We covered a lot of ground with what we shipped in July. If you missed our weekly recaps on LinkedIn and X, here’s what happened:
The week of July 20: We beat baremetal interconnect speeds
We launched the ability to connect S3 buckets while working on the Lightning Cloud. We’ve optimized the path for data loading, read, and write for reliable pre-training and post-training performance.
As labs and enterprises are experimenting, smaller GPUs are essential for smoke testing ahead of longer training runs. So we’ve added T4s, L4s and RTX Pro 6000s to the Lightning Cloud.
Multi-node training is also in early access! I’m super proud of the team for not just matching, but actually exceeding baremetal interconnect speeds on the Lightning Cloud.
We’re actively onboarding labs and enterprises, so reach out to our team here.
Hugging Face’s bad day makes the case for open-weight models
I’ve been talking a bit about owning your intelligence the last few weeks since the launch of the Lightning Cloud. Separately, but still so related, an AI agent broke into Hugging Face this month and their solution was so illuminating.
The Hugging Face team had to feed a model the raw evidence to reconstruct the attack and their inteligence was throttled by OpenAI and Anthropic. Enter OSS - they fed the real commands, exploit payloads, C2 artifacts into a self-hosted GLM 5.2, an open-weight model. The model did the job. No data ever left their environment.
Their blog post is great, but this isn’t just a clever solution for our customers. It’s the only path. Enterprise security postures will never allow shipping that extent of sensitive internal information outside the org.
The week of July 13: GB300s - live from Chicago
Our team brought our first GB300 cluster live in Chicago for our friends at Reflection.ai. We’re grateful to all of our partners at Nvidia, Dell, and of course the team at Reflection for their support. The team even got a special shout out from Nvidia and Michael Dell on LinkedIn:
We have additional capacity coming online in Colorado Springs and Vancouver later this fall into early 2027, don’t hesitate to get in touch.
Thomas Chaton, from our global compute team, also broke down how they stood up the Lightning Cloud in under 3 months. I’m super proud of our team for unlocking baremetal multi-machine training performance on the Lightning Cloud, fully on demand. This is filled with lower level technical detail, learn more here.
The week of July 6: The Lightning Cloud is live
We launched the Lightning Cloud. It ships with instant H100 access, high performance storage, fully on demand. Yeah…soak that in.
Our global compute team have been cooking to get this live our customers and they’ve delivered reliability and interconnect speeds that no other managed platform can match because we own and operate every layer of the stack. I suspect that many other platform and infra businesses will be doing the same in 2026 into 2027.
Our friends at Archetype AI are already training on the Lightning Cloud and we’re proud to power their training and experimentation teams. Keep an eye out for their flagship model, Newton!
If Lightning Cloud sounds like a good fit, we’re onboarding teams now. Get in touch with our team here.
Week of June 29: Time to sock shop because our footprint just got bigger
It was a short week in the US with the 4th of July, but a productive one nonetheless.
We added another data center in Colorado Springs to our map. COS1 brings 6 MW of NVIDIA’s newest Vera Rubin platform online, built on our Data Center 3.0 blueprint, for pre- and post-training. That comes just weeks after we locked in an 8MW facility in Vancouver, Canada.
We're not fully sold out, yet. Our sales team can give you more details if you’re interested.
Also … one of Anthropic’s most capable models is now running on Lightning in a single line of code. Claude Fable 5 verifies its own work and keeps context. But you don’t have to take my word for it → Run it on Lightning now.
Until next month
-Neil