I keep a spreadsheet of every AI-adjacent job I've applied to since 2024. It has 187 rows. Eleven of them are for teams that do not exist anymore, and I only found that out because I went looking. Nobody emails you when the org you interviewed with gets folded into another org.
If you've been reading coverage of Amazon AI strategy the way I was reading it — as a map of where the work is moving, and therefore where you should be pointing your next six months — that number should sit badly with you. It did with me. I had been treating headcount announcements as job security signals. They are not the same thing, and the gap between them is the whole lesson.
The myth: getting acquired is a promotion
Here's the version you've heard, probably from someone who meant well. A strong team builds something ahead of the market. A hyperscaler acquires them, or licenses their tech and hires the people, which is the same thing wearing a different hat for the regulators. Now the team has compute, capital, and distribution. Their roadmap becomes the acquirer's roadmap. The mission continues, with a bigger engine.
The career conclusion people draw from that is: attach yourself to the researchers. Find the team with the best names on the door, because the names are the asset, and assets get protected.
Smart people believe this. It's how the deal memo reads. It's how the blog post reads. It is not how the org chart resolves eighteen months later.
Did Amazon close its San Francisco AGI lab?
Yes. Amazon confirmed it is closing the San Francisco site of its AGI organization, after The Information reported the move, and the company said its work on frontier models continues elsewhere in the org. Both halves of that sentence are true, and the reason this piece exists is that most people only read the first half.
The history, as reported: in mid-2024 Amazon licensed Adept's technology and hired much of its team, including co-founder David Luan. The San Francisco lab was formally stood up around December 2024 and grew to roughly 80 people at its peak. Pieter Abbeel, who arrived through Amazon's deal with Covariant, was among the leadership. By the time the closure was reported, more than a dozen of the people who had come in through Adept had already gone — some to other labs, some to their own things.
So the lab lasted somewhere in the neighborhood of a year and a half as a distinct place with a distinct address. The people who staffed it were, by any measure, the opposite of the reader this newsletter is written for. Berkeley professors. Founders with exits. Names that open doors before the resume loads. They networked. They had the brand. They got restructured anyway.
I want to be careful here, because it would be easy to turn this into a revenge story, and it isn't one. Nobody in that building is having the week you're having after 200 applications. But if the thing you were told is "the researchers are the asset, so be near the researchers," the closure is evidence that assets and org charts are governed by different rules.
What actually kept shipping
This is the part that got compressed into a clause in most coverage. While the site closed, the AWS agentic AI products that the work fed did not close. Nova Act, the browser-driving agent, stayed available and in customer hands. Bedrock AgentCore — the runtime, memory, identity, and gateway plumbing for running agents in production — kept expanding. The agentic lineup got wider, not narrower, and Amazon has continued to put reported billion-dollar-scale commitments behind that push as of this writing.
Read those two facts next to each other. The research site closed. The billable surfaces the research fed kept shipping, kept getting docs, kept getting release notes.
That is not a contradiction. That is the mechanism.
The mechanism: a SKU outlives an org chart
A research lab is a cost center with a thesis. A product surface is a revenue line with a pager attached.
When a reorg lands, the question executives are answering is not "is this work important." Everyone's work is important; that's what makes reorgs hard. The question is narrower and more mechanical: who gets billed when this works, and who gets paged when it breaks. A managed service with metered pricing has an unambiguous answer to both. A frontier lab has a mission statement, a compute allocation, and an argument.
The argument usually wins for a while. It rarely wins forever, and it almost never wins during a cost review.
Here's why that matters for you specifically, and not only for people tracking Amazon's AI investments from the outside. Job requisitions inherit the same physics. A req attached to a service that customers are being invoiced for survives a reorg, because killing it means telling a customer no. A req attached to a strategy deck evaporates in a quarter, and the recruiter who was emailing you weekly goes quiet and you assume you did something wrong in the phone screen. You didn't. The line item moved.
It also explains something about resumes that took me an embarrassingly long time to see. "Member of Technical Staff, AGI Lab" is a prestige signal that decays the moment the lab's name stops meaning anything to the person reading it. "Took the agent runtime from preview to GA, 40 customers, on-call rotation of six" does not decay, because it describes a thing that existed and had users. The second one is available to you at a company nobody's heard of. The first one isn't available to you at all.
How to tell if a product will outlive its reorg
Before you spend six months learning a platform — as a customer betting an architecture on it, or as an engineer betting a resume line on it — spend forty minutes on this. All of it is public.
| Signal | Where to check | Why it matters |
|---|---|---|
| Metered pricing exists | The product's pricing page | Someone's revenue number depends on this staying alive |
| Generally available, not preview | Docs header or service status badge | Previews get retired quietly; GA services get deprecation notices and migration paths |
| Named in customer case studies | The vendor's own customer page | Named customers are contractual gravity |
| Recent, dated release notes | The "What's New" feed, filtered to the service | A living changelog means a staffed team |
| Support tiers and SLAs | Support documentation | An SLA means an on-call rotation, which means headcount that's hard to cut |
A service that clears four of five will still be here when you finish learning it. One that clears one or two might be excellent technology and still be gone.
Your concrete next step: pick one agentic surface that clears the table — the runtime, not the research demo — and build one small thing on it that a real person uses, even if the real person is your former coworker. Then write the resume line the way the second example above is written: what shipped, who used it, what you were responsible for when it broke. That line reads the same whether you went to Stanford or finished a bootcamp in Tulsa, which is the entire reason to prefer it.
What I couldn't answer here
Three things, and I'd rather name them than paper over them.
I don't know where the frontier research actually landed inside Amazon, or whether it landed anywhere with the autonomy it had in San Francisco. "Work continues" is a company statement, not an org chart. I don't know how the people who left have done since; departures from a closing site get written up as a trend and lived as 80 individual situations. And I don't know whether agentic products will earn back what's being spent on them — nobody does, including the people approving the spend.
Where to look next, in the order I'd look: the AWS "What's New" feed filtered to Bedrock and its agent services, because release cadence is the least spinnable signal there is; the pricing pages, because a price is a commitment; Amazon's quarterly earnings call transcripts, where capital expenditure and agent revenue either get mentioned by name or conspicuously don't; and the team names buried in job requisitions, which tell you what's being staffed six months before any announcement does.
The lab was the story. The SKU was the job.
