The FT reports that the cost crunch isn't about AI being expensive in the abstract — it's that the shift from chat-style usage to agentic/embedded tools (Copilot, Cursor, RAG bots) blew up budgets calibrated on human-typed prompts. A single developer in an agentic IDE can outspend an entire marketing team's monthly ChatGPT usage in one afternoon, leaving finance teams 5–20x over plan.
By submitting the FT piece and driving it to 97 points, fandorin amplifies the framing that enterprises were caught off guard by the cost gap between chat and agentic workloads. The 'we created a monster' quote resonated specifically because it captures the surprise that pilots gave no signal of production costs.
The editorial argues this is the second act every infrastructure shift goes through — cloud hit the same wall in 2015–2017 before FinOps emerged as a discipline. Enterprises aren't churning off AI; they're imposing per-seat token caps, routing cheap tasks to cheaper models (Sonnet over Opus, mini over full), and pulling CFOs into procurement calls that used to be CTO turf.
The editorial contends that AI's cost mechanics are uglier than EC2's because there's no visible 'instance' to reason about — an agent autonomously decides to re-read a repo, re-embed a doc, or retry a tool call, producing emergent spend that finance teams can't forecast. This makes traditional capacity planning inadequate and forces new gateway/routing infrastructure as the control plane.
The Financial Times published a piece this week with a quote that's going to outlive the article: "We created a monster." That's an executive describing what happened after their company opened the AI faucet to employees and watched the invoice arrive. According to the FT, a growing number of enterprises — including names in financial services, consulting, and law — are now actively rationing AI usage after pilot-era budgets collapsed under real workloads.
The specifics matter. The story isn't "AI is too expensive" in the abstract. It's that the unit economics flipped the moment companies moved from chat-style usage (a human typing one prompt at a time) to agentic and embedded usage (Copilot, Cursor, internal RAG bots, automated triage agents). A single developer running an agentic IDE can burn through more tokens in an afternoon than the entire marketing team did in a month of ChatGPT-style queries. Finance teams that budgeted on the chat-era assumption are now looking at line items 5–20x over plan.
The FT names the obvious culprits — Anthropic, OpenAI, Microsoft — but the more interesting detail is the corporate response. Companies aren't churning. They're throttling. Per-seat token caps. Model downgrades (Sonnet instead of Opus, mini instead of full). Internal gateways that route "cheap" tasks to cheaper models. CFOs sitting in on procurement calls that used to belong to the CTO.
This is the predictable second act of every infrastructure shift, and it's arriving on schedule. Cloud went through the exact same cycle around 2015–2017: unlimited adoption, surprise bills, and then the rise of FinOps as a discipline. AI is now hitting that wall, just compressed into 18 months instead of seven years.
The cost mechanics are uglier than cloud's were, though. With EC2 you could at least see the instance and reason about its cost. With agentic AI, the cost is emergent: an agent decides to re-read the repo, re-embed a doc, retry a failed tool call, or expand a chain-of-thought trace. Every one of those decisions is a billing event the user never sees. Cursor's pricing drama earlier this year — where power users suddenly hit usage walls mid-feature — was the leading indicator. The FT story is the trailing one, told from the buyer's seat.
The community reaction on Hacker News is worth reading, because it splits cleanly. One camp says this is just rational repricing: pilots were free-tier theater, real usage costs real money, welcome to capitalism. The other camp — and this is the more uncomfortable one — argues the vendors knowingly underpriced inference to capture seats, and the "true cost" pricing now landing on enterprise invoices is the markup, not the math. Anthropic's and OpenAI's gross margins on API are not public, but the recent round of price hikes and rate-limit tightening suggests the honeymoon discount is being clawed back.
The deeper issue is that nobody — vendors included — has a clean answer for how to price a non-deterministic product whose cost-per-task varies by 50x depending on how the model decides to "think." Fixed per-seat pricing transfers risk to the vendor. Pure metered pricing transfers it to the buyer, who can't forecast it. Hybrid caps (the current default) just create a worse user experience for the heavy users who are, not coincidentally, the people generating the most value.
If you're a platform engineer or eng leader, three things are about to land on your desk, if they haven't already.
First: an AI gateway is no longer optional. The companies surviving this with their budgets intact are the ones that put a proxy between their developers and the model APIs months ago. LiteLLM, Portkey, Helicone, Kong's AI plugin, or a homegrown gateway — it doesn't matter which, but you need one. Without it you have no per-team attribution, no model routing, no prompt-cache enforcement, and no kill switch. With it, you can do the boring FinOps work: route summarization to Haiku, route reasoning to Opus, cache aggressively on system prompts, and bill back to the team that incurred the cost.
Second: prompt caching is now a load-bearing optimization, not a nice-to-have. Anthropic's prompt cache cuts input costs by ~90% on cache hits. If your RAG pipeline or coding agent isn't structured to maximize cache hit rate — stable system prompt at the top, volatile content at the bottom, deterministic tool definitions — you are leaving real money on the table. Audit your prompt structure. Measure hit rate. Teams I've talked to are seeing 60–80% cache hit rates after one afternoon of restructuring, which translates to 50%+ bill reductions on the same workload.
Third: expect finance to start asking about tokens the way they ask about AWS spend. Get ahead of it. Build a dashboard before you're asked for one. Tag every API call with team, project, and use case. The companies in the FT piece got blindsided because they had no visibility — by the time the bill arrived, the horse was out of the barn. Don't be that team.
The pullback is not the end of enterprise AI; it's the end of enterprise AI's pilot phase. What comes next is the boring, profitable middle: governed usage, measured ROI, smaller models for the 80% of tasks that don't need a frontier model, and a generation of FinOps-for-AI tooling that will be acquired by Datadog and AWS within 24 months. The companies that thrive in this phase won't be the ones with the biggest AI budgets — they'll be the ones with the best per-task cost discipline. The monster in the FT headline isn't AI. It's unmetered AI. Meter it, and it goes back in the box.
We're in a dangerous valley where AI is _just_ good enough to fool some otherwise very smart people. Similar to the old adage of "a little bit of information is a dangerous thing." Lots of CEOs got duped into thinking that model capabilities were far ahead of where they actually were.
> Since the start of the year, Chinese AI models have overtaken their US counterparts in token consumption, according to data from OpenRouter, an aggregation platform that allows users to access multiple AI models.That's a bit of a dodgy statistic. OpenRouter only tracks their own users - th
I genuinely have no idea how some of these companies got so far over their skis on AI. It simply does not make sense to me.
> The ride-hailing company has introduced usage caps, limiting employees to $1,500 in monthly token spending on individual AI tools, after blowing through its entire AI 2026 budget by April.Right, because they set their 2026 budget in 2025. And in 2025 nobody could predict how good (and token-hun
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I'm so frustrated by both the zealous AI bulls and the blind AI opposition.There's a lot of issues, ranging over technical, cultural, environmental, and moral problems. But there's also obvious value. To say otherwise tells me you haven't actually tried to make use of these tools