CROSSPOST: HORACE DEDIU: The Most Brilliant Move in Corporate History?
Horace’s tagline: ‘An allergy to centralized computing. Based on Tweets by MilkRoadAI’. As Facebook, Amazon, Google, and Microsoft together plan to spend $650 billion of their roughly $1.6 trillion...
Horace’s tagline: ‘An allergy to centralized computing. Based on Tweets by MilkRoadAI’. As Facebook, Amazon, Google, and Microsoft together plan to spend $650 billion of their roughly $1.6 trillion in annual revenue on the build-out, Apple says “no thanks”. It believes it can buy whatever it turns out to need in terms of cloud-datacenter “AI” on the cheap from one of those four—or from Anthropic or OpenAI. Could the world work out so that it cannot do that, and its device-sales profits melt away as someone else offers sufficiently better natural-language interfaces or other game-changing capabilities? I find it hard to see how. Maybe someone smarter than I am, however, could…
Horace Dediu’s piece argues that Apple’s restraint on AI capex may be the most brilliant corporate move of this cycle. Where Amazon, Google, Microsoft, and Meta are together spending about $650 billion a year on AI data centers—94% of operating cash flow, financed increasingly with debt—Apple has kept capex to roughly $14 billion and refused to hand its cash flow to Nvidia. The hyperscalers’ AI services currently generate only about $35 billion in revenue, a small fraction of what they are spending on infrastructure. They think they have to spend not so much to make money as to guard against Christensenian disruption of their current platform-monopoly profits. Apple does not see that as a risk—Apple Silicon made by TSMC and the promise of on-device low-latency low-infrastructure cost inference are sufficient protection. And I am not ingenious enough to see how they could be likely to be wrong:
Horace Dediu: The Most Brilliant Move in Corporate History? <https://asymco.com/2026/03/10/the-most-brilliant-move-in-corporate-history/>: ‘The 3rd? most valuable company on Earth watched as its rivals lit $650 billion on fire and did nothing. The biggest cash bonfire in history, by far, eagerly fed by all the usual suspects but one.
It’s either the dumbest or the most brilliant move in corporate history.
Apple used to be the biggest capex spender, mainly because it paid for most of the property plant and equipment in the factories that made its phones and computers. The so-called tooling or equipment spending plus its leased store renovations and some data centers were costs no other tech giants had and so Apple was an outlier. It was, as I wrote at the time, equivalent to buying a few aircraft carriers every year.
But that all changed with AI. Amazon is spending $200 billion this year on AI data centers. Google, $185 billion. Microsoft, $114 billion. Meta, $135 billion. Combined: $650 billion. [Not including OpenAI, Anthropic and SpaceX/XAI.] That is like buying the US Navy every year. And yet Apple’s capital budget is still a modest $14 billion, oscillating with new hardware tooling cycles.
Apple is refusing to transfer its cash flow to Nvidia. Curiously, it believes that its cash flow belongs to its shareholders, not to Nvidia’s.
The hyperscalers are now spending 94% of their operating cash flows on AI infrastructure. Amazon is projected to go negative free cash flow this year with as much as $28 billion in the red. Alphabet’s free cash flow is expected to collapse 90% from $73 billion to $8 billion. These companies used to be the greatest cash machines ever built. Now they’re borrowing money to keep the data center lights on.
The Big Five raised $121 billion in bonds in 2025 alone. Morgan Stanley projects $1.5 trillion in tech debt over the coming years. For the first time in history, hyperscalers hold more debt than cash. Perhaps this is why their P/E ratios slumped from mid thirties to mid twenties.
And what are they getting for that $650 billion? AI services generate roughly $35 billion in total revenue or 5% of what’s being spent on infrastructure. There are dreams of more of course, but the business models of AI have yet to resonate, especially for consumers.
Now here is where Apple’s bet becomes genius. AI models are commoditizing faster than anyone predicted. Software and hardware both have tendencies to commodify. Protections exist but they have to do with integration and distribution. DeepSeek built a model for $6 million that matches systems costing $100 million. Open source models now power 80% of startups seeking VC funding. The moat these companies are spending hundreds of billions to build is evaporating.
Apple understood this before anyone else. It didn’t build its own AI model, it licensed Google’s Gemini for about $1 billion a year. Why spend $100 billion building a factory when outsourcing costs a billion? And if a better model appears next year, Apple just switches vendors.
But Apple is not sitting still. It just dropped the M5 chip with a 16 core Neural Engine and Neural Accelerators built into every GPU core. It runs 70 billion parameter AI models locally, eventually even on your phone. The M5 delivers 4x the AI performance of the M4 and Apple doesn’t need $200 billion in data centers.
Because Apple turned 2 billion devices into the data center. Every iPhone, Mac, iPad gets distributed AI at a scale no server farm can match. While its rivals burn cash, Apple is doing the opposite. $90.7 billion in stock buybacks last fiscal year. Its competitors? Combined buybacks collapsed 74% from their peak.
Apple didn’t miss the AI revolution. It just bet that the winners won’t be the ones who build the infrastructure. They’ll be the ones who own the customer and no one else on Earth owns the best customers.
What do I think?
I think Horace is probably right.
The view that Anthropic and Claude are going to conquer the digital world with AI runs into the fact that Anthropic’s current run-rate revenues are $14 billion a year (at least half of OpenAI’s), while Facebook, Amazon, Google, and Microsoft together have $1.6 trillion a year in revenue and have every incentive, if push comes to shove, to cross-license their AI technologies so that nobody can take their platform-monopoly profits from them via superior enterprise or consumer natural-language or agentic-’bot interfaces. 100x scale has a logic of its own. And as long as the sun shines on TSMC, Apple will pay whatever is necessary to keep Apple Silicon-based on-device inference with its latency and infrastructure cost advantages at least the equal of what can be done in the cloud. Apple could fail to execute—but so could everybody else.
The others are making the big $650 billion betsnot so much to make money as to guard against Christensenian disruption from someone else’s AI-cloud. But that is really not a big risk Apple thinks it needs to buy insurance again, and I am not invntie enough to see how it could be likely to be wrong.
I think this is likely the case, but with some notes. There’s a difference between consumer-facing and the background stack.
First, it’s worth noting that Apple coming out well once the dust settles is largely an accident of historical position, not design. They failed to execute on their own AI strategy. They prefer doing things in house (until their hand is forced) and have been unable to execute on Siri or LLMs in a decade.
On the other hand, I suspect LLM “AI” and this round of hype will go the way of every other AI boom. We’ve had roughly seven machine learning “AI” hype cycles since Babbage and five since the 1960s, giving rise to the oldest joke in computing: if it works it’s computer science; if it doesn’t work it’s AI.
The data centres will largely be useful, maybe not at the scale needed to make untold trillions from digital gods, but everyone will forget about them. Like the fibre optic cables.
One reason you still need the data centres: once you have an LLM doing inference runs to build code and plan things, token processing explodes. It’s not like using a chatbot. Token consumption for coding and similar tasks is 50-100x or more. Then going forward as this gets baked into operating systems and UI, broadly for getting your computer to build more complex things, it’ll go exponentially(?) higher. Even if you stopped at the uses we have today, and it becomes part of the background of coding tools and ui you need a lot of computing power to run systems that operate through natural language - even if most of it is edge.
However, unitary models, one giant general-purpose model handling everything, I don’t think are the way this plays out for most use cases. What you actually want is your existing tools getting smarter within their own domain. Pages doesn’t need a model that can debate philosophy. It needs a constrained model specifically designed for document manipulation, one that understands structure, style, and reorganisation (maybe even constrained ui/ux manipulation). That model can be small, fast, and run on-device most of the time, with access to a more powerful cloud model when the task demands it (eg ui personalisation).
I expect it’ll shake down to an orchestration layer, a router that sits between you and the models which is mostly edge, picking what goes to the cloud. You make a request in natural language and the router decomposes it: some subtasks go to small expert models on-device, some go to domain-specific models in the cloud, and the big planning or research tasks, anything requiring long multi-step reasoning, go to the big frontier general-purpose models in the data centre. The results get stitched back together and you never see the plumbing. This is essentially how cloud computing already works. Microservices, each handling a specific function, with intelligent routing underneath. Nobody builds monolithic applications anymore, and I can’t see why AI deployment will be different.
Why would it not just follow previous patterns… I’ve not seen exceptional evidence for the exceptional claims. So… Edge computing on-device will be good enough for most people most of the time, just as an iPhone is good enough for most things. But as soon as you want anything more powerful that requires building over long autonomous runs, planning, building, testing, you need the big data centres again. Where exactly the lines fall between edge and cloud, between expert models and frontier models, who knows. But the structural principle, orchestration routing to a mix of specialist and generalist models, is just another computer engineering cycle. Custom mainframes to general-purpose servers to virtualised instances to containers to serverless functions… currently we’re pretty much at general purpose servers and edging into virtualisation.
The market for AI products seems sure to be very large, but not all-consuming in the way that enthusiasts assert. It seems to me that it is unlikely to be great enough to provide a good return for the frenzied investments in hyperscaling resources by Microsoft, Alphabet, et al.. It's one of those situations where the winners may wind up envying the losers, and thus Apple's "losers" strategy looks prudent to me. Even if the hyperscalers' strategy works out better than I fear, there are worse things than becoming a well-resourced fast follower. Obviously Apple must keep very well informed about the technology and the market and be ready to jump if it turns to be necessary.