Original reporting and analysis on the model labs, the inference economy, and the people building the future faster than they can explain it.
Nebius's $643M purchase of 20-person Eigen AI, valuing inference optimization at $32 million per engineer, and CoreWeave's $21 billion Meta deal signal that the neocloud race now centers on extracting maximum tokens per GPU rather than GPU count.
Anthropic's six-week, $35 billion compute procurement sprint, capped by a lease of SpaceX's 220,000-GPU Colossus 1 data center, signals a scramble for inference capacity that is reshaping who builds, pays for, and controls AI infrastructure.
With Goodfire's Silico debugger, AI lie detectors nearing production, and new safety fellowships at Anthropic and OpenAI, the field is building real-world infrastructure while the crucial question of what these evals actually measure persists.
On a single April afternoon, OpenAI's loss of a chief product officer, research head, and enterprise CTO underscores an accelerating talent churn that is rewriting the org charts of foundation-model labs faster than executive search firms can adapt, raising the stakes for those who stay.
Hyperscalers are pouring $700 billion into 2026 data center capex while the cooling market grows 19.2 percent annually, reshaping per-token energy costs for inference.
After years as a niche research discipline, mechanistic interpretability is now spawning startups, fellowship programs, and off-the-shelf debugging tools, though the hardest problems remain unsolved.
A cascade of executive exits and strategic pivots at OpenAI, Anthropic, and DeepMind is redefining leadership in frontier AI, with consequences that extend far beyond corporate hierarchy.
Nebius's $643M acquisition of a 20-person MIT inference-optimization spinout resets neocloud valuations as CoreWeave books $27B in weekly deals and xAI builds its own chip fab.
Google’s decision to release Gemma 4 under the permissive Apache 2.0 license reshuffles the open-weights landscape, putting immediate pressure on Meta’s Llama and any other lab still shipping models with restrictive usage terms.
As exploit windows shrink to hours, AI red teaming shifts from quarterly checkpoints to continuous automation, yet blind spots in the methodology remain that tools alone cannot fix.
Apple’s spatial reasoning research contrasts with the stalled Siri overhaul, exposing a multi-billion-dollar research-to-product gap that DeepSeek, Mistral, and Google are rapidly closing.
DeepL's launch of real-time spoken translation marks a moment when the time between an AI research breakthrough and a shipping product compresses to weeks, reshaping enterprise expectations.
As CoreWeave locked $21 billion in Meta and Anthropic deals in 48 hours and xAI may become a neocloud in orbit, the per-token economy is reshaping who builds, who pays, and who captures the margin in AI infrastructure.
From SWE-Bench Pro to the Stanford AI Index, AI benchmark leaderboards now drive billions in investment and geopolitical posturing, yet the mechanics behind the numbers are more fragile than the scores suggest.
While reserved H100 contracts surge to $2.35 per hour, spot pricing hits $3.80, fueling a widening spread that's reshaping the AI infrastructure stack as enterprise GPU fleets languish at 5% utilization.
A new SRE benchmark reveals a 29% pass rate for frontier models, as classified US government testing and the Mythos security scare redefine what it means for AI to be ready for release.
Reserved H100 pricing surged 38% in six months, but the widening spread between spot and reserved GPU instances reveals a deeper supply-demand imbalance in the AI infrastructure market.
From OpenAI's sudden leadership exodus to Anthropic's reorg and Meta's talent raid, the foundation-model labs of the 2020s are becoming the AI platform companies of the 2030s.
While reserved H100 pricing surged nearly 40% in six months to $2.35/hour, spot GPU markets are reshaping AI compute economics via idle capacity, forward curves, and a 5% utilization crisis.
As hyperscale AI capex tops $200B in 2026 and cooling infrastructure grows at 19.2% annually, the true per-token energy cost reveals a hidden margin shift that few invoices disclose.
New alignment reading lists and literature surveys are reshaping the field's self-definition, while battles over which papers make the cut expose deeper fractures in AI safety research.
The agreement to run Claude on Elon Musk's Colossus supercomputer caps a six-month period where every major foundation model lab reshuffled its cloud relationships, leaving the infrastructure map transformed.
From drug discovery to code generation, benchmark leaderboards have become the scorecard for AI progress, but a rash of contamination scandals, domain mismatches, and license shell games is forcing the community to confront what these rankings actually measure.
In just three weeks, four landmark agreements—including Anthropic's SpaceX deal—have reshaped the foundation model compute market beyond recognition from the cloud market of 2024.
Five stories. Four minutes. Zero hot takes. Sent at 7:00 a.m. local time, every weekday.