AI Eats The World

Platform Shifts

Every 10–15 years, a platform shift reshapes technology. We are currently living through the deployment phase of Generative AI, marked by massive capital expenditure, uncertainty, and the search for value.

I'm riffing on former a16z Partner Benedict Evans' 2025 update, AI eats the world.

The Next Platform Shift

We have seen this pattern before. Mainframes, PCs, the Web, and Smartphones each represented a sigmoid curve of adoption. Now, Generative AI attempts to start a new curve. However, unlike previous shifts where we knew the physical limits (screen size, bandwidth), we do not yet know the ceiling of LLM capability.

What happens in a platform shift?

Who is affected, and how much?

Dominance in one era does not guarantee survival in the next. Microsoft dominated the PC era with a 90%+ share of computing, but when the center of gravity shifted to smartphones, its share of total computing devices collapsed.

Microsoft OS share of global computer unit sales

Furthermore, early leaders often disappear. The "first" to market was not the winner in PCs, browsers, search, social, or smartphones. We should expect similar turbulence in the AI model layer.

Global PC unit sales share

How will the new thing work? We don't know

For every new platform, we forget how many ideas failed and how unclear everything was.

Similar to how the internet eventually became the default medium for meeting partners, AI adoption may follow a slow-then-explosive trajectory.

When things are exciting, people get excited. This often leads to bubbles.

But when the dust settles, the world has changed

US heterosexual couples who met online, by year of meeting

The internet has gone from the New Thing to a basic part of daily life.

New platforms mean new tools (and new revenue)

SaaS means the typical large enterprise in the USA now uses 4-500 apps

One way this platform shift is different, though

For PCs, the web or smartphones, we knew the physical limits of what could happen next year

With LLMs, we don’t know how much better this could get 🤯

“The race to AGI is afoot”

— Sergey Brin

Another platform shift, or more?

We know this will get better, but we don’t know how much

So how will the new thing turn out?

If this is ‘only’ as big as mobile or the internet, that seems like enough

The Capex Boom

The race for AGI has triggered an investment cycle without precedent. The "Big 4" (Microsoft, Meta, Alphabet, Amazon) are projected to spend over $330 billion in 2025 alone on AI infrastructure. This level of spending is driven by FOMO and the existential threat of missing the next platform.

"The risk of under-investing is significantly greater than the risk of over-investing."

— Sundar Pichai, Q2 2024

FOMO drives a capex surge

~$400bn in 2025 for the big four alone (for comparison, global telecoms is ~$300bn).

Big Tech Capex ($bn)

Those four companies alone are expected to devote more than $380 billion combined to capital expenditures in their current fiscal years, with most going to chips, servers and other data center-related expenses. That’s a more than 1,300% jump from a decade ago. And they’ve all pledged to spend significantly more in the year after that.

"The very worst case would be that we have just pre-built for a couple of years."

— Mark Zuckerberg, Q3 2025

A new investment cycle

This spending is reshaping the physical world. In the US, construction spending on data centers is rapidly approaching spending on office buildings.

US construction value (2025 $bn, seasonally adjusted annual rate)

Nvidia can’t keep up

While Nvidia is leading the AI chip market, it can’t keep up with the demand. They're trying to build a new Sun Microsystems (though China and the hyperscalers’ own chip designs are coming up close behind).

Nvidia vs Intel

This demand flows to one bottleneck: Nvidia. While Intel struggles to maintain relevance, Nvidia's revenue has exploded, decoupling entirely from traditional semiconductor cycles.

TSMC can't keep up either

TSMC is unwilling (or unable) to expand capacity fast enough to meet Nvidia's book of business.

Quarterly revenue ($bn)

The Constraints

We are seeing wild estimates for future capacity. Some forecasts suggest we will need to triple global data center capacity by 2030, costing trillions of dollars. However, real-world constraints like power, chips, and permits are becoming major bottlenecks.

Utility power availability is now the primary bottleneck in the US, where power demand growth has historically been flat (around 2%) but is now spiking due to AI.

US power backlogs becoming a major issue

US power demand growth is ~2%, and AI might add 1% that’s hard to build fast (this is not an issue in China)

Main constraints to data centre construction, USA (February 2025)

"It’s been almost impossible to build capacity fast enough since ChatGPT launched"

— Kevin Scott, Microsoft CTO

Data centre capacity triples?

Some very large numbers (although some of these ‘bragawatts’ may be more performative than real).

Global data centre capacity estimates, H1 2025 (GW)

“Three trillion dollars!”

Annualised AI capex aspirations are a similar magnitude to mature global capital-intensive industries

The hyperscalers can afford it…

The hyperscalers have massive cash flows, which generally cover their capital expenditures.

Annual Capex and Free Cash Flow, 2010 to 2025 ($bn)

… up to a point

However, when we include leases, the total burden is much higher, especially for Amazon and Microsoft. Capital leases are not new, but they’ve grown much larger.

Annual Capex and Free Cash Flow, 2010 to 2025e ($bn)

Hyperscalers add leases and debt, while some analysts suggest Oracle’s cloud capex might be >100% of revenue.

Circular Revenue

Without its own cashflows, OpenAI partners with Nvidia, Oracle, Softbank, and petrodollars.

OpenAI is buying Nvidia chips with Nvidia's cashflow…

Which comes from the hyperscalers…

and using Nvidia's cash to turn AMD into an Nvidia competitor, and pay Broadcom to make its own chips…

Yes, but where has all this money got us?

After three years, lots more science and engineering, but no real clarity on the shape of the market

Every week introduces new models, new (problematic, gamed, and saturated) benchmarks, new acronyms.

Model Proliferation

Meanwhile, the models themselves are converging. On general benchmarks, the gap between the leading model and the followers is shrinking, suggesting that raw intelligence might become a commodity.

Models converge and leaders change weekly

Dozens of (saturated) benchmarks to choose from, but on the most general, the leaders are very close.

Tech versus brand versus distribution?

The models may be close to commodities (especially for general use), but market position is not, so far.

Weekly active users of generative AI tools as share of the population, June 2025

Adoption is also uneven globally. The US leads significantly in the usage of generative AI tools, while developed markets like Japan lag significantly behind.

3. Usage and Adoption

Despite the hype, we are early. ChatGPT has roughly 800 million weekly active users, but usage intensity varies. Surveys suggest that for many, AI is a "sometimes" tool rather than a daily habit.

ChatGPT global weekly active users (millions)

New tools like Cursor are seeing explosive growth, challenging traditional IDEs.

Cursor Growth (ARR $m)

Still more experimentation than daily use

So far, many more people use chatbots occasionally than make them part of their daily lives.

How many people use generative AI chatbots in the USA?

Most data shows the same picture

Surveys are early, scattered and inconsistent, but an engagement gap seems clear.

How many people use generative AI chatbots in the USA?

Is this just early? Or a harder problem?

Why do most users of ChatGPT only use it a little bit?

People don’t know what they want until you show it to them.

You’ve got to start with the experience and work backwards to the technology

— Steve Jobs

When the dust settles, however, new technologies fundamentally change society. The internet went from a novelty to the primary way couples meet, displacing friends and family.

How Couples Meet

So where does a model lab compete?

Where is the value capture for a research-heavy, capital-intensive commodity?

Microsoft's shift away from network effects?

From competing on network effects to competing on access to capital?

Microsoft's capital intensity is spiking as it builds out the infrastructure for this new era.

Microsoft capex/sales

For OpenAI, “yes!” to everything

Everything, everywhere, yesterday (on other people’s balance sheets), before the market slips away

Automation and Unbundling

AI is often described as "infinite interns." This leads to the Jevons paradox: as the cost of intelligence drops, we will consume vastly more of it. We saw this with the steam engine, and we saw it with retail automation.

We are also seeing the unbundling of software. Startups are flooding into every vertical, attempting to unbundle Excel, Google, and traditional SaaS using AI.

The coming wave of AI startups trying to unbundle Google, Excel, email and Oracle… and ChatGPT.

Y Combinator startups by field

AI coding as the new AWS

“Vibe coding” as the new abstraction layer, after AWS, libraries, operating systems…

A new step change reduction in software creation costs.

How do you know what to automate?

Step one: ask your local systems integrator.

Accenture reported new quarterly ‘generative AI’ contracts ($m)

How do you know what to automate?

Step two: buy some SaaS from Dr Evil.

Palantir quarterly revenue by segment ($m)

Pilots come first and deployment takes time

“Agentic!’ is 2025’s buzzword, but deployment takes longer

Pilots vs Deployment

Not everything works? Welcome to tech

“Why did our AI pilot fail?” That’s a CTO question, not an AI question.

The future can take time

A quarter of CIOs have launched something - but 40% don’t plan anything until at least 2026

CIO expected timing for first LLM projects in production, September 2025

But the future always takes time

Cloud is old and boring - but still only 30% of workflows.

Enterprise workloads in public cloud

Sometimes ‘automation’ alone is a big deal

Automation also allows us to handle complexity. The introduction of barcodes in the 1970s allowed supermarkets to handle 5x more SKUs than before.

Average SKUs per supermarket, USA

What next?

What comes after automating the obvious, easy things?

How do we use automation that makes ‘mistakes’?

We have no indication that error rates will go away, so where’s the human in the loop?

300m interns? Jevons paradox at work

In 1865, William Stanley Jevons first described a paradox. He maintained that more efficient steam engines would not decrease the use of coal in British factories but would actually increase it. As the fossil fuel became cheaper, demand for the resource would grow, leading to the construction of more engines.

Satya Nadella

@satyanadella

Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of.

1.6K

1:15 PM · Feb 7, 2025

The paradox applies to various technological shifts. For example, Los Angeles had 10,000 horses in 1900, but by 1950, there were 1 million cars. Similarly, despite the rise of fuel-efficient and electric vehicles, total miles driven has increased, and energy-efficient computers have led to widespread smartphone use and data centers that now consume 1.5% of global electricity.

Steam engines gave Britain the equivalent labour of (very roughly) 5x its total population by 1900.

UK population and steam engine labour unit equivalent (m)

Where’s the human in the loop?

All recommendation systems today work by driving, capturing and analysing user activity.

Value to capture!

Brands spend a trillion dollars a year to talk to consumers - plus rent, shipping, marketing, returns…

Global ad revenue, 2024 ($bn)

This is still early days

We are seeing AI used primarily to expand capabilities and test new ideas, rather than to replace existing processes.

US consumer search preference (September 2025)

The web has been dying since 1997

Wired: March 1, 1997: “You can kiss your web browser goodbye” – Kevin Kelly and Gary Wolf, The Big Story

But where are we going?

What if recommendations go from correlation to an understanding of what those SKUs really represent?

Don't forget: The old stuff from before ChatGPT is still here

E-commerce as % US retail sales

And all the other new stuff

After a decade of promises and tens of billions of dollars, ‘automatic cars’ might be starting to work.

Monthly ‘robotaxi’ trips in California

But general-purpose robots are harder. This is "Moravec's Paradox": high-level reasoning (chess, poetry) requires very little computation, but low-level sensorimotor skills (walking, holding an egg) require enormous computational resources.

We are seeing progress in "Embodied AI"—robots that learn to move by themselves—but we are still far from a Rosie the Robot in every home.

So what?

We are in the deployment phase. The technology works, the capital is flowing, and the models are getting better. Now comes the hard work of figuring out the product-market fit, the business models, and the societal impact.

Consider the elevator operator. In 1950, there were over 100,000 elevator operators in the US. Automation (buttons) didn't just remove the job; it made the elevator invisible.

Elevator attendants, USA

“AI is whatever machines can’t do yet”

— Larry Tesler, 1970

Watch Benedict Evans deliver this presentation at SuperAI Singapore 2025: