Is "Digital AGI" Dead? Amazon's Internal Restructuring Tears Off the Last Fig Leaf of the Pure Algorithm School

[Introduction] Rejecting mediocre dialog boxes, Amazon is tearing apart the dimensional wall of AI with its own hands! "Engineering Tsar" DeSantis joins forces with "Robot Master" Abbeel to brutally stitch together chips, models, and steel shells. This isn't just about updating software; it's about forging the physical form of AGI! The giant has awakened, the closed-loop is complete, and the era of physical AI has officially arrived!

For a long time, the public's imagination of AGI has been compressed into a dialog box.

Just yesterday, Amazon CEO Andy Jassy announced a rare organizational adjustment: integrating the Nova large model team, the self-developed chip department, and quantum computing research into the same system.

Almost simultaneously, Pieter Abbeel, a representative figure in the field of robotics and reinforcement learning, officially joined Amazon's AGI system.

Rather than continuing to stack parameters in the cloud, Amazon has chosen to incorporate computing power, models, and the physical world into the same blueprint.

In Amazon's view, AGI shouldn't just be a chatty intellectual but a "physical entity" with chips as its bones and algorithms as its soul.

Computing power is no longer a resource; power is

Amazon has chosen veteran Peter DeSantis to oversee this round of integration.

DeSantis is well-known within the company for his strong engineering delivery capabilities. He is not only the helmsman of cloud infrastructure but also the mastermind behind Amazon's self-developed chip plan.

This appointment means that AGI is no longer just a research topic for the model team but is regarded as a systematic project that requires long-term engineering investment.

In the eyes of the outside world, forcibly integrating the AGI team, the self-developed chip department, and quantum computing seems bloated. However, in Amazon's view, this is a decision related to computing power sovereignty.

In the second half of the AGI competition, the lead in algorithms often only has a window period of a few months, while the lead in cost is a permanent moat.

For a long time, Silicon Valley giants have been deeply troubled by Nvidia's production capacity. Amazon is trying to break this passive situation through vertical integration.

Its self-developed Trainium 2 series of chips has already demonstrated its ambition in training performance and energy efficiency.

The combination of the AGI team and the chip team enables every line of underlying code of the Nova model to achieve "pixel-level" optimization on self-developed chips.

This deep adaptation is like Apple's unique M series of chips, allowing Amazon to drive the complex Amazon Nova model at a much lower cost than its competitors.

For Amazon, computing power is not just a resource issue but a strategic boundary.

Pieter Abbeel and the next form of AGI

If the architectural restructuring is the "skeleton," then Pieter Abbeel's joining is like injecting the "soul" into this body.

In the past two years, the improvement of large model capabilities has increasingly relied on scale expansion rather than paradigm shifts.

The reason why Amazon invited him with a large sum of money is that the AGI competition is facing a huge hidden bottleneck: the high-quality text data on the Internet is almost used up.

Pieter Abbeel is one of the key figures in the fields of robot learning and reinforcement learning.

As a professor at UC Berkeley, he has long-term academic accumulation in the directions of robot learning and reinforcement learning and has trained many important figures who influence contemporary large model research.

While most models are still spinning in the virtual world of characters, Abbeel brings the key to the "real world."

Abbeel believes that the ultimate evolution of AI must be achieved through real-time interaction with the physical world. This is the so-called "embodied intelligence."

Coincidentally, Amazon has the world's largest logistics warehousing and thousands of Proteus autonomous mobile robots.

Different from research robots in the laboratory, these systems operate in a highly standardized but extremely large-scale real environment.

For Abbeel, these are perfect samples for collecting physical law data.

In his research framework, the ultimate form of AGI is not limited to the screen.

In Amazon's plan, AGI shouldn't just be a secretary who polishes emails but a "super labor force" that can understand vague instructions, sense gravity and friction, and accurately grab items in a chaotic warehouse or even a home environment.

When the logical reasoning of the Nova model is combined with Abbeel's physical manipulation technology, AGI will evolve into an "omnipotent craftsman."

This closed-loop from the cloud to the fingertips gives the AGI that Amazon is defining an unprecedented sense of solidity.

It no longer floats in the digital world but takes root in steel and algorithms.

While others are still competing on models, Amazon has switched to a different track

Looking at the larger industry landscape, Amazon's choice is particularly different.

One is the "cloud intelligence" school represented by Microsoft and OpenAI. They have the strongest logical deduction ability but are always trapped behind the screen and lack the ability to interact with the physical world.

The other is the "all-rounder" school represented by Google. Although it has the technological accumulation of DeepMind, it is troubled by the fragmentation of its internal organization and has difficulty achieving a real closed-loop between cutting-edge algorithms and hardware manufacturing.

What Amazon is taking is the third path: forcing the model with scenarios and reshaping the ecosystem with a closed-loop.

The premise of this path is whether there are enough, real, and long-term operating physical scenarios.

Compared with pure software giants like OpenAI, Amazon's greatest advantage lies in that it has more than 750,000 industrial robots globally.

In Silicon Valley, training data often needs to be purchased and cleaned. In Amazon, the physical world itself is a continuously operating data source.

While opponents are still using posts on Reddit to teach AI to understand the world, Amazon can already use the data of gravity, resistance, and friction in the real world through its global automation centers to train an AGI that truly understands physical laws.

This path shows Amazon's commercial arrogance and is also its ultimate confidence.

If AGI must ultimately serve the real economy, then those who master the physical scenarios will master the power to make rules.

Just the high threshold of "heavy assets" competition puts startups that simply rely on third-party computing power and lack physical implementation scenarios at a disadvantage in this long-distance race.

AGI is entering the era of large-scale industry

The appointments of Pieter Abbeel and Peter DeSantis are not just ordinary personnel changes but also foreshadow the power shift that will occur in the AI industry.

For a long time, the progress of AI has heavily relied on individual heroism and algorithmic ingenuity in the laboratory.

However, with the popularization of industrial-grade models like Amazon Nova and the demand for large-scale computing power bases like Trainium 2, the threshold of the AI competition has undergone a qualitative change.

We are witnessing the end of the "laboratory era" and the beginning of the "era of large-scale industry."

In the future, AGI will only belong to a very small number of giants that can achieve a "full-stack closed-loop." This closed-loop is not only a closed-loop of code but also a physical closed-loop of "computing power - brain - body."

Amazon's bold bet defines a new and harsh reality: if the second half of AGI is embodied intelligence, then those who have the cheapest computing power sovereignty and the most real physical interaction data will hold the ticket to the future.

From this moment on, the competition of AGI is no longer a game that all companies can participate in.

Rather than repeatedly discussing whether AI will replace a specific job, it's better to focus on another thing:

When AGI puts on steel armor and walks out of the laboratory, it will completely reshape the relationship between humans and the physical world.

And that day is approaching rapidly with Amazon's major power restructuring.

Reference materials:  

https://www.bloomberg.com/news/articles/2025-12-17/amazon-names-new-ai-chief-amid-battle-to-take-on-tech-rivals?srnd=phx-technology 

https://x.com/bookwormengr/status/2001353147055509881https://www.cnbc.com/2025/12/17/amazon-ai-chief-prasad-leaving-peter-desantis-agi-group.htmlhttps://www.reuters.com/technology/amazon-taps-veteran-peter-desantis-lead-ai-chip-quantum-push-2025-12-17/https://www.businessinsider.com/amazon-agi-executive-leaving-ai-models-rohit-prasad-2025-12 

This article is from the WeChat official account "New Intelligence Yuan", edited by Qingqing. It is published by 36Kr with authorization.