I had 41 tabs open, every one an AI job posting, every one saved because somebody had told me the same thing: go where the money is going. Then the Amazon layoffs landed on the AGI group — the people doing post-training, RLHF, the unglamorous work of shaping how a model behaves — during the same stretch the company was guiding toward a capital spending number with twelve digits in it. Nine of my 41 tabs went dead inside a month. The spending went up. The seats went away. Both things were true, and neither one was a mistake.
The verdict, in one sentence: the advice is directionally right and geographically wrong — the capital in AI is real, but at Amazon it is flowing to the deployment layer rather than the research layer, so if you have six months and no pedigree, the boring middle of the AWS AI stack gives you better odds than the frontier does.
The advice, stated fairly
The advice is not stupid. It goes: AI is where the capital is, capital eventually becomes headcount, so point your learning hours at AI and wait for the money to reach you. People who say this are usually pointing at model work — training, alignment, evals, the stuff with papers attached. It is the version of "learn to code" for people who already code.
And for a while it described reality well enough. If you had shipped anything model-shaped between 2022 and 2024, recruiters found you.
Where the advice is roughly right
The buildout is not a story. Amazon's AI business inside AWS was reported at roughly a $15 billion annualized run rate around the time of these announcements, growing at a rate that made analysts revise AWS forecasts upward. Capital expenditure guidance was up better than half year over year. The company has its own training silicon in Trainium and has been loud about total-cost-of-ownership advantages against Nvidia's newest parts — those specific comparisons age fast, so treat the claim as a directional one, not a spec sheet you can quote next year.
So: anyone telling you AI hiring is a mirage is wrong. The money exists. It is moving. That part of the advice survives contact.
Where it breaks down
The advice breaks because it treats "AI spend" as one bucket, and inside a company like Amazon it is at least two ledgers that behave differently.
One ledger is capital expenditure: land, buildings, power contracts, chips, networking, and the engineers who make all of that run at scale. That number went up. Capital budgets are announced to shareholders and defended for years, because half-built data centers do not generate returns.
The other ledger is research payroll. That is where the AGI cuts came from. It is small relative to the capex line, it is measured quarterly, and it is the easiest thing in the building to reduce without breaking a customer promise.
Read the two together and the strategy stops looking confused. Amazon's own leadership has publicly conceded that its models have not been at the very frontier — an unusual thing for a company to admit while spending at that scale. A company that says that out loud and keeps writing the checks is telling you it does not believe the margin lives in having the best model. It believes the margin lives in being the road every model drives on. Nova exists to be good enough and cheap to serve on Amazon's own silicon, not to win a benchmark thread.
Here is the part nobody attaches to the advice when they give it to you: research is the most credential-gated door in this industry. It wants a PhD, a publication record, or a referral from someone who was in the room. That door was narrow when it was hiring. It is narrower now that the job cuts have put people with actual frontier-lab experience back into the market, competing for the seats you were aiming at.
How I'd actually decide
| Layer | Realistic time to employable | Best at | The honest negative |
|---|---|---|---|
| Application (Bedrock, retrieval, evals, agents) | 3–6 months | Most open listings, least gatekeeping, easiest to demo | Commoditizing fastest; a thousand people are building the same RAG demo you are |
| Deployment and MLOps (SageMaker, inference cost, serving, observability) | 6–9 months | Sits directly on the capex; hard to fake in an interview | Ops-shaped career: on-call, thankless, rarely the name on the launch |
| Systems and silicon (Trainium, kernels, distributed training) | 12+ months | Fewest competitors, most defensible skill in the stack | Narrow employer set and vendor-specific; if the silicon bet cools, you carry the sunk cost |
| Frontier research and post-training | Years, plus credentials | Highest ceiling if you get in | These are the seats being cut, and the bar assumes a pedigree you may not have |
On cost, because this matters when you are between jobs: the certification path is the cheap end. Associate-level AWS exams have run about $150 and professional-level about $300 as of writing, and most of the hands-on learning can happen inside free-tier limits if you are disciplined about tearing down what you spin up. Bedrock and inference charges are metered — set a budget alarm the same day you set up the account, because the first surprise bill is usually the last time someone touches the console.
What a certificate buys you is a keyword filter pass, not a job. What the teardown-and-rebuild reps buy you is the ability to answer "what did inference cost you per thousand requests, and how did you get it down" without stalling. Interviewers at Capital One, Wayfair, and JPMorgan ask a version of that question. They do not ask you to derive an attention mechanism.
Who this is for, and who it isn't
This is for you if you are mid-career, employed or recently not, with maybe ten focused hours a week and no lab on your resume. Take the deployment layer. It is closest to the money that cannot be cancelled in a quarter, and it is the least credentialed door in the building.
This is not for you if you already have the frontier credential. If you have published, or you spent two years on a post-training team, the layoffs made your market uglier but they did not disqualify you — do not retrain into MLOps out of panic.
And if you are early enough to be choosing a specialization rather than switching one, the silicon layer is the least crowded room in the industry. It will cost you a year before anyone pays you for it. That is the trade.
Learn the thing that gets bought with capital, not the thing that gets paid out of headcount — capital survives the quarter that headcount doesn't.
