AI Hiring Pulse — Week of 2026-06-08

Window note. This pulse covers a wider, uneven window — 2026-06-08 → 2026-06-17 — so the counts reflect up to ten days of postings rather than a clean week. For that reason we're setting aside volume rankings and week-over-week deltas this issue and reading only composition: the mix of stages and roles within each company, which is unaffected by the window length. The usual net open/close tables return next pulse.

📊 Pattern of the week — the dividing line is compute, not size

Across 1,398 AI roles at 155 companies this window, the obvious hypothesis is wrong. You might expect the big, well-capitalized incumbents to hire "deeper" in the stack than smaller application companies. They don't. Split the universe into frontier/big-tech (Google, Amazon, Microsoft, Meta, Apple, NVIDIA, OpenAI, Anthropic) versus everyone else, and the upstream share — roles at Data, Pretrain, Post-train, or Serve — is 30% versus 27%. Effectively identical. Size is not the axis.

The axis is compute ownership. Sort companies by their upstream share and a clean cleavage appears: the firms that own silicon or train their own models hire overwhelmingly upstream, and the firms that build products on rented foundation models hire almost entirely at the Agent station.

This is the same "owns the stack versus rents the stack" distinction we flagged when Oracle reversed two weeks ago — but now the full universe is in frame, and it resolves into the cleanest structural split in the index.

🪞 The two poles, same week, same index

The clearest way to see it is to put the extremes side by side. Meta and Salesforce are both enormous, both flush with cash, both "AI-first" by their own description. Their hiring could not look more different.

Meta — 23 roles: 17 upstream, exactly one at the Agent station. The top archetype is researcher (8 of 23). The titles are unambiguous: "AI Research Scientist — Language — Meta AI," "Staff Research Scientist, FAIR (RL / LLM's)," "Research Engineer (Technical Leadership), FAIR Data — Meta Superintelligence Labs," plus a cluster of "Software Engineer, Systems ML" and an ML-accelerator "Design Verification Engineer." Meta is hiring people to build models and the systems that train them. Its single Agent role this window is a Technical Program Manager for AI-generated-media provenance — governance, not product.

Salesforce — 35 roles: 31 at the Agent station, 2 upstream. The top archetype is builder (26 of 35), and the modal title is some variant of "Agentforce Forward Deployed Engineer" or "AI Engineer — Forward Deployed Engineer." Salesforce is hiring people to deploy agents at customers on top of models it does not train. There is no Pretrain, no Post-train cluster, no model-layer signal at all.

Same week, same index, opposite ends of the stack. Neither is wrong — they're rational responses to different positions. Meta's bet is that the model is the product and the moat. Salesforce's bet is that the model is a commodity input and the value is in the agent that sits on the customer's data. What determines which bet a company makes is whether it controls the compute underneath.

🚨 What the silicon and full-stack cohort confirms

1. The chipmakers hire the deepest of anyone. AMD is 86% upstream — 13 of 15 roles at Serve, and the titles are pure metal: "Lead GPU Kernel Optimization Engineer," "Principal Software Development Engineer — Compiler & ML Acceleration," "AI Systems Architect — GPU Software/Hardware Co-Design." NVIDIA is 62% upstream with a full ladder — Data through Post-train, then 18 roles at Serve. If your product is the compute, you staff the layer that squeezes performance out of it. There is no clearer "owns the stack" signature in the data.

2. A regulated incumbent can be more upstream than the SaaS leaders. Capital One is 55% upstream — 19 of 34 roles at Data / Pretrain / Post-train / Serve, including a dense "Data Scientist" and "Gen AI Platform" engineering cluster. A bank out-hiring Salesforce and Adobe on model-layer depth is the counterintuitive proof that this is about owning your data-and-compute foundation, not about being a "tech company." It's the same full-stack profile we've flagged for Capital One before, now sharpened against the rent-the-model cohort.

3. The rent-the-model cohort is consistent and large. Salesforce (5% upstream), Adobe (5%, 12 of 17 builders at Agent), Uber (8%), CrowdStrike (20%) — the application incumbents are uniformly concentrated at the Agent station. UiPath is the limit case: across its entire history in our index, zero infrastructure or researcher roles, ever. This is the dominant mode for enterprise software — most companies are renting the foundation and pouring headcount into agents on top.

4. OpenAI is the exception that confirms the rule. Only 7% upstream — which looks like a "renter" until you see that 58% of its roles are Eval or Ship. OpenAI isn't renting a model; it's productizing one it already trained. The compute-ownership lens predicts exactly this: once you own the model, your marginal hire moves to evaluating and shipping it, not building it again.

🔮 Prediction watch

Last pulse's predictions on Google, Oracle, Deloitte, and the consumer cohort can't be graded on this uneven window and roll forward to next pulse, when a clean weekly baseline returns. New predictions from this issue:

• The compute-ownership split is structural, not a window artifact. Prediction: next pulse, with a clean weekly window, Meta + NVIDIA + AMD again post a majority of new roles upstream (Data / Pretrain / Post-train / Serve), while Salesforce + Adobe again post a majority at Agent. Cashable 2026-06-22. If the split collapses, this week was an accumulation artifact; if it holds on clean data, it's the real structure.
• Capital One stays the most upstream non-tech incumbent. Prediction: over the next four pulses, Capital One's upstream share stays above 40% and above every consumer-facing SaaS company in our index. Cashable 2026-07-13. The full-stack-bank thesis lives or dies here.

📰 Reading list

1. Meta in our index — read the FAIR / Meta Superintelligence Labs research titles directly against the Salesforce Agentforce Forward Deployed Engineer list. The two title sets are the whole thesis in raw form.
2. AMD in our index — the GPU-kernel and compiler titles are the cleanest "we own the compute, so we hire the serving layer" signal in the data.
3. Capital One in our index — the bank that out-hires the SaaS leaders on model-layer depth. The counterexample that proves the axis is compute, not industry.
4. /insights/saas-vs-labs — the cohort comparison, now with the full universe restored. The compute-ownership split is visible across the whole board, not just the labs.