Amazon Internal Documents Reveal Major Restructuring of Indiana AI Data Center
Claire Weston
Amazon is merging multiple data centers in Indiana into a single "AGI SuperCluster" to train its next-generation frontier AI model — a move that fuses its custom-chip and in-house-model efforts into one push aimed squarely at Google's and Microsoft's large-scale training capabilities.
What is the "AGI SuperCluster," and why merge several buildings into one system?
Amazon is consolidating multiple standalone data centers on its Indiana campus into one unified compute system, internally dubbed the "AGI SuperCluster."
This means → the buildings used to operate independently; now Amazon is re-engineering the networking, storage, and fiber infrastructure so thousands of servers work together like a single giant machine.
Some existing buildings will become "annexes" — no longer self-contained, but sharing core networking gear with adjacent data centers.
How urgent is this — and why the year-end deadline?
Internal documents include an "urgent request": deploy more than 6,000 Trainium-powered AI servers, with the timeline pulled forward by several weeks.
The target milestone is Amazon's re:Invent conference in early December — the company wants its next-generation frontier AI model ready for training by then.
This means → re:Invent is both a product-launch window and a public showcase of AI capability; showing up without a new model would put Amazon on the back foot in the AI race narrative.
Custom-chip upgrade — what does swapping Trainium 2 for Trainium 3 mean?
Documents show some facilities plan to replace older Trainium 2 servers with newer Trainium 3 servers.
In plain terms = Trainium is Amazon's in-house AI training chip — its homegrown alternative to Nvidia GPUs. Moving from gen 2 to gen 3 means faster silicon running larger models.
Amazon disclosed last week that its custom-chip business (Trainium AI chips + Graviton processors) is on track for annualized revenue above $25 billion, up from a $20 billion forecast the prior quarter.
Anthropic's project sits on the same campus — will they compete for resources?
The Indiana campus also hosts Project Rainforest — Amazon's Trainium-powered AI supercomputer built primarily for Anthropic.
Amazon stated explicitly: the new AGI SuperCluster will not draw on Project Rainforest's existing servers; the two tracks run in parallel.
This reflects Amazon's dual bet in AI: building its own frontier model while backing external ally Anthropic — with no intention of scaling back either side.
Where does the money come from — how does $220 billion in capex account for this?
Amazon recently raised its 2026 capital-expenditure outlook from $200 billion to $220 billion; this restructuring is part of that spend.
CEO Andy Jassy's logic: data centers last decades, while servers and networking gear can be upgraded over time — so heavy upfront investment yields attractive long-term returns.
Documents also show the expansion will add storage capacity for multimodal AI training — training that processes text, images, and video simultaneously — which requires handling large volumes of high-resolution images and saving model checkpoints frequently.
The AGI org just had layoffs — has this plan really not slowed down?
The project falls under a broader internal initiative called "AGI Pivot," led by Amazon's AGI organization.
People familiar with the matter say that despite recent layoffs in the AGI org — and the discontinuation of the former Nova model series — the frontier-model push has "not slowed down."
This means → Amazon cut the model lines it judged to be dead ends, but concentrated resources on next-generation infrastructure. Layoffs and doubling down are happening at the same time — the direction is narrowing, not shrinking.
Content is for reference only, not financial advice.