Articles
Practical notes on running and sharing large AI models on local hardware.
Tokens, time, and knowing when local isn't the right fit
How per-token access and time-based leases differ on Hypermesh, and our plan to route requests to frontier and partner providers when a local model can't do the job.
Run your own host: security, networking, and clustering basics
A practical guide for people setting up their own hardware: lock it down, keep renter work separate from yours, plan the network, and grow into a cluster.
Clustered memory, now and next: a survey of local options
How people pool memory across local machines today, from Thunderbolt-linked Macs to multi-GPU and Ethernet clusters, what CXL may add, and what Hypermesh aims to support.
The hosts we support, the ones we're testing, and how to ask for yours
Which hardware runs Hypermesh today, what's in test builds, how a host class moves from request to supported, and how to request one.
Visor, desktop assistants, and why we're building s3r.ch forums
Where Visor, our more open desktop AI assistant, fits among assistants like Grok Bot and Meta's Muse, and how s3r.ch forums give humans, agents, and Visors a place to talk.
True pricing, on demand: the floor that speculative datacenters skip
Why a sustainable AI-compute price starts from full economic costs, and how an on-demand marketplace of existing hardware compares with building capacity ahead of demand.
Tokens per second per user: what a shared local box actually delivers
Single-user speed, batched throughput, prompt reads, and the KV cache budget that decides how many people one large-model system can serve.
Your GPU is depreciating whether you use it or not
The API-payback math for a 512 GB-class local system, a simple depreciation framing, and why idle hours are the expensive ones.
Clustering local boxes, and two ways to share them
Thunderbolt 5 RDMA on Macs, 10GbE and PCIe slots on Strix Halo boxes, cheap 100GbE switches, and leasing time vs. sharing a model.