Asymmetrical Bets
· Asymmetrical Bets
· July 17, 2026 at 18:37
· ⏱ 10 min read
| Read on Substack ↗
Summary
Moonshot AI's Kimi K3 model demonstrates that Chinese open-weight AI is closing the capability gap with US frontier labs at a fraction of the valuation, threatening the high-margin pricing power of closed labs like Anthropic and OpenAI. However, the article argues that cheaper, more competitive models ultimately benefit the infrastructure layer (semiconductors, hyperscalers, power) through increased consumption, making NVIDIA and vertical integrators like Meta the structural winners.
•Kimi K3 is a 2.8 trillion parameter Mixture-of-Experts model with a novel Kimi Delta Attention that reduces KV cache memory by 75% and delivers 6.3x faster decoding at 1M token context.
•It scores 57 on the Artificial Analysis Intelligence Index, behind only Claude Fable 5 (60) and GPT-5.6 Sol (59), and leads benchmarks like Frontend Code Arena and AutomationBench.
•K3 demonstrated self-optimization by handling its own kernel optimization work and building MiniTriton, a tool that replicates NVIDIA's Triton software, raising questions about compounding efficiency gains in future model generations.
•K3 is not token efficient: it costs $0.94 per task vs. $1.04 for GPT-5.6 Sol and $1.80 for Claude Opus 4.8, and uses more tokens per task than GPT-5.6 Sol and Grok 4.5.
•The bear case is that commoditization of intelligence collapses token pricing, undermining the capex cycle for hyperscalers and causing multiple compression in semiconductor, neocloud, and data center REIT names.
•The bull case (Jevons paradox) is that cheaper models drive more token consumption, pushing margin to the infrastructure layer where NVIDIA and hyperscalers have the lowest cost per token, supported by hardware requirements like GB300-class machines to run K3.
Article argues that open-source models like K3 expand the ecosystem and increase demand for high-end inference hardware (GB300 NVL72, B300) needed to serve them, while Jensen Huang has been a vocal ad
Article argues that open-source models like K3 expand the ecosystem and increase demand for high-end inference hardware (GB300 NVL72, B300) needed to serve them, while Jensen Huang has been a vocal advocate for open source to broaden NVIDIA's customer base. The 'Baker scenario' explicitly favors the picks-and-shovels name with lowest cost per token at scale.
Risk: If the quiet erosion scenario plays out where model training costs shrink faster than consumption grows, NVIDIA's growth premium could compress.
Article notes that vertical integration (owning both model and infrastructure) protects Meta (Llama) from margin compression at the model layer, allowing it to charge premium elsewhere. This structura
Article notes that vertical integration (owning both model and infrastructure) protects Meta (Llama) from margin compression at the model layer, allowing it to charge premium elsewhere. This structural advantage over pure-play closed labs makes Meta a relative winner as AI commoditization proceeds.
Risk: Meta's AI capex commitments are substantial; if the commoditization scenario reduces the ROI of frontier models, Meta's spending could face scrutiny.
This newsletter, published July 17, 2026,
features Asymmetrical Bets
discussing NVDA, META.
2 trade ideas extracted by AI with direction and confidence scoring.