Everyone is building data centres.
Apple is building the exit.

A finance video crossed my feed this week claiming Microsoft and Google have signed $2.1 trillion of “compute mortgages” and that a crash is counting down. The number is real. The direction is backwards. And underneath the theatrics sits the most interesting argument in technology right now, which almost nobody in Malaysia is having.

The money is not a capital budget. It is a loan book.

Start with the number. $2.1 trillion is the aggregate contracted backlog across the four big platforms — Microsoft, Google, Amazon, Oracle. It is money owed to them, not borrowed by them. The viral version got that exactly inverted.

Strip the arrangement to its bones and it is lending. A lab signs a take-or-pay contract for compute it has not used yet. The provider books the promise, borrows against it, pours concrete and racks chips. The building is the collateral, the contracted payments are the debt service, and capital expenditure is origination volume. As an independent research note put it in July: the market is celebrating loan growth and calling it revenue growth.

The concentration is what makes it interesting. Roughly half of that backlog — about $1.05 trillion — is owed by just two AI labs. Oracle’s book is 54% those two names. Amazon’s is 51%, Microsoft’s 49%, Google’s 43%.

The trigger nobody watches

Here is the mechanism, and it is worth holding whichever side you end up on.

American subprime defaults did not begin when house prices fell. They turned upward in 2006, while prices were still at record highs. Prices had simply stopped rising fast enough for borrowers to keep refinancing into the next cheap rate. The structure needed acceleration, not appreciation.

Apply it here. Hyperscaler capital spending grew roughly +51%, +81%, +77%, then +52%. The level is at a record. The growth is still enormous. The acceleration has already turned negative. And the largest borrower in the chain services its payments not from earnings but from each fundraising being larger than the last — round-over-round step-ups of 1.83×, 1.91×, 1.67×, 1.70×, and roughly 1.23× implied at its next mark.

Nothing has to collapse. It only has to slow.

The constraint is a watt, not a chip

This is where I part company with most of the commentary, including the bearish kind, which assumes surplus chips simply become cheap.

They don’t, because electricity is the binding constraint and it is fixed in the short run. Large power transformers now carry three-to-five-year lead times, up from 24–30 months before 2020. Medium-voltage switchgear is effectively sold out through 2028. Grid connection approvals run 24–36 months. Gartner expects 40% of AI data centres to be power-constrained by 2027.

So depreciation is not a schedule. It is queue position for a watt. A chip dies economically when its output per watt falls below whatever else wants that same connection — because nobody pays the same electricity bill for less work.

That cuts in both directions, and it relocates the asset. The scarce thing is neither the building nor the silicon. It is the energised grid connection: the item with a four-year queue in front of it and no substitute. Which extends rather than replaces the argument I made about the landlord layer and the silicon layer being two different businesses wearing one name.

The case that the data centre wins

Now the other side at full strength, because it is strong.

Demand still exceeds supply everywhere. As of February this year, Amazon says it has never retired an A100 — a chip from 2020. Nothing is sitting idle. If everything built is running, nothing is stranded.

Efficiency has historically increased consumption, not reduced it. Cheaper tokens invite more use. Every previous efficiency gain in computing ended with more total compute consumed, not less.

The frontier is getting heavier. As small models absorb the easy questions, serious work moves to long-horizon agentic tasks — hours of tool use, enormous context. That is more compute-hungry, not less.

Training never leaves. Whoever trains the model that runs on your phone needed a data centre to do it.

And the phone has a real ceiling. A flagship today runs a 3.8-billion-parameter model at roughly 13–18 tokens per second using about 2.7GB of memory. Genuinely useful; not a frontier system. The limit is memory bandwidth, which does not improve as fast as raw chip speed.

Quantum is not the escape hatch either. This year was a real one for it — 96 logical qubits from 448 physical at one lab, 48 from 98 at another. But a quantum processor does optimisation and sampling, not the matrix multiplication that inference actually is. It arrives beside the GPU, not instead of it.

The case that it does not

Efficiency is the fastest lever in the system, because it is software and has no lead time. No transformer queue, no grid approval, no concrete. Distillation, sparsity, quantisation, routing simple questions to small models. A 3-billion-parameter model with the right inference strategy now outperforms a 70-billion one. And Chinese open-weight models have taken more than 60% of the tokens on one major routing platform, at a fraction of the price.

Which produces the sentence at the centre of this whole argument: what threatens the data centre is not less AI. It is more efficient AI — and improving output per watt is the industry’s own central engineering project. They are building the thing that strands their own collateral.

Supply is already in the ground and arrives on a lag. Capacity contracted in 2024 and 2025 energises through 2026 and 2027. That is the ordinary property cycle: supply lands after demand turns, every time, because building takes years.

And optics quietly attacks the moat. Co-packaged optical switches ship this year claiming 3.5× lower networking power and tens of megawatts saved at a large site. The direct saving is modest — networking is a minority of the draw. The real effect is geographic: light lets you stitch a cluster across distance, so compute can go where the power already is instead of demanding a gigawatt at one substation. That relieves the transformer bottleneck — and in doing so it erodes the scarcity premium of the site that was valuable precisely because it held the connection.

Why Apple is the tell

Watch incentives rather than announcements.

Google, Microsoft and Amazon each have a meter to protect. Every query that moves onto your device is revenue leaving their building. Apple is the only one of them with no cloud rent to defend. It sells the silicon in your pocket. Every inference pushed on-device converts a competitor’s operating cost into Apple’s hardware differentiator — and it gets to call the whole thing privacy.

Its architecture says exactly that: a roughly 3-billion-parameter model on the device, a second larger on-device model for higher-end hardware handling text, images and speech, a framework letting developers run their own local models on Apple silicon, and a private server tier for the queries that genuinely need one.

Note the honesty in that last part. Even the champion of on-device keeps a server tier. Apple is not claiming the data centre dies. It is claiming most of what you ask does not need one.

The part that decides it

Here is the question I would put to both camps.

The bulls are right that a glut heals. Demand normally grows back into empty buildings — that is how every property cycle ends. But the recovery requires the demand to still exist when the buildings are cheap.

If routine everyday inference migrates onto devices people already own and power themselves, it does not come back. On-device is not the trigger. It is the reason there may be no recovery.

And the two clocks are mismatched in the wrong direction. Phone fleets turn over in three to five years. The financing matures sooner. Which means the write-off, if it comes, arrives before the on-device era does — and the buildings get repriced by people who cannot yet see what replaced their tenant.

What it means from here

For Malaysia the exposure is specific. Johor’s pitch was geology and cheap power, which I have argued was inherited rather than engineered. Add the tenant question: a lease is only as good as whoever sits at the end of the chain, and that chain increasingly terminates at borrowers with no operating income. When a hyperscaler protects its credit rating by cutting the marginal gigawatt, it cuts the most discretionary site — newest, furthest from head office. Racks cannot move, but tenancy can.

And if the on-device case is right, national AI capability is the phones in people’s hands, not a building. Which makes RM2 billion for the cloud against RM15 million for the people look worse in hindsight rather than better — the small number was the correct instinct, badly sized.

For an ordinary business, none of this is a reason to wait. If the pessimists are right, compute gets cheaper and somebody else absorbs the loss. If the optimists are right, compute stays plentiful. Either way, do not build a company that resells tokens. Own the part that survives both outcomes: the data you capture, the standard you capture it in, and the person who signs off on the result.

The bottom line

Both cases are honest. The bulls are watching demand, and demand is real. The bears are watching acceleration, and acceleration has turned. One camp is looking at how big it is; the other at how fast it is still getting bigger. They are not disagreeing about AI. They are disagreeing about arithmetic.

I will not pretend to know when, and neither does anyone writing these reports. I am a technology CEO, not a financial adviser, and nothing here is a view on any company’s share price. You should also discount me: my business sells AI capability, so I benefit from you believing the efficiency case.

But if you asked me where to place a bet with a ten-year horizon, I would put it on the model becoming small enough and smart enough to live in the device already in your hand. And I would put it there for one unglamorous reason.

Apple is a smart company. And it is the only one in this race that makes money if the data centre turns out to be unnecessary.