Goldman's 25K vs 9K: The Sourcing Map Hiding in the AI Layoff Math

Goldman Sachs's economists just gave the doom crowd a headline: AI is shaving roughly 16,000 US jobs from monthly payroll growth, a number now anchoring this week's labor debate alongside Q1 2026's 78,557 tech layoffs. Read it as a sourcing map and the math flips. Twenty-five thousand engineers a month are losing seats. Nine thousand new seats are opening on the augmentation side. Your job is to move the right 9,000 before the rest of the market notices the pattern.

The Goldman math, read as a pipeline

Goldman's framework is regression-based, built on payroll data, job postings, and revenue at firms that have not replaced workers, using the IMF's AI-substitution index. The number you've seen quoted everywhere (a 16,000-jobs-per-month drag on payroll growth, plus 0.1 points on unemployment) is a net. The gross is more useful: about 25,000 monthly displacements from substitution, offset by about 9,000 monthly additions from augmentation.

Goldman also flagged something most coverage skipped. The +9,000 is undercounted, because the model doesn't fully capture data center construction, grid build-out, or the incremental labor demand from AI-driven productivity gains. The visible augmentation roles (AI Engineer, ML Engineer, Applied Scientist, MLOps, AI Platform, Forward Deployed Engineer, AI Governance) are the tip. The actual hiring footprint is wider.

25,000 vs 9,000

Who's actually getting hired on the +9K side

Goldman's published winners list (education workers, judges, construction managers) is useless for technical recruiters. The real augmentation roles cluster in five buckets, and they pay.

AI engineers average $170,750, which is 17.7% above non-AI peers. ML engineers average $186,067. Cloud architects average $158,000. AI job postings sit 134% above 2020 levels. In January 2026 alone, 275,000 US postings required AI skills. AI governance demand is up 150%. AI ethics is up 125%. Prompt engineering is up 90%.

The five buckets that absorb the +9K, in order of headcount:

1. AI Engineer / ML Engineer / Applied Scientist. Our index counts roughly 10,500 US professionals in these titles, concentrated in the Bay Area and NYC. Distyl AI, Amazon, Scale AI, and Modal are the names actively building benches.
2. Forward Deployed Engineer / AI Solutions Architect / AI Platform Engineer. A narrower cut, around 1,578 people. Palantir (which invented the role), Salesforce, ElevenLabs, Scale AI, and Modal are the top employers. NYC leads the Bay Area 2:1 here.
3. MLOps and AI Platform. Adjacent to bucket 2, often filled by displaced senior backend and infra engineers from the layoff wave.
4. AI Governance, Risk, Safety. New, growing fast, often filled from security and compliance, not from research.
5. Data center and grid adjacent. Not on most recruiters' boards because the titles look like "electrical engineer" or "power systems engineer," but Goldman explicitly calls this out as the undercounted offset.

If you're only sourcing bucket 1, you're competing with everyone. Buckets 2 through 5 are where the arb lives.

The 25K is not who you think it is

Cognizant's Chief AI Officer Babak Hodjat told Nikkei in April that the industry is still in a transitional stage and that much of the current workforce reduction is motivated more by anticipations regarding AI than by actual efficiency improvements. He runs AI at a company that is itself reportedly cutting 12,000 to 15,000 people under "Project Leap," with $230M to $320M earmarked for severance. When the CAIO of the firm doing the cuts says the cuts are running ahead of the capability, take the note.

That means a meaningful slice of the 25,000 monthly "displaced" are not obsolete. They are politically displaced, ahead of a productivity curve their employer is betting on. In 6 to 12 months, smaller AI-native competitors will hire them back, often cheaper than they could have hired equivalent talent in 2024. The recruiters who pipeline these engineers now (not when they're sitting in a Discord asking for referrals) win the next cycle.

Recent named displacement events worth indexing this quarter:

• Cognizant Project Leap. 12,000 to 15,000 globally, IT services skew, heavy on 3 to 8 YOE engineers.
• Block. Jack Dorsey cut nearly half the workforce, attributed to AI automation. Payments and fintech depth.
• Atlassian. 10% cut to fund AI investments. Strong on platform, dev tools, Jira-adjacent infra.
• Meta. Reported plans to cut 20% of a roughly 79,000-person workforce.

The buyers of those engineers are not (mostly) FAANG. They're Distyl, Scale, ElevenLabs, Modal, Palantir, and the long tail of Series B AI-native firms whose hiring volume is too small to show up in BLS data but adds up to the +9K.

The first rung is gone, so stop sourcing it

Stanford HAI's AI Index 2026 reports that entry-level software developer employment among workers aged 22 to 25 is down nearly 20% since 2024. Mid-career and senior positions in the same fields are flat or growing. The career ladder still exists; the first rung has been pulled up.

For recruiters this is operationally simple: junior reqs are sourcing into a vacuum. The viable pipeline for augmentation hires is 3 to 7 YOE engineers from the displacement events above, plus 8 to 15 YOE staff and principal engineers who can credibly run multi-agent workflows. If your hiring manager is still insisting on "new grad with AI fluency," you're going to spend the quarter losing.

Junior reqs are sourcing into a vacuum. The first rung was pulled up in 2025 and nobody put it back.

The behavioral tell: "I assist AI"

Andrej Karpathy described his workflow flip publicly in late 2025: from 80% human, 20% AI to 20% human, 80%+ delegated to agents. He calls himself a director of agentic systems rather than a direct author of code. Garry Tan used the phrase "cyber psychosis" for the same pattern, sleeping four hours a night because he can't stop building with Claude Code.

This matters for sourcing because the +9K augmentation hires are identifiable by behavior, not by resume keywords. The signals:

• Multiple Claude Code, Cursor, or Aider sessions running in parallel, treated like junior reports.
• GitHub commits with agent attribution or co-author lines.
• Public writing or talks about multi-agent orchestration, eval pipelines, or agent observability.
• Side projects that ship fast and have visible AI scaffolding in the repo (CLAUDE.md, .cursorrules, agent prompts in /docs).
• A history of moving from "AI assists me" framing in 2024 posts to "I assist AI" framing in 2026.

None of those appear on a LinkedIn keyword search. They appear in commits, README files, conference talks, and the gap between someone's 2023 resume and their 2026 GitHub. This is exactly the kind of cross-source behavioral pattern keyword filters miss, which is why we built Refolk around plain-English search across GitHub, LinkedIn, and the open web instead of boolean strings against a single index.

Geography: Bay Area supply, NYC demand

Our index shows the Forward Deployed Engineer / AI Solutions Architect cut concentrates in NYC at roughly 2 to 1 over the Bay Area. The displaced tech labor pool from Meta, Block, and Atlassian still skews Bay Area. That's a textbook arb: move Bay engineers into NYC-headquartered AI-services firms (Palantir, ElevenLabs NYC, Scale AI's NYC presence), or build remote-FDE benches and let geography stop mattering.

Two practical moves:

1. Build a "displaced Bay senior, open to NYC or remote" segment. Pitch it to Palantir-style buyers as a 4 to 6 week pipeline they can close before the engineer signs with the nearest Series B.
2. Track AT&T, Amazon, and JPMorgan internal mobility. Amazon spent $1.2B on Upskilling 2025 and moved 100,000 employees into higher-skilled roles. JPMorgan commits $600M annually to training. AT&T put $1B into shifting 140,000 employees from legacy telecom into software and data. Corporate AI upskilling is now a $32B global market. Some of those people are about to be the cheapest mid-career AI hires on the board, and they're not on the layoff trackers.

$32B

A 30-day playbook for the +9K

If you have one sourcing quarter to act on the Goldman math, here's the order of operations.

Week 1: Name the displacement events you care about

Pick three. Cognizant Project Leap, Meta, and Atlassian are the highest-volume English-speaking pools right now. Block if you're hiring payments. Save the names of the laid-off engineers as they appear (LinkedIn "open to work" banners, X posts, GitHub bio updates) into a single segment.

Week 2: Filter by behavior, not title

Inside each displacement pool, find the 10 to 20% who already show the augmentation behavioral signals: agent-orchestration repos, multi-tool workflows, public writing on AI-native development. Refolk's plain-English search is built for exactly this kind of cross-source query ("ex-Meta backend engineers shipping multi-agent side projects since January"), so you don't have to maintain a boolean monster across three platforms.

Week 3: Match to the +9K buckets

Bucket each candidate into one of the five augmentation categories above. The 3 to 7 YOE backend cohort goes to MLOps and AI Platform roles. Senior infra engineers go to FDE and AI Solutions Architect. Compliance and security backgrounds go to AI Governance. Don't push everyone toward "ML Engineer," that's the most competitive bucket and the worst placement rate.

Week 4: Outreach with the actual job, not "AI roles"

The engineers in this pool have seen 40 generic "AI Engineer" InMails this month. The ones that get replies name the company, the agent stack, the team's current eval problem, and the comp band. If you're sourcing for a Palantir-style FDE seat, say so. If you're sourcing for a $1B AT&T reskilling pod, that's actually interesting and recruiters underuse it.

For sustained pipeline work across all four weeks, the bottleneck stops being where to look and starts being how fast you can describe what you want. That's the part Refolk is built for: ask in plain English, get the right people across GitHub, LinkedIn, and the open web, every time.

What technical recruiting looks like in 2026

The honest summary: technical recruiting AI 2026 is no longer about whether candidates use AI. It's about which side of the Goldman split they sit on, and whether you can tell the difference from their commits before your competitor does. The 25K and the 9K are not separate populations; they're the same population sorted by who got to the augmentation behaviors first.

Sourcing AI-displaced engineers is the most leveraged work a recruiter can do this year, because the supply is high, the buyers are concentrated, and the matching problem is behavioral. Goldman Sachs AI labor displacement headlines will keep coming. The recruiters who treat each one as a pipeline event, not a news event, will own the next two cycles of AI-augmented engineering roles.

FAQ

How do I verify the Goldman 25K vs 9K split is real and not a model artifact?

Goldman's own writeup flags the limits: it's a regression on payroll data, job postings, and revenue at firms that haven't replaced workers, using the IMF's AI-substitution index, and it explicitly undercounts data center build-out and productivity-driven labor demand. Treat the directional split as solid (more substitution than augmentation in the visible labor market, by roughly 2.8x) and the absolute numbers as a floor for the +9K side, not a ceiling.

Which displacement events should I prioritize this quarter?

Cognizant Project Leap (12,000 to 15,000 cuts, mid-career IT services skew) is the largest English-language pool with the most usable engineers. Meta's reported 20% cut against 79,000 employees is the highest-signal pool if you're hiring for AI infrastructure or platform. Atlassian (10% cut to fund AI investments) is the best fit for dev tools and platform roles. Block is the play if you're hiring payments or fintech.

Are the laid-off engineers actually obsolete, or are they hireable?

Cognizant's own CAIO is on record saying current cuts are running ahead of real productivity gains and are motivated more by anticipation than efficiency. A meaningful share of the 25K monthly "displaced" engineers are politically displaced, not technically obsolete. They will be re-hired by smaller AI-native firms within 6 to 12 months. Pipeline them now while they're available and price hasn't reset.

What's the single biggest mistake recruiters are making on AI augmentation jobs?

Sourcing on titles instead of behaviors. The 10,500 US "AI Engineer" titles are the most competitive segment of the +9K, and most candidates in it are already pipelined. The actual leverage is in the 3 to 7 YOE engineers from displacement events who show augmentation behaviors (multi-agent workflows, Claude Code or Cursor in their commit history, agent-orchestration side projects) but whose resumes still say "Senior Backend Engineer." Behavioral search beats keyword search by a wide margin in this market.