Daily AI Briefing: Amazon's $42.4B AI spend and more
Amazon burns $42.4B on AI; Chinese labs allegedly distilled GPT-6 Astra; Mistral opens a 1T model.
Amazon's negative free cash flow is a wall
Distillation accusations put model provenance in the spotlight
Open weights versus gated security for cybersecurity AI
Five stories from today's AI briefing converge on one theme: the industry's biggest bets are now visible in cash flows, model provenance, and access decisions. For developers and technical decision-makers, each one is a signal worth reading before choosing a cloud, a model, or a vendor strategy.
Amazon reported negative free cash flow of $42.4 billion for the 12 months through September, against $14.8 billion positive a year earlier. Capital expenditures rose 82% to $210.7 billion, primarily on AI, even though operating cash flow grew 29% and AWS grew 39%. That divergence is the story: the business is healthy, but AI infrastructure is consuming every dollar and more. With Alphabet already down 13% and Microsoft now the lone big cloud provider with positive free cash flow, the market is no longer ignoring the bill. The tension is whether AI demand grows fast enough to turn this into a profitable loop, or whether vendors keep borrowing against a future that hasn't arrived. For cloud customers, that means watching pricing and commit terms closely.
DeepSeek and Moonshot AI reportedly relied heavily on OpenAI's GPT-6 Astra outputs to build their latest models, according to The Information's anonymous sources. The technique—training a new model on another model's answers—is ordinary inside a lab, but contested when used against a rival without permission. The report lands a month after OpenAI tied an extraction campaign involving over 16,000 attempted requests to people linked to Moonshot. If the claims hold, they suggest account bans and post-hoc enforcement aren't slowing distillation. The lack of published evidence means the central claim remains unverified, and OpenAI's response will matter. Either way, power users should treat model performance claims with attention to how and on what data a model was built.
Mistral open-sourced Large 4, aka Le Chonk, a roughly one-trillion-parameter model with 52 billion active parameters per token and a one-million-token context window. Its Artificial Analysis Intelligence index score of 26 makes it the best Western open model, by Mistral's own admission, while leading Chinese open models score higher. Google went the opposite direction with Gemini 4 Carbon, scoring 59, which it limits to vetted defenders. Both approaches are defensible: open weights cannot be cut off, but a capability that finds software holes can also be abused. Security teams now face a genuine trade-off between transparency and containment, and the debate is not going away.
Elsewhere in the briefing, OpenAI's free ChatGPT now writes shorter, less flattering answers with more factual errors—average HealthBench scores dropped about 12%. Reddit also says Google's AI Mode costs it over one million visits per week, a reminder that AI assistants are starting to reshape referral traffic. The practical takeaway for developers: track AI spending signals, verify model provenance, and evaluate open and gated models on the task at hand, because these decisions affect reliability, cost, and security.
A useful daily snapshot for developers tracking AI's cost, provenance, and openness debates.