The AI Realist Podcast
AI without the hype: what it costs, who profits, and what the filings say.

The audio edition of The AI Realist, Julien Simon's newsletter on the AI industry. Each episode is one written piece, narrated by an AI voice: GPU and cloud economics, inference costs, open-weight models, chips and export rules, AI regulation, and the business models of the AI labs.
14 episodes
· 13 min
Mistral Large 4: Where Did the 41 Go?
Artificial Analysis removed the run behind Mistral's headline coding score. Mistral's release page, model card, and technical report still cite it.
Read the written piece, with every source· 11 min
Mistral Large 4 Weights: The Score Comes With the Harness
Mistral Large 4's weights landed October 29th, 2026. Its Terminal-Bench 41 comes from Mistral's own coding agent; in a generic agent it scores 25.
Read the written piece, with every source· 15 min
GeForce vs Workstation GPUs: NVIDIA's Half Rate, Priced
NVIDIA sells one chip at two training speeds, and the unlocks prove it's a setting. Even capped, GeForce wins per dollar to buy.
Read the written piece, with every source· 30 min
Exploring Apple's On-Device Model: Twenty Billion Parameters in Flash
Apple Intelligence's 20-billion-parameter model lives in flash and wakes a few experts per prompt. Your Mac gains memory and battery, not math.
Read the written piece, with every source· 21 min
Ticking Boxes, Shipping Models: Doctolib, Amazon Web Services, Nvidia
Doctolib ticks the health-data compliance boxes with US clouds and models. For now, its public promises say more than its documents show.
Read the written piece, with every source· 35 min
Skill Issue: Mistral AI, GPUs, Kolibri
Mistral can build the model. Does it know how to launch one?
Read the written piece, with every source· 27 min
Kolibri Is Large Enough: GPUs, Mistral AI, AI chips
American chips, Chinese teachers, about $3 million for the training run. What Germany's new open model adds is eligibility.
Read the written piece, with every source· 37 min
Eight Changes to Transformer Attention, and What Each Costs: Nvidia, AI chips, GPUs
The newest open models from Alibaba, Moonshot, DeepSeek and Z.AI no longer read every earlier token. Eight designs, what each costs, and what to test.
Read the written piece, with every source· 11 min
An Act of God, With Letterhead: Oracle, Stargate, AI data centers
Oracle reportedly called its power problem an act of God. Its annual report had already listed it.
Read the written piece, with every source· 28 min
Vent Mauvais: Mistral AI, Microsoft, open-weight models
Verlaine’s ill wind carries the leaf where it will. States bought Mistral without a tender; where free choice can be measured, it goes elsewhere.
Read the written piece, with every source· 12 min
At Least 45 Days: Amazon Web Services, Microsoft, open-weight models
Almost every model on Bedrock gets six months' notice before it can be pulled. The one Washington just accused gets 45 days, and no stated reason.
Read the written piece, with every source· 26 min
Selective Availability: Anthropic, OpenAI, Google
Three agencies advised American AI providers to degrade answers for accounts they suspect of malicious distillation, subtly enough to “avoid triggering obvious alerts”.
Read the written piece, with every source· 24 min
The Cache Is the Price: Anthropic, OpenAI
Routing a job to a cheaper model looks like free money. Whether it is depends on how the frontier bills the handover.
Read the written piece, with every source· 22 min
Event Horizon: Nvidia, OpenAI, Anthropic
Nobody could pay to keep Hugging Face neutral. Nvidia is paying to own it.
Read the written piece, with every source