Who’s minding the store? Google and no one else

Let’s start at the very beginning. A very good place to start:

There are only a handful of companies that actually provide AI at scale — the firms that own or control the data centers, power, chips and routing that make large models possible. Everyone else, no matter how famous the model, runs on top of them.

There are only four true AI infrastructure providers

• AWS (Amazon)
• Azure (Microsoft)
• Google Cloud
• Oracle

Everyone else:

• rents from them, or
• partially owns clusters but still depends on power, GPUs and facilities they don’t control.

This concentrates economic power at the infrastructure layer.

LLM companies are not software businesses

• They look like software.
• They price like software.
• Investors treat them like software.

But economically they behave like industrial compute businesses with linear marginal costs:

• Every unit of usage costs real money.
• Scale does not create operating leverage.
• It increases absolute losses unless pricing changes.

OpenAI is the best possible test case for an LLM as a business — and it fails the test, despite…

• the largest consumer user base ever
• the strongest brand
• first-mover advantage
• deep subsidy from Microsoft
• enterprise traction
• premium pricing vs competitors

And yet:

• it is not clearly profitable
• margins are flat to negative
• costs scale with usage
• feature expansion raises cost per user
• adoption is no longer accelerating

If this configuration can’t clear the math, weaker ones won’t.

“Free” consumers are not neutral — they are a liability

95% of ChatGPT users pay nothing, but still incur…

• inference cost
• power cost
• infra cost

This is not like Google Search, where marginal cost trends toward zero, because free usage of ChatGPT permanently increases losses.

Ads do not rescue LLMs because

• most users are low-intent
• sessions are episodic
• targeting is weak
• ad systems add compute cost
• trust caps ad density

ChatGPT is a terrible ad surface compared to Search or Social.

Why no one is talking about this

Because no one is incentivized to integrate the whole system. Most observers are trapped in silos:

• Technologists focus on model quality and benchmarks
• Product teams focus on engagement and features
• Investors focus on adoption curves and TAM
• Media focuses on headlines, not unit economics

Very few people are paid to ask a simple but fatal question:

Do the economics clear when usage scales?

Even fewer are willing to answer it publicly.

There is also powerful SaaS muscle memory at work.

For two decades, software followed a reliable pattern: scale users first, margins expand later. AI looks like software, so that pattern is being applied reflexively — even though AI behaves more like industrial compute, where marginal cost remains real and persistent.

Finally, subsidies distort perception. As long as hyperscalers, strategic partners and deep-pocketed incumbents absorb losses, the system appears healthy. Losses are real, but they are hidden — deferred, shared or silently absorbed.

The result is a collective blind spot: activity is mistaken for viability, and usage is mistaken for profit.

Why Google is different — and why Gemini doesn’t have to “work” like ChatGPT

Google is the only major player for whom AI does not need to be a profit center.

For Google, Gemini is defensive infrastructure, not a standalone business.

Its purpose is not to monetize AI directly, but to protect Search, which remains one of the most profitable businesses ever created.

Google understands the risk clearly:

Most users now approach AI as if it were search, even though it is not. If another LLM became the default interface for everyday questions, Google would risk losing ownership of user intent — and with it, its advertising engine.

Gemini exists to prevent that displacement.

Crucially, Google does not need Gemini everywhere, all the time. It can — and does — gate its use:

• Traditional search handles the majority of queries cheaply
• Gemini is invoked selectively, where it adds defensive value
• Ads and commercial intent are preserved structurally
• Cost is controlled through routing, optimization and hardware ownership

Google is willing to burn 10% of Search profit to run Gemini if it protects the remaining 90%. That trade is rational. Losing Search entirely would be catastrophic.

This is the asymmetry that breaks the comparison with OpenAI and Anthropic. They must make AI profitable on its own. Google does not. It only needs AI to ensure users never leave Google to ask their questions elsewhere.

Gemini is not there to win.
It is there to prevent losing.

So — is Google the clear winner?

Not so fast.

Gemini’s problem is not intelligence. It is trust.

Hallucinations rarely surface in demos. They surface only when users move beyond simple retrieval and begin relying on AI for reasoning, judgment and synthesis — the moment AI stops behaving like search and starts behaving like cognition.

Independent testing shows that when Gemini 3 is pushed into reasoning tasks, hallucination rates become extraordinarily high — as high as 88% in controlled evaluations. This is not obvious in casual use. It only appears after users begin to trust the system — once.

And trust, once broken, does not come back easily.

Let me put it another way.

If someone is honest 99% of the time, we don’t call them “mostly honest.”
We call them a liar.

That is how users react the first time an LLM confidently lies to them.

• That is what users said about Gemini 1.0.
• That is what they said about Gemini 1.5.

And that is what they will say about Gemini 3.0 after it earns trust — briefly — and then loses it.

• Better models will not solve this problem.
• Faster models will not solve it.
• More demos will not solve it.

Because the root issue is structural.

LLMs do not remember in a loss-less, durable way. They are constrained by context windows and probabilistic recall. “Persistent memory” keeps trying — but it is not the same thing as indelible memory.

Until AI systems can remember reliably — not approximately, not statistically, but correctly — trust will remain fragile.

Google knows this.

Which is why Gemini is gated, constrained and selectively deployed. And it is why, despite Google’s unique strategic position, the solution to AI’s biggest problem has not been deployed – yet.

That capability is not a feature.

It is not a demo.

And it is still conspicuously absent from the roadmap.

– Published Thursday, January 8, 2026