The Grid Interconnection Queue Is the Real AI Compute Constraint
2,060 GW queued, a five-year median wait, and a 13 percent historical completion rate. Why the map of where large-scale compute exists in 2030 is being drawn by grid geography.
Emerging Technology.
Compute, infrastructure, policy and the technologies arriving next — assessed on evidence rather than launch videos.
LLM Token Pricing Is a Facilities Question
Output costs five times input because decode is memory-bound. Batch costs half because utilisation is the provider's central problem. Beneath both sits a floor made of power contracts and grid queues.
AI Rack Cost: How to Price Accelerated Compute End to End
Capital, energy times PUE, and overhead, divided by capacity times utilisation. The framework, the multipliers that matter, and the five places the arithmetic reliably breaks.
AI Inference Chips: Why Memory Bandwidth Beats Peak FLOPS
Decode reads the whole model to produce one token. Working from NVIDIA's own published rack figures, here is the arithmetic that explains why serving throughput ignores the number on the box.
Confidential Computing Attestation: What It Proves, and What It Doesn't
Attestation proves genuine hardware and a known measurement. It does not prove the code is safe, the data protected either side of processing, or that anyone checked the report — and 2026 research…
EU AI Act Deployer Obligations: What Actually Applies Now
The Digital Omnibus moved the high-risk deadlines to December 2027 and August 2028. What binds an organisation deploying AI today, what was postponed, and why the delay is engineering time rather…