Six angles this week

Data centers were on the ballot this week, and they lost almost everywhere.

Nearly every local measure opposing them passed. Texas and Ohio — two of the biggest states for planned AI capacity — elected governors who campaigned on pausing construction. And here's the part that should stop you: both parties ran against data centers. This isn't a pendulum that swings back in two years.

For the past two years the story of AI has been a story of constraints falling one after another. First we were short on compute. That ramped. Then memory. That will ramp. Then power. That's being solved. Every constraint had an engineering answer.

The newest constraint doesn't. It's political. You can't engineer your way past a community that votes no.

What this means for anyone building on AI: stop assuming intelligence becomes free. The commodity narrative — cheap, abundant compute forever — may simply be wrong. We could be looking at real, persistent marginal costs for a long time.

That changes how you architect. Right-size your models. Push inference to the edge where it makes sense. Build cost-per-token discipline into the design, not as an afterthought. Lock in your compute relationships while you can.

The systems that survive the next five years are the ones that still make sense when every AI call costs something. The ones that only work if compute is free are already in trouble.

Plan for scarcity. Hope for abundance.

Intuit's CEO was asked how many accountants there will be in five years. His answer: "Fewer. Many fewer."

But the more interesting part came next. Today an accountant spends roughly 90% of their time on manual work — entering data, reconciling accounts, categorizing transactions — and 10% on judgment. He thinks that inverts. In a few years it's 90% judgment, 10% manual.

That reframes the whole AI-and-jobs conversation. The question isn't "will AI replace the accountant." It's "what is the accountant's job once the manual work is gone." And the answer is the part that was always the hard part: is this right, are the books clean, what decision should you make next.

This is the template for every judgment-heavy role in our industry. Underwriting. Claims. Actuarial review. The reconciliation and the data gathering get automated. What's left is the call only a human should make — and that part becomes more valuable, not less.

The design principle that follows is clear: automate the drudgery, keep the judgment visibly human. Don't build systems that try to replace the expert. Build systems that give the expert their time back for the decisions that actually require them.

Intuit's warning to accountants applies to all of us: the people who go all-in on automation win. The ones who cling to the manual work because it's familiar are the ones who get disrupted.

The job isn't disappearing. It's being distilled down to the part that mattered all along.

Every few months someone points out that AI hasn't shown up in the productivity statistics yet, and concludes the whole thing is overhyped.

Here's why that's the wrong lens.

In the PC era, it took fourteen years for computers to move the headline productivity numbers. But individual companies were reporting real gains years before that. The aggregate is always the last thing to turn — it's noisy, it's lagged, and by the time it's obvious the winners have already pulled ahead.

So stop watching GDP. Watch what specific firms are actually doing.

A major law firm just committed 3% of its billable time to AI — about $1.1 billion over five years. Their stated ambition isn't efficiency. It's "redesigning the service completely." That's the tell. They're not bolting AI onto the old workflow. They're rebuilding the workflow around it.

That's the difference between the firms that will get results and the ones that won't. Deploying a tool into an unchanged process gives you a rounding error. Redesigning the process around the tool gives you something real.

My advice to anyone waiting for proof: don't. Pick one process. Set a measurable target. Redesign it end-to-end. Budget real money against it the way that law firm did.

The companies that move now will be reporting gains while the skeptics are still refreshing the aggregate statistics, waiting for permission.

The payoff is already here. It's just not evenly distributed yet.

Intuit made a mistake worth learning from.

They launched a payments agent, a payroll agent, an accounting agent. Each did its slice well. But the CEO admitted the combination "didn't actually deliver a great experience end-to-end." The agents worked. The system didn't.

The reason is the most important architecture lesson in AI right now: they weren't designed as a system. They were designed as slices. And a collection of good slices is not a good experience.

What they're building instead is telling. One unified platform — all the data, all the data models, all the AI models on a single foundation — serving both sides of a network. Accountants on one side, businesses on the other. And critically, they've shifted from treating accountants as a sales channel to treating them as customers, because each side makes the other more valuable.

That's where the moat is. Not the model — anyone can access a good model. The defensibility comes from two things: a unified data layer that lets agents span the entire workflow, and a network where the parties genuinely need each other.

For anyone in insurance, this maps directly. Carrier, agent, policyholder — three parties who depend on each other. The temptation is to bolt a clever agent onto each siloed system and call it progress. That's the Intuit mistake, pre-correction.

When you evaluate an AI investment, ask one question: does this deepen a shared platform and strengthen a network, or is it another disconnected agent a competitor can copy next quarter?

Models are rented. Architecture and networks are owned.

There's a lot of excitement about outcome-based and usage-based pricing for AI. The theory is elegant: you pay for what you get, or you pay for what you use.

Intuit's CEO just poured cold water on it, and he runs one of the most sophisticated software businesses on earth.

His finding from actually testing it with customers: they're not interested. Not in outcome-based pricing, not in token-based pricing. The reason is simple and human — they can't control what they'll end up paying, and unpredictability is something buyers hate when real money is on the line.

What they want is a fixed number they can plan around.

Now layer that against the other thing happening in AI right now: compute is getting harder to build, not easier, and inference is likely to carry a real marginal cost for a long time. Put those two facts together and you get an uncomfortable conclusion.

If your costs move with consumption, but your customers insist on fixed prices, then the volatility doesn't disappear. It just moves onto your margin instead of theirs.

That's the risk a lot of AI business models are quietly carrying without pricing it in. They assume they can pass variable compute costs straight through to the customer. The customer, it turns out, has other plans.

If you're modeling AI economics for your organization, build this in now. Assume you'll eat more of the cost volatility than you'd like. The companies that get caught out are the ones who assumed the customer would absorb it.

Predictability is a feature customers will pay for. It's also a risk you'll carry.

The AI buildout is quietly becoming an insurance story, and most of the industry hasn't noticed yet.

Look at three things that landed in the same few weeks.

Forty percent of planned US data center capacity sits in states that now want to pause construction. That's correlated risk — construction delay, political risk, business interruption — concentrated in a handful of hostile jurisdictions. If you're underwriting any of it, the geography just got a lot more dangerous.

Second: a single stolen key made 1.7 million calls to Amazon's AI models in thirteen hours. That's a new kind of cyber exposure. Not data theft in the old sense — automated, metered consumption of expensive AI infrastructure on someone else's credentials. Most existing cyber policies never imagined it.

Third: the companies in the middle of the AI economy — the ones holding other people's keys and borrowing other people's money to sell compute — are starting to show credit strain. That's counterparty and surety and D&O risk, bleeding in from a sector that didn't exist five years ago.

Here's the point for anyone in P&C. The AI economy is manufacturing new categories of risk faster than the market can price them. That is exposure sitting silently in books today. It's also the biggest coverage opportunity in a decade.

The carriers who actually understand the technology will price these risks first — and most profitably. The ones who treat AI as someone else's problem will discover they were exposed all along.

Start mapping where AI risk already lives in your book. It's further in than you think.