Jim Beasley (@JimBeasleyCA) header
Jim Beasley (@JimBeasleyCA) profile picture

Regional Sales Manager at WEKA. My tweets are my own.

San Francisco, CA
Joined June 2011
Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)Photo by Jim Beasley (@JimBeasleyCA)
Great discussion with @weka's Liran Zvibel on @business’s Tech Disruptors podcast about how AI is changing infrastructure, why memory and storage are converging, and what comes next for inference. Worth a listen: spr.ly/6015ByjYJW
3
Not every workload needs a dedicated KV cache tier yet. @weka lays out three checks to run this quarter before adding one, and why long running agents that refill context all day are the first to need it. spr.ly/6012ByjMAE
5
GPUs that drop KV cache recompute work they already did, up to 60% of their compute. @weka field CTO Alan McSeveney with @VentureBeat: on OCI, shared NVMe offload took tokens from 200,000 to 2 million vs DRAM. spr.ly/6018ByZjWm
8
451 Research projects inference at 46% of AI infrastructure spending by 2030, with agents driving a growing share of tokens. More GPUs alone will not fix the memory bottlenecks that come with that. New Pathfinder paper via @weka: spr.ly/6013Bynxnl
1
18
A dashboard can show GPUs 95% busy while agents make them rebuild context they already computed. That repeated work reads as a healthy fleet and lands on the invoice. @weka on the new math of inference: useful work per watt, per rack, per dollar. spr.ly/6017ByTvRX
7
On HBM alone, 60% of prefill work in an eight hour run went to recomputing context the cluster already had. With a shared context tier: 6.8%, and 3.7x the output tokens per GPU-hour on the same B300 node. @weka Liran Zvibel on the new inference scorecard. spr.ly/6015ByR9mh
19
9,000 concurrent jobs and 880 GB/s from 16 rack units of storage, independently audited on SPECstorage AI_IMAGE. Average latency moved from 1.0 to 1.3 ms as the load rose tenfold. @weka spr.ly/6016ByKwue
8
Training is like Formula One, where spending is barely the question. Inference is about economics: power, capital and how many tokens you get from the same GPUs. @weka Liran Zvibel with Steve Hou of Silicon Data at the NeoCloud Summit. spr.ly/6019ByHxix
12
A rack of GPUs already draws hundreds of kilowatts, and the latest generation is heading toward a megawatt per rack. @weka Val Bercovici on Deep Geeks Ep 1: GPU deals are now announced in gigawatts, roughly one nuclear plant each. spr.ly/6016By8wNu
5
Excited to see what HTX, @Singtel RE:AI, and @weka have accomplished together! Achieving 99.49% GPU utilization with multitenancy enabled is a great example of what’s possible when infrastructure is built to work as a system. Get the full story ⬇️ spr.ly/6018By6uQM
38
US datacenter construction declined in 2025 for the first time since 2020, just as AI demand accelerated, and high-voltage transformer lead times now run five years. @weka Liran Zvibel on why the capital exists but new capacity will not arrive in time. spr.ly/6013BHhI47
18
Households served by a Nebius datacenter spent 10% less on heating in 2025. Nebius recovers 20 to 30% of its electricity use as heat and gives it back to the local energy system. Daria Mukhortova on @weka Deep Geeks Ep 1. spr.ly/6012BHCEPn
1
30
Morgan Stanley projects a 49 GW US power shortfall through 2028, and grid connection queues in major markets now run four to seven years. @weka Liran Zvibel on why every rack unit and watt already deployed has to produce more inference. spr.ly/6014BHNzv0
1
14
What does it really cost to run AI at scale? It’s a question that’s becoming increasingly important as AI adoption accelerates. This @weka article explores seven forces shaping AI economics, from GPU pricing and memory to energy, financing, and demand. spr.ly/6013BHFcMV
4