Investors can’t survive AI’s one-two punch

LLMs are not SaaS like…

• Oracle databases
• Microsoft’s Word+Excel+Powerpoint
• Adobe’s Photoshop+InDesign+Illustrator+Acrobat

Those products follow software economics. LLMs do not.

For two decades, investors have believed in one rule:

• Build once
• Sell forever
• Marginal costs fall toward zero.

That rule trained markets to ignore physics.
LLMs reintroduce cost as a factor.

LLMs deliver a one-two punch investors aren’t equipped for:

1. Extreme capex up front
2. Relentless opex every time it’s used.

Why traditional SaaS works

• Capacity is provisioned once, then amortized over time
• Serving another user costs almost nothing
• Usage growth improves margins
• Scale is the business model

That’s how Oracle licenses, Office subscriptions and Adobe seats print money.

AI breaks that model immediately

Capex: expensive to build

• Frontier models require scarce GPUs with long lead times
• Datacenters must be purpose-built, not generic cloud
• Power contracts, cooling and networking are non-optional
• Capital is committed years before revenue is proven
• Once spent, it’s largely irreversible

You can’t A/B test a datacenter.
You can’t pivot half-built infrastructure.

Opex: expensive to operate

• Every prompt consumes real compute
• Every response burns power and silicon time
• Inference never becomes free
• There is no “zero marginal cost” user

With SaaS, growth lowers unit cost.
With LLMs, growth multiplies cost.

This flips the economics

• Revenue does not automatically outpace usage
• Margins can worsen as adoption increases
• Scale becomes a liability, not a moat
• Pricing mistakes compound instantly

That’s the punch investors missed.

Why the market is waking up now

• Capex is ballooning faster than revenue
• Datacenter delays destroy IRR instead of deferring it
• Physical constraints replace cloud flexibility
• LLMs look less like software and more like industry

Steel, power, cooling and permitting don’t care about narratives.

Why comparisons keep failing

• Social networks were cheap to serve
• Cloud software reused shared infrastructure
• Mobile apps rode existing networks

LLMs must build their own factories.
Then pay to run them every second.

The repricing isn’t about sentiment

• It’s about math becoming unavoidable
• It’s about costs investors were told would disappear
• It’s about business models behaving nothing like SaaS

LLMs may still change everything.
But investors can’t survive pretending they have software economics.

The winners will price compute like the scarce industrial resource it is.
Everyone else will learn that capex and opex don’t forgive belief.