Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Labs may corner compute, but AI hiring stays dispersed

Posts warned of hyper‑centralization and data center backlash; our tracker shows broad, growing AI hiring led by consultancies and enterprises, not just the frontier labs.

Read the original at AI Jobs Report
10 min read0 viewsBy AI Jobs Report

Posts warned of hyper‑centralization and data center backlash; our tracker shows broad, growing AI hiring led by consultancies and enterprises, not just the frontier labs.

The week’s split-screen: centralization, backlash, and retrenchment claims

Across the past two days, several practitioners drew a picture of AI consolidating around a few labs and running into growing public resistance. Dwarkesh Patel framed the stakes bluntly, writing that Anthropic and OpenAI are “on track to control most of the world’s usable FLOPs within the next few years (because they can monetize compute better and thus outbid everyone).” Zvi Mowshowitz, surveying siting fights and local politics, put it more culturally: “The American People Really Hate Data Centers.” Ethan Mollick relayed a different view from @randomwalker: “state-level data center bans have very little impact overall on AI progress due to rising efficiency.”

At the same time, Gary Marcus argued that frontier-lab ambitions have raced ahead of reality, calling out customer pullbacks: “Meanwhile, back on Planet Earth, Thomson Reuters has become the latest company to pull back on its Claude usage.” Jack Clark highlighted where AI is and isn’t speeding up science: “AI is accelerating some types of progress but not others:” with “Cyber vulnerabilities: Major acceleration.” and “Mathematics research: Minor acceleration, but harder to measure.” Azeem Azhar flagged signals of automation and strain in the labor market, noting “The canary keeps coughing.” Simon Willison’s changelog showed the day-to-day plumbing work continues: “Release: llm-anthropic 0.27” and “OpenAI made the same change in their v3.0.0 release two weeks ago.”

We tested these claims against our employer-side tracker of AI jobs.

Our tracker counts 4,248 open AI roles across 174 employers, taken from their own job feeds.

Hyper-centralization of compute vs dispersion in hiring

Patel’s argument is clear: “how Anthropic and OpenAI are on track to control most of the world’s usable FLOPs within the next few years.” He also raises the financing stakes: “whether the >$10T of total AI capex we’ll see by the end of the decade will cause a sovereign debt crisis, where hyperscaler debt raises interest rates, drives non-AI exposed countries into bankruptcy, and crashes non-AI equities.”

Our hiring data shows strong demand at the frontier labs, but not dominance in jobs. OpenAI lists 153 open roles, with 66 opened in the last 30 days. Anthropic lists 118 open roles, with 35 opened in the last 30 days. Together that is 271 of 4,248 open roles. In other words, even if compute centralizes, the hiring market remains broad. The largest single employer in our dataset is Accenture with 612 open roles (222 opened in 30 days), followed by Capital One at 183, Amgen at 174, and PwC at 115. Waymo (100) and Databricks (88) are also significant. That distribution suggests talent demand is dispersed across consultancies, financial services, pharma, and platform companies, not concentrated solely at the labs.

By role family, the spread is similarly broad: Modelling and engineering accounts for 1,656 open roles across 145 employers; Data has 820 across 120 employers; Infrastructure 396 across 88; Research 222 across 50; Product and design 103 across 51; and Evaluation and safety 43 across 19. The breadth of employers and functions does not look like a hiring market dominated by two labs, even if capital expenditure and compute access may be concentrating.

Data centers: public resistance vs macro consequences

Mowshowitz emphasizes falling support, including the line: “Support for data centers keeps cratering.” He also characterizes the mood with a thesis header: “Thesis: People Mostly Dislike Data Centers Because They Dislike and Distrust AI, Tech Companies, Big Money And Building Things.” Mollick, in contrast, shares a view from @randomwalker that the macro trajectory will not hinge on any single jurisdiction: “state-level data center bans have very little impact overall on AI progress due to rising efficiency. If your concern is AI development, concentration of power, future use, or anything else other than "is a datacenter near me?" this is not a substitute for policy.”

Our hiring figures align more with Mollick’s forwarded claim than with a picture of activity grinding to a halt. Infrastructure roles total 396 open, with 76 opened and 21 closed in the last 30 days, across 88 employers. That is active staffing for the teams that build, run, and optimize the compute backbone. Hiring does not adjudicate public sentiment, and Mowshowitz focuses on politics, not headcount. But on the question of whether resistance is already throttling the build-out, we do not see it in employer postings. Companies are still adding infrastructure roles net of closures.

Where the growth is coming from

Clark’s summary of a METR analysis notes that different fields are moving at different speeds: “AI is accelerating some types of progress but not others:” with “Cyber vulnerabilities: Major acceleration.” and “Mathematics research: Minor acceleration, but harder to measure.” Our tracker is not segmented finely enough to isolate cybersecurity-specific AI roles, so we cannot validate a corresponding uptick there. What we can see is that employers are adding across the core build functions. In the past 30 days, postings opened outpaced closures in every tracked family:

  • Modelling and engineering: 432 opened vs 112 closed
  • Data: 221 opened vs 31 closed
  • Infrastructure: 76 opened vs 21 closed
  • Research: 37 opened vs 5 closed
  • Product and design: 27 opened vs 2 closed
  • Evaluation and safety: 11 opened vs 2 closed

That aggregate shows net additions across the stack. If acceleration in some domains is pulling investment, it is showing up as broader hiring appetite rather than a narrow spike.

Agents rise, juniors wobble, and what postings do and don’t show

Azhar reports a striking set of usage and labor signals. On models: “In our latest inference token update, the share of open-weight tokens has doubled in the last twelve months.” and “While we are approaching a 1:1 closed-to-open token ratio, the number of closed-weight tokens grew sevenfold over the same period.” On automation: “AI agents used more tokens than humans in February this year; agents are now using 14x that amount, while human token usage grew only 2.8x.” And on labor: “The canary keeps coughing.”

Our hiring data neither confirms nor refutes shifts in token mix or agent usage by itself. On the labor point, Azhar cites a below-trend figure for young workers in AI-exposed occupations. Our tracker does not measure employment by age, only the flow and stock of employer job listings. What we can say is that employers are expanding AI teams right now: across the six families, 804 roles opened and 173 closed in the past 30 days. That is a net positive momentum in postings. Whether those roles ultimately go to younger workers is outside our data, but we do not see a pullback in advertised demand.

It is also notable that Product and design (103 open) and Evaluation and safety (43 open) remain much smaller than Modelling and engineering (1,656) and Data (820). If agents are doing more work and systems are becoming more autonomous, some organizations may still be prioritizing core engineering hires over end-user experience or governance staffing. That is a plausible reading of postings, though it is not proof of how teams are allocating responsibility once hires are in place.

Adoption: fantasy, friction, and the plumbing work

Marcus casts cold water on frontier exuberance. He opens with a cautionary read on Anthropic’s ambitions and ends with customer friction: “Meanwhile, back on Planet Earth, Thomson Reuters has become the latest company to pull back on its Claude usage.” He closes: “The race between fantasy and reality continues.” Our tracker cannot confirm or deny usage shifts at individual customers. We can, however, check if the labs themselves are slowing hiring. They are not. Anthropic lists 118 open roles, with 35 opened in the last 30 days. OpenAI lists 153 open roles, with 66 opened in the last 30 days. That looks like continued capacity-building, not retrenchment, at least in headcount plans.

On the day-to-day integration front, Willison’s note is a reminder that a lot of work is maintenance and compatibility: “This release of the Anthropic plugin for LLM mainly provides compatibility with the recently released anthropic v1.0.0 Python library, which switches from httpx to httpx2.” and “OpenAI made the same change in their v3.0.0 release two weeks ago.” The hiring data is consistent with that ongoing grind: Modelling and engineering is the largest family by far, spread across 145 employers. The posting volumes suggest many firms are still in the thick of building, upgrading SDKs, and stitching models into products and pipelines.

Cutting through “superhuman” framing

Emily M. Bender pushes back on exaggerated capability claims: “Everything you talk about tech as "superhuman" at anything you sound as ridiculous as NBC here. But the robots slamming into the padding at the end and falling down is a nice metaphor for something.” Our postings data lines up with a pragmatic buildout rather than a leap to “superhuman.” Employers are mostly hiring engineers and data practitioners, with a long tail of research and some product roles. Evaluation and safety roles are a small but real slice (43 open across 19 employers), which indicates some investment in oversight, though not at the scale of engineering.

Bottom line

  • Patel’s centralization thesis may yet describe compute, but our hiring data shows a dispersed labor market: 4,248 open roles across 174 employers, led by consultancies and enterprises rather than labs.
  • Mowshowitz’s public-opinion slump does not show up as a hiring freeze. Infrastructure roles are growing, which aligns more closely with the Mollick-forwarded view that local bans have limited macro impact.
  • Clark’s “accelerating some types of progress but not others” is compatible with what we see: broad-based net new postings, not a collapse or single-domain spike we can isolate.
  • Azhar’s warning light on young workers cannot be tested in our dataset; what we do see is employers adding AI roles net of closures.
  • Marcus points to customer friction, but lab hiring remains strong. And Willison’s update reflects the plumbing that our largest role family is being hired to do.

Disagreements abound in the commentary. In the postings, one pattern is clear: companies are still staffing up to build, operate, and integrate AI systems at scale.

What we read

Every quote above is taken verbatim from one of these posts.

More on this

More from ANALYSIS