We build photonic based computing systems using a disaggregated architecture and software control plane.
Joined June 2021
- Tweets64
- Following54
- Followers22
- Likes8
What if the computer you needed didn't have to fit inside the server you bought?
Mercury Rack composes pooled GPUs, xPUs, SSDs & FPGAs over native PCIe-over-optical into right-sized blocks: one 32-GPU domain, two 16s, or four 8s.
corespan.ai/resources/blog/m…
In the Lab: 8x AMD Radeon AI PRO R9700, One Fabric
AMD specs the R9700 for 4-way. We ran 8 as one fabric.
Every GPU-to-GPU pair: ~111 GB/s bidirectional, line-rate PCIe 5.0 x16. Flat topology, no cross-domain cliff.
ROCm + PyTorch. No NVLink, no CUDA. ~2× what NVIDIA targets for the RTX PRO 6000.
Full breakdown corespan.ai/resources/blog/8…
The asset a lender repossesses isn't a GPU. It's a rack. And most AI racks financed at scale are HGX boxes: 8 accelerators welded to one baseboard, one topology, one cooling design, one generation.
New from Bill Koss on why bankable compute runs through the rack, not just the runtime: corespan.ai/resources/blog/e…
The most expensive GPU is the one you can't fully use.
Wrong ratio, and half the rack sits idle waiting for a job shaped like the box.
Scale-across composes the exact compute, memory & I/O each workload needs — then reclaims it.
Utilization is the multiplier.
corespan.ai/resources/blog/s…
The inference era won't be won by the biggest boxes. It'll be won by the fabric that composes GPUs, memory & I/O around each workload, then reclaims it for the next one.
Tokens per dollar, not GPUs per rack.
corespan.ai/resources/blog/s…
We were excited to see this development...let the software teams cook!
🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta!
🔷 We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. Check out the massive performance leap below! 👇
🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!
Check out the configuration details in our official API docs: api-docs.deepseek.com/quick_…
Corespan Systems retweeted
🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta!
🔷 We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. Check out the massive performance leap below! 👇
🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!
Check out the configuration details in our official API docs: api-docs.deepseek.com/quick_…
Introducing Corespan Mercury Rack
corespan.ai/resources/blog/m…
Using the PRU 2500 with RTX 5090 GPUs and Kimi K3.
corespan.ai/resources/blog/f…
@totheagi
@Michaelzsguo
From Trophy to Utility:
corespan.ai/resources/blog/f…
In the lab with Graid:
corespan.ai/resources/blog/i…
Does it make sense to use a gaming GPU?
corespan.ai/resources/blog/t…
Before there was Kimi, there was this:
corespan.ai/resources/blog/c…
Disaggregated SSDs beats putting drives in servers.
corespan.ai/resources/blog/c…
How photonic switching enters the data center.
corespan.ai/resources/blog/t…
Catching up on some prior blog posts. Here is when we introduced the PRU 2500 with RTX 5090 GPUs:
corespan.ai/resources/blog/a…
AI in energy doesn’t fail in the lab — it fails in the field. Legacy infrastructure and data bottlenecks slow real-world AI.
Our new blog looks at how composable GPU + NVMe infrastructure helps AI scale where operations actually run.
corespan.ai/resources/blog/d…