An Orientation Map for School of Engineering Students
Artificial intelligence is not one industry. It is five stacked ones — and almost every engineering program touches more than one of them.
Written for third and fourth year students — the point where "specialise, or stay broad?" stops being a theoretical question and starts costing you a semester.
Figures current as of August 2026 · Check before reuse
Ground
These layers sort work, not companies. Google appears at all five — it buys power, designs its own chips, runs data centres, trains models and ships products. The names above are examples of who does that layer's work, not a filing of who belongs where. So when you look at a job, the question is not which company sits in your layer. It is which layer's work you would actually be doing there.
A point of view, not a finding
Everything above this line is a map built from cited sources. This part is an argument. The evidence is sourced; the conclusion is a judgement, and you are entitled to a different one.
Two things are happening at the same time. Hiring has tightened at the top of the stack, where nearly every student trains. And one layer down, employers cannot find people at all.
The layer where hiring has tightened is the one nearly everyone trains for. The layers that cannot find people are the ones almost nobody trains for.
Large technology companies are cutting headcount while raising capital spending. Around half of 2026 layoff announcements name AI as a factor, but that share moved from roughly 7% in January to about 40% by May — AI capability did not improve fivefold in four months, which tells you something about how the word is being used.
Economists are openly divided on cause. New York Fed research attributes most of the recent rise in graduate unemployment to remote and hybrid work, treating generative AI as a smaller layered influence; a National Bureau of Economic Research working paper found 90% of executives reporting no AI employment impact at their own company. The effect on entry-level hiring is real regardless: employers increasingly ask for three to five years of experience on roles once open to graduates, and one venture firm's tracking found large technology companies cut new-graduate hiring by about a quarter in a single year.
On the other side, Uptime Institute found roughly half of data centre operators struggling to recruit qualified candidates, with skilled electrical trades and senior facility operators hardest to fill. Estimates of the 2026 shortfall in data centre construction labour run into the hundreds of thousands. Schneider Electric's assessment is that talent, rather than power or semiconductors, may become the primary barrier to scaling AI.
India shows the same split differently: NASSCOM and Deloitte project the AI talent pool roughly doubling to over 1.25 million by 2027 while demand grows faster still, and the Ministry of Electronics and IT estimates only about 16% of Indian IT professionals are AI-skilled today.
The Stanford study that found the sharpest effects on young workers named the mechanism: AI is strong at codified knowledge — learned in school, applied to routine tasks — and weak at tacit knowledge, earned by doing. That is an exact description of what a fresh graduate has and does not yet have. The part of your education that can be written down is the part that just became cheap.
Before committing a year to something, ask: will this still be true in ten years? If yes — power systems, semiconductor physics, distributed systems, control theory, thermodynamics — go deep. If no, it is a tool. Learn it in a weekend and do not build an identity on it.
Choosing a specialisation is a congestion game: your payoff depends not on which route is objectively best, but on how many others choose it. Everyone reasoning independently toward "AI is hot, so do AI" produces a crowded road and a thin payoff, while four other roads sit empty.
So the question is not which layer is best? It is which layer is best, given where everyone else is heading? And then: if my whole class reads this page, does my answer change?
Crowding tells you where the empty roads are. It does not tell you which one you would actually walk down.
So don't ask what you are interested in — look at what you already do unprompted. Nobody had to tell you to do it, which is what makes it evidence.
One weekend does not count. Look for the thing you kept coming back to — one afternoon of curiosity is noise, three years of returning to the same thing is a signal. Then notice which layer it was already in:
It was not a hobby. It was an unpaid apprenticeship in a layer.
If your evidence points at an empty road, go early. If it points at the crowded one, go anyway — but only the deep end pays. Being one of ten thousand people who can call a model through an interface is not a position. Being one of fifty who can make inference cheap at scale is. Pick one layer, go deep enough that someone would let you be accountable for something in it, and get literate in the layer next door. Breadth has never been cheaper to acquire later. Depth has never been harder, and there is no tool for it. Spend the years you have left on the expensive half.
If you agree with any of the above, it should change five things before the semester starts. None of them require permission. Where this map comes from
In March 2026, NVIDIA's chief executive Jensen Huang published an essay arguing that AI should not be understood as a clever app or a single model, but as infrastructure — closer to electricity or the internet than to software.
He described it as a five-layer cake: energy at the bottom, then chips, then infrastructure, then models, and applications at the top. He had made the same argument in January 2026 at Davos, calling the build the largest infrastructure project in human history.
Most public conversation about AI happens at the top two layers — the chatbot you can see and the model behind it. The five-layer view makes three things visible that the chatbot view hides.
Read this with one eye open
This framework comes from the chief executive of the company that sells much of layers 2 and 3. It is a genuinely useful map, and it is a map drawn by an interested party.
It also leaves things out: no data layer, no talent layer, no capital layer, no regulation layer — all of which shape this industry as much as megawatts do. What it adds, and why it is worth using, is putting energy at the bottom, where most technology framings forget to look.
Pick a program or a field
Fields that cut across layers
Programs
Nothing selected. Every layer is shown.
A second cake, not a sixth layer
These are not a layer of the AI-factory stack. They are a parallel stack, usually called physical or edge AI, with its own five layers: batteries instead of grid power, low-power inference chips instead of data-centre accelerators, on-device runtimes instead of clusters, small compressed models instead of large ones, and machines instead of software.
This matters locally. Mindgrove and InCore in Chennai build chips for exactly that second stack, not for the data-centre one. It is a real fork in the road, and worth choosing deliberately rather than drifting into.
Every figure on this page carries three things: what it measures, what period it covers, and who published it and when. That discipline matters more than the numbers. The same organisation publishes different figures across a single year as it revises — Gartner notes it widened the scope of its AI forecast between vintages, so comparisons across vintages are not meaningful.
Gartner places AI in what it calls the trough of disillusionment through 2026, and says predictability of return on investment must improve before enterprises can truly scale AI up. The gap between what is being spent on infrastructure and what applications actually earn is the central open question of this cycle. A page that showed only the upside would be teaching you the wrong instinct.
India's data centre capacity is a good worked example. CBRE reports stock crossing 1,700 MW at the end of 2025. Wood Mackenzie puts operational capacity at 2.2 GW in the same year. Both are defensible — they measure different things, third-party colocation stock versus total operational capacity. When two credible numbers disagree, the definition is usually the reason.