Five angles this week

A senior Amazon employee said something this week that should worry every technology leader: "It's difficult to figure out how much anything AI-related costs."

That's after a single Claude project ran $1.8 million over five months.

Here's what I keep seeing. Teams stand up an AI workflow, it works, everyone moves on. Nobody notices it's running the most powerful and most expensive model available for a task a far cheaper one handles perfectly. The bill just grows quietly in the background.

One of the newsletters I read described an agent that blew through $500 a day before anyone caught it. After an audit, same work, better output, six dollars a day. The difference was entirely model selection and routing.

This is the cloud cost story all over again. In 2015 we learned that unmanaged infrastructure spend surprises you. AI spend is the same lesson wearing a new coat.

For those of us in insurance, this isn't abstract. A claims triage or underwriting-support workflow runs millions of times a month. Which model you pick is a margin decision, not a technical footnote.

Three things worth doing now: attribute cost to each workload, route tasks to the cheapest model that clears the quality bar, and make someone accountable for token economics.

Model prices are falling fast. The real risk is paying last year's prices for this year's routine tasks. Audit before finance does.

How are you tracking your AI spend today?

Stripe just agreed to pay more than seven billion dollars for a company that never trained an AI model.

OpenRouter routes requests across 400-plus models for millions of developers, handling 55 trillion tokens a week, and takes 5.5% of everything that passes through. Stripe didn't buy a model. It bought the toll booth between developers and every model.

That's the lesson worth sitting with. In a market where the "best model" changes every quarter, the durable position isn't picking the winner. It's owning the connective tissue everyone has to pass through regardless of who wins.

There's a direct read-across for anyone running enterprise architecture. If you wire your core systems straight into one provider's API, you've quietly inherited their pricing, their outages, and their roadmap. When a cheaper or better model shows up next month, you can't move.

The alternative is an abstraction layer of your own between your applications and the model providers. One place to swap models as prices drop. One place to enforce cost controls, logging, and governance. One place that survives when the leaderboard reshuffles.

Stripe paid 50 times revenue for exactly this insight. You can apply the same thinking internally for the cost of an architecture decision.

Don't bet the enterprise on a single model vendor. Build the switch that lets you change your mind.

Are you architecting for model choice, or hardwired to one provider?

The clearest AI experiment I've seen this year quietly answered the question everyone keeps arguing about.

Prime Intellect locked 18 frontier models in offline sandboxes for eight days and had them run 153 autonomous research experiments. The best model closed 82% of the gap to a human record that took months of expert work to set.

Impressive. But here's the part that matters: not one model invented a method that wasn't already sitting in the published literature. On the grinding work of running experiments and re-testing dead ends, the machines were extraordinary. On the single new idea, they scored zero.

The framing I keep coming back to is genius as one percent inspiration and ninety-nine percent perspiration. AI is now devastatingly good at the ninety-nine. It cannot yet take the one.

That should change how you deploy it. Point AI at the perspiration: the variations, the dead ends, the tireless re-testing no human sustains. Don't expect it to originate the novel underwriting insight or the genuinely new product structure. It recombines what exists; it doesn't leap.

The practical move is to redesign roles around this line. Route the grind to the machine. Free your best people for the one percent only they can reach.

The organizations that win won't be the ones that automate the most. They'll be the ones that put scarce human judgment exactly where the machine falls short.

Where are your best people spending their time right now?

The most famous consumer AI product in history just told investors that consumers aren't the main event anymore, and said it so casually nobody asked a follow-up.

OpenAI's CFO confirmed enterprise revenue has overtaken ChatGPT's consumer business, months ahead of schedule. Enterprise is growing 32%. The chatbot everyone downloaded was the billboard. The real store is the API, selling intelligence to companies by the unit.

Two things follow from this that matter if you buy AI for a living.

First, the frontier labs are now optimizing for you. Their pricing, their reliability commitments, their roadmaps will increasingly reflect what large enterprises need rather than what consumers want. Expect better service terms and more stability. But also recognize you're now a target market with real leverage. Don't accept consumer-grade terms for production workloads. Negotiate like the customer they actually need.

Second, and easy to miss: the same CFO declared "tokenmaxxing" dead. The era of throwing maximum compute at every problem is ending. Efficiency and fitness-for-purpose are the new frame. The biggest lab in the world is telling you the winning move is no longer the most expensive model on every task.

That lines up with what teams are learning in practice. The right model for the job beats the most powerful model for everything.

The vendors have quietly shifted their center of gravity toward the enterprise. Shift your negotiating posture and your architecture to match.

Are you still buying AI on consumer terms?

An insurtech quietly launched something in California this week that deserves more attention than the headline gives it.

bolt's new "Connected Distribution" is being described as an AI-powered distribution platform. But read past the label. The actual pitch is combining customer data, workflows, and market access into a single operating model across all lines. AI isn't the product. It's the connective layer.

That's the part worth sitting with, because distribution is exactly where P&C has historically hidden its fragmentation. Separate systems for data, for quoting, for market access, held together by human effort and process. It works, mostly, until someone offers your distribution partners something cleaner and faster across every line at once.

The uncomfortable truth is that AI creates the most value precisely where your underlying architecture is already unified, and it exposes you fastest where it isn't. You can't sprinkle intelligence over a pile of disconnected point systems and expect a coherent experience to come out the other side.

So the defensible response to a platform like this isn't buying an AI feature. It's collapsing your own data, workflow, and market-access silos into something coherent enough that intelligence has clean ground to stand on.

This is the real bridge between AI and core platforms. The AI conversation and the architecture conversation are the same conversation. Treating them separately is how incumbents get outflanked on their own distribution.

Is your distribution stack a platform, or point systems held together by process?