On July 9, 2025, Bloomberg reported that Meta had paid Ruoming Pang, the engineer who ran Apple's foundation models group, a compensation package worth over $200 million to join a new Superintelligence team. Then, over the summer, Entrepreneur reported that Mark Zuckerberg had offered $1.5 billion to a single person and been turned down. The numbers were meant to be signals. They signaled that AI research talent had priced itself into a category with two entries above the CEO of Apple and none below.
The recruiting response was predictable. Every enterprise procurement deck for the last four quarters has a slide with the words "hire AI engineers" on it. Fortune 500 heads of engineering nod. A search opens on Workday the same afternoon. The market then opens two very different rooms behind the same door.
The market is two markets
Robert Half and Kore1 have both published 2026 AI engineer compensation data this year. The two data sets converge on the same picture. A senior AI engineer in the Bay Area averages about $252,000 in base pay. Total cash and equity for a working AI engineer at a funded enterprise runs $250,000 to $500,000. Then the curve breaks. A staff engineer at OpenAI or Anthropic pulls $795,000 and up. Frontier-lab principals clear a million. And Meta will pay a full nine figures for the specific engineer who has trained a general-purpose foundation model and knows how the loss curves broke on the last one.
Two markets. One job title.
The confusion is not accidental. Recruiters price on scarcity, and the scarce role is the one Meta priced. A mid-market retailer writes "Senior AI Engineer" on a requisition, gets a shortlist priced against the frontier lab market, and either overspends by 3x or stalls the search for eight months. Neither outcome produces the team the retailer needs. What the retailer needs sits on the other rung. That rung is trained on different problems and paid from a different budget.
The Rung You Can Reach is the rung below Meta's poach. Different job, same $170,000-to-$310,000 band as any senior software role at a serious company.
What each rung actually does
The lab-caliber rung, the one at $600K to a million and above, trains models. It ports architectures across accelerator generations, shepherds a fifty-billion-parameter training run past a loss spike at week six, runs the ablation that decides whether attention should be windowed or global at 2 million tokens, and reads the resulting loss curves for signal the checkpoint could hide. There are maybe eight thousand people on the planet who have shipped that work. OpenAI is trying to hire about 3,500 of them into an 8,000-person company by year-end, per its own filings covered widely through 2026. Meta is trying to hire the top three hundred. That is the pool your recruiter's shortlist is drawn from.
The applied rung composes models rather than training them. Its engineer writes the evaluation harness that decides which model to route which class of query to, builds the tool interface the agent uses to read your CRM without hallucinating a customer that does not exist, debugs why the ranking function put the wrong answer in slot one, tunes the prompt, adds the retriever, benchmarks the fix, ships the change on Wednesday, and reads the trace on Thursday morning. This is the work that puts an AI feature in front of a paying customer. Engineering work, judged by engineering metrics.
Both rungs answer to "AI engineer" in a Workday req. The applied one is the one you can actually hire.
The poach doesn't reach you
Say the words out loud. Meta cannot poach the person you need. The Superintelligence team is shopping in a different market entirely, one you are not competing in. Your competitor down the street is not there either, and neither is the boutique consultancy that quoted your CIO $2 million for a pilot. The applied rung is a real market, deep, growing 4.1 percent by tier in 2026 per Robert Half data, and it clears at roughly $170,000 at the 25th percentile and $235,000 at the 75th. You can win in that market with a competitive offer and a real problem to solve. You cannot win at the frontier lab tier because the frontier lab tier is not shopping.
That should be a relief. Instead, three quarters of enterprise hiring plans I read conflate the two rungs on purpose. It sounds better in the board deck to say "we are competing for the same talent as OpenAI" than to say "we are hiring a small team of engineers who will make the model do a specific thing our business needs."
The first sentence never survives a budget cycle. The second one ships.
The team is smaller than the org chart wants
A functioning internal AI team, at a company that is not itself a foundation lab, fits in a room. The composition maps to the surface of the AI product being built rather than to a symmetrical org chart.
A retailer building agent-driven customer service needs about five people. One lead applied engineer with model integration experience. One evaluation engineer who owns quality gates. One data engineer who owns the retrieval store. One domain product owner from customer operations who can read a trace. One MLOps engineer who owns the observability of every request. Five people. The applied rung stays thin because you do not need it to be thick.
The MIT NANDA finding from July 2025, which every enterprise IT outlet republished through August and September 2026, put a number on the failure mode. Ninety-five percent of enterprise generative AI pilots produced no measurable P&L impact. Enterprise spend crossed $37 billion. The five percent that shipped put small teams close to a specific business surface, gave those teams the authority to change the surface, and evaluated the result against a metric the finance team already tracked. Nothing about that requires poaching from Meta.
The Klarna reversal is a hiring lesson
In February 2024, Klarna's CEO Sebastian Siemiatkowski told the market that an OpenAI-built agent had done the work of 700 human customer service agents in a month, cutting resolution time from 11 minutes to under 2. In May 2025 he told the press the same firm was hiring humans back, because "lower quality" service was showing up in retention numbers his original dashboard didn't measure. Through 2025 and into 2026, Klarna rebuilt a hybrid team.
The internet framed this as evidence that AI is not ready. The framing is wrong. Klarna's agent did the routine work exactly as promised. The reversal happened because no internal team was owning the escalation path or reading the tickets the agent misclassified. Klarna outsourced a capability that needed to be internal. When the capability broke, the vendor could not repair the breakage, because the vendor did not sit inside the retention conversation. Klarna rebuilt the missing rung.
If you take one thing from Klarna, take this. The applied AI team owns the loop between what the model does and what the business measures. Vendors can supply the model. The loop stays inside.
Compensation, and why the board deck is wrong about it
The compensation numbers people quote are usually the wrong numbers. When a Bloomberg piece runs the Pang $200 million figure, it becomes the anchor for every internal budget conversation for six months. Executives who need to hire an applied engineer walk into the meeting expecting to have to break a bank. Then they see Robert Half's median at $200,000 all-in, and something unhelpful happens. They decide the market must be wrong, because it can't really be that cheap if Meta pays what Meta pays.
The gap between the two prices two different products. The Meta package buys optionality on the future of general intelligence. The $200,000 applied hire buys the person who will read a trace on Thursday morning and know why the model chose the wrong customer segment. Both are valuable. Only one is what your operation actually needs.
The variance inside the applied rung is real and worth respecting. Kore1 puts the range at $145,000 to $310,000. The top end prices scarcity within the applied rung, where engineers have already shipped an evaluation harness against a real business metric. Every enterprise wants that person. Almost no enterprise has trained one, because training one requires giving them ownership of a real business metric first. Two moves. Write the job description around ownership of one measurable business surface. Pay the top of the applied band for that ownership. Everything else is recruiter theater priced against a market you are not shopping in.
Where the team sits on the org chart
The applied AI team does not sit inside a central "AI Center of Excellence" if you want it to work. Every post-mortem of a failed enterprise AI program I have read contains the phrase "the AI CoE was too far from the business." The CoE model is a structural apology for not knowing where to put the function. Put it next to the surface it modifies.
A customer service AI team belongs inside customer service, answering to the retention number rather than a central AI leader. The same rule applies to finance (owned by the CFO, judged by the close cycle) and to product (owned by the head of product, judged by activation).
The exception is the platform layer. If more than one surface team needs the same retrieval store or the same evaluation infrastructure, that shared plumbing wants its own two or three engineers behind an API. Same shape as any platform team inside a mature engineering org. Infrastructure with a specific mandate.
Two structural rules cover most of the failure modes.
The first rule: every applied AI engineer reports through the surface they modify. If they report to a central AI leader who does not answer to the surface's business metric, the loop between model output and business outcome breaks. Klarna is the canonical example.
The second rule: shared plumbing lives on a platform team. If your "AI leadership" spends its week writing decks about AI strategy and holding office hours with vendors, the applied work is happening despite them. Move the surface engineers back into the surfaces, and let the plumbing team live behind an API.
The hiring interview that filters correctly
The interview is where the two rungs sort themselves. A frontier-lab interview asks about attention variants and how you would rescue a stalled training run past week six. An applied interview asks about evaluation. Specifically, it asks the candidate to bring a real evaluation harness they have shipped against a real metric, and to explain a failure mode the harness caught and a failure mode it missed.
Almost no one can answer the second question honestly, because most enterprise AI work is still shipping without evaluation harnesses. The candidate who can walk through a real failure mode the harness missed is the candidate you want. They have already learned the expensive lesson on someone else's payroll.
The corollary is that the applied engineer you hired six months ago will not interview well against a research candidate today. That is fine. You are not hiring for research. Filter your recruiting funnel accordingly and stop letting your recruiter show you profiles priced against a market you are not shopping in.
The build-versus-buy question is a distraction
Every AI vendor is selling a "Center of Excellence in a box." Every consultancy is offering to staff your applied team. Every hyperscaler is bundling an implementation partner into the model contract. All of that is real, some of it is useful, none of it substitutes for owning the loop.
You will spend money on vendors regardless. The real question is which parts of the stack you own outright. The model itself is a rental at every scale below a foundation lab. The eval harness, the routing logic, the tool interface into your business systems, and the observability of every trace against a business metric are the differentiated part of the stack, and none of those are rentals. Own them. Staff them on the applied rung, and let the vendor supply the model.
The mistake to avoid is the mirror-image mistake. Some companies pull all of the applied work in-house and end up staffing a small foundation lab by accident, chasing the frontier rung they cannot afford. Others outsource the evaluation harness to a consultancy, which is the same as outsourcing the loop, which is Klarna. The two mistakes look opposite and produce the same outcome: neither company owns the surface the model touches.
What the next twelve months look like
The September 2026 hiring data shows layoffs rising 199 percent month over month, driven by internet companies restructuring around AI. OpenAI is doubling headcount, and Meta is still poaching. The applied rung, meanwhile, deepens at 4.1 percent per tier per year, which is more than double the average tech salary trend of 1.6 percent.
The signal in that data is that the frontier lab market is thickening at the top and the applied market is deepening across the middle. Both markets are functional. The failure mode is executives who read the frontier headlines and try to hire from that market to solve a problem that lives in the middle market. They do not find anyone, they spend the search budget, and the initiative quietly stalls. Then the executive blames the technology. The technology is fine. The rung was the wrong one.
Every quarter of that stall costs measurable ground. Salesforce is embedding Agentforce inside Claude and Claude inside Salesforce, per Salesforce's September 2026 announcements with Anthropic. ServiceNow, SAP, Microsoft, and Google Cloud (through Accenture's Gemini Enterprise partnership) are pushing the same shape from the other side. Gartner's forecast for end of 2026 puts 40 percent of enterprise applications inside a task-specific agent, up from under 5 percent in 2025. The applied team turns those agents from a vendor demo into a working part of the operation. You either have that team by early 2027 or you are running your operations on someone else's default configuration.
The architecture question
The reason this piece is not titled with a compensation number is that hiring is downstream of architecture. The applied team is only useful when the surfaces it will modify have been architected to accept its modifications. That architecture work runs before the first requisition opens. Which business surfaces will the models touch. Which evaluations will decide whether the touch was correct. Which vendors supply the model, and where in the pipeline is the seam between rented and owned. Which surfaces get a full loop and which get a single-turn call.
None of those questions have generic answers. Every real answer is specific to the operation, the data, the customers, and the risk appetite of the enterprise. Off-the-shelf blueprints exist and they get most of the answers wrong for any specific business, because they cannot afford to know the operation. The applied team can only start once someone has answered the architecture questions specifically enough to write a job description that means something.
This is the part where you either do the work of architecting the surface, or you skip it and end up back in the ninety-five percent that produced no P&L impact. Every consultancy that promised you a third path was selling a pilot budget dressed as strategy.
Agor AI Advisory does the architecture. We sit with the operator, name the surfaces, write the evaluations, specify the seams, and hand back a working design a five-person applied team can implement. Then we help you hire that team against the design, at market rates on the rung you can reach, not the rung your recruiter wants to sell you.
Sources
- Meta Poached Apple's Ruoming Pang With Pay Package Over $200 Million, Bloomberg, July 9, 2025
- Mark Zuckerberg Reportedly Made One Person a $1.5 Billion Job Offer, Entrepreneur, 2025
- AI Engineer Salary 2026: $145K, $310K (Real Offer Data), Kore1
- AI/ML Engineer Salary (Updated for 2026), Robert Half
- AI Compensation Benchmarks 2026: The AI Hiring Bubble, Pin
- Salesforce CEO confirms 4,000 layoffs 'because I need less heads' with AI, CNBC, September 2, 2025
- Klarna Reverses AI Push, Says Customers Prefer Human Support, Forbes, May 18, 2025
- AI Skills Shortage: How Enterprises Are Closing the Gap, AI Conference London, August 14, 2026