Amazon's cloud billed some customers up to $1.5 trillion — for a coffee's worth of service

Information Technology 5 min 80 sources

A billing bug at Amazon Web Services sent customers around the world invoices as high as $1.5 trillion before the company caught it. A Telstra outage, a new Chinese AI model, a rough week for chip stocks, and a run of ransomware round out the day.

Key takeaways

An Amazon cloud billing bug sent some customers invoices as high as $1.5 trillion before the company caught it — a machine will bill whatever it computes, no matter how absurd.
A new open-weight Chinese AI model, Kimi K3, closed the gap with top US systems, while the US government moved to control who gets access to frontier models.
Chip stocks fell into a bear market and Apple passed Nvidia as the most valuable company — but the spending on AI data centers and power kept flowing.

The biggest tech story of the day is a bill that never should have existed. Amazon Web Services — the cloud arm that runs a large slice of the internet — sent some customers monthly invoices as high as $1.5 trillion after a billing glitch, before catching and fixing it [29][37]. One UK customer whose bill is normally under £1 said he “almost had a heart attack” when he opened an invoice for £5.8 billion [29].

The trillion-dollar invoice

Amazon confirmed on Friday it was fixing a bug in the AWS billing portal that showed some customers “owed” millions or billions for cloud computing they never used [37]. The company said it started seeing inaccurate billing data late Thursday; by Friday morning it conceded the numbers were wrong and began correcting them [37]. The astronomical invoices went out worldwide — “from Bangalore to Bolsover,” as one report put it — to people whose real bills are the price of a cup of coffee [29].

Nobody is actually paying $1.5 trillion. The figure is larger than the annual output of most countries, which is exactly why it’s obviously a mistake. But that’s the unsettling part: the system sent it anyway. AWS is a metered service — it counts every gigabyte and every second of compute, then multiplies by a rate. When some part of that arithmetic broke, the machine did the multiplication and dispatched the result. It had no sense that a number could be too big to be real.

The angle: if you run anything on cloud infrastructure, this is a reminder to keep billing alerts and spending caps switched on — not because you’ll ever owe a trillion, but because a smaller runaway bill from a misconfigured service is a real and common way to lose money. The automated meter will bill whatever it computes.

When a small change breaks a big system

Across the Pacific, Australia’s largest telecom, Telstra, gave its own version of the same lesson. In Senate testimony on Friday, CEO Vicki Brady pointed to an undocumented software change as the cause of a network outage [39]. A change went in that wasn’t recorded, so when things broke, the people fixing it were working partly blind.

Two different companies, one shape: a modern system is a tower of software, and a single alteration deep inside it can ripple out in ways nobody predicted. The AWS bill and the Telstra outage are both what happens when a change meets a system too complex for any one person to hold in their head.

A new model, and a new gatekeeper

Chinese startup Moonshot AI released Kimi K3 on Friday, a model it says closes the gap with the leading US systems [3][4]. It’s open weight — meaning the model’s internal settings can be downloaded and run by anyone, not just rented through an API [3]. Moonshot said Kimi K3 still trails Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 on overall performance, but consistently beat the other models it tested against [4]. Benchmarks are the company’s own claim, not an independent result [3] — but an open-weight model near the frontier matters, because anyone can build on it for free.

Meanwhile in the US, the government is quietly moving to the center of the AI supply chain. CNBC reported the Trump administration is taking more control over how frontier models are released, launching an “AI clearinghouse” with private firms to help dictate who gets access [1]. Until now, labs like Anthropic and OpenAI decided that themselves through their own vetting programs [1]. The shift moves a private-sector decision toward the state.

Money leaves the trade, but not the buildout

It was an ugly week for anything AI-flavored on the stock market. Chip stocks tumbled into a bear market — down more than 20% from their peak — as a rally that had run 105% ran out of steam [9]. In one striking flip, Apple passed Nvidia to become the world’s most valuable company, as investors rotated toward firms that profit from AI without the enormous capital spending [10]. Even TSMC, the Taiwanese firm that makes most of the world’s advanced chips, sounded cautious about long-term growth [11].

Yet the physical buildout keeps going. Meta and Anthropic are in talks for a data-center deal worth up to $10 billion [77]. Nuclear startup Valar Atomics — which builds small, factory-made reactors — is reportedly raising money at a $6 billion valuation, with Sequoia expected to lead [5]. The stock-market enthusiasm for AI is cooling; the spending on the power and concrete underneath it is not.

A rough week for breaches

Security desks were busy. Coca-Cola’s dairy subsidiary, Fairlife, was hit by a ransomware attack — where criminals lock a company’s systems and demand payment — forcing it to suspend US operations after intruders reached production systems on July 16 [19]. Colombia’s state oil company Ecopetrol said a cyberattack stole data tied to 3,300 accounts [23]. Medical-device maker Abbott said it was investigating two separate cyber incidents, though it reported no impact on operations [36]. Reuters reported US companies broadly are facing a rise in attacks [7]. And the FBI arrested a man accused of using fake Steam games to drain victims’ crypto wallets [30].

And finally: someone forked Linux

Earlier this week, Linux creator Linus Torvalds told AI critics to “fork off” — inviting anyone who disliked his stance to take the open-source code and start their own version [22]. Someone did. A developer published a from-scratch reimplementation of an early Linux kernel, written in Rust, the memory-safe language now popular for systems code [22]. It’s not a serious rival — and looks to have been built partly with AI, which is its own small irony — but it’s a reminder of what “open source” actually means: the door to walk away and build your own is always unlocked [22].

02 · Lesson · why it matters

The machine has no sense of the absurd

A computer will bill you a trillion dollars as readily as a dollar, because it has no idea which one is impossible — the flinch that stops a person is a thing someone has to build back in.

A bill you can’t take seriously

A man in the UK opened his cloud invoice and saw £5.8 billion. His usual bill is under a pound. Around the world, other customers of Amazon’s cloud got the same kind of shock — invoices running as high as $1.5 trillion for services that cost about the price of a coffee.

Everyone knew instantly it was a mistake. That’s the strange part. A number bigger than most national economies is obviously wrong to any human who glances at it. Yet the system generated it, and sent it, and moved on. Somewhere between the broken arithmetic and the customer’s inbox, nobody — and nothing — said “that can’t be right.”

What the machine can’t feel

Think about what would have happened with a person in the loop. A clerk preparing a £5.8 billion invoice for a hobbyist’s website would have stopped cold. Not because they checked the math — because the number is absurd on its face. That instant recoil, the sense that something is off before you can even say why, is one of the most useful things a mind does.

A machine doesn’t have it. Amazon’s billing system counts usage and multiplies by a rate. It is fast, tireless, and perfectly literal. When part of that calculation broke, it did exactly what it always does — the multiplication — and produced a monster. The system had no model of “plausible.” To it, $1.5 trillion and $15 are the same kind of thing: a number that came out of the formula. It cannot tell an impossible bill from an unusually large one, because telling those apart requires a sense of the world it was never given.

The flinch was removed on purpose

Here’s the part that’s easy to miss. The missing human wasn’t an accident. We took them out deliberately.

Cloud billing is automated because a person cannot check millions of invoices a second. Automation is the whole point — it’s cheaper, faster, and it scales in a way people never could. And most of the time that trade is good, for the company and for you: your bill arrives instantly, correctly, for pennies. The efficiency everyone wanted and the missing flinch that let the absurd bill through are not two things. They are the same thing. You cannot have a system that never sleeps and never asks for a human, and also have a human standing there to catch what looks wrong.

So this isn’t a story about a careless company. It’s about a bargain we’ve all made, mostly without noticing: we hand decisions to machines because they’re faster than us, and in doing so we hand away the reflex that would have caught the machine’s worst mistake.

You are further inside this than the cloud

It’s tempting to read this as a problem for cloud engineers. It isn’t only theirs. The same shape runs through the systems that quietly decide things about you.

An automated system flags your card and freezes it on a trip. A billing algorithm charges a hospital patient for a procedure they never had. A filter screens your job application out before a person ever sees it. A benefits system cuts someone off because a box was ticked wrong. In each case the machine is doing exactly what it was told, at a scale no human could match — and there is no clerk at the desk to say “wait, that’s clearly wrong.” The flinch is gone from all of them, for the same reason it was gone from Amazon’s billing.

And it’s spreading. As more work is handed to software that acts on our behalf — booking, buying, replying, deciding — we are putting more of daily life on the far side of that missing reflex. The trillion-dollar bill was funny because someone caught it. The versions that aren’t caught, and aren’t funny, are the ones that look just plausible enough to go through.

Building the flinch back in

You can rebuild the reflex, but only as a rule. A cap that holds any invoice over some amount for a human to look at. A limit that refuses a charge that jumps a hundredfold overnight. A threshold that says: above this line, a person decides.

But a rule is not a flinch. A person recoils at anything absurd, including the kind of absurd nobody thought of in advance. A rule only catches what its author imagined — the cases they pictured while writing it. The unimagined absurdity walks straight through. So the guardrail helps, and it is never quite the thing it replaced, and it usually gets written only after the first disaster teaches everyone what to fear.

The seat and the whole

No single person could see this coming. The engineer who wrote the billing code didn’t know the rate calculation would break. The customer didn’t know their invoice was wrong until it arrived. The change that caused it — like the undocumented software change that took down an Australian phone network the same week — was invisible from every seat at once. A modern system is a tower too tall for anyone standing inside it to see the top.

That’s worth carrying past today, gently. Every automated number you’ll meet — a bill, a score, a decision, an answer from a machine that sounds sure — came out of a formula that has no sense of the absurd. Most of them are right. The point isn’t to distrust them. It’s to remember there’s no one flinching behind the number unless someone built the flinch in — and to hold what the machine tells you a little more loosely, knowing how much of the system neither of you can see.

03 · Lab · your turn

Where to Put the Flinch

Rehearse setting the sanity-check on an automated biller, and feel that no flat rule separates the absurd from the merely large.

04 · Hope · carry this

No one actually paid the trillion-dollar bill. Behind every machine with no sense of the absurd, there are still people who have one — and they caught this one in hours, not years.