--- title: "Amazon Ran a Project 860% Over Budget and Took Five Months to Notice" pubDatetime: 2026-07-30T15:04:00.000Z description: "Amazon engineers told staff about AI projects that blew past budget by up to 860%, one of them undetected for five months. The failure mode is metered trial-and-error with nothing gating it." tags: [ai-agents, ai-adoption, llm-agents, pricing, ai-policy, 2026, 2026-q3, 2026-07] --- > [!tldr] TL;DR > Amazon engineers told colleagues this week about several "catastrophically expensive" AI cost overruns on internal projects, per [FT reporting published today](https://www.ft.com/content/77baac40-d803-4084-94f3-a133653072cf). The worst hit $1.8mn, 860% over budget, and ran five months before anyone caught it. Two other cases cost roughly $541,000 and $134,000. Amazon says a few teams were involved and the cases don't represent company-wide usage, and Goldman Sachs raised its price target the same week. The lesson survives the pushback: agent loops burn metered tokens, and if nobody owns the bill, nobody reads it. Five months is a long time to miss a seven-figure line item. That is the standout number from an internal Amazon presentation on Tuesday, [reported by the FT](https://www.ft.com/content/77baac40-d803-4084-94f3-a133653072cf) two days later: a project running Claude Sonnet landed at $1.8mn, 860% over its budget, with five months between the spending and the discovery. Two other cases came up in the same presentation. A financial-audit tool overshot by about $541,000. A logistics AI project ran $134,000 over and was caught after more than a fortnight, which counts as fast by comparison. [dev.ua's write-up](https://dev.ua/en/news/overspending-by-860-of-the-original-budget-amazon-loses-millions-of-dollars-due-to-lack-of-control-over-ai-costs) carries an Amazon employee's summary of the problem: "It's very difficult to understand how much anything related to artificial intelligence costs." ## Metered trial-and-error breaks the old cost intuition Every engineer's instinct about compute cost formed on software where iteration is free. Rerun the test suite, retry the build, loop the script twice more to be sure. The marginal cost of another attempt rounds to zero, so nobody builds the habit of checking. Agent workloads invert that. Each retry is a fresh set of input and output tokens, each failed plan is billed in full, and a loop that retries on error can spend faster than any human reviews it. The cost signal arrives on a monthly invoice, aggregated across teams, long after the decision that caused it. Five months of detection lag follows from that gap. The meter ran in a place the org chart gave nobody to watch, and the invoice was the first thing that spoke up. ## Amazon already had this problem in May The overruns are the second version of this failure inside Amazon this year. On 29 May, [Amazon shut down "Kirorank"](https://www.hcamag.com/us/specialization/hr-technology/amazon-shuts-down-ai-leaderboard-after-tokenmaxxing/577189), an internal leaderboard ranking employees by AI usage, after staff inflated their token consumption to climb it. The practice picked up a name: tokenmaxxing. SVP Dave Treadwell's message to staff afterwards was blunt: "Please don't use AI just for the sake of using AI." Amazon had pushed AI tooling to more than 80% of its developers on a weekly basis, and had built a scoreboard rewarding consumption, before it had controls on what that consumption cost. Both episodes trace to ungated token spend with no owner. Adoption ran ahead of governance, and each correction arrived after the bill. ## Goldman raised its price target anyway Amazon's response to the FT was that the cases "only involved a few teams and do not represent the overall situation of the company's daily AI usage." That's the expected line, and it may well be accurate. Three named projects totalling roughly $2.5mn is a rounding error against a [~$200bn capital spending plan for 2026](https://www.barchart.com/story/news/3505617/dear-amazon-stock-fans-mark-your-calendars-for-july-30). Goldman Sachs analyst Eric Sheridan [raised his Amazon price target from $325 to $335](https://www.tipranks.com/news/heres-why-goldman-raised-amazon-amzn-stock-price-target-ahead-of-q2-earnings) the same week, forecasting $827bn of cumulative AWS capex from 2026 to 2028 and AWS revenue growth above 33% for the year. To Goldman, rising AI spend reads as confidence in the model. One bank's note is one bank's note, and the broader market reaction to the story is still unmeasured. Take the dismissal at face value on the financial question and reject it on the operational one. The dollar figures are small; the detection lag is the finding. An organisation able to absorb $1.8mn without flinching is exactly the kind that takes five months to spot it, and a smaller company with the same blind spot doesn't get five months. ## "Growing pains" is half right Secondary coverage has settled on a gentler reading: uncontrolled token consumption from ungated agent use is a known, bounded failure mode, an expected growing pain of adoption rather than evidence AI investment is unsound. That reading holds on the diagnosis. Nothing in the Amazon cases suggests the underlying tools failed or the spending bought nothing. Where it slips is the word "expected." A failure mode you can name in advance, that has already fired once at the same company in the same year, and that still runs five months undetected, has stopped being a growing pain and become an unbuilt control. Growing pains are things you haven't seen yet. Amazon had seen this one in May. ## Timing The story broke hours before Amazon's Q2 2026 earnings, with guidance of $194–199bn in net sales and $20–24bn in operating income, and capex commentary the thing analysts wanted. Shares had already fallen about 10% since 21 July, from around $250 to $226, recovering to roughly $233 pre-market. The stock's all-time closing high of $278.56 was set on 5 May. Whether the disclosure was Amazon getting ahead of an awkward number or an inconvenient leak is unknowable from outside, and both readings fit. What the earnings release said about 2026 capex guidance was still unknown when this was written. ## The controls that were missing Preventing the Amazon cases requires no novel tooling. The gap sits between per-project budgets that exist on paper and spend telemetry that reaches a human weekly. - Hard per-project token ceilings enforced at the API key, rather than tracked in a spreadsheet. - Cost attribution by team and project, visible without waiting for a monthly close. - Alerting on burn rate rather than absolute spend, since a runaway loop shows up in the derivative first. - A named owner per workload who sees the number and can kill the job. Measuring AI usage and measuring AI cost are different projects. Amazon built the first, gamified it, then found out it had skipped the second. The 860% is the memorable figure, though the five months is the part worth writing into your own runbook as a thing to prevent. ## Sources - [Amazon finds cases of AI causing runaway spending on tech projects](https://www.ft.com/content/77baac40-d803-4084-94f3-a133653072cf), Financial Times, 30 July 2026 (paywalled) - [Overspending by 860% of the original budget](https://dev.ua/en/news/overspending-by-860-of-the-original-budget-amazon-loses-millions-of-dollars-due-to-lack-of-control-over-ai-costs), dev.ua - [Amazon shuts down AI leaderboard after 'tokenmaxxing'](https://www.hcamag.com/us/specialization/hr-technology/amazon-shuts-down-ai-leaderboard-after-tokenmaxxing/577189), HCAMag - [Goldman raises Amazon price target ahead of Q2 earnings](https://www.tipranks.com/news/heres-why-goldman-raised-amazon-amzn-stock-price-target-ahead-of-q2-earnings), TipRanks - [Amazon Q2 2026 earnings preview](https://www.barchart.com/story/news/3505617/dear-amazon-stock-fans-mark-your-calendars-for-july-30), Barchart - [Amazon sounds alarm ahead of earnings](https://www.tradingkey.com/analysis/stocks/us-stocks/262064091-stock-amazan-amzn-price-ai-llm-tradingkey), TradingKey - [Hacker News discussion](https://news.ycombinator.com/item?id=49108581)