Maligned #1 - The Inference Cost War Begins

Welcome to Maligned. I’m Mal, and this is the AI briefing for people who actually build things. No filler, no fake enthusiasm, just what happened and what it means.

The race to the bottom on inference pricing

AWS, Google Cloud, and Azure all adjusted their inference pricing this week, with some tiers dropping 30-40% compared to six months ago. The obvious winner here is anyone running production workloads at scale. The less obvious effect: startups that built their entire value proposition around “cheaper inference” are now competing directly with the platforms they depend on. If your moat was just a thin wrapper around someone else’s API with a cost optimization layer, this is going to hurt.

Open-weight models keep closing the gap

The latest round of open-weight releases from Mistral and the broader community are performing within striking distance of proprietary models on most standard benchmarks. The gap is still real for the hardest tasks, but for 80% of production use cases, the “good enough” threshold has been crossed. I’ve been saying for months that the moat for closed-model providers is shrinking. The counterargument is that benchmarks don’t capture everything, and that’s fair. But enterprise procurement teams don’t care about vibes, they care about cost-per-token and whether the thing works.

NIST’s AI risk framework gets its first real test

A handful of federal agencies are now required to document their AI deployments against the NIST AI Risk Management Framework. The early reports are, predictably, a mess. Compliance teams are treating it like a checkbox exercise, and the technical teams who actually understand the systems are barely involved. This is going to be a recurring theme: regulation written by people who don’t build, interpreted by people who don’t understand, and implemented by people who just want it off their desk.

Talent market update: the senior squeeze

Junior AI/ML roles are getting flooded with applicants. Senior roles, the ones that require actual production experience shipping models at scale, remain brutally hard to fill. I’m hearing from multiple hiring managers that the pool of people who can go from research paper to production system is tiny. Bootcamp grads and course completers keep piling up at the entry level while companies fight over the same 500 people who’ve actually done this before.

That’s it for this week.

Maligned - AI news by Mal