There’s a chart circulating that wants you to feel a specific kind of vertigo. It plots “annualized AI spend per employee” for the top 1% of firms, draws a line for the average software engineer at $192K a year, and lets the curve arc up to meet it by the end of 20261. The implication is left tastefully unstated: the machine is about to cost as much as the person, and then more.
I use these tools every day, at the level where their value is real and not theoretical, and I do not believe that chart. Not because the spend isn’t rising; it is, but because the chart measures the wrong thing, extrapolates it dishonestly, and then invites you to compare it against a number it has nothing to do with.
Start with what the number is
The endpoint everyone reacts to is a projection. The observed figure, from the Ramp AI Index, is about $7,449 per employee per month for the top 1% of firms2; annualize it, and you get roughly $89K, still less than half of what a software engineer costs. Ramp’s own lead economist says it plainly: despite the proclamations that you should spend as much on AI as on an engineer, almost no one is actually doing it3. The terminal point on that chart is a 14.1% month-over-month growth rate extended forward as if a compounding early-adopter exponential never has to bend. It always has to bend.
The chart’s scary number and any sane number about developer productivity aren’t in conflict. They’re different denominators wearing the same axis label.
And that denominator matters more than the curve. “AI spend per employee” is total firm AI spend divided by total headcount. For the companies at the top of that distribution, the AI-native ones, the bulk of that spend is the product itself: inference served to their own customers. It is cost of goods sold, not a productivity multiplier bolted onto a developer. When you isolate the thing the chart makes you afraid of, a coder with an agent, the number collapses.
Microsoft found its own engineers running $500 to $2,000 a month on Claude Code4, which annualizes comfortably below any reasonable ceiling. The frightening headline and the boring reality are simply not measuring the same activity.
The metric that would actually matter is the one nobody puts on a slide: fully-loaded cost per resolved task. Token spend is an input. A firm burning $7,449 an employee could be resolving fewer tasks per dollar than one spending a fraction of that with disciplined agent architecture; an orchestrated agentic workflow runs roughly thirty times the cost of a simple linear call.
High spend is as easily a signal of bad engineering as of high output. The chart can’t tell the difference, because it’s counting the fuel and calling it distance.
The number that actually tells you something
So if the spend curve is noise, where’s the signal? It’s in what the people who build these models are doing with their own money.
On July 2, Microsoft stood up the Microsoft Frontier Company5: $2.5 billion and 6,000 employees whose job is to embed inside customers and build and run AI systems on-site6. This is forward-deployed engineering, and it is not a Microsoft quirk. Amazon put a billion dollars behind its own version. Anthropic and OpenAI both stood up FDE groups in the spring. Every frontier player is, at the same moment, spending serious money to put humans into the room.
Read the stated reason and the whole thesis falls out. Companies bought the tools, ran the demos, and discovered the payoff was harder to capture than the demo suggested — because deploying AI inside a real business, with its own data and rules and entrenched habits, is difficult. As Goldman’s Marc Nachmann put it, “having the model alone doesn’t change your workflows”; you need people who can fuse the technology to what the business actually does. The scarce input isn’t the model. It’s judgment.
You do not spend $2.5 billion on human implementers if you believe the next checkpoint dissolves the problem.
This is the part I want to sit on, because it inverts the reflex. The reflex says: the tools underperform in most hands, so we need stronger models to close the skills gap. But look at what the labs are doing. If they believed a stronger model closed this gap, the rational move is to ship the weights and wait.
Instead, they are hiring armies of people to sit inside customers and do the integration by hand. That is the single loudest piece of evidence available that the gap they face is not a model-capability gap. The enterprise consensus has already turned the same corner: the running theme of 2026 is that better outcomes don’t require a bigger model so much as better orchestration into real workflows7.
“Wait for the next model” is simply a more sellable story to a customer than “you have a judgment problem,” so it’s the story that gets told.
Which gap, and for how long
There are two different skills getting collapsed here, and the distinction is the whole game.
- The first is the low-level operating skill: coaxing, prompting, working around the model’s dumbness. That one genuinely commoditizes. Every capable release makes the tool more forgiving, and the narrow craft of getting the machine to cooperate erodes on schedule. If your edge is that you’re good at driving Copilot, your edge has an expiration date.
- The second is the meta-skill: knowing what to build, decomposing it into pieces an agent can actually execute, and, the part that keeps getting underpriced, verifying the result.
Stronger models don’t retire this skill. They raise its premium.
More capable models produce subtler errors and more plausible wrong answers, and agentic systems widen the blast radius of any bad call. The scarcer thing in that world is the person who can look at confident output and know it’s confidently wrong. That person gets more valuable as the model gets better, not less.
The window closes fast for people who are good at driving the tool. It stays open for years for people who know what to build.
That’s why “short-term advantage” is only half right. The advantage bifurcates. It commoditizes quickly for boilerplate and well-trodden patterns. It stays open, for a long time, at the level of architecture, novel domains, and integration — precisely the work the FDE armies exist to brute-force with bodies, because you don’t send six thousand people to do something you already know how to encode in software.
The gap doesn’t close when the model gets stronger. It closes, slowly, when tacit judgment gets institutionalized, written into playbooks, frameworks, and eventually tooling. That is a human and organizational process, and it moves at the speed of humans, not the speed of a release cadence.
I watch a smaller version of this every week. A team gets faster with the tools and its defect rate doesn’t move. The capability is present; the value isn’t landing; and the missing piece is judgment that no model upgrade supplies.
That is the Frontier Company problem at the scale of a single standup.
The distributional wrinkle is the uncomfortable one, and I’ll only gesture at it here. If the surplus from all this compression flowed to the skilled operators creating it, this would be a story about raises. It mostly won’t, because the surplus flows to whoever holds pricing power, and a 5x that becomes the baseline stops commanding a premium. The durable edge, then, isn’t being 5x. It’s being the person whose judgment other people are currently paying $2.5 billion to rent, and, eventually, being the person who can encode that judgment so it stops needing to be rented at all.
The chart wants you to watch the spend line. Watch the hiring instead.