dave blundin

Ternary/binary quantization (1.58 bits) with <5% accuracy loss enables frontier models on smartphones. Sub-1-bit quantization via sparsity and low-rank factorization going mainstream in 12 months. Photonic computing substrates next. This multiplies with algorithmic improvements for million-fold effective compute gains.

SpaceX's S-1 frames a $28.5T addressable market (Starlink, AI infra, Macrohard digital labor) and reveals Anthropic paying $15B/yr for Colossus compute, signaling a pivot to owning the orbital infra layer (Dyson swarm) and application layer (Macrohard) — a 'Microsoft in space' model. Starship V3 hourly launch cadence enables packet-switched solar system transport.

Opus 4.8's dynamic workflows enable hundreds of agents to self-fork with full conversation history, eliminating the 20-30 minute context-rebuild penalty. This unlocks billion/trillion concurrent worker scale for software and science, with assimilation quality finally matching parallel generation. Investment implication: agent orchestration layers, context management infrastructure, and multi-agent frameworks become critical moats.

Google's 6x CapEx to $185B, OpenAI's $120B raise, and Cerebras' $20B+ OpenAI deal signal that AI compute infrastructure spending is entering a multi-year supercycle driven by inference demand, not just training.

The next platform war is owning the primary agent (Jarvis) that coordinates all other agents; whoever controls this interface captures all consumer AI value and data.

AI has infinite appetite for compute - every new GPU creates pure value (curing diseases, feeding people). Corporate America unaware that surplus compute era is over; compute auctions per token will determine allocation. Even Terrafab coming online won't meet demand.

The core government concern is not cybersecurity but models that can answer 'help me build yourself' — enabling recursive self-improvement escape velocity; China may already have reached this independently.

As model intelligence becomes 'too cheap to meter' (99.95% cost drop in 3.5 years), value migrates up to application-layer platforms with distribution (Meta 3.5B users, Google 2B) and down to compute infrastructure (Nvidia GPUs, TSMC fabs, memory). Foundation model companies face innovator's dilemma: cheaper open alternatives (Kimi K3, Llama) erode API pricing power. Netscape/Mozilla analogy: browsers went free, value shifted to Google/Amazon.