The AI bill came due

The meter is running. As of today, GitHub Copilot’s token billing meters usage by AI Credits rather than a flat request rate, and developers report bills jumping from $29 to nearly $750, one from $50 to roughly $3,000. The cheaper fallback model is gone. Contrast that with Pinterest’s 90% cost cut, which swapped a frontier vision layer for in-house embeddings across 620 million users, dropping latency from 1.2 to 0.3 seconds. The bill is only fixed if you stop paying it.

Why the meter runs so high sits upstream. Samsung’s 12-layer HBM4E started shipping to Nvidia, AMD, and Google — 48GB stacks at 3.6 TB/s, more than 20% faster than HBM4 with 16% better efficiency. The demand shows in the orders: Dell’s $43.8B quarter grew revenue 88%, with AI servers up 757% to $16.1 billion and a $51.3 billion backlog. Every token developers burn traces back to silicon someone is fighting to ship first.

Yet the cost reckoning hasn’t slowed the megabets. SoftBank’s €75B France bet targets 5 gigawatts of data centers across Hauts-de-France by 2031, starting with a €45 billion phase for 3.1 GW. The buildout is reaching new domains, too: OpenAI’s Rosalind biodefense model is opening to Lawrence Livermore, Johns Hopkins APL, and CEPI, after an earlier autonomous lab cut protein-synthesis costs about 40%. The bill came due, and the spending answered by getting larger.