economics of AI

the $10 trillion Question

In early 2023, a leaked internal memo from Microsoft revealed something nobody wanted to admit: ChatGPT was haemorrhaging money. analysts estimated OpenAI was burning through $700,000 per day just to keep the lights on.

that's roughly $4 per ChatGPT Plus subscriber per month—on a $20 subscription. the math didn't math.

but Microsoft wasn't panicking. they doubled down, committing another $10 billion on top of their initial investment.

Google declared a "code red," pulling founders Larry Page and Sergey Brin back from retirement. Meta's Mark Zuckerberg pivoted from the metaverse, announcing that they'd spend $ 30 billion or more on AI infrastructure in 2024 alone.

here's what makes it insane: nobody knows if this bet pays off. these models cost $100 million to train. they require data centres that gulp enough electricity to power small cities.

the H100 chips they run on cost $30,000 each, if you can get them.

NVIDIA's revenue quintupled over the last two years.

and yet, despite the gold rush, most AI companies are still hunting for a business model. It's as if we've built the world's most expensive infrastructure for an economy that doesn't exist yet.

or maybe one that's about to obliterate the old economy entirely.

the economics of AI are a paradox wrapped in a revolution wrapped in a really expensive electric bill.

fixed costs, winner takes most, and the knowledge deflation bomb

AI breaks the traditional software economics playbook.

for decades, software had low fixed costs and near-zero marginal costs.

build once, distribute infinitely.

not anymore.

the fixed cost monster

training GPT-4 reportedly cost OpenAI over $100 million. Anthropic's models? similar.

each new frontier model requires thousands of GPUs running for months, burning through electricity like there's no tomorrow. one study estimated that training a large language model produces as much CO2 as five cars over their entire lifetimes per training run.

but that's just training!

Inference actually running these models costs real money every single time someone asks ChatGPT a question. OpenAI's inference costs are estimated at 10-20 cents per conversation. multiply that by hundreds of millions of users, and you see the problem.

this creates a bizarre economic structure: massive fixed costs (training, infrastructure) plus non-trivial variable costs (inference). It's like running an airline where you not only have to build the plane, but each flight still costs serious money.

winner take most dynamics

standard economic theory says competition drives prices toward marginal cost. but

AI economics favour concentration. here's why:

first, data moats. the best models train on the most data. the most-used models generate the most user data. this creates a flywheel.

second, compute scale. only a handful of companies can afford $100 million training runs. third, talent concentration. top AI researchers cost $1-5 million per year in compensation. they cluster at well-funded labs.

the result? a handful of frontier models (GPT, Claude, Gemini) with everyone else licensing or building on top.

It's not quite a monopoly, but it's an oligopoly at best.

the deflationary spiral

here's the part that terrifies and excites economists: AI is massively deflationary for knowledge work. tasks that cost $50-200 per hour (writing, coding, analysis, design) now cost pennies in compute. a lawyer billing $400/hour for contract review versus AI doing it for $0.40.

This creates a paradox.

companies spending billions on AI might reduce total economic activity in knowledge sectors. If you can replace $100 billion in labour with $10 billion in compute, GDP actually shrinks even as productivity explodes.

economists call this "Jevons Paradox" in reverse. usually, making something cheaper increases total consumption. but AI might be so deflationary that it crushes pricing power across entire sectors faster than new demand emerges.

the Infrastructure capture

meanwhile, NVIDIA mints money. They sell the picks and shovels the GPUs and don't care who strikes gold.

this is classic infrastructure economics. In gold rushes, the miners often went broke.

the guys selling supplies got rich.

the same pattern plays out in energy, data centers, and networking.

the infrastructure layer captures value while the application layer fights over scraps. It's the opposite of the 2010s, when software ate the world and infrastructure was commoditised.

who's winning, who's bleeding

NVIDIA: from gaming GPUs to economic kingmaker

In 2022, NVIDIA's market cap was $360 billion. its over $4 trillion right now.

CEO Jensen Huang became the unlikeliest kingmaker in tech.

the economics are simple: AI models need parallel processing. GPUs do parallel processing. NVIDIA owns 90%+ of the AI chip market.

their H100 chips became the new oil, scarce, expensive, and essential.

but here's the twist: NVIDIA's gross margins are 70%+. they're printing money while everyone else burns it.

they've captured more value from the AI boom than arguably every AI company combined.

OpenAI/Microsoft vs. Google: the search economics disruption

Google's search business prints $200+ billion in revenue annually at 50%+ margins. It's the greatest business model ever created. then ChatGPT threatened to obsolete it.

the economics are brutal: Google search costs fractions of a penny per query. AI-powered search costs 5- 10x more. Google's margins compress. Microsoft, with nothing to lose in search, happily trades lower margins for market share.

this is classic disruption economics: the incumbent can't afford to cannibalise their cash cow. the challenger doesn't have a cash cow to protect. Microsoft can lose money on AI search to grab 5% market share that's $10 billion in revenue they didn't have. for Google, that's a $10 billion loss plus compressed margins on what remains.

AI wrapper trap

Jasper, an AI writing tool, raised $125 million at a $1.5 billion valuation in 2022. They built a sleek interface on top of OpenAI's models. revenue hit $90 million.

then OpenAI released ChatGPT.

FOR FREE.

with a better interface.

Jasper's revenue reportedly dropped 50%+ within months. why pay $50/month for a tool when you can use the underlying model for $20 or free?

this exposes the economic vulnerability of AI wrappers. With no proprietary models, data moats, or distribution advantages, you're one API change away from obsolescence. the marginal cost of competition is near zero when everyone uses the same model.

the companies surviving? those with proprietary data (Bloomberg's financial AI), specific workflows (Harvey for law, Cursor for coding), or embedded distribution (Microsoft copilot in office).

takeaways

1. capital beats distribution (for now)

for the first time in software history, the companies with the most money might actually win. access to $10 billion in compute matters more than growth hacking or viral distribution. If you're betting on AI startups, bet on the ones with access to serious capital or proprietary infrastructure.

2. Infrastructure captures value in buildouts

during infrastructure build-outs, the infrastructure layer wins. don't just look at AI companies look at NVIDIA, data center REITs, energy infrastructure, and networking.

3. the wrapper threat Is real

building on someone else's models without differentiation is economic suicide. you need proprietary data, unique workflows, or distribution moats. otherwise, you're renting your existence from OpenAI, Anthropic, or Google and they can change the terms anytime.

4. winner takes most means bet on the giants or niche

the middle is dying. either you're a frontier model with billions in capital, or you're a niche specialist with proprietary advantages. the "good enough AI company" will get crushed between free models above and specialized solutions below.

the economics of AI aren't inevitable

the winners will be those who understand the incentives, follow the money, and see which way gravity pulls.