![\"Empty](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/amazons-ai-retreat-what-nova-inline-1.jpg\")
<\/figure>\nWhat Amazon Actually Cut, and What It Kept<\/h2>\n
The cuts follow a pattern worth studying, because it’s the same triage decision sitting in your 2026 budget.<\/p>\n
The models that went to ‘keep the lights on’<\/h3>\n
Per Bloomberg, the text-based Nova models went to minimum-resource maintenance. So did the video generators and the image generators. They are still supported in the sense that a server somewhere is running, but no serious engineering capacity is attached to them anymore.<\/p>\n
Look at what those three categories have in common. None of them answers a specific operational question. They are capability demonstrations, built on the assumption that if you make a model good enough at everything, the commercial use case shows up later. Amazon has spent years waiting for that use case and has now stopped waiting.<\/p>\n
The 80-person San Francisco office working on artificial general intelligence closed outright. That team wasn’t underperforming on a deliverable. There was no deliverable. The further a project sat from a measurable outcome, the faster it went.<\/p>\n
Why the Covariant robotics bet survived the cut<\/h3>\n
One effort got the people and the compute: a single frontier-model push under Pieter Abbeel, director of the Berkeley Robot Learning Lab. Amazon acquired his robotics company, Covariant, in 2024. Abbeel’s team was building AI models for robots, which lines up precisely with Amazon’s ambition to automate as much of its physical operation as it can.<\/p>\n
That alignment is the whole explanation. Robotics models have an owner inside the business, a process they plug into, and a number they move. Warehouse throughput is measurable this quarter. “Human-level reasoning” is measurable never.<\/p>\n
Amazon didn’t retreat from AI. It retreated from general-purpose AI and concentrated everything behind AI attached to a concrete operational outcome it already understands. For a manufacturer, that distinction is the entire lesson of this episode: the surviving projects were the ones closest to the physical work the company actually gets paid for.<\/p>\n
General-Purpose Models vs. Operational AI: The Real Dividing Line<\/h2>\nWhy commodity capability races have no defensible ROI<\/h3>\n
Nova was built to compete in a race that already had three recognised leaders. OpenAI, Anthropic, and Google were all shipping general-purpose models to the same buyers, and Amazon entered with nothing a customer could point to as a reason to switch. When your product is interchangeable with three others, price is the only lever you have left. That is a terrible place to spend billions in compute.<\/p>\n
The same logic applies at plant scale, just with smaller numbers. A project framed as “we need to understand what generative AI can do for us” has no comparison baseline, no owner on the P&L, and no defined finish line. It survives exactly as long as the enthusiasm of the person who championed it. The first time budgets tighten, it goes to keep-the-lights-on status, which is corporate for dead.<\/p>\n
What it looks like when AI attaches to a cost line<\/h3>\n
Pieter Abbeel’s frontier effort got the people and the compute because robotics connects directly to the two numbers Amazon tracks obsessively: labour cost per unit shipped and throughput per facility. A model that moves either one pays for itself in a way a finance team can verify. That is the whole difference. One bet had a measurable denominator, the others had a press release.<\/p>\n
Copy the test, not the technology. Before you fund anything, name the process it touches, the current cost or cycle time of that process, and the number you expect it to hit. Deviation investigations that take six days instead of twelve. Incoming inspection reports that take twenty minutes instead of two hours. CAPA closure rates, scrap percentage, first-pass yield.<\/p>\n
If you cannot write the baseline figure down today, you do not have a project. You have an experiment, and experiments should be funded as experiments: small, time-boxed, and with a defined kill date. Amazon just demonstrated what happens when nobody sets one.<\/p>\n “Keep the lights on” is a specific status, and it is worth translating into operational terms. It means no new capability, no parity with whatever ships next quarter, and a quiet clock running toward deprecation that nobody will announce until the notice period starts. If a model on that list sits inside your inspection workflow or your document review process, your capability stopped improving the day the reorg memo went out.<\/p>\n This is not an argument for building everything yourself. It is an argument for knowing what a switch would cost you before someone else forces the timing.<\/p>\n Open a spreadsheet and list every AI capability currently running in your operation. Next to each one, write the provider, the specific model or endpoint, and what breaks if it disappears in 90 days. Most teams discover three or four dependencies they had forgotten about, usually inside tools a vendor bolted AI onto without telling anyone.<\/p>\n Then flag the rows where model output feeds a quality or compliance decision. Those are your priority cases, because a model swap changes behaviour, and changed behaviour in a regulated process means requalification. Anything touching CAPA triage, deviation classification, or supplier documentation belongs in that group.<\/p>\n Portability comes from architecture, not from vendor promises. Keep your prompts, business logic, and evaluation sets in your own repository rather than inside a provider’s console. Route model calls through one internal layer so replacing a provider is a configuration change instead of a rebuild across six systems.<\/p>\n The part most teams skip is the evaluation set. Collect a few hundred real examples from your own process with known correct answers, and keep them current. That set is what lets you test a replacement model in an afternoon and prove to an auditor that output quality held. Without it, every vendor change becomes a months-long argument you cannot settle with evidence.<\/p>\n![\"Split](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/amazons-ai-retreat-what-nova-inline-2.jpg\")
<\/figure>\nWhat This Means If Your AI Stack Depends on a Vendor’s Roadmap<\/h2>\n
A one-hour audit of your AI dependency exposure<\/h3>\n
Designing workflows that survive a vendor’s reorganisation<\/h3>\n