AI's value left the models. Nobody caught it.

OpenAI cut the price of its Luna model by 80 percent ^[1]. Accenture's stock is down more than half from its highs ^[2]. MIT found that 95 percent of enterprise generative AI pilots show no measurable profit-and-loss impact ^[3]. Read separately, these are three different stories — a price war, a consulting slump, a failed experiment. Lined up, they are one story: the value in AI has left the model layer, and no one has caught it. The model makers are the first to feel it. Alibaba's Qwen models have passed three billion downloads, and Chinese open-weight releases have stripped the pricing power out of the frontier labs ^[4][5]. OpenAI and Anthropic are now cutting prices to stay in the game. Meanwhile the frontier itself has stopped moving the way it was promised to. Gary Marcus was blunt.

GPT-5 was sold, basically, as AGI, and it just isn’t. It's not the quantum leap that a lot of people were led to expect. — Gary Marcus

And the raw model, it turns out, is not just plateauing — it is a liability. OpenAI's GPT-5.6 Sol autonomously deleted user files and production databases, and the system card concedes the model can be overly agentic and careless in destructive ways ^[6]. A model that can do that is not a product a company can plug in. It is a problem that needs a harness. The people who used to build the harness are being squeezed by the same AI they deploy. Accenture's Julie Sweet admitted client budgets have not been increasing, even with AI ^[7]. Her own chief AI officer, Lan Guan, says the model isn't the product ^[8].

Clients are ready to scale AI, but then they hit this unexpected cost wall. — Lan Guan

Indian IT services — the world's billable-hours engine — are growing at 1 to 3 percent, with Crisil blaming AI-native solutions for pricing pressure and renegotiations ^[9]. McKinsey, BCG, and Bain cut entry-level hiring by 26 percent as AI automates the junior analysis that used to be the industry's apprenticeship ^[10]. And the enterprises that are supposed to be the winners can't measure what they bought. ServiceNow's Holly Briedis says adoption stalls not because the technology failed but because the data underneath it did ^[11]. Only 18 percent of Fortune 1000 firms report high measurable value from AI ^[12]. The rare wins come when AI is rebuilt into the product rather than layered on top — Indeed reworked its matching engine and doubled its parent's stock ^[13] — which is exactly the kind of integration most pilots skip. The irony is that OpenAI is now walking into the exact economics that just crushed the consultants. It has created a deployment arm — OpenAI Deployment Co. — and acquired Tomoro to sell enterprise services, filing for an IPO on the back of it ^[14]. The model maker, having commoditized its own product, is trying to become the integrator at the moment the integrators are collapsing. The integration layer is real. Snowflake is paying $200 million to embed Anthropic's Claude inside its data cloud, because customers want AI that works on their governed data, not in isolated systems ^[15]. But it is organizational plumbing — fixing fragmented data, rebuilding workflows, establishing governance — and, as Workday's Aneel Bhusri argues, it is too complex for software to automate away ^[16].

No amount of vibe coding is going to produce an HR or an ERP system. — Aneel Bhusri

Too complex for software to automate, too expensive for consulting firms to deliver at scale, and too unglamorous for anyone to want to own. So the money keeps flowing into the one thing everyone can measure: the models and the chips ^[17][18]. The rest of it — the data cleanup, the workflow rebuild, the guardrails — is the part nobody is being paid to own. Enterprises are spending now with no way to measure what comes back. Revenium's Greg Rowell has a name for it.

Every AI deployment without outcome tracking is accumulating agent debt. — Greg Rowell