Topic · 5 articles

AI.

LLM infrastructure and economics for 2026: compute costs, model revocability, hallucination benchmarking, multi-provider fallback, and browser-native MCP — by Indra Gusti Prasetya.

When a model vendor shuts down, every app hard-coded to that endpoint breaks. These articles cover the infrastructure decisions that keep AI systems running across the full model lifecycle — from procurement economics and fallback routing to hallucination benchmarking and the emerging browser-native MCP layer.

  • Economics & procurement — compute cost curves, power constraints, and stack decisions for 2026
  • Reliability — multi-provider fallback routing and what the Fable 5 shutdown taught us about model revocability
  • Quality — what the 86% hallucination figure actually measures and why it is two different clocks
  • Interfaces — WebMCP, Chrome origin trials, and where browser-native MCP is heading

GPT-5.5 Hallucination Rate: 86% Means It Never Abstains

GPT-5.5 scored 57% on knowledge benchmarks and 86% hallucination rate in the same run. Here is what that number actually counts.

Read more

AI Compute Costs in 2026: Power and the Serving Stack

AI compute cost in 2026 is decided by tokens per watt, the serving stack, and the grid, not the chip. Here is the math that survives finance.

Read more

Multi-Provider LLM Fallback Done Right in 2026

Anthropic retired Claude Opus 4 on June 15 and a directive killed Fable 5 in hours. Your model ID runs on two clocks, and a naive fallback fails both.

Read more

Fable 5's Shutdown Means Your Model Is Revocable

A US directive disabled Anthropic's Fable 5 and Mythos 5 in hours on June 12, 2026. The operator lesson is supply-chain continuity, not the jailbreak.

Read more

WebMCP Hands Your Site's Tools to AI Agents, and Chrome 149 Just Made It Real

Chrome 149's origin trial ships WebMCP, the Google, Microsoft API that lets sites declare callable tools to AI agents instead of being screenshotted and guessed at.

Read more

Browse other topics

All writing

Subscribe to the Technical Journal

Deep dives into infrastructure, security, and technical leadership. No noise, just engineering rigor. Subscribe and grab the 2026 AI-agent & infrastructure security checklist.