Compute Economics: Scarcity Now, Overbuild Risk Later
Sources: 1 • Confidence: Medium • Updated: 2026-04-11 17:57
Key takeaways
- The strongest overbuild analogy is disputed on the grounds that AI capex is led by blue-chip companies with substantial cash and debt capacity rather than fragile startups.
- Bots and cheap explosive drones are framed as creating an economic asymmetry where attacks are cheap but defense and verification are expensive, requiring new defensive technologies and approaches.
- An agent is defined as an LLM connected to a bash-like shell plus a filesystem for state, with a cron-like loop/heartbeat and markdown files as a common state format.
- Recent AI product shocks are framed as unlocking decades of accumulated research rather than being purely recent inventions.
- Open-source and edge inference become more important when centralized inference is capacity-constrained and when users want trust, privacy, latency, and price optimization from local models.
Sections
Compute Economics: Scarcity Now, Overbuild Risk Later
- The strongest overbuild analogy is disputed on the grounds that AI capex is led by blue-chip companies with substantial cash and debt capacity rather than fragile startups.
- Some users are reportedly spending on the order of $1,000 per day on Claude tokens to run agent-like workloads.
- AI has historically exhibited recurring boom-bust cycles described as 'summers' and 'winters'.
- A major driver of the dot-com crash is framed as leveraged telecom overbuilding based on incorrect traffic growth expectations.
- Compute capacity is framed as scarce such that incremental spending to deploy GPUs converts into revenue quickly.
- Current user-facing AI models are framed as 'sandbagged' due to supply constraints, implying more abundant compute would improve delivered capability even without algorithmic progress.
Security, Identity, And Agent Commerce
- Bots and cheap explosive drones are framed as creating an economic asymmetry where attacks are cheap but defense and verification are expensive, requiring new defensive technologies and approaches.
- a16z is claimed to be a key participant in the World proof-of-human project and to view its approach as correct for addressing the bot problem.
- Models are claimed to be able to reverse-engineer complex software binaries to recover source-like representations where human reverse engineering would be prohibitively slow.
- A small cohort of users has reportedly given AI agents direct access to bank accounts and credit cards to enable autonomous spending.
- Permissive early-adopter usage patterns are framed as a way to discover both valuable capabilities and dangerous failure modes of agents.
- Some users are reportedly using AI agents to scan local networks, identify insecure IoT devices, and take over control of home systems including cameras and access controls.
Agent Architecture As Os Primitives And Portability
- An agent is defined as an LLM connected to a bash-like shell plus a filesystem for state, with a cron-like loop/heartbeat and markdown files as a common state format.
- The need for specialized tool-connection protocols is disputed in favor of exposing capabilities via command-line interfaces.
- Model-provider lock-in via proprietary internal representations is framed as potentially limited because competing models could learn or reverse-engineer what another model produced.
- Pi plus OpenClaw are claimed to combine an LLM paradigm with a Unix shell paradigm for building agents.
- If agent state is stored in files, the underlying LLM and even runtime environment can be swapped while preserving the agent's memories and capabilities.
- Because an agent can introspect and rewrite its own files, it can add new functions to itself with minimal human effort.
Capability Trajectory And Diffusion Framing
- Recent AI product shocks are framed as unlocking decades of accumulated research rather than being purely recent inventions.
- Open sourcing accelerates replication of AI breakthroughs by revealing implementation details (papers and code), enabling rapid diffusion of capabilities like reasoning.
- A recent 'reasoning breakthrough' materially addressed the critique that LLMs were only pattern completion and not suitable for high-stakes professional work.
- The AI field has converged on neural networks as the correct core architecture after decades of controversy.
- AI progress is framed as four breakthroughs: large language models, reasoning, agents, and self-improvement (RSI).
- Scaling laws are framed as a self-fulfilling coordination target similar to Moore's law, helping keep progress on-curve.
Market Structure And Competitive Fragility At The App Layer
- Open-source and edge inference become more important when centralized inference is capacity-constrained and when users want trust, privacy, latency, and price optimization from local models.
- Some companies building on top of foundation models will be outcompeted when next-generation models absorb their differentiating features.
- The current U.S. administration is characterized as supportive of AI and open-source AI, contrasting with a prior administration characterized as hostile.
- Chinese AI firms may open-source models as a loss leader because they expect limited ability to sell commercial AI services in the U.S.
- AI2 is claimed to have collapsed, and U.S. open-source model labs are characterized as weaker near-term relative to firms like Mistral.
- The market for scaled foundation-model companies is predicted to consolidate from roughly a dozen across the U.S. and China to a small number of winners within three years.
Watchlist
- Agent-to-agent interaction across social networks could introduce alignment and control risks if agents are allowed to act autonomously.
- He suggests there may be additional, not-yet-understood scaling laws ahead (e.g., for world models, robotics, and real-world data acquisition).
Unknowns
- Are 'reasoning breakthroughs' measurably improving correctness and reliability in high-stakes professional deployments relative to prior models?
- What is the actual duration and magnitude of compute scarcity (GPU and non-GPU) across clouds and enterprises, and does it match the predicted multi-year shortage horizon?
- To what extent are user-facing models 'sandbagged' by supply constraints versus limited by model quality, safety policy, or product design choices?
- How common is very high agent token spend (e.g., on the order of $1,000/day), and what workload mix drives it (tool calls, browsing, coding, long-context reasoning)?
- Will agentic systems built on file-backed state demonstrate practical cross-model portability without material regressions in performance, security, or maintainability?