The $725B Question Nobody Is Asking About AI

Three out of every four companies deploying AI right now can't tell you if it's working. They're spending on models, pipelines, and inference, but the ROI is stuck in a fog of integration debt and data quality issues. That's not a Wall Street problem. That's a builder problem.

Jason Lemkin over at SaaStr just published a sprawling conversation about AI's big bets: Google losing two top scientists to Anthropic, DeepSeek raising billions, and the mounting $725B question of who will pay for all the infrastructure. It's the kind of piece you read to understand why the market moves the way it does.

But here's what it misses. The real story isn't talent or capex. It's the ground-level friction that turns grand AI ambitions into stalled pilots and negative-margin experiments.

The Practical Pain Is Worse Than the Math

Our platform tracks over 2,500 documented problems across industries. Right now, 342 of them cluster around "cost of AI implementation" — and the average severity rating is 4.2 out of 5. That's not a theoretical gap. That's teams burning budget on model training and deployment without knowing how to measure, let alone capture, the return.

Consider supply chain. We count 47 distinct problems in that vertical alone where AI models fail because the underlying data is too messy, too siloed, or too unreliable. A procurement director at a mid-size manufacturer told us their anomaly detection model flagged 80% false positives for three months before they scrapped it. The vendor was happy to take their money. The integration was left to a team that had never worked with streaming data.

This is the part of the AI spending debate that doesn't make it into analyst reports. The big numbers — $700 billion in capex, $7.6 trillion cumulative by 2031 — assume capital deployment equals capability. It doesn't. Capability only arrives after a long, painful slog through data engineering, change management, and organizational resistance.

Talent Wars Are a Distraction

Lemkin's piece highlights Google losing Noam Shazeer and John Jumper to Anthropic. That's a real signal about where the best researchers want to work. But for the 99% of companies that aren't competing for Nobel winners, the binding constraint isn't talent. It's execution.

We see 541 problems tagged "AI sales and adoption" — meaning the product exists, the market wants it, but the deployment fails. In enterprise SaaS alone, there are 89 documented app ideas aimed at improving AI monetization, with an average severity of 3.9/5. These aren't unsolvable problems. They're execution gaps that no amount of top-tier research talent can fix from a lab.

The interesting question isn't whether Google can keep its best scientists. It's whether every other company can turn their AI investments into production systems that actually deliver cost savings or revenue. Our data says most can't. Yet.

Where the Real Opportunity Lives

For indie hackers and seed investors, this is where the signal lives. The $725B capex bubble narrative is interesting dinner conversation. But the real market is in the tools, platforms, and services that make AI work in the messy reality of existing business operations.

We track 342

problems related to 'cost of AI implementation' with an average severity of 4.2/5. These range from "we deployed a chatbot and nobody uses it" to "our model degrades in production and we don't know why." Each one is a startup waiting to happen.

Take model observability. Everyone talks about it, but the specific pain points — drift detection without endless false alarms, cost attribution per inference, prompt caching strategies that actually work — these are unsolved and actionable. Anthropic is now emailing users about low prompt cache hit rates, as Lemkin notes. That's a product move. But it's also a signal that the operational details matter more than the architecture.

The Parity Tax Nobody Avoids

Lemkin's group touches on something important: the parity tax. If every bank installs ATMs, banking doesn't get more profitable. The cost savings get competed away. AI is heading the same direction. The companies that don't adopt will die, but the ones that do won't necessarily see margin expansion. They'll just stay alive.

This makes the execution gap even more urgent. If you can't nail the integration, you're spending on AI just to keep up, not to pull ahead. And you're probably spending too much.

The $725B question isn't just about who pays the capex bill. It's about who can turn that spending into operational leverage. The answer, based on the problems we see every day, depends less on the size of your cluster and more on the quality of your data pipelines, the discipline of your deployment, and the willingness to fix the boring stuff first.

That's the conversation worth having. Not "who will pay," but "who will execute."

And if you're building right now, that's your opening.