The AI labs' pivot to boring software is a countdown to 2027

In late February, Anthropic plugged Claude into Excel, PowerPoint, and Slack, and its enterprise chief, Scott White, drew the line ^[1].

It’s not a product that’s trying to own every workflow. — Scott White

In late March, OpenAI failed to convert its non-binding chip orders into actual purchases because of financing difficulties, helping set off a sell-off that took 19.7% off Samsung and 7.5% off Nvidia ^[2]. On August 19, OpenAI pivoted toward AI agents and cut prices on two new models inside the same report that showed growth decelerating to 18% quarter over quarter and losses widening to $12.3 billion ^[3]. On September 29, TypeSafe shipped Jev — a model that skips open-ended text for cheap, structured decisions — and its developer-relations head asked an uncomfortable question ^[4].

What if this path that we've been so horse-blinders on, the LLM path, what if that's not the only path? — Allie Laabs

Within seven days, OpenAI answered with its own Decisions API, priced at $0.042 per million input tokens against Jev's $0.10 ^[5] — in the same week its revenue came in roughly $20 billion below the $70 billion previously reported for the year ^[6]. Adjacency is not causation. A calendar that lines up this neatly proves nothing by itself — the people shipping products and the people minding the balance sheet could be running on unrelated schedules. So the motive has to come from what the labs themselves said, and OpenAI said it plainly. The August report stated the link: OpenAI pivoted toward AI agents and reduced prices in response to decelerating growth and widening losses ^[3]. Analysts in that same report began treating frontier models as interchangeable commodities, comparing the rivalry to Coca-Cola versus Pepsi — which is what happens to a premium product once the premium is gone ^[3]. OpenAI's chief financial officer has spelled out the destination's economics. Automating her own finance department's credit checks cut the cost from $200 to roughly seventeen cents, and she framed the trade-off in one line — pay a junior analyst to do the work, or let a model do it for a fraction of a cent ^[7]. Routine decisions are the product now, and their price has collapsed a thousandfold. The catalog confirms the costume change: Anthropic's autumn releases were dashboards, slide tools, and motion graphics ^[8], and its July enterprise beta was a compliance checklist — single sign-on, admin policy keys, a Microsoft 365 connector ^[9]. OpenAI hired two senior Cursor sales executives in two months ^[10], and a five-month-old shop of ex-OpenAI people raised $60 million to build agents that own business outcomes ^[11]. The frontier labs are stocking the shelves of a software company. Friar has also named the deadline.

The IPO is not a finish line, it is a milestone, another fundraise. — Sarah Friar

Both labs have confidentially filed to go public — OpenAI, by its CFO's account, with 2027 on the clock ^[12]. Meanwhile the buildout's bills are being carried on someone else's books. Broadcom negotiated a debt package that could reach $100 billion, structured through a special-purpose vehicle that buys Broadcom's own hardware and leases it to tenants like Anthropic — a setup analysts call phantom leverage ^[13]. Public markets have shown what a conditional window looks like: data-center developer Firmus scrapped its Australian IPO after institutions refused to pay an A$43.7 billion valuation against A$50 million in annual revenue ^[14]. And the money behind the machine has set its own clock — Temasek's chief investment officer, Rohit Sipahimalani, calls the unwinding of the AI trade the biggest risk of 2027, while Panmure Liberum's Joachim Klement, a strategist reading the same trade, puts the bubble's burst in 2027 or 2028 ^[15]. Nvidia's CFO still projects hyperscaler spending reaching $1.3 trillion in 2027 ^[16]. None of it is failing. OpenAI's revenue is still growing 77% year over year, and its enterprise revenue is up 107% in the year through October ^[6]. More than ten thousand organizations signed onto the Decisions API within days of launch ^[5]. Anthropic recorded its first operating profit — $559 million — and overtook OpenAI in quarterly revenue, $11.6 billion against $6.7 billion ^[3]. Microsoft, the only big-three hyperscaler running on positive free cash flow, funds its AI spending from operations on a hybrid of subscription fees and usage-based agent revenue — and that is how the labs' buildout is being paid for while everyone waits ^[17]. The year has set up an arithmetic face-off: a per-task price that collapsed a thousandfold, against a buildout whose billing year tops a trillion dollars. Two clocks, both set to 2027. Neither has rung yet.