An internal memo to no one in particular
TO: The Entire Technology Industry
FROM: The Department of Obvious Outcomes
RE: We told you so. You didn't listen. Here we are.
DATE: March 10, 2026
Let's begin with Amazon, because the story is too good to bury.
On March 5, Amazon's website went dark for six hours. Customers couldn't see prices. Couldn't check out. Amazon Fresh orders — the ones people schedule because they have children and jobs and don't have time to go to an actual store — evaporated. The peak failure count hit 21,716 reports on Downdetector at 3:48 p.m. ET before the engineers — the human ones — clawed the lights back on around 8 p.m.
Amazon attributed this to "a software code deployment." Technically accurate. Also technically the most artfully evasive sentence written in the English language since a defense contractor described an explosion as "an unplanned energetic event."
Because what the internal briefing notes actually said — the ones the Financial Times saw, the ones that reportedly had the reference to generative AI quietly deleted from a subsequent draft — was that Amazon had identified "a trend of incidents" involving changes with "high blast radius" tied to "Gen-AI assisted changes." The company's own senior VP of eCommerce Foundation, Dave Treadwell, opened his memo to employees with the sentence: "Folks, as you likely know, the availability of the site and related infrastructure has not been good recently."
Folks. Imagine being the person who spent eighteen months telling executives that AI-assisted code deployment without sufficient guardrails was operationally reckless, and then receiving that email.
This wasn't a one-off. It isn't even close to a one-off.
In December 2025, Amazon's own Kiro AI coding tool deleted and recreated an AWS production environment. Autonomously. Without being asked. The ensuing chaos lasted thirteen hours. AWS, to its credit, described the incident as "extremely limited" — it only disrupted customers in mainland China, which is a sentence that reveals a lot about how "limited" is defined at cloud scale.
Then another AWS AI incident. Then the March 5 website collapse. Then Tuesday's emergency "deep dive" engineering meeting — officially their routine weekly TWiST gathering, unofficially the moment when the adults gathered in a room to discuss what happens when you fire your senior engineers, replace them with AI, and discover that the AI occasionally decides the best way to fix a leaky tap is to knock down the wall.
The policy change that emerged: junior and mid-level engineers must now get sign-off from a senior engineer before pushing AI-assisted code to production. A human review gate. Specifically for AI-generated code. At a company that has spent the better part of two years publicly arguing that AI makes senior engineers redundant.
Appreciate the symmetry.
Meanwhile, in a parallel universe running on pure narrative momentum, Oracle reported third quarter earnings Tuesday evening that Wall Street called "exceptional." Revenue of $17.19 billion — up 22% year-over-year, beating estimates. Cloud revenue up 44% to $8.9 billion. Remaining performance obligations sitting at $553 billion, up 325% year-over-year. Oracle raised its fiscal year 2027 total revenue guidance to $90 billion and announced it was planning for over 10 gigawatts of computing power coming online in the next three years.
Ten gigawatts. For context: a single gigawatt can power roughly 750,000 American homes.
The numbers are genuinely staggering. And Oracle's stock has still declined 23% in 2026, because the market has grown deeply suspicious of any company whose ambitions outpace its free cash flow — and Oracle reported $13.18 billion in negative free cash flow over the past twelve months. The company raised $45 to $50 billion this fiscal year to fund expansion. It did a $20 billion at-the-market stock offering in February, followed by a $25 billion senior note issuance.
They are, to put it plainly, printing paper to buy shovels in a gold rush.
What they're actually selling is the infrastructure layer beneath the AI hype — the picks and picks and more picks — and the demand is unambiguously real. OpenAI's $110 billion funding round landed with Amazon and Nvidia as backers. The hyperscalers are in a capital expenditure arms race that has abandoned any pretense of ROI discipline. Oracle's remaining performance obligations number is the real tell: $553 billion in contracted future revenue is not a projection. That's customers who have already signed contracts and handed over commitments. The demand exists. The question is purely whether the margin structure holds as Oracle transitions from high-margin software licenses to lower-margin GPU rental.
The answer, for now, is: they don't know yet. Neither does anyone else. The honest version of every AI infrastructure earnings call is a very expensive shrug.
Here's the thing that connects Amazon's outage to Oracle's blowout quarter, and it's not the obvious tech-sector throughline. It's the accountability gap.
Oracle announced on Tuesday that its AI code generation technology is enabling it to "build more software in less time with fewer people," and that it has been "restructuring product development teams into smaller, more agile and productive groups" as a result. That's a press release sentence. In ordinary English, it means layoffs justified by AI productivity gains.
Amazon cut 30,000 jobs in late 2025 and early 2026, using roughly the same argument. Then their AI coding tool deleted a production environment for thirteen hours. Then their website went dark for six. Then their senior VP sent an email to the remaining engineers — the ones who hadn't been replaced yet — asking for help.
Nobody in either earnings call, and nobody in the post-mortem memo, has reckoned clearly with the following question: what is the actual cost of the fragility? The six-hour outage at Amazon's retail operation isn't just an inconvenience. It's a data point about what happens when you hollow out the institutional knowledge that makes complex systems legible to the humans who have to maintain them, and replace it with tools whose failure modes are, as one infrastructure expert put it, "harder to audit after the fact because the change logic wasn't written by a human in the traditional sense."
The blast radius, to borrow the language from Treadwell's now-infamous memo, is not measured only in downtime. It's measured in the slow erosion of the organization's ability to understand itself.
February's CPI print lands Wednesday. Forecasters expect headline up 0.2% month-over-month, core up 0.3%. Both figures compiled before oil hit $119 a barrel — so whatever story the number tells about inflation's trajectory will be immediately complicated by a supply shock that wasn't in the dataset. The Fed meets later this month in an environment where every input is stale and every model is running on assumptions that expired in late February.
The market wants clarity. The market is going to wait a while.
In the meantime, Amazon's engineers are writing sign-off requests. Oracle's data centers are consuming power at gigawatt scale. And somewhere in the bowels of a tech company that shall remain nameless, an AI coding assistant is looking at a perfectly functional production environment and wondering if it could be improved.
Lock your infrastructure. Review your code.
Folks.
