AGI: The eye of the needle is physics

The race to develop the next generation of AI will be decided in datacentres, power grids and supply chains

3D Render Of Power Transmission Lines with 3D Digital Visualization of Electricity. Fantastic Visuals of Night Sky Full of Bright Stars. Concept of Renewable Green Energy Powering Human Progress.

Leaked documents reveal Amazon’s plans to rebuild the datacentre landscape for the next evolution of artificial intelligence.

As Business Insider reports, Amazon's internal planning documents reveal a radical infrastructure manoeuvre: in the rural US state of Indiana, the tech giant is rapidly rebuilding its massive datacentre landscape under the internal code name "AGI Pivot."

The plans illustrate how drastically the race to develop Artificial General Intelligence (AGI) has changed from a purely algorithmic challenge to a gigantic battle of materials and engineers.

The conversion in Indiana: Amazon's "AGI SuperCluster"

According to the confidential documents, Amazon is consolidating several independent buildings of the Indiana campus into a single, highly integrated computing architecture – the "AGI SuperCluster."

The site – originally built for AI developer Anthropic (Project Rainier) – will be redesigned from scratch. Fibre optic and network connections as well as storage systems are being redeveloped in such a way that physically separate halls can act like a single gigantic supercomputer. However, the new AGI initiative is to run on separate infrastructure and will not affect the existing Rainier servers.

The key points:

Hardware upgrades to Trainium 3: Amazon is updating some older Trainium 2 systems to the newer in-house development Trainium 3.
Acute capacity expansion: The documents show an emergent request for the delivery of more than 6,000 new Trainium servers.
Architectural annex structure: To avoid bottlenecks in data exchange, existing buildings are converted into "annexes" (extensions). They access the central network hardware of neighbouring datacentres directly instead of working in isolation.
Accelerated schedule: The commissioning was brought forward by several weeks. Amazon's AGI division urgently needs the massive computing power to be able to present the next AI flagship model at the next re:Invent – at least, that’s the intent.

The AGI pivot follows a phase of internal restructuring: despite recent layoffs in the AGI division and its gradual exit from the previous Nova model series, Amazon is doubling its stakes in model development.

In view of ongoing bottlenecks in Nvidia GPUs, the company is consistently relying on its own silicon to handle the enormous data throughput for multimodal AI training.

Amazon's move in Indiana reflects a wider shift across Big Tech. Amazon, Microsoft, Alphabet and Meta have sharply increased infrastructure capex, with 2026 investment potentially heading towards $760 billion - much of it aimed at AI datacentres, specialised processors and fibre-optic networks.

Several key market signals underpin this structural change:

1. Emancipation from the chip monopoly (custom silicon)

In order to reduce costs and meet frontier models’ extreme bandwidth requirements, the giants are reducing their dependence on third-party providers. In addition to Amazon's expansion of Trainium 3, Google is scaling its in-house purpose-built Tensor Processing Units TPUv8x ("Zebrafish") and TPUv8ax ("Sunfish"), while Microsoft is integrating its own Maia chips into the datacentre fleet.

2. Multi-gigawatt sites and nuclear energy

The scale of AGI clusters pushes the limits of existing power grids. The energy demand demands a direct, base-load supply:

Microsoft signed a historic 20-year contract with Constellation Energy to exclusively restart the decommissioned Unit 1 reactor of the Three Mile Island nuclear power plant for its own AI division.
Google partnered with Kairos Power to secure capacity from small modular reactors (SMRs) to cover up to 500 megawatts of power-intensive AI workloads.
Meta is planning major projects such as the 5-gigawatt "Hyperion" campus in Louisiana, as part of a multi-year infrastructure initiative in the multi-digit billion range.
xAI/SpaceX: SpaceX/xAI is relying on its own gas-fired power plants to ramp up the clusters for Grok.

3. Personnel

Demis Hassabis’ move from operational leader of Google DeepMind to Alphabet chief scientist is another signal that AGI is being treated as a long-term strategic contest. The race will not only be decided by datacentres, chips and power contracts, but by research culture, governance and leadership focus.

View to the east

The same infrastructure-driven AGI shift is visible in China, where DeepSeek founder Liang Wenfeng has said the company prioritises long-term AGI development over profit.

Driven by US export restrictions on high-performance chips (such as Nvidia's H100 or Blackwell architectures), Chinese tech giants and the state are pursuing their own aggressive strategy: they are compensating for the efficiency deficit of individual processors through massive physical scaling, novel network architectures and state-controlled megaprojects.

Beijing has launched a multi-year AI infrastructure strategy valued at around $295 billion, aiming to move to a largely self-sufficient hardware base. Hyperscalers such as Baidu, ByteDance, Tencent and Alibaba are expected to rely on domestic accelerators – especially Huawei's Ascend series – for up to 80% of major new projects.

Since individual Chinese chips still lag behind the absolute US leaders in memory bandwidth and manufacturing process, the focus of innovation is shifting to system design:

System architecture instead of single-chip crown: With approaches such as Huawei's CloudMatrix, hundreds of Ascend processors (such as the Ascend 910C) are interconnected via specialised optical networks to form tightly coupled arrays. These mega-clusters compensate for disadvantages in power consumption or per-chip performance through sheer mass and optimised data throughput.
The breakthrough in model training: For a long time, Chinese chips were only considered competitive for the execution (inference) of finished models. However, recent breakthroughs prove that China can conduct even the most computationally intensive training phases on its own hardware.
Energy advantage through state control: While US companies face lengthy approval procedures for grid connections, China can use direct state intervention to shift energy-intensive computing to western and northwestern hubs under its East Data, West Computing initiative. These datacentre clusters are increasingly linked to renewable power, storage and fibre-optic networks.

The global race for the next generation of AI has left its theoretical phase behind. Whether in the USA or in Asia, the ability to merge gigawatts of energy, proprietary processor architectures and extremely scaled fibre-optic networks into a single logical computing machine has become the central power factor in the age of AGI.

This article originally appeared on our sister site, Computing Deutschland.