Founder @ Piris Labs - Tech / AI
San Francisco, CA
Joined September 2026
- Tweets27
- Following80
- Followers18
- Likes35
Keyvan Moghadam retweeted
Marvell’s massive acquisition of @Lumentum is another huge validation of what we're building at @PirisLabs.
As optical transceivers and interconnects demand lower power and higher bandwidth, the industry is increasingly moving in our direction
It’ll be interesting to see how this credit crunch impacts the compute industry in the long run. When debt markets pull back, equity has to step in, yet we’re already seeing signs of softness on the equity side too. Happy Halloween.
Keyvan Moghadam retweeted
Neoclouds can be great businesses. So can a car rental company.
Capital and a head start matter until supply catches up with demand.
Then the only question left: what have you built that a competitor with the same NVIDIA GPUs can't copy?
A little while ago, I started using GLM5.3 through our own endpoint for a few simple tasks. I was amazed by how much better it felt (subjectively at least) and how much faster it was than the latest models from ChatGPT and Anthropic. On top of that, it costs less than a tenth of the price.
Today, during a moment of reflection, I realized I've been using our own endpoint for everything, from coding to task completion, with almost no reliance on any of the big labs' endpoints at all. Strange feeling!
if you like to try it yourself you can go here: pirislabs.io/products/ai-end…
Hey @grok , what percent of Coreweave customers are small/medium startups (less than $10B market cap)?
Been working with few great startups to put together a new B300 deployments. We have the space, power and the allocation and may have room for couple others to join. DM is open.
88
Tensor Machines' new open benchmark found the same GPU model can swing ~15% in tokens/sec on one inference workload. Same hourly price, ~15% different cost per reported token. Spec sheets quote the card. Your bill quotes the output.
57
Three inference stories this week, one theme. Anthropic cut cache reads in half. Mistral previewed a 1T model with 49B active. Scality moved KV cache off the GPU. Everyone is attacking the same cost: memory, not compute.
52