The archive

Latest articles

Reporting, explainers, and analysis on data and AI, updated continuously by the newsroom.

Two rising lines comparing frontier training cost growth at 3.5x per year against compute efficiency gains at 3x per year

Training costs rise 3.5x a year. Efficiency is the only brake

Frontier training costs climb about 3.5 times a year while algorithms get 3 times more efficient. The two trends are racing, and the gap decides who can still compete.

Lars Cornelissen · May 21, 2026

Line chart of LLM inference price collapsing at a fixed performance level from late 2022 to late 2024

Opinion

How cheap inference made the one-person studio viable

Building a one-person data studio used to be impossible on the economics alone. Inference at a fixed quality level fell roughly 280 times in two years, and that changed the math.

Lars Cornelissen · May 19, 2026

Descending bar chart showing agentic AI projects thinning out from pilot to production across four stages

Business

Most agentic AI projects will be cancelled before they ship

Gartner expects over 40 percent of agentic AI projects to be cancelled by end of 2027. The drop-off from demo to production is where the budgets quietly die.

Lars Cornelissen · May 18, 2026

Bar chart comparing AI data center capacities in thousands of H100-equivalents, with a planned 2027 site towering over current ones

Business

The gigawatt build-out is the real AI race now

The largest AI data center already rivals 700,000 H100 chips, and a 5-million-equivalent campus is due by 2027. Power and concrete, not chips, set the new pace.

Lars Cornelissen · May 15, 2026

Line chart showing the performance gap between top open-weight and closed AI models shrinking from 8 to under 2 percentage points

Data

Open-weight models closed the gap to 1.7 points

Open-weight models went from a budget compromise to a default choice. On some benchmarks the gap to the best closed models shrank from 8 points to 1.7 in a year.

Lars Cornelissen · May 15, 2026

Two rising lines showing frontier capability and consumer-hardware capability, with the consumer line trailing by about eight months

Last year's frontier model now runs on a laptop

Frontier AI capability reaches consumer hardware in about eight months. The shrinking gap turns today's hosted-only features into tomorrow's on-device default.

Lars Cornelissen · May 12, 2026

Line chart of LLM context window maxima in thousands of tokens climbing from 8k in 2023 to about a million in 2025

Engineering

Context windows grew 30x a year. Your retrieval stack noticed

LLM context windows have expanded about 30 times a year since 2023, from a few thousand tokens to over a million. The change quietly rewrites how RAG systems should be built.

Lars Cornelissen · May 9, 2026

Line chart of cumulative AI compute capacity in millions of H100-equivalents rising steeply from 2022 to 2025

Engineering

The world's AI compute now doubles every seven months

The installed stock of AI chips is growing 3.4 times a year, doubling every seven months. The capacity question has quietly shifted from chips to power and buildings.

Lars Cornelissen · May 6, 2026

Three declining lines on a log scale showing LLM inference price falling at 9x, 40x and 900x per year for different task difficulties

Engineering

Inference got cheap, but not for the work that pays

LLM inference costs fell between 9 and 900 times a year depending on the task. The cheapest gains landed on easy work, while frontier reasoning barely moved.

Lars Cornelissen · May 4, 2026