HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

CoreWeave CloudCoreWeave—CoreWeave CloudCoreWeave—

CoreWeave Cloud

by CoreWeaveUS

ScalingStage 5 of 5

Aims for more than 1.85 GW of active power by the end of 2026.

Updated 11 Aug 2026Checked 25 Sep0 updates this week

Milestones

Next · More than 1.85 GW active power
  1. First GB200 NVL72 cloud instances generally availableFeb 2025Complete.
  2. More than 1.85 GW active powerTarget Dec 2026Current milestone.

Most important updates

  • 11 Aug 2026
  • 25 Sep 2025
  • Sep 2025
  • 28 Mar 2025
  • Mar 2025

Upcoming

  1. Dec 2026More than 1.85 GW active powerin about 7 weeks (next)

Current obstacles

  • Leverage and customer concentrationGrowth is funded with heavy debt, backed by contracts from a handful of AI customers.
  • Power and construction delaysDelays at outside data-center builders have pushed back capacity and revenue.

Physics limits

  • Light is too slow for spread-out clustersTraining jobs synchronise tens of thousands of GPUs many times a second. Light in fibre takes about 5 µs per km, so one training cluster must sit on one campus, and the slowest link sets everyone's pace.
  • Failures scale with sizeWith 100,000 GPUs, some part fails every few hours. Jobs must save checkpoints and restart, so reliability, not peak speed, often limits useful throughput.
  • All the power becomes heatAn AI rack draws over 100 kW, and nearly all of it ends up as heat that liquid cooling must carry away. Available power and cooling, not chips, often cap how fast capacity grows.

How it works

3 parts
Rows of NVIDIA-GPU racks in Summit, an Oak Ridge supercomputer, not a CoreWeave site: the kind of hall a GPU cloud rents out by the machine
Rows of NVIDIA-GPU racks in Summit, an Oak Ridge supercomputer, not a CoreWeave site: the kind of hall a GPU cloud rents out by the machinePhoto: OLCF at ORNL · CC BY 2.0 (opens commons.wikimedia.org)
Hardware

Newest NVIDIA racks first

CoreWeave builds clusters of the latest NVIDIA systems (GB200, GB300) with InfiniBand networks, and is often first to offer them.

Software

Managed bare metal

Customers get near-bare-metal performance run through Kubernetes, with tools that detect failing GPUs and swap them out automatically.

Contracts

Contracts fund the GPUs

Multi-year contracts with customers such as Microsoft, OpenAI and Meta serve as collateral for loans used to buy GPUs and build sites.

Update log

6 updates

Tue 11 Aug

  • Minor: PressCommercial

Thu 30 Oct 2025

  • Minor: FilingOther

Thu 25 Sep 2025

  • Major: BlogCommercial

Sep 2025

  • Major: FilingCommercial

Fri 28 Mar 2025

  • Major: BlogFunding

Mar 2025

  • Major: BlogCommercial

About CoreWeave

The team behind CoreWeave Cloud

CoreWeave

GPU cloud, AI data centers

CoreWeave is the largest specialist GPU cloud, renting NVIDIA GPU clusters to AI labs and big tech. It started as a crypto miner in 2017 and listed on Nasdaq in March 2025.

  • Order backlog about $104B at end of June 2026. 2026 revenue guidance is $12.4-13.2B.
  • Big customers include OpenAI, Microsoft and Meta.
  • Plans $35-39B of 2026 spending, funded largely with debt.

coreweave.com (opens coreweave.com)Company profile →

Founded
20179 yrs
Headquarters
United States
Status
Public
Listed
CRWV (opens google.com)NASDAQ
Valuation
$48Bmarket capAug 2026
Last round
IPO · $1.5BMar 2025
Works in
ComputingData centers & cloudAI
Coverage
1 program · 6 updateslatest 11 Aug 2026checked 25 Sep
Customers
OpenAIMeta
Suppliers
NVIDIA
People
Michael IntratorCo-founder and CEO
Primer