stocks. SK Hynix () and Samsung account for nearly 14% of the ETF. It also helps explain why the sector outperformed during the first half of the year.
Tweet of the Day
