The exciting layer runs on a boring one
Intelligence gets the headlines. Underneath it sits a layer nobody writes about: electricity, transformers, cooling water, fibre, physical security, and the people who keep all of it running at three in the morning. That layer is not optional, and unlike software it cannot be scaled by spending alone. This page is about the constraints — and about the difference between a sector mattering and a share price rewarding you for noticing.
It is electricity, and it has a queue
The head of the company selling the chips said it plainly on an earnings call: if we had more compute, we could generate more profitable tokens. Compute is the constraint. But compute has a constraint of its own, and it is not fabs — it is power, and the land and grid connections that power arrives through.
| What is documented | Figure |
|---|---|
| Global data-centre electricity use, 2025 | ≈ 485 TWh |
| IEA projection for 2030 | ≈ 950 TWh, about 3% of world electricity |
| AI-optimised data-centre demand by 2030 | More than quadruple |
| US share of global data-centre electricity, 2024 | ≈ 45% |
| Projected growth in US data-centre demand to 2030 | ≈ +130% |
Source: International Energy Agency, Energy and AI. Projections are scenarios, not promises, and the IEA revises them. Check the current edition before quoting these numbers.
Two features of that demand make it harder to serve than the raw number suggests. It concentrates. Electric vehicles spread load across a country; a gigawatt-class cluster lands in one county. And it arrives faster than grids are built — connection requests queue, planning permission takes years, and the transformers and turbines have order books measured in the same units.
"Just one of these covers a significant part of the footprint of Manhattan."
That is the sentence to keep. Whatever happens to any model company, the concrete, copper and generating capacity have to exist first, and they take years that money cannot compress.
What has to be kept running, and who keeps it
A useful way to think about the next few decades is to ask which parts of the system fail if neglected for a week. Those are the parts that get maintained forever, and maintenance revenue behaves very differently from product revenue: smaller, duller, and far more repeatable.
| Layer | What it actually involves | What it connects to |
|---|---|---|
| Generation | Gas turbines for firm capacity, nuclear including the small-reactor projects now being signed against data-centre demand, and renewables plus the storage that makes them dispatchable | The utilities, turbine makers and EPC contractors — old industries with long order backlogs and, usually, regulated returns |
| Transmission and equipment | Transformers, switchgear, high-voltage cable, substations. The bottleneck that rarely makes headlines: lead times for large transformers are measured in years | Grid equipment makers. Japan and Korea both have real positions here, in heavy electrical machinery and cable |
| The building | Data-centre shells, power distribution, and cooling — increasingly liquid, because the chips now exceed what air can carry away | Construction, HVAC and thermal-management suppliers, plus water rights, which is becoming a genuine siting constraint |
| Security | Both kinds. Physical access control at sites that are now national infrastructure, and the cyber layer covered in AI & Stocks | Security vendors and integrators. Note that Japan's Cyber Response Capability Enhancement Act takes effect 1 October 2026 and is aimed squarely at critical infrastructure operators |
| People | Electricians, linesmen, HVAC technicians, site engineers. The least discussed constraint of all, and the one that cannot be solved by a purchase order | Staffing and industrial services, and a labour-cost pressure that runs through every line above |
The industry stopped being unbuyable — which starts a harder argument
For years the honest answer about SpaceX was that you could not own it, so it did not belong in a plan. That changed on 12 June 2026, when SpaceX listed on Nasdaq under SPCX in the largest IPO on record: priced at $135, first day closing near $161, a market capitalisation around $2 trillion. Since February 2026 the same entity also contains xAI, so one listed company now spans launch, satellite internet and frontier AI models.
The question this creates is not whether the engineering is extraordinary. It is the question this site asks about everything: at that price, what does the business have to deliver for the buyer to do well? Put your own assumptions into the decomposition tool and see what they require. A great company bought at a great price and a great company bought at any price are the two different stories that Japan in 1989 and Cisco in 2000 already tell on this site.
There is also a record worth knowing before buying the theme rather than the company. The wave of pure-play space companies that reached the market through SPACs in 2020–21 was, for shareholders, close to a catalogue of disasters — several fell more than 90%, and some no longer exist. The technology thesis was not wrong. The vehicles were early, unprofitable, and bought at prices that required everything to go right.
One more practical point that is new: SpaceX now files. Revenue, margins, segment detail and risk factors are public for the first time. Anyone forming a view should read the filings rather than the mythology — the company has spent two decades being discussed without numbers, and the numbers are now available.
SDGs: separate the engineering from the label
There are two quite different things wearing the same word. One is an engineering and industrial reality: grids are being rebuilt, heat is being electrified, efficiency is being retrofitted, and water is becoming a siting constraint for the data centres in the section above. The other is a labelling industry — funds, ratings and scores.
The record of the second is genuinely mixed. Broad ESG funds have often charged more than plain index funds while holding much the same large-cap companies; ratings from different providers frequently disagree about the same firm; and regulators in several jurisdictions have had to act against overstated environmental claims. None of that means the underlying transition is fake. It means the label is not the investment, and a fee paid for a label is a fee paid for nothing — which the cost section shows compounds against you for thirty years.
Where the physical work is
Transmission and distribution equipment, transformers, cabling, storage, heat pumps, insulation, grid software, water treatment. Unglamorous, capital-intensive, frequently regulated — and the actual content of the transition.
What to check before paying for a theme
The expense ratio against a plain index fund, what the fund actually holds when you open the factsheet, how often constituents change, and whether you already own most of it through a broad index. Very often the honest answer is that you do.
A long horizon is the right frame for this transition, because it takes decades. That is an argument for patience, not an argument for paying more.
The Musk map: what each business does, and what it touches
Elon Musk's companies are unusually useful as a teaching example, because between them they sit on almost every layer this page describes — energy, compute, connectivity, robotics and launch. Here is the map as it currently stands, and what a long-horizon investor can and cannot actually buy.
| Business | What it does | Which layer it connects to | Access |
|---|---|---|---|
| Tesla | Electric vehicles; Megapack grid-scale battery storage; solar; driver-assistance software; the Optimus humanoid robot programme; in-house AI silicon | Energy and robotics. The storage business connects directly to the grid constraint above — batteries are how intermittent generation becomes dispatchable. Optimus connects to the labour constraint | Listed (Nasdaq) |
| SpaceX | Launch (Falcon, Starship) and Starlink satellite internet. Since February 2026 it also contains xAI | Launch, connectivity and now frontier AI in one entity. Starlink is infrastructure for places fibre does not reach; the merger put a capital-hungry AI lab inside the same balance sheet | Listed since 12 June 2026 (Nasdaq, SPCX) |
| xAI | The Grok models and the Colossus training infrastructure | Compute, and therefore power. Frontier training is exactly the demand the electricity section describes | No longer separate — inside SpaceX |
| X | The social platform, formerly Twitter | Distribution and data. Its role in the group is as a channel and a data source for the models | Not separately listed; sits under the same umbrella |
| Neuralink | Implanted brain–computer interfaces, with higher-volume production and more automated surgery stated as 2026 goals | Medical devices. Long regulatory timelines; the layer least connected to the rest | Private — not purchasable |
| The Boring Company | Tunnelling, so far proven mainly on a limited Las Vegas system | Civil infrastructure. The least commercially established of the group | Private — not purchasable |
Group structures change, and this one has changed twice in the past year. Now that SpaceX is listed, verify the current structure in its filings rather than in coverage.
What the people building it are saying
Statements from the executives actually allocating this capital, each with a date and a source you can check. The point is not to follow them. It is to have a dated record of what was claimed before the outcome was known — which is the only way to find out, years later, whose forecasts were worth anything.
Quotations are reproduced from the linked sources for commentary. Where a statement was made in another language, the original is shown alongside. The reading under each quote is this site's own and is not endorsed by the speaker.
Four things that survive being wrong about the details
- Check what you already own before adding anything. A broad global index fund already holds the utilities, the grid-equipment makers, the security platforms and the hyperscalers. For most readers the honest answer to "how do I get exposure to this" is that you have it.
- Distinguish "will be busy" from "will earn." Reliable demand attracts capital and competitors. Ask which businesses in the chain can raise prices, and be suspicious of any answer that assumes a commodity supplier keeps a windfall margin.
- Treat thematic funds as a fee question first. Energy-transition, space and AI-infrastructure ETFs turn over holdings heavily and cost more than a plain index. Open the factsheet and look at what is actually inside.
- Keep the schedule. None of this changes the mechanism that does the work: regular contributions into a broad, low-cost core, held through the decades in which the story changes three times. The calculator is still the page that matters most.
Related pages
Being right and losing anyway
Cisco, Intel and Nokia: the technology thesis was correct and the shareholders still lost.
The security layer
Why the demand is written into law, and why that still does not settle the returns.
Sizing a conviction
How large a satellite position can be before it starts deciding your retirement.