Capital commitments have already hit hundreds of billions of dollars and are forecast to top a trillion by 2029, larger than the GDP of most countries.
asserted
commitments → hit → countries
Massive campuses spread over thousands of acres are planned.
asserted
campuses → spread → acres
As a microchip executive recently put it, “This is the largest scale infrastructure build-out in the history of humanity.”
asserted
This → put → humanity
The question is, are we building the right thing?
asserted
we → build → thing
The entire build-out rests on an unexamined bet: that the majority of growth in artificial intelligence will happen inside centralized data centers that must continue to expand if AI is to progress.
asserted
AI → rest → centers
By pretending it is, we’re locking ourselves into massively expensive infrastructure based on a snapshot of today’s technology, while the technology itself is moving toward smaller, local, specialized models.
asserted
technology → pretend → models
In the first inning of the AI age, centralized, cloud-based systems such as those at OpenAI and Anthropic have made headlines and have reached enormous valuations.
asserted
systems → base → valuations
But the technology is still new, and the architecture that ultimately wins this race hasn’t been decided yet.
asserted
that → win → race
There are many problems with the centralized cloud model, even aside from the fact that few of AI’s biggest companies are profitable.
asserted
few → be → companies
For one, asking a single model to answer every question in every situation is neither practical nor cost-effective.
asserted
asking → ask → situation
Is it better to have one intelligent generalist who answers every question and tackles every problem, no matter what field or discipline it’s in?
asserted
it → have → field
The world is vastly complex, and the intelligence built to navigate it should be too.
asserted
intelligence → build → it
That’s why specialization produces our greatest scientists.
asserted
specialization → ’ → scientists
Katalin Kariko spent four decades researching a single molecule — messenger RNA — during which time she was demoted and her research was defunded.
asserted
research → spend → molecule
Decades later, her research formed the basis of the COVID-19 vaccine that saved hundreds of millions of lives.
asserted
that → form → lives
Gregor Mendel spent his life studying pea plants and unlocked inheritance and population genetics.
asserted
Mendel → spend → inheritance
The pattern holds in engineering, too.
asserted
pattern → hold → engineering
Nobody designs a jet engine from a survey course.
asserted
Nobody → design → course
It takes years inside a system to understand it, let alone design a new one.
asserted
It → take → one
Expertise comes from depth.
asserted
Expertise → come → depth
Asking a single model to do everything means settling for mediocrity everywhere.
asserted
model → ask → mediocrity
At the same time, centralizing AI compute in a few massive data centers makes critical infrastructure vulnerable to cloud outages or targeted cyberattacks.
asserted
infrastructure → centralize → outages
Renting rather than owning its AI infrastructure makes each company beholden to someone else’s operations, pricing, and priorities.
asserted
company → rent → operations
Critical industries, from aviation to energy to defense and healthcare, depend on secure data that live as close to their operations as possible.
asserted
that → depend → operations
Feeding proprietary operational data into servers hundreds of miles away is not only impractical but also often illegal, a direct conflict with the data sovereignty and compliance rules these industries already operate under.
asserted
industries → feed → rules
And who wants to risk putting their custom and copyrighted trade secrets into an AI system that could be jailbroken by competitors or even geopolitical enemies?
uncertain
that → want → competitors
As AI becomes increasingly sophisticated, it will become less centralized.
asserted
it → become → ?
When television was first introduced, everyone watched the same few channels, until the technology matured and shows became more audience-specific, with thousands of offerings to choose from.
asserted
shows → introduce → offerings
AI is already moving in that direction, with smaller, more efficient models tailored to a specific product, company, or industry.
asserted
AI → move → product
Running on a computer or a phone, rather than at a distant data center, they leverage existing hardware to make AI setup minimal and cheap.
asserted
setup → run → hardware
It’s natural for companies to want to own their own intelligence.
asserted
companies → ’ → intelligence
In a free market, who wants to run the same AI model that everyone else has?
asserted
everyone → want → that
In the early days of computers, organizations would lease mainframes from companies such as IBM until costs came down enough to buy PCs and servers outright.
asserted
costs → lease → PCs
Owning the technology makes fine-tuning and repeated iteration possible, which companies can control and use to their advantage.
asserted
companies → own → advantage
Bespoke models run at a fraction of the cost, because they don’t need to be capable of every task imaginable.
asserted
they → run → task
Overcoming the tyranny of distance, latency all but disappears as real-time processing becomes the default instead of the exception.
asserted
processing → overcome → distance
A hospital intensive care unit’s monitoring system, for instance, needs to flag a patient crash in milliseconds, a task that can’t tolerate a round-trip to a distant data center, let alone an outage.
asserted
that → need → center
It also depends on sensitive patient health data that are safer staying within the hospital’s own walls than on outside servers.
asserted
that → depend → servers
Of course, there may always be some demand for large models for specific tasks that require generalized intelligence.
uncertain
that → require → intelligence
But the majority of commodity work will not run through data centers.
asserted
majority → run → centers
…and 4 more, not listed.
