AI Infrastructure Drives Market Divergence; Local Models and Decentralized Networks Gain Traction

Key Takeaways

1. AI infrastructure stocks dominate market gains in 2025, while software companies lag, signaling a shift in investor focus to AI spending accountability.
2. Open-source local models are predicted to reach sub-$10K desktops by 2026, with infinite tokens disrupting compute economics.
3. Nostr-based decentralized networks show vulnerability resistance after India's ban, highlighting a new censorship-proof paradigm.
4. Geoffrey Hinton warns that LLMs' unique failure modes due to massive data compression pose undetectable risks in critical deployments.

1. AI Infrastructure Boom vs. Software Bloodbath

1. The S&P 500's top 9 YTD gainers are all AI infrastructure beneficiaries, while 17 of the worst 30 performers are software and services companies. Investors are pricing AI as an existential threat to traditional software models. — via 1
2. Q2 2025 net profit margins for S&P 500 are projected at 15.7% (record high), with earnings growth of 38% YoY, driven by mega-cap tech EPS expansion. — via 1 2
3. Apple's $129B free cash flow vs. Oracle's $24B burn explains their stock divergence (+56% vs -52%), signaling investors are drawing a line on AI spending — cash flow eventually matters. — via 1

2. Local AI Models on the Horizon

1. @jason predicts that infinite-token context windows and local model desktops under $10K will arrive by 2026, with open-source pricing pressure cutting SLM costs by another 50%, aiming for a 1TB local desktop below $10K. — via 1 2
2. The trend challenges cloud-dependent AI providers and could shift compute economics toward edge devices for many workloads.

3. Decentralized Networks Show Censorship Resilience

1. The Nostr protocol is gaining traction: two of the top three GitHub trending projects are Nostr-related, per @LynAldenContact. — via 1
2. After India blocked the Bitchat source code, copies proliferated across Radicle, Codeberg, and other decentralized platforms, demonstrating anti-censorship resilience. — via 1

4. Hinton Warns on LLM Risk

1. Geoffrey Hinton notes LLMs use only ~1% of human brain's neural connections but have thousands of times more experience, leading to unique failure modes from over-compression — not human-like errors. He worries about deploying such unexaminable systems (with trillions of parameters) in hospitals and courts. — via 1