AI in the data centre or on the device: who wins, who loses, and who pays

In the three months to the end of July, NVIDIA sold $89.0 billion of data centre products.¹ Across its whole 2025 financial year, Apple spent about $12.7 billion on capital investment of every kind, from shops to servers.² Put those two numbers side by side and you can see that the biggest names in technology are not running the same race at all.

One camp thinks AI will live in enormous buildings full of rented chips, paid for by the hour or by the word. The other thinks most of it will run on the phone in your pocket and the laptop on your desk, with the big servers called in only when a job is too heavy for the device. Both camps are putting serious money behind their view. They can’t both be completely right.

Why should a business owner in Glasgow or Valletta care? Because whichever camp is right will set your AI bill for years. It decides who sees your data while a model is working on it, and how painful it is to walk away from a supplier. Even your next phone is part of this. It may cost more, and some of its cleverest tricks may stop working the moment you lose signal.

So this read does two things. It follows the money through both models, with every figure sourced. Then it gives you a plain test for deciding where a particular AI job should run in your own organisation.

The cloud model: build first, bill later

The cloud model is easy to describe and very hard to pay for. Amazon, Microsoft, Alphabet and Meta order chips in quantities that used to belong in national budgets. They put them into vast halls and sell access back to us by the hour, by the seat or by the word. When your company pays for Copilot, that is the money you are touching.

How big is it? Amazon told investors in July that it would now spend $220 billion this year, $20 billion more than its earlier plan, and that it still couldn’t meet demand.³ Google’s parent has put its own figure at $180 to $190 billion. Meta says $125 to $145 billion.⁴ Microsoft’s number, about $175 billion, came down a little, but only because it now spreads long leases over more years.⁵ Apple, remember: $12.7 billion.

Add Oracle to the four and FactSet gets to more than $690 billion for 2026. In 2020 the same group spent around $95 billion.⁴ I’m less struck by the size than by the financing. Most of these firms are expected to end the year with free cash flow at or below zero, and borrowing has gone from about a tenth of their capex in 2024 to roughly a third in the year to June. In July, S&P moved Oracle down to BBB minus, one notch above junk.⁴

That doesn’t make the bet wrong. It does mean that for the sums to work, someone has to buy an awful lot of AI, at a good price, for years. That someone is you.

Capital spending: the cloud builders against Apple (US$ billion)

Show the data

Company	Capital spending
Amazon	$220bn
Alphabet	$180bn to $190bn
Microsoft	$175bn
Meta	$125bn to $145bn
Apple	$12.7bn

Where NVIDIA’s risk really sits

Sitting at the centre of all that spending is NVIDIA. Its business model is almost too simple: it designs the chips everyone wants, it sells them faster than anyone can make the buildings to hold them, and it keeps about three quarters of every dollar as gross profit.¹ In the spring of 2025 its data centre arm was bringing in $39.1 billion a quarter. By this summer it was $89.0 billion, and for the quarter we’re in now the company expects about $108 billion in total sales.¹

IDC’s estimate gives NVIDIA about 81 per cent of the AI chip market, which is the kind of number that rivals and regulators both notice.⁶ A position like that rarely lasts untouched, and the challengers are not small. Google has started selling time on its own chips, called TPUs, to outsiders: Anthropic agreed in October 2025 to use up to one million of them, a deal it described as worth tens of billions of dollars.⁷ Broadcom designs custom chips for the large cloud firms, AMD sells its own data centre accelerators, and Qualcomm has launched chips aimed at running models rather than training them.^{6, 8}

So NVIDIA’s risk isn’t really a better chip next year. It is that its biggest customers, who are also its fiercest rivals, slowly learn to need it less. Every dollar Amazon puts into its own Trainium chips, or Google into TPUs, is a dollar of margin they keep for themselves. NVIDIA’s other exposure is the one in the previous section: its sales depend on its customers being able to keep borrowing and building.

NVIDIA data centre revenue per quarter (US$ billion)

Show the data

Quarter ending	Data centre revenue
Apr 2025	$39.1bn
Jul 2025	$41.1bn
Oct 2025	$51.2bn
Jan 2026	$62.3bn
Apr 2026	$75.2bn
Jul 2026	$89.0bn

Apple’s bet: let the customer pay for the chip

Apple has taken almost the opposite road, and for a while it looked like the company had simply missed the moment. Its answer, now in its third year, is a layered system. Small models run on the iPhone or Mac itself, on the Neural Engine built into Apple’s own chips. When a request is too big, it goes to what Apple calls Private Cloud Compute, servers running on Apple silicon that are designed so your data is processed and then thrown away, not stored or seen by Apple.⁹

The surprise came in January 2026, when Apple confirmed that the new Siri would be built in collaboration with Google’s Gemini models. Tim Cook was careful to say those models would still run on the device and in Private Cloud Compute, not on Google’s servers.¹⁰ In plain terms, Apple rents the brains and keeps the customer.

Look at what that does to the cost structure. Apple doesn’t need to pour $200 billion into concrete. The hardware cost of running AI is paid by you, the buyer, every time you buy a new phone with a faster chip and more memory, and Apple earns its normal margin on that phone. Its capital spending runs at about 3 per cent of revenue, against roughly 35 per cent at Microsoft.²

The weaknesses are real. Apple was late, and it now depends on a rival for its most visible models. Some features have daily usage limits because they lean on server models, and the new Siri won’t be available at first on iPhones in the EU.⁹ Being careful about privacy is a strategy. Being slow is just slow.

Everyone else is hedging

The rest of the field is less tidy, and that is the interesting part. Most of the large firms are quietly backing both models at once.

Microsoft is the biggest cloud bet after Amazon, with Azure growing 43 per cent last quarter.⁵ It also wrote the rulebook for the AI laptop: to carry the Copilot+ label, a Windows PC needs a chip that can run more than 40 trillion operations per second locally.¹¹ Google may be the best hedged of all. It owns the TPU, the Gemini models, the cloud that rents them out, and Android, where a small model called Gemini Nano runs on the phone itself and only falls back to the cloud for harder jobs.¹²

Amazon is the purest landlord. It sells capacity through AWS, designs its own chips, and has very little presence on your device. Meta spends heavily, but mostly to run its own apps and advertising, and it has released many of its Llama models openly, so other firms can run them wherever they like. Qualcomm comes at it from the other side. Its chips are inside a large share of Android phones and a growing number of Windows laptops, and since October 2025 it has been trying to sell into data centres too.⁸

Then there is OpenAI. In February it said 900 million people were using ChatGPT every week and about 50 million were paying, yet it rents most of the computing it runs on.¹³ Its Stargate project, announced in January 2025 with SoftBank and Oracle, aimed at up to $500 billion of US data centre building.¹⁴ It is the most exposed name in the industry: the most users, and the biggest bill.

Two limits nobody can wish away: power and memory

Under all the strategy there are two hard limits that no press release can talk round: power and memory.

Start with power. The International Energy Agency put data centre use at about 415 terawatt hours in 2024. That’s roughly one and a half per cent of the world’s electricity, and on the agency’s central case it more than doubles to around 945 by 2030. That is a bit more than the whole of Japan uses today.¹⁵ Power is already short in several markets, and a building without a grid connection is just an expensive shed.

Electricity used by data centres worldwide (terawatt hours)

Show the data

Year	TWh
2024	415
2030 projection	945
2035 projection	1,200

Then memory, which is where ordinary buyers start to feel it. AI servers need huge amounts of fast memory, and it comes off the same production lines as the memory in your phone and laptop. So when the cloud firms hoover it up, everyone else pays. Last December IDC said PC makers were getting ready to raise prices by as much as 8 per cent this year, and it thought phones could go up by 3 to 8 per cent.¹⁶ Amazon even pointed to memory costs when it raised its own spending forecast.³

Meanwhile the devices keep getting smarter. Gartner reckoned about 16 per cent of PCs shipped in 2024 had a dedicated AI chip, 31 per cent in 2025, and it expected around 55 per cent this year.¹⁷ Plenty of buyers won’t care. They’ll end up with the chip anyway, and every job it does locally is a job nobody has to pay a server to do.

Share of PCs shipped with a dedicated AI chip (per cent)

Show the data

Year	Share of PC shipments
2024	15.6%
2025	31.0%
2026 forecast	54.7%

Who wins, who loses

So who comes out on top? This part is my own judgement, and I could be wrong. Still, when I put the numbers next to each other, I keep coming back to three possible endings.

The cloud could simply keep winning. If the best models stay far too big for any phone, companies will get used to renting intelligence the way they rent office space, and NVIDIA, Amazon and Microsoft will keep collecting the rent. Apple, in that version, becomes a superb phone maker that resells Google’s models.

My own bet is on a split. The dull daily jobs, meeting notes, sorting the inbox, cleaning up a draft, drift onto the device, since it’s quicker, it’s cheaper and the data stays in the building. The hard reasoning stays in the data centre. Google, Apple and Qualcomm do well out of that. The firms that only resell cloud capacity get squeezed from both sides.

Or the builders could get ahead of themselves. If prices fall faster than the interest bills, some of those halls will sit half empty, and the people who borrowed to build them will feel it. Buyers would enjoy that for a while. I’d watch the most indebted builders closely, and the AI firms that rent everything they use.

There’s one group I can’t see winning in any of these: companies that signed long, expensive deals on the assumption that today’s prices would last.

Ending	What happens	Likely to gain	Likely to lose
The cloud keeps winning	The best models stay too big for devices; firms rent intelligence like office space	NVIDIA, Amazon, Microsoft	Apple's own AI ambitions
The work splits (my bet)	Everyday tasks move onto devices; hard reasoning stays in data centres	Google, Apple, Qualcomm	Cloud resellers with nothing on the device
The builders overreach	Prices fall faster than interest bills; some capacity sits idle	Buyers, for a while	Indebted builders, AI firms that rent everything

My own judgement, set out so you can disagree with it. Not a forecast.

What it means for consumers

For consumers, the trade-off comes down to three things: price, privacy and dependence.

Price first. Running a model in the cloud has become dramatically cheaper. Epoch AI found that the cost of reaching a given level of performance fell by anywhere between 9 and 900 times a year, depending on the task.¹⁸ But that saving doesn’t always reach you, because it comes packaged in a monthly subscription that rarely falls. On the device, by contrast, the cost is hidden in the price of the phone, and that price is going up because of memory.

Privacy is the clearest difference. A task done on the phone never leaves it. A task done in the cloud depends on how well the provider keeps its promises, and on which country’s law applies to the server. For anyone in the EU or the UK, that is not an abstract point.

Dependence is the one people miss. Cloud AI works only while you are online, while you pay, and while the provider keeps offering the service on the same terms. Device AI works on a plane. My advice to anyone buying a phone or laptop in the next two years is simple: buy enough memory, because that is now what decides how much of the AI runs on the machine you own.

What a business should do now

Who wins the race is the wrong question for a business owner. The useful one is where each of your AI jobs ought to run, and whether you’re already paying for the same thing twice. I’d guess most firms haven’t looked. One team takes out a cloud subscription, another team buys a different tool, and the new laptops with their AI chips sit on the desks doing nothing special.

Start with a list. Write down every AI task people actually do, plus the ones you’d like them to do, and run each through the test below. You’ll find three piles. The small, frequent, private jobs, dictating, summarising a meeting, tidying an email, can usually run on the device. Big documents and hard analysis go to the cloud. A few things shouldn’t go near AI at all, because the risk or the regulation isn’t worth the twenty minutes saved.

Then treat your AI contracts like any other large supplier contract. Prices have been dropping fast, and a three-year deal at this year’s rate is a wager that they’ll stop. Find out where the data is processed, whether it’s used to train anything, and how you’d get it back. If the supplier goes vague, you have your answer.

Most firms count how many people use AI. I’d rather know what each useful task costs, and where the data went while it was being done.

← Back to Writing