We buy datacenter-grade CXL memory pools, virtualize every rack into many memory machines, and rent them to the AI industry. The yield flows back to holders. Every share is a slab of real memory — live, bound, and earning.
Training is bound by GPUs. Inference is bound by memory: the model has to fit, the conversation cache has to persist, the long context has to hold. The industry calls it the memory wall. When compute is short you queue — when memory is short, it simply does not run.
HBM is eating the production of Samsung, SK hynix and Micron. Server memory is under-supplied and climbing. On the demand side, global inference volume multiplies every year. Supply contracts while demand detonates — memory has become the hard currency of the AI world.
The CXL interconnect standard let memory leave a single server for the first time — pooled, allocated on demand, billed by the gigabyte-hour. Microsoft Azure shipped the first commercial CXL memory service; Samsung and Micron are in volume production. Memory turned from a part soldered to a board into an asset you can own, operate, and make yield on. The first time in its history.
We buy proven CXL memory chassis (H3 Falcon class) and enterprise memory modules from Samsung and Micron, and run them in core datacenters. On top of the hardware we run MemVM — our memory-virtualization layer that carves every physical rack into many independent memory machines, rented by capacity and by time.
When a customer's server runs short, it draws capacity straight from the core and pays by usage. No buying for peak, no idle silicon.
GB · HOURInference companies park their massive conversation caches in the core. Far cheaper than recomputing, and over five times faster than spilling to SSD.
5× VS SSDA whole slab, rented exclusively — carrying oversized model inference and in-memory databases that no single box can hold.
DEDICATEDFor teams running models, the core beats building your own: more elastic, higher performance, and none of the datacenter to babysit. Capacity on tap.
MemVM is the control plane we built on top of the CXL fabric. It carves every physical rack into independent memory machines, surfaces each one to a tenant as standard system memory, and meters delivered service down to the gigabyte-hour. No application changes on their side. Full accounting on ours.
EXPLORE THE ARCHITECTURE →You are purchasing yield rights to memory machines. Neutron operates the machines and delivers the service; rental income is distributed per share on a defined cadence and recorded on-chain.
A neutron star sweeps its beam past us on a millisecond beat. Neutron pays rental yield to every holder with the same regularity.
Memory production can't keep up; rental pricing moves with the market.
A coming flood of models will live on MaaS — more elastic, higher performance, and simply easier. The demand side keeps widening.
The memory machines Neutron holds appreciate as the underlying silicon does.
Yield tracks real utilization and market pricing. Distributions are settled on-chain from metered rental income.
Collapse the world's memory into Neutron. Build the densest, largest memory mine on Earth and become the gravity well of the AI economy. Density creates gravity — let the world's AI inference be captured by the core.