Two things happened at Amazon in the last ten days, and the company would very much prefer you only noticed one of them.
The first: on July 30, Amazon reported that AWS grew 37% to $42.2 billion in the second quarter — its fastest cloud growth since late 2021 — and raised its 2026 capital expenditure plan to $220 billion, up from the $200 billion it had already guided to. Total net sales hit $200.6 billion. Operating income rose 43%. The stock jumped in after-hours trading.
The second: in roughly that same stretch of days, Amazon confirmed layoffs inside its AGI organisation, moved four of its flagship Nova AI models into what employees call "KTLO" mode, and confirmed it is closing the San Francisco research site it had built its artificial-general-intelligence ambitions around.
Those are not contradictory results. Read together, they are the clearest statement yet of what Amazon has decided it actually is in the AI economy — and what it has decided to stop pretending to be.
Which Models, and What Actually Changes
According to reporting from Business Insider and Bloomberg, four in-house models have been moved to that footing: Nova Premier, Amazon's high-end flagship; Nova Omni, its multimodal model; Reel, its video generator; and Canvas, its image generator.
If you are an AWS customer with one of those wired into production, nothing breaks this week. They remain available. What ends is the roadmap.
What Amazon has not done is exit the model business, and coverage suggesting otherwise is overstating it. The Nova 2 generation — Nova 2 Lite, the cheap reasoning model, and Nova 2 Sonic, the speech-to-speech model, both generally available on Bedrock since December — continues. So do Nova Forge, the six-figure programme that hands enterprises pre-trained and mid-trained Nova checkpoints to customise, and Nova Act, the browser-automation agent platform.
| Model | What it was for | Status |
|---|---|---|
| Nova Premier | Flagship high-end reasoning model | KTLO — runs, no longer developed |
| Nova Omni | Multimodal across text, image, audio | KTLO — runs, no longer developed |
| Reel | Video generation | KTLO — runs, no longer developed |
| Canvas | Image generation | KTLO — runs, no longer developed |
| Nova 2 Lite | Fast, low-cost reasoning | Continues on Bedrock |
| Nova 2 Sonic | Speech-to-speech conversational AI | Continues on Bedrock |
| Nova Forge | Customisable model checkpoints for enterprises | Continues |
| Nova Act | Browser-automation agents | Continues |
The pattern is not hard to read. The models Amazon is keeping are the ones that make AWS stickier — cheap, fast, embedded in someone's workflow. The models it is putting down are the ones that were supposed to compete head-on with OpenAI, Anthropic and Google on raw capability.
The Lab That Didn't Work Out
The organisational story is blunter than the product story.
Amazon's AGI Lab was founded in December 2024, built around several dozen employees the company brought over from the startup Adept — including Adept's co-founder and chief executive, David Luan. At its peak the team reached roughly 80 people, according to The Information. More than a dozen of those Adept hires have since left. Luan departed in February.
Above him, Rohit Prasad — the head of Amazon's AGI organisation, and before that the executive behind Alexa — left the company late last year.
This week Amazon confirmed it is closing the San Francisco site that housed that work. It confirmed the layoffs on July 22, without giving a number.
An acquisition, a lab, a founder, an executive, an office. Eighteen months, and all of it has now been unwound.
Every clause of that is true. It is also the standard shape of a statement issued when a strategy has not worked and the company has decided to stop funding it.
One Bet Instead of Four
What survives is a group called Frontier Model Research, led by Pieter Abbeel — the UC Berkeley professor who came to Amazon in 2024 when it licensed the technology and hired the team from Covariant, the robotics company he co-founded.
Abbeel's mandate is a single flagship foundation model, expected to debut at Amazon's re:Invent conference this autumn. Instead of separate specialised systems for text, image, audio and video, one frontier-scale model meant to stand against the best that OpenAI, Anthropic and Google can field. Reporting suggests it may still carry the Nova name.
Concentration is a defensible response to Amazon's actual constraint. The company has enormous capital and a finite supply of frontier-calibre researchers, and it was spreading them across four product lines that were each losing to a competitor. Consolidating scarce talent and scarce priority onto one shot is what you do when the diversified portfolio has already failed.
It is also a genuine bet. There is no fallback flagship if the re:Invent model lands short. That is the entire point of the reorganisation — and the entire risk of it.
The Landlord's Arithmetic
Here is the part that makes the whole thing rational rather than simply humiliating.
Amazon does not need to win the model race to win the AI buildout. Nobody's Nova subscription is what drove AWS to a 37% growth quarter. What drove it is that everyone else's AI runs on Amazon's hardware.
Amazon has committed roughly $8 billion to Anthropic, whose models are sold through Bedrock and trained on Amazon silicon. In November it signed a reported $38 billion compute agreement with OpenAI. Bedrock's whole proposition is that it will rent you whichever frontier model you want, including ones built by Amazon's supposed competitors. On trailing twelve months, property and equipment purchases reached $169 billion, up 64% year over year.
That is a landlord's balance sheet, not a model lab's. And landlords collect rent whoever wins the race upstairs.
The uncomfortable question is whether that position is as safe as the quarter makes it look. Renting capacity to your competitors is excellent business right up until one of them has enough leverage to renegotiate, or builds its own silicon, or decides the margin you are taking is worth eliminating. Google is the counterexample Amazon has to keep looking at: it builds the chips, the models and the cloud, and it does not have to ask anyone's permission at any layer. Amazon owns two of those three and has just conceded ground on the third.
What to Take From This
- Nothing you use stops working. KTLO means maintained, not deleted. If Nova Premier, Omni, Reel or Canvas is in your stack, plan a migration on your own timeline — but there is no cliff this quarter.
- Amazon's AI business is doing fine; Amazon's AI models are not. These are separate scoreboards and this week they moved in opposite directions. Conflating them is the most common error in coverage of this story.
- Four years of frontier-model effort has been compressed into one unreleased system. re:Invent this autumn is now the single event on which Amazon's credibility as a model builder rests.
- The consolidation trend is the real signal. Building frontier models is expensive enough that a company spending $220 billion a year on infrastructure decided it could not afford to do it in four directions at once. That says something about the cost curve that no benchmark chart does.
The Takeaway
Amazon spent eighteen months and an acquisition trying to be a frontier AI lab. It hired Adept's founder, opened a San Francisco AGI site, and shipped models across text, image, audio and video. This week it closed the site, absorbed the departures, and moved four of those models into maintenance.
In the same week it posted its fastest AWS growth since 2021 and raised its infrastructure budget by another $20 billion.
Amazon has looked at both sets of numbers and drawn the obvious conclusion: it is winning decisively as the place AI gets built, and losing as a builder of AI. So it is consolidating everything that remains of the second thing into a single model due this autumn, and pouring the money into the first. Whether that reads as clear-eyed discipline or a quiet surrender depends almost entirely on what walks onstage at re:Invent.