The Money Moved to Whoever Can Make AI Actually Work

For a couple of years the whole conversation was about models. Which one is smartest. Who trained the biggest thing. That race is not over, but the money just moved to a different question: who can take these models and actually make them do work inside a real company. Let me catch you up on what shifted this week, because it changes how you should think about build versus buy.

The bet is on the plumbers, not the model makers

Anthropic and Blackstone just put a name on it. Ode with Anthropic is a $1.5 billion company built to send engineers into enterprises and wire AI into their core work. Blackstone dreamed it up after watching consultants and small shops fumble AI across its portfolio companies. It bought Fractional AI to be the core. OpenAI has the same idea with The Deployment Company.

The reason is blunt. Chris Taylor, who runs Ode, says the goal is projects that are "the top one or two priority for the CEO." His co-founder Eddie Siegel puts model choice in its place: "It's one ingredient in a system that has to be engineered. I would not define an enterprise transformation in terms of whether they choose Python or Java." Translation for you: picking the model is the small decision. Making it fit your business is the hard one.

Pilots die in the gap between demo and production

The pain these companies are selling into is real, and you have probably felt it. A pilot looks great in a demo and then never ships. Ode's founders built their pitch around exactly that failure, the reason so many AI pilots never reach production.

That gap is where the work lives. Deployment, hosting, testing, debugging, and the messy job of rewiring a real business process. Anthropic's own team stays on the "strategic, mission-aligned" jobs and hands the rest to Ode's 100 engineers. If your team keeps producing impressive prototypes that stall, you are staring at the same gap these firms just raised billions to fill.

Buy the model, own the control layer

JPMorgan drew the sharpest line here, and it is worth copying. The bank is building an AI control layer in Seattle that decides when to route work to its own data centers, when to hit public cloud, and when to use cheaper specialty chips. Global CIO Lori Beer says they are being "careful about lock-in, strategic risk, financial risk, all those things."

The split is clean. JPMorgan builds and owns the software that runs its AI agents, and treats the underlying models as interchangeable. The agents are specific to the bank. The models are commodities you swap as the market moves. There is a cost angle too: Beer notes engineers reach for the newest, priciest model even when a cheaper one works fine, so the control layer routes work to the right model for the job.

A crowd of startups selling "we'll run it for you"

Everybody sees the same opening. Apptio's founders reunited to launch Thira, which builds agents to run back-office work like resetting a locked account or closing an IT ticket across ServiceNow and Jira. Madrona's Matt McIlwain calls it the biggest opportunity he has chased with these founders "by far." India's Emergent hit a $1.5 billion valuation selling "an engineering team in a box" to small firms, at a $120 million run rate.

And Oak raised $60 million to fix identity when agents, not just humans, need access to your systems. Notice the pattern. None of these are selling you a smarter model. They are selling the operational layer around it: execution, access, and getting the thing into production.

The deep cut

The reusable asset is the layer you build on top of the models, not the models themselves. JPMorgan owns its agents and rents its models. That is the move to steal. Before your next build-versus-buy call, split your AI work into two buckets: the general-purpose model, which you should keep swappable and never marry, and the agent logic, routing, evals, and access controls that are specific to your business, which you should own. If a vendor wants to own that second bucket for you, understand you are handing them the part that compounds. And demand what Ode runs internally: constant evals that measure real business impact, not demo polish.

Three questions for your team

1. Which of our AI wins are locked to one model, and how fast could we swap that model out without breaking the workflow?
2. Where are our pilots dying between demo and production, and is that a talent gap we fill with hires or a vendor like Ode fills for us?
3. Do we own the agent and routing layer, or are we about to hand the compounding part of our AI stack to someone else?