<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Economics &amp; Capacity on URE</title><link>https://ure.us/pillars/economics--capacity/</link><description>Recent content in Economics &amp; Capacity on URE</description><generator>Hugo -- 0.164.0</generator><language>en-us</language><lastBuildDate>Mon, 22 Jun 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://ure.us/pillars/economics--capacity/index.xml" rel="self" type="application/rss+xml"/><item><title>In the Long Run, Economics Wins</title><link>https://ure.us/articles/in-the-long-run-economics-wins/</link><pubDate>Mon, 22 Jun 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/in-the-long-run-economics-wins/</guid><description>The local AI machine versus the cloud token factory is the wrong fight for a CTO. The future is the model that fits your economics. Here is where to gate it.</description></item><item><title>Frontier AI Is a System, Not a Model</title><link>https://ure.us/articles/frontier-ai-is-a-system-not-a-model/</link><pubDate>Fri, 19 Jun 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/frontier-ai-is-a-system-not-a-model/</guid><description>Frontier AI is a system, not a model you download. The cheapest layer you can own is GPU vector search. What NVIDIA cuVS actually does at 14 million vectors.</description></item><item><title>Visual RAG Beats the Vision Model</title><link>https://ure.us/articles/visual-rag-beats-the-vision-model/</link><pubDate>Fri, 12 Jun 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/visual-rag-beats-the-vision-model/</guid><description>A distilled retrieval corpus beat a 3B vision model on accuracy at 180x the speed, on the hardest images online. Why engineering beats brute-force inference.</description></item><item><title>NVFP4: What 4-Bit Really Costs on Blackwell</title><link>https://ure.us/articles/benchmarking-nvfp4-blackwell/</link><pubDate>Mon, 08 Jun 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/benchmarking-nvfp4-blackwell/</guid><description>An independent 16-arm benchmark of FP8, INT4-AWQ and NVFP4 vs BF16 on a 96 GB Blackwell workstation: quality, throughput, and cross-validation vs NVIDIA.</description></item><item><title>GPU Fleet AIOps: 7 LLM Backends, 6 Failure Scenarios</title><link>https://ure.us/articles/gpu-fleet-aiops-llm-backend-benchmark/</link><pubDate>Tue, 07 Apr 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/gpu-fleet-aiops-llm-backend-benchmark/</guid><description>Benchmarking seven LLM backends as autonomous operators for an 8,000-GPU cluster with six realistic failure scenarios and deterministic checklist scoring.</description></item><item><title>292x: Why Batch Inference Breaks on API Pricing</title><link>https://ure.us/articles/292x-why-batch-inference-breaks-on-api-pricing/</link><pubDate>Thu, 02 Apr 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/292x-why-batch-inference-breaks-on-api-pricing/</guid><description>One rented B200 GPU processed a million documents in 11 hours for $70. The same workload through API providers costs up to $20,419 and takes 144 days.</description></item><item><title>Local LLM Bench: Scaling Swarms Beyond Four</title><link>https://ure.us/articles/best-local-llm-scaling-coding-swarms/</link><pubDate>Mon, 09 Mar 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/best-local-llm-scaling-coding-swarms/</guid><description>Per-task throughput plateaus at four concurrent agents and holds flat through eight. Agents five through eight are free. The contention wall is a floor.</description></item><item><title>Local LLM Bench: Best Model for Coding Swarms</title><link>https://ure.us/articles/best-local-llm-coding-agent-swarm/</link><pubDate>Sat, 07 Mar 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/best-local-llm-coding-agent-swarm/</guid><description>MoE is 4.9x faster than Dense when four coding agents share one GPU. We ran the concurrent-load benchmark nobody published - single-request numbers lied.</description></item><item><title>The Heat Nobody Counts - PUE Ends at the Meter</title><link>https://ure.us/articles/the-heat-nobody-counts/</link><pubDate>Sat, 07 Mar 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/the-heat-nobody-counts/</guid><description>PUE measures the data center envelope. It ignores gigawatts of waste heat from on-site power generation. Space data centers won&amp;#39;t fix it.</description></item><item><title>Local LLM Bench: MoE vs Dense on One RTX 3090</title><link>https://ure.us/articles/best-local-llm-agentic-coding/</link><pubDate>Fri, 06 Mar 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/best-local-llm-agentic-coding/</guid><description>Real benchmarks on dual RTX 3090: the best local setup for agentic coding is one GPU and an MoE model. 168 tok/s, NVLink optional. Data and recommendations.</description></item><item><title>The Concorde Problem in AI Infrastructure</title><link>https://ure.us/articles/the-concorde-problem-in-ai-infrastructure/</link><pubDate>Wed, 25 Feb 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/the-concorde-problem-in-ai-infrastructure/</guid><description>The Concorde was a triumph of engineering and a failure of economics. The 747 won by collapsing cost, not speed. AI infrastructure is replaying the same bet.</description></item><item><title>Why Foreign AI Specialists Keep Failing</title><link>https://ure.us/articles/why-foreign-ai-specialists-keep-failing/</link><pubDate>Mon, 02 Feb 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/why-foreign-ai-specialists-keep-failing/</guid><description>AI context is now portable - but domain gravity and the human translation layer still decide who wins when building real products in real markets.</description></item><item><title>AI and Society: Three Phases of Tech Adoption</title><link>https://ure.us/articles/ai-and-society-the-three-phases-of-technological-adoption/</link><pubDate>Tue, 27 Jan 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/ai-and-society-the-three-phases-of-technological-adoption/</guid><description>A three-phase model of tech adoption through infrastructure economics: disruption follows collapsing constraints in cost, capacity, and reliability.</description></item><item><title>The Entropy of Sovereign AI: Map vs. Territory</title><link>https://ure.us/articles/the-entropy-of-sovereign-ai-why-the-map-is-not-the-territory/</link><pubDate>Mon, 26 Jan 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/the-entropy-of-sovereign-ai-why-the-map-is-not-the-territory/</guid><description>Sovereign AI isn&amp;#39;t a policy memo - it&amp;#39;s a moving contest of leverage, export controls, incentives, and real infrastructure built under shifting rules.</description></item><item><title>It Took a Pandemic to Learn Why Standards Failed</title><link>https://ure.us/articles/it-took-a-pandemic-to-learn-why-standards-failed/</link><pubDate>Fri, 23 Jan 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/it-took-a-pandemic-to-learn-why-standards-failed/</guid><description>Outside-in SOPs drift, create friction, and weaken shared fate. Resilient standards are generated in workflow by the people who operate them.</description></item><item><title>Why GPU Fleet Control Starts with a Map</title><link>https://ure.us/articles/why-gpu-fleet-control-starts-with-a-map/</link><pubDate>Wed, 07 Jan 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/why-gpu-fleet-control-starts-with-a-map/</guid><description>GPU operations starts with footprint truth: a living map of where compute really is, across sites, standards, and drift.</description></item><item><title>Project Atlas: Technical Stack</title><link>https://ure.us/articles/project-atlas-technical-stack/</link><pubDate>Mon, 05 Jan 2026 00:00:00 +0000</pubDate><guid>https://ure.us/articles/project-atlas-technical-stack/</guid><description>Technical architecture of Atlas: Redpanda, TimescaleDB, Hasura GraphQL, and deck.gl for capacity planning and workload placement.</description></item><item><title>AI Infrastructure Placement Is a Business Decision</title><link>https://ure.us/articles/ai-infrastructure-placement-business-decision/</link><pubDate>Thu, 11 Dec 2025 00:00:00 +0000</pubDate><guid>https://ure.us/articles/ai-infrastructure-placement-business-decision/</guid><description>AI compute can&amp;#39;t be cached at the edge. Every inference needs real GPU cycles. Latency rings reveal which populations get responsive AI -and which don&amp;#39;t.</description></item></channel></rss>