Apple's $9bn Answer to Microsoft's $80bn AI Arms Race
Microsoft said in January 2025 it was on track to spend roughly $80bn on AI-enabled datacentres in fiscal 2025, the whole infrastructure bill rather than an AI-only budget, with the servers inside it depreciating over six years. Apple's option to wait costs nothing to carry, and that accounting gap, not model quality, is the real story of the AI arms race.
Microsoft's stake in the AI arms race is not an expense. It is an asset, and assets depreciate. In January 2025 the company's president Brad Smith wrote that Microsoft was on track to spend about $80bn in fiscal 2025 building AI-enabled datacentres: that is the company's whole bill for the infrastructure, land and buildings included, not a pure AI line. Capital spending on that scale does not pass through the profit and loss account and vanish. It is capitalised into property and equipment and written down over the assets' useful lives. For the servers that do the actual work, the schedule is six years: Microsoft's fiscal 2022 annual report records its decision to "increase the estimated useful lives of both server and network equipment from four years to six years", and it has depreciated its server estate on that basis since. Apple's fiscal 2024 results show $9.4bn of capital expenditure and $31.4bn of research and development for the entire company. Those are fiscal 2024 figures set against Microsoft's fiscal 2025 plan, a mismatch of a year worth stating plainly; it changes nothing material, because the gap is close to an order of magnitude and no single year of Apple capex growth closes it. Apple is not carrying the asset. It is carrying an option on the same outcome, and the option costs nothing to hold.
That asymmetry does the work the usual "Apple is behind" story cannot. A datacentre must earn back its own depreciation before it earns anything else, and if AI hardware ages faster than the accounting assumes (GPU generations are turning over rather quicker than six years), the write-downs arrive early. An option to integrate whichever model wins carries no such clock. Its holder needs one thing to be true: that somebody wins. Worse for the builders, the strike price is falling, because every pound a frontier lab spends makes the model layer more contested and cheaper to buy access to. Waiting is not just free for Apple; it is getting more valuable.
Should your business join the AI arms race?
The same asymmetry decides the boardroom version of the question, so start there rather than with Apple. The pressure to announce large AI investment is real: it signals seriousness to investors, customers and staff. But signalling is precisely the wrong reason to put a depreciating asset on the balance sheet. Three tests separate spend that digs a moat from spend that decorates a press release. First, does the money buy proprietary advantage (your data, your workflow lock-in, your customer relationships) or generic capability that is commoditising underneath you? Second, what does the switching cost look like in two years if you defer: are you locked out, or can you integrate the winner on better terms than today's? Third, who captures the margin in your value chain once the capability is cheap? Applying the tests honestly is the substance of serious technical strategy work, and before any build decision it pays to pressure-test whether your organisation is ready to build at all.
Run the tests on a composite of the kind of decision playing out across the UK mid-market. Consider a regional freight forwarder, buried in customs paperwork since Brexit, quoted a seven-figure programme to build and fine-tune its own document-processing model, hosted in-house for control. Test one: the proprietary asset is its archive of cleared customs entries and the broker relationships built on fast clearance, and neither requires owning a model, because a rented frontier model with that archive behind it delivers the same advantage. Test two: deferring the build costs almost nothing, since model prices are moving in the buyer's favour and an integration layer written against one vendor can be pointed at another in weeks. Test three: the firm's margin lives in clearance speed and client trust, layers no model vendor can reach, while any margin in the model itself is being competed away. The build fails all three tests, and the seven-figure programme would have sat on the balance sheet depreciating exactly as Microsoft's datacentres do, only without Microsoft's revenue behind it. The right spend is the boring one: cleaning the data, wiring the workflow, and renting the intelligence on a contract the firm can walk away from.
Is Apple actually behind in AI?
If the game is building the most capable model, then yes: Apple is nowhere, and no serious observer claims otherwise. But apply the third test to the industry itself and watch what is happening to the price of the thing everyone is racing to build. Stanford's 2025 AI Index found that the cost of running a model performing at GPT-3.5 level fell from $20 per million tokens in November 2022 to seven cents by October 2024, a decline of more than 280-fold in under two years, while the gap between the best model and the chasing pack narrowed sharply. A product whose price collapses that fast while its substitutes converge is a commodity in the making. When a layer commoditises, the economics migrate to whichever layer stays scarce, and what stays scarce is distribution: the device in the pocket, the default screen, the billing relationship. Apple holds a formidable version of all three.
Pioneers tend to build the category; fast followers with distribution tend to take the money. The pattern is not a law, but it recurs often enough that "behind in AI" should be tested against it before being accepted as a verdict. The depreciation charge sharpens it: the pioneers are not just spending to build the category, they are booking the cost against their own future earnings, on kit that may be economically obsolete before the accounts say it is worthless. Whichever lab wins still has to reach users, and there are only a handful of doors worth knocking on. The owner of the biggest door sets terms without ever having bid in the auction its rivals are funding.
What would change my mind
The fast-follower case rests on one load-bearing assumption: that no single lab opens a durable, compounding capability gap. If one model pulls decisively ahead and stays there, the calculus inverts. The winner gains pricing power, the depreciating datacentres become the toll booths of the industry, distribution owners become sharecroppers on their own platforms, and having waited becomes expensive. Watch three signals. Pricing power at the model layer: if a leading lab can raise prices without losing enterprise customers, the Stanford cost curve has broken and the layer has stopped commoditising. Capability convergence: if independent evaluations stop showing rival models trading places at the top, a gap is forming. Hardware independence: if a new AI device sells in volume to people who then leave the phone at home, Apple's distribution is less of a chokepoint than the option pricing assumes. None of the three is visible today, and I would update quickly if any appeared.
Until then, the accounting tells the story. The builders must win the model race, then win deployment, then defend the position against well-funded rivals making the identical bet, all while the depreciation clock runs. Apple needs somebody, anybody, to win. Most businesses cannot copy Apple's distribution, but every business can copy the discipline of asking, before signing the cheque, whose moat the money is actually digging and whose balance sheet carries the write-down if the answer is nobody's. For AI capability without that bet, practical AI under human control is where the thinking starts.
Questions people ask
What is a fast-follower strategy in AI?
It means letting others fund the race to frontier capability, then integrating the winner once the technology and its price have settled. It works when the capability layer is commoditising and you own something scarce that the builders need, usually distribution, proprietary data or the customer relationship.
How do I know whether our AI investment is building a moat?
Ask whether a competitor could rent the same capability from a vendor within a year. If they could, the spend is buying capability rather than advantage. Moats come from what the spend makes uniquely yours: data nobody else holds, workflow lock-in, or a customer relationship the capability deepens.
Is it better to build or buy AI capability in 2026?
For most organisations, buy, integrate, and keep switching costs low, because rival models are close enough in capability that prices are moving in the buyer's favour. Building makes sense only where you hold proprietary data or distribution that converts the investment into something rivals cannot replicate, and where you can carry the depreciation if the bet sours.
Related
- Epic v Google: The App Catalogue Remedy That Hands Rivals the Moat
- Memory Prices Turned. Renegotiate the Supply Contract Before the Next Refresh.
- How AI Makes Bad Bosses Worse: The Agreeability Machine in the Corner Office
- Digital Business
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.