Thoughts about the AI industry, AI in general, inference economics, and the servescale.ai thesis: why enterprise AI needs a private inference cloud inside their control boundaries.
RSS feed
Kubernetes is a strong deployment scheduler, but LLM inference needs continuous control of live state, KV cache, HBM, compute, bandwidth, and request placement below the pod boundary.
Read post →
The room is the architecture. Enterprise inference depends on power, topology, cache locality, utilization, and disciplined cost control.
Read post →
Production AI needs right-sized models, routing, validators, tools, and escalation paths instead of sending every task to the largest model.
Read post →
AI is moving from developer experimentation to operational platforms owned by CIO and CTO organizations.
Read post →
The inference market is splitting between rented-capacity velocity and enterprise sovereignty, governance, locality, and permanence.
Read post →
Enterprise AI pilots need unified model access, governance, policy, infrastructure economics, and accountability to become durable platforms.
Read post →
Agentic AI does not need fake guardrails. It needs anti-capture architecture, promise discipline, and systems that metabolize uncertainty.
Read post →
A power-noir argument that AI economics are now shaped by scarce electricity, cooling limits, PUE, reliability, and energy per token.
Read post →
The AI revolution is becoming a physical contest for coolable, controllable megawatts, water, siting, and energy sovereignty.
Read post →
Inference demand, power constraints, hardware depreciation, and token-margin pressure can turn GPU clouds against their own balance sheets.
Read post →
A systems argument for treating agentic AI trust as an ecology of promises, constraints, observability, and negotiated behavior.
Read post →
AI infrastructure should be financed like power and shipped like product: disciplined, metered, reliable, and built for production utility.
Read post →
A founder’s report from Italy on startup geography, local ambition, and what Silicon Valley thinking looks like far from Silicon Valley.
Read post →
Identity-first trust is cracking under agentic AI. Production systems need explicit promises, context, verification, and observable behavior.
Read post →