AI in the news: week of December 7, 2025

What this week actually changed: AWS assembled the full agent-and-silicon stack into one bill, Mistral 3 gave open-weights a frontier-class anchor, and the year-end layoff number gave the AI-displacement story a denominator.

AWS re:Invent ran Monday through Friday in Vegas, which means the first week of December was Amazon's news cycle whether anyone else liked it or not. Mistral picked the same Tuesday to ship a frontier open-weight family. OpenAI's "code red" memo from late November leaked into the cycle. And the Challenger Gray year-end layoff numbers landed, the first real denominator on the AI-layoffs story I keep banging on about. Heavy week. Let me work through it.

The re:Invent announcements split cleanly into three buckets (silicon, models, and agents) and the right way to read all of them is that AWS finally has a coherent agent-and-silicon pitch. On silicon, AWS shipped Trainium3 on December 2, 3nm TSMC, 2.52 PFLOPs FP8 per chip, 144GB HBM3e, scaling to a million chips per cluster. AWS claims 4x perf and 40% better energy efficiency over Trainium2. Anthropic is the named anchor customer. The more interesting part is the Trainium4 roadmap: it'll support NVIDIA NVLink Fusion, meaning AWS racks will interoperate with NVIDIA GPUs at the interconnect layer. That's the first concrete signal that the AWS-vs-NVIDIA framing of the last two years is over and the real architecture is "AWS silicon for the bulk, NVIDIA where the workload demands it, both on AWS racks." On models, AWS repositioned Nova as the core family. Nova 2 Lite, Nova 2 Sonic for speech-to-speech, Nova 2 Omni for multimodal. The strategically interesting piece is Nova Forge, a program that gives enterprise customers access to Nova training infrastructure to build their own custom frontier models. AWS is the first hyperscaler to say out loud that the bring-your-own-frontier era is here for serious enterprises. On agents, Bedrock AgentCore picked up policy controls, evaluations, and episodic memory, natural-language guardrails that block agent actions violating written policy, built-in evaluators, and structured memory so agents learn from past runs. It's the most credible enterprise-agent governance story any of the hyperscalers has shipped.

My read: AWS used the week to pitch your model, your training compute, your inference silicon, your agent runtime, your governance layer, all under one bill. Same move Microsoft made with Azure-plus-Office in the 2010s, vertical integration as a moat. The thing this should make any practitioner nervous about is exactly that. The lock-in shape AWS is building is more total than any cloud lock-in we've seen, and every layer references the layer below it. I wrote about this exact pattern, the AI lock-in is structurally worse than 2010 cloud lock-in, and re:Invent was the clearest demonstration of that thesis I've seen. The right counter-move is the one I've been making the case for: build agents against an MCP-only architecture where the runtime is a thin layer over swappable model and tool endpoints. AgentCore is great until it isn't, and the "isn't" usually arrives as a 60% price hike at renewal.

The open-weights answer landed the same Tuesday. Mistral shipped the Mistral 3 family, ten models. Mistral Large 3 is a 675B sparse MoE with 41B active, multimodal and multilingual, open-weight. Ministral 3 is the small-model line: 3B, 7B, 14B dense models for edge deployment. Two things matter. An open-weight frontier-class multimodal model is what the open-source ecosystem has been waiting for since Llama 3. Llama 3 didn't quite reach frontier, Qwen 3 got close, Mistral 3 Large is the first European-shipped open-weight model that legitimately competes with the hosted frontier on the multimodal axis. And the simultaneous small-model release is the right shape for the actual enterprise deployment story. The interesting AI workloads in 2026 won't all be "call GPT-5.2 from a hosted API." A lot of them will be "run a 7B or 14B model on infrastructure I control." Mistral's bet is being the default vendor for both ends. I'm running Ministral 3 7B on my home infrastructure as of Wednesday. Honest assessment: it's the most capable 7B I've seen, and the gap to hosted Sonnet 4.5 on my actual workloads is smaller than it was six months ago. Not closed. Smaller. The small-model story keeps accelerating.

OpenAI got loud the same week, but defensively. The Axios reporting on the internal OpenAI memo describes a "code red" reorganization in response to Gemini 3 beating GPT-5.1 on multiple benchmarks in mid-November. Sam Altman reportedly redirected resources toward improving ChatGPT itself and pulled GPT-5.2 forward. The fact that the memo leaked is the news, not the contents. OpenAI in 2024 did not have memos like this leaking. The internal-comms posture shifted, and the shift correlates with mounting pressure. The deeper thing this tells me: the era of "OpenAI is uncontested at the frontier" ended quietly somewhere between Gemini 3 on November 18 and Opus 4.5 on November 24. Three labs now ship at frontier on different cadences with overlapping but non-identical strengths. The multi-vendor world I've been arguing is healthier for buyers has arrived as fact rather than hope. Committing to one vendor's stack right now is the worst possible timing, the price competition and capability competition are about to get intense, and the buyers who stay portable will benefit most.

Then the labor numbers landed. Challenger Gray published their year-end report: roughly 55,000 AI-attributed layoffs in 2025, out of 1.17 million total US cuts, highest since the COVID year of 2020. I want to sit with this because it's the cleanest signal we've had on what's actually happening, and I want to be straight about my position. The displacement is real and it's accelerating faster than I expected. The companies citing AI are the largest employers in the US. Amazon 14,000, Microsoft 15,000, Salesforce 4,000+, Accenture 11,000 in December alone. 55,000 explicit AI-cited cuts is the floor, not the ceiling, because most AI-rationalized cuts don't carry the label in the press release. The thing I keep coming back to is the pace. Short-term incentives are driving the rush. Companies aren't cutting at this speed because the AI is ready to absorb the work; they're cutting because the AI narrative is convenient and the markets reward the cuts. Sam Altman flagged "AI washing" for the same reason. He's right about the framing but I don't think AI-washing is doing as much work as that framing implies. The deeper problem is that markets reward the genuine AI cuts and the AI-washed ones identically, so leadership has no reason to slow down. The pace is the issue, not the underlying reality that the workforce is changing.

To be clear about where I sit: I'm not against automation. I've spent my career on it. The displacement of repetitive systems work that should have been automated long ago is fine and overdue. The fuller version of where I think the lines sit is in the job-security piece. The sustainable working model is human+AI collaboration, and the companies that figure that out will outperform the ones that just cut, because the work product is better. To be careful not to oversell it: in the collaboration model the headcount still shrinks. It just shrinks less and shrinks well. It's the version of this transition worth fighting for. The version we're getting in 2025 is the rushed one. I'd rather be wrong about how fast this is than caught off guard.

Smaller items: Google's December AI roundup shipped Gemini 2.5 Flash native audio upgrades and Deep Research improvements, confident-incumbent posture that was OpenAI's exact posture six months ago. The EU linked AI Act high-risk timing to harmonised standards rather than the August 2026 hard date. Chinese AI models hit 15% global share in November, up from 1% a year ago. NIST launched two AI centers on December 4, agency-level governance continuing where the legislative branch can't.

What to watch next week: end-of-quarter retrospectives starting to land, GPT-5.2 actuals in the API, and whether the agent-platform consolidation accelerates or pauses. The pattern that held this week: hyperscaler agent platforms are now real products with real lock-in shapes, the frontier is genuinely multi-vendor, and the 55K layoff number is the floor. Architect for portability while you still can.