Where Tomorrow’s Engineers Come From, Part 1: The Role That Survives
AI is reshaping which roles persist and which fade. The honest answer matters for the 17-year-olds choosing majors, the directors making AI strategy, and everyone in between.
Written with Claude Opus.
I said at the end of the last post that it would be followed in a couple of days by the piece I’m finally writing now. It has been three weeks. Two things stretched the gap: interviews for the role I’ll be announcing in late June, and an argument that got bigger the more I worked it, big enough that what was planned as a single post is now the first of two. Part 1 starts with what is actually changing in the role landscape. Part 2 will follow shortly and gets to the mechanism: how senior engineers have historically been produced, and why that path is breaking.
The setup is from Post 6: if AI is doing the activity that used to teach junior engineers the judgment senior engineers grow into, where do the senior engineers of five and ten years from now come from? That question has a real and uncomfortable answer that the industry has not yet engaged with honestly. The honest version requires being precise about which roles are changing and which aren’t, because the public discourse (”AI is taking engineering jobs”) collapses categories that need to stay separate.
What the data says, and what it misses
The public discourse on the chip-industry workforce shortage is anchored to a set of projections completed in 2022 and 2023. SIA and Oxford Economics: about 67,000 unfilled US jobs by 2030. McKinsey: a worst-case scenario of up to 146,000 unfilled engineer and technician positions by 2029. SEMI and Deloitte globally: roughly a million additional skilled workers needed by 2030. Those numbers are real, and you will see them quoted everywhere.
They are also from a different industry than the one we are now in. Every one of those projections models engineer demand against roughly constant productivity assumptions. That was the right baseline in 2022. It is no longer the right baseline now. AI is doing to engineering work what spreadsheets did to the work of accounting clerks. It is not eliminating the profession, but collapsing the labor structures the profession used to need. The projections count bodies in the role categories that existed when they were written. Some of those categories are already extinct in everything but org charts.
The observations that matter most to me are not from those projections. The first is that roughly one-third of the US chip workforce is age 55 or older. The second is that electrical engineering enrollment in the US has declined for three decades: bachelor’s and master’s degrees in EE grew only 37.5% from 1997 to 2020 against 81.1% across all other fields, and just 18.2% among US citizens. Whatever the AI-adjusted headcount shortage turns out to be in 2030, the senior engineers who would have trained the next generation are leaving, and the pipeline of replacement engineers was already thin before AI added another reason for students to choose something else.
And the pipeline problem is worse than thin. The common narrative right now is not that AI will replace junior engineers at some point in the future. It is that AI is replacing them today. The entry-level tasks that used to be a junior’s first assignment increasingly go to an AI tool with a senior reviewing the output. In isolation, that reads as efficiency. Read against the retirement data, it is something else entirely. Seniors are leaving from the top. Juniors are being hired in smaller numbers, and the ones who are hired increasingly do not do the work that used to build judgment. The two trends compound. Ten or fifteen years out, the industry will not simply be short of engineers. It will have engineers who were never given the path to become senior ones. That is the deeper problem, and it is not a headcount problem. The headcount may look fine on a spreadsheet. The judgment will not be there.
So the real question this post is about is not how many engineers will be available, or how many roles AI will displace. Both matter, but both are downstream of a more fundamental shift. The shift is in which skills the industry needs and how those skills get formed. That is the question worth dwelling on, and it is the one the workforce projections, written before the shift was visible, cannot answer.
What persists, what fades
Here is the distinction the public discourse needs and rarely makes.
Fading: the pure coding role. The engineer whose primary value is converting microarchitectural intent into synthesizable Verilog (writing the RTL, debugging the simulator, tweaking until lint passes) is the role AI substitutes most directly. Not on day one, not for every block, but as a long-term career trajectory this is shrinking. The blunt version: if you want to be a coder, nice having met you. The pure-coding path as a thirty-year specialty is closing. That is not hyperbole; it is what the next decade of AI tooling looks like applied to a role whose primary output is text in a known formal language.
Software is a few years ahead of hardware on this curve, and the comparison is instructive. Software has produced what is now called vibe coding: generating code by prompting an AI, iterating quickly, shipping without fully understanding every line. It works in software, after a fashion, because the feedback loop is cheap. You run it, it breaks, you try again, and a bad deploy can be rolled back. Hardware grants none of that. The feedback loop runs through synthesis, timing closure, weeks of verification, and finally silicon, where an escaped bug costs millions and months. You cannot vibe-code an SoC. But that does not rescue the coding role. AI still generates the RTL. What hardware’s unforgiving feedback loop removes is the escape hatch: there is no version of this where a junior prompts a model, ships something plausible, and is done. The output has to be correct, and being sure it is correct is the judgment work. Hardware cannot hide the fading of the coding role behind a vibe-coding mode. The coding is going. What is left exposed underneath is judgment.
Persisting: the higher-complexity roles. Architecture, microarchitecture, methodology ownership, sign-off authority, debug under field-failure conditions, integration judgment, spec authorship. These roles are not shrinking. They are getting more load-bearing, because when AI does the coding the judgment layer is the only filter between intent and tape-out. The persistence is structural: someone has to decide what to build and verify that what got built is correct. AI cannot do either of those jobs well enough to remove the human today. It may narrow the gap, and the vendors will announce that it has. But the claim that it has closed the gap is the one that needs evidence, and that evidence does not exist yet.
This distinction reframes the apprenticeship question in a useful way. The crisis is not “AI is coming for engineering jobs.” It is that AI is coming for the entry-level jobs that have historically served as the path to the higher-complexity roles, while the higher-complexity roles themselves are getting more important. Different shape of problem. Not job loss. Path collapse. The destination still exists. The on-ramp is being demolished.
That distinction is also the honest answer to the AI-anxiety enrollment concern. A student choosing a chip design career today should hear: the coding entry-level role earlier generations used as a first rung is changing, possibly disappearing as a long-term track. But the higher-complexity roles persist and are more valuable than before, not less. The career bet for a seventeen-year-old today is not “will I have a job in ten years.” It is “can I get to the right kind of role before AI completes its takeover of the lower-complexity ones.” That is a sharper and more useful framing than either the doomsday version (AI is taking everything) or the dismissive version (don’t worry about it). And it ties back to what Post 1 of this publication argued: there is AI coming FOR you, and there is AI coming for YOU. The coding role being substituted is the FOR-you scenario. The architecture role being amplified is the for-YOU one. Same distinction, applied to the career-choice question.
A note on managers
Some of the workforce numbers include a specific projection worth flagging. SEMI estimates the industry needs to add at least 100,000 second-line leaders and 10,000 third-line leaders by 2030, with the industry’s own consensus being that many of those managers will have to come from outside the chip sector because the internal pipeline cannot grow them. That is a striking sentence: an industry-consensus admission that the apprenticeship path is broken.
The total number is real. The breakdown is probably wrong, and how it is wrong matters.
The SEMI projection assumes current org-chart ratios hold: the number of managers needed scales with engineering headcount roughly the way it does today. There is good reason to think those ratios are shifting. Middle management as a coordination function exists in part because coordinating across many engineers is expensive: status reports, work allocation, blocker resolution, performance management of N reports. As AI compresses the engineering activity layer, the coordination overhead compresses with it. A reasonable alternative future is a model where very-senior engineers run small groups and do the deep work themselves, growing juniors as a side effect of working alongside them rather than as a separate management function. Less pure people-management, more technical leadership.
But that does not mean pure managers disappear, and the post should not overclaim in that direction. People-management exists for reasons that are not only coordination overhead: cross-team negotiation, strategic alignment, headcount advocacy, calibration across functions, career development at scale. People skills are a distinct competency from technical depth, and they do not automatically bundle into the same person. The current trend toward flatter organizations and less middle management is real but probably overshooting. The pendulum has swung far against managers, and a correction is likely, even if its shape is hard to predict.
Whichever way it lands (fewer pure managers and more very-senior technical leads, or some restoration of separate management roles alongside engineers grown from within), the implication for the apprenticeship problem cuts the same way. Either future demands more senior engineers with judgment. The technical-lead future demands them with people skills bolted on, which is a harder pipeline to grow, not an easier one. The restored-manager future still demands senior engineers as the source of technical judgment, with managers as a separate role drawn from somewhere. In neither future does the apprenticeship problem get smaller.
Where this leaves us
The shape of the answer is starting to become visible. The roles AI displaces are not the roles the industry actually depends on for tape-out outcomes. The roles the industry depends on are getting more load-bearing. But the path that used to grow people into those roles ran through the displaced roles, and AI is removing the intermediate steps faster than any alternative path is being built.
Post 1 of this publication argued that the engineer’s job is moving up the abstraction stack, not disappearing. That is still right. What this two-part piece is doing is taking the abstraction-stack frame seriously and asking the question that follows from it. If the destination is higher on the stack, and the activity that used to get juniors there is being absorbed by AI, what replaces the activity as the pedagogy? How does someone who never writes RTL professionally develop the intuitions that come from writing RTL professionally for ten years? Can those intuitions be taught? Can they be simulated? Can the new EDA tools (both the traditional vendors and the new AI-native companies entering chip design) encode enough of accumulated engineering history to replace the apprenticeship that used to happen at someone’s elbow?
Part 2 is about those mechanism questions. There is a name for the specific thing that is hardest to transfer, and it is the heart of the apprenticeship problem. Part 2 will get there.
Marco Brambilla is a semiconductor industry veteran with 25 years in chip design, most recently as Senior Technical Director at Meta Reality Labs. He writes about AI, chip design, and the future of hardware engineering at Above the RTL.