Today we announced our Q3 earnings:
🔹$3.7B revenue, up 168% YoY
🔹$379M adjusted operating income, well above guidance
🔹Double-digit adj. operating margins, a quarter ahead of plan
🔹1.7GW+ of active power, doubling our active footprint YTD
🔹Roughly $9B in customer prepayments and deposits in this quarter
Congratulations to our team, who remain focused on executing for our customers.
Read the full Q3 2026 results. https://t.co/WLTRurYFSQ
The @NVIDIA Vera CPU is coming soon to CoreWeave, and our sandbox team already has receipts.
Diana Aung, Senior Product Manager at NVIDIA, shares how we measured more than 3x faster agent sandbox startup times than on an alternative x86 CPU.
Here is why we are excited over here at CoreWeave about the first CPU purpose-built for AI agents: https://t.co/4pyyKETyoZ
What happens when @CrowdStrike and CoreWeave join forces?
CrowdStrike CBO Daniel Bernard sits down with AITIAFY CEO Ritu Jyoti to share how the partnership creates a multiplier effect, helping enterprises adopt AI faster and more securely.
Full episode here: https://t.co/ggiOtp0E1C
Las Vegas GP is this week 🏁
I decided to build my own AI Race Engineer to help cut down on my terrible lap times in Vegas.
This is an open-source models running locally, live telemetry streamed between two machines, and every radio call traced in @CoreWeave Agent Lens to catch the answers that go wrong.
A 32B model was cold-started in just 1.1 seconds on our infrastructure.
Numerata’s Lupine engine moves live GPU state onto a new GPU, so nothing gets rebuilt.
Their Qwen3-32B cold start time on vLLM dropped from 37.76 seconds to just 1.11 seconds.
Storage usually eats that win, but Lupine built its storage tier on CoreWeave Dedicated VAST Storage, so checkpoints move over a 400 Gb/s RDMA fabric in the same data center as the GPUs.
That makes scaling an endpoint to zero practical.
Read how they did it: https://t.co/J0ReIUqFn7
"50% of our clients in the first three quarters of this year are not AI labs,” @CoreWeave CEO Mike Intrator tells @azeem.
“They are representatives from across the economy that are consuming compute because they need that to drive their business to be more effective.”
Only 35% of the 252 enterprises in a new Humatica study have a formal operating model for AI.
That's a lot of room to pull ahead.
The study, sponsored by CoreWeave, spans 14 industries and lays out five practices to turn AI ambition into lasting value: https://t.co/zS5PlRmxPq
Yes, our AI cloud is competitively priced. The performance clears too 👀
A @Signal_65 study found a token served on CoreWeave costs up to 65% less than on three hyperscalers.
On price alone, we're 22% under the most competitively priced hyperscaler.
Performance does the rest. Our GPUs run on bare metal with no hypervisor, over a 1:1 nonblocking fabric, so each one can serve more tokens.
For a team serving DeepSeek-R1 on @NVIDIA GB200 NVL72, $1 million a year buys up to 2.9x as many tokens.
See why we're built for it: https://t.co/xIGBtSz9og
A pit window can open and close in under 30 seconds, and no one can sift through 40 radio channels in that time.
Our engineers worked with @AstonMartinF1 to build AI that transcribes and sorts every channel by topic within five seconds of capture.
Clarity of communication, powered by AI.
Our Las Vegas GP livery celebrates @CoreWeave's integration across the team, including AI that makes race radio searchable.
Find out more: https://t.co/Q0nZkNqzKw
Security and speed go hand in hand when scaling AI. @CrowdStrike CBO Daniel Bernard sits down with Ritu Jyoti to unpack how the CoreWeave partnership unlocks secure computing for enterprise AI
Full episode here: https://t.co/CCoDrsnnYJ
Since May, @WaldenRobotics robots have worked full shifts beside people on a Toyota line. When one needs help, it calls a person, and that call feeds the next training run.
Building factories that learn is a human pursuit. We’re built for yours.
https://t.co/zvvMUgN1Rd
40 channels of radio chatter. One question: mediums or hards? AI Research with CoreWeave & @AstonMartinF1 reads every word on the radio in near real time, so the pit wall hears what the field is saying and turns noise into insight 👇
Pioneering the Possible, in practice.
Our Las Vegas GP livery celebrates @CoreWeave's role in connecting the CoreWeave AIR Tunnel and track data to support our engineers.
Find out more: https://t.co/Q0nZkNqzKw
Getting an AI agent working once is not the goal. Keeping it reliable in production means running the whole AI loop: tracing real failures, turning them into evals, and proving a fix actually works before you ship.
Most teams lose momentum gluing all of the tools together. We built CoreWeave Forge to keep the entire loop working together.
@AbhijayArora explains CoreWeave Forge in the video below.
Small models have a habit: they repeat whatever they just heard.
So @TheCodingSoup built a tiny town of @liquidai LFM2.5 agents, traced every line into @CoreWeave Agent Lens, and tagged the repeats.
No fine-tuning, just a harness, and the traces proved it worked: repeated lines dropped from 58% → 13%.
Best part: it's all plain OpenTelemetry from Swift. No Python or TypeScript SDK needed.
MLPerf® Training v6.1 results are out, and we're the first cloud provider to submit multi-rack @nvidia Vera Rubin results ⏱️
The benchmark times how long it takes to train a model to a target quality.
DeepSeek-V3 671B ⇢ 17.65 minutes on 256 NVIDIA Rubin GPUs, 2.7x faster than our NVIDIA GB200 NVL72 result.
Llama 3.1 8B ⇢ 7.16 minutes on one NVIDIA Vera Rubin NVL72 rack, the fastest 72 GPU submission in the round.
Every number came off the same production infrastructure available to our customers.
See for yourself here: https://t.co/zoWnxmIA8z
CoreWeave Forge's first customer was CoreWeave.
Building CoreWeave ARIA and a speculative decoding workflow, @GoldbergChen's team hit the same problem twice. Too much of the work was stitching tools together by hand.
So we built CoreWeave Forge to run the whole AI loop in one place.
Watch @CoreySandersWA demo it live here: https://t.co/P3CboyLIfQ