The eight kinds of employer, and how many of each

The picture in most people’s heads is the smallest box

Ask someone where AI jobs are and they name four or five companies that train frontier models. Those organisations together employ a few thousand technical staff worldwide, many with doctorates (Anthropic says about half of its technical staff have one), nearly all in a handful of cities. That is a real employer category. It is also the one you are least likely to work for and the one that shapes almost none of the market you will actually apply into.

Here is the fuller list. Read it as a map of where to spend your applications.

1. Frontier labs and big-tech research

Tiny, famous, PhD-heavy, publication-driven. Salaries at the extreme top. If you are not already publishing, this is not a plan.

2. Big-tech product teams

Large and real: the recommendation, search, ads, fraud and infrastructure teams. They hire steadily but skew senior, and their junior intake mostly comes through internships and graduate programmes with fixed windows. If you are outside that pipeline, entry is hard and lateral entry at three years’ experience is much easier than entry at zero.

3. IT services and outsourcing

Infosys, TCS, Wipro, HCL, Cognizant, Accenture, Capgemini, Globant, and hundreds of smaller firms. In India, the Philippines, parts of Latin America and Eastern Europe this is the single largest employer of junior technical people, by a wide margin.

The economics are worth understanding because they change your strategy. These firms bill clients for your time and often bill more for certified staff, which is why a cloud certification has genuine commercial value here and almost none elsewhere. You may sit “on the bench” between projects. Training is real but generic. Pay starts low and rises slowly.

The honest read: it is a legitimate and very wide door, and a large number of working data engineers walked through it. Take the job, get two years, then move.

4. Product startups

Twenty to three hundred people. They read applications, they hire on evidence, and there is no qualifications filter because there is no HR department to run one. This is where a portfolio does the most work per hour you put into it. Against that: less mentorship, more chaos, and a real chance the company is gone in eighteen months.

5. Enterprises with internal data teams

Banks, insurers, telecoms, retailers, airlines, hospitals, manufacturers, logistics firms, utilities. This category is where most analysts and data engineers on earth actually work, and it is almost never discussed, because a regional insurer does not run a developer marketing campaign.

They have decades of data, real problems, formal processes, and hiring that runs on qualifications and references. Slower, steadier, and far less competitive per opening than a known technology brand.

6. Agencies and consultancies

Small firms doing project work for other companies. Broad exposure, many stacks, fast learning, and constant context-switching. Good for building range early. Hard to go deep.

7. Public sector, universities, NGOs, research institutes

Census bureaus, health ministries, transport authorities, agricultural research bodies, international organisations. Lower pay, slow processes, interesting data, and unusually low competition because almost nobody applies. Also unusually good for a first portfolio-driven hire, because these organisations publish their data and you can build something useful before you ever apply.

8. Yourself

Freelance and contract. Covered in its own lesson. Not a fallback: a category with its own economics.

The ratio nobody prints

For every organisation that trains its own foundation models, there are thousands that merely run software and have accumulated data they cannot use. Categories 3, 5 and 7 hold most of the jobs and generate the least conversation.

If you are optimising for a first job, apply where the jobs are, not where the news is.

The “AI team” that is not one

A caution you will meet quickly. Many companies now advertise AI roles while having no model in production and no data warehouse. The actual job is spreadsheets, one integration with a hosted model API, and a dashboard. This is not fraud; it is where those companies are. But go in knowing it, because it decides what you learn in your first year. The interview lesson on questions to ask exists mostly for this.

Finding these employers without paying anyone

The paid tools (LinkedIn Recruiter, Sales Navigator, Apollo, paid job aggregators) sell you a filtered list of companies. You can assemble the same list free:

• LinkedIn’s free search, filtered by industry and employee count, in your city.
• Company career pages directly. Most enterprise roles are posted there and syndicated late or never.
• Your national employment portal, industry association member lists and chamber-of-commerce directories.
• Government tender and procurement portals, which name every firm doing data work for the state.
• Wellfound and Y Combinator’s job board for startups, and the careers pages of the companies whose open models and tools you already use.

An afternoon spent building a list of forty local employers in categories 3, 5 and 7 is worth more than any subscription.