ai-wars

Colossus

Source-backed tracker for Colossus, GPU/power claims, and the AI infrastructure race around xAI and partner capacity.

Status: Source-backed expansion Last checked: 2026-10-19 Fresh

Colossus is where AI Wars meets physical infrastructure: chips, power, cooling, land, capital, and partner access. Readers understand chatbots. Fewer understand that capacity can decide which model companies ship next.

Core framing

Separate these claim types every time:

Dimension	Why it differs
GPU count announced	Marketing/ops claim; generation mix matters (H100 vs H200 vs GB200)
Power (MW)	Electrical capacity ≠ usable training FLOPs
Delivered vs contracted	“Agreement” is not “running at full load”
Training vs inference	Different utilization patterns
Owner / operator / tenant	xAI, SpaceX, partners may share language differently
“World’s largest”	Comparison claim — needs date + comparison set

What is source-backed from company pages

From the paired May 2026 announcements (re-check originals when updating):

Claim	Label	Posture
Colossus 1 tied to Anthropic compute access	Official company announcements	xAI + Anthropic both published
Over ~220,000 NVIDIA GPUs	Official company claim	Attribute + date; do not invent chip mix
>300 MW new capacity (Anthropic wording)	Anthropic official claim	Attribute to Anthropic
Claude product limits raised after deal	Anthropic official claim	Product ops claim
Interest in orbital AI compute	Company-stated future interest	Speculative / future — not deployed

See full product-impact detail on Anthropic / Claude compute.

Open questions (research queue)

Who controls scheduling priority across workloads?
What share is training vs inference vs reserved partner capacity?
Which capacity is online vs announced?
Power purchase, grid interconnection, and local permitting status
How “largest” claims compare to other public/private clusters on the same date

Why Muskularity owns this lane

If the site only reviews chat demos, it is a fan blog. If it tracks compute + energy + legal + product limits with receipts, it becomes a research asset.

Renting inference is a different sentence from a cluster count

On October 19, 2026, at 03:05:35 UTC, Mr. Musk replied to his own post from 21:21:41 UTC the day before:

This describes the conversation SpaceX has had when we have tried to rent inference compute from AWS, GCP or anyone at, so that we can shift load off our training clusters!!

The parent post, status 2111930742615802163, says only: “What it’s like trying to rent AI/SI compute.” It has a video this desk did not transcribe.

Label: His claim about conversations. “Training clusters” is his phrase for the capacity he says he wants to spare. It is not a GPU count, and it is not a contract with AWS or Google. The orbital design is Starmind. A direct fetch of x.com returned 403. Both posts were read through fxtwitter.

Source anchors

Muskularity note

Independent research map. Company capacity claims are not independent audits.