Memory overtook compute as the cost
The economics of serving have inverted. Low-power DRAM, the commodity memory in every phone, now reportedly costs more per gigabyte than the high-bandwidth memory bolted to a flagship accelerator. That is not a temporary dislocation; it is demand from AI infrastructure arriving in a supply chain built for handsets.
When bytes are scarce and expensive, anything that multiplies the useful bandwidth per byte, or compresses the state a session occupies, is worth more rather than less. That is the entire commercial premise of this architecture, and rising memory prices strengthen it instead of weakening it.
One buyer is absorbing the supply
A single accelerator vendor is reported to be on track to become the largest purchaser of low-power DRAM in the world: on the order of six billion gigabytes in 2027, possibly more than half of global supply. Module capacities have reportedly been cut for want of parts.
A market where one participant can absorb half the supply of a commodity input is a market where efficiency per byte stops being an optimisation and becomes an access strategy.
The capacity problem is publicly unsolved
The industry's answer to decode latency has been to move state into very fast, very small memory. It works, and it does not scale to a fleet of agents.
A recently productised decode architecture places roughly 128 gigabytes of on-die static memory in a rack. That is extraordinarily fast, and on reported figures it holds on the order of forty uncompressed sessions.
Forty is not a fleet. It is a demonstration. The gap between forty resident sessions and the tens of thousands an agent platform needs is precisely the space this architecture is designed for, and it is a gap the industry has been closing with software tiering rather than with memory that understands what a session is.
The category is being valued in public
We make no claim about our own valuation and offer nothing for investment. These are reported third-party transactions in adjacent inference silicon, included because they establish that the category is real and liquid.
Where each part sells
One architecture reaches several distinct buyers. They are not the same market, they do not buy in the same unit, and their demand is driven by different things. This is how we think about each.
Production agent fleets
Industry: cloud and AI infrastructure. Buyer: platform teams at AI providers, neoclouds and large enterprises. Sold as: a rack-scale system or a module for an existing chassis.
Where it is used: a support platform holding two million customer conversations open overnight so none has to be re-read in the morning. A coding assistant keeping a repository's context resident between a developer's questions. An insurance back office running claim-handling agents that wait minutes on external systems. A research tool where each analyst's thread stays live for a week.
Sizing
Per million concurrent sessions: 900 TB resident, about 75 nodes, $375–750M of infrastructure at reported comparable rack costs.
Scales linearly. Ten million concurrent sessions is $3.8–7.5B.
Rollout-heavy post-training
Industry: AI model development. Buyer: research infrastructure teams at model developers and enterprise fine-tuning groups. Sold as: a cluster, fabric-wired to RM-1 rollout nodes.
Where it is used: a lab running preference optimisation where thousands of rollouts branch from one prompt prefix. A robotics team doing agentic self-play across long episodes. An enterprise fine-tuning a support model on its own transcripts and needing a run to be reproducible bit for bit when a regulator asks.
Sizing
Sized as a fraction of the same opportunity: an RL bundle is RM-1 rollout nodes plus RM-T. Rollouts dominate reinforcement-learning wall-clock, so the attachment rate to serving hardware is high.
We publish no separate figure.
Assistants that remember across days
Industry: consumer devices, robotics, automotive, industrial. Buyer: device manufacturers and system integrators. Sold as: a system-on-chip.
Where it is used: a phone assistant that still holds last week's conversation after a hundred sleep cycles. A warehouse robot resuming an interrupted task with its state intact. A vehicle keeping a driver's context across ignition cycles. A set-top box or home hub holding a household's history without sending it anywhere.
Sizing
Unit economics rather than rack economics: device volume × attach rate × silicon content. Device volumes are very large and the content per unit is small.
We publish no price, so we publish no figure.
Session semantics inside someone else's silicon
Industry: semiconductor. Buyer: phone and device SoC designers, and other accelerator vendors. Sold as: an IP licence with a per-unit royalty.
Where it is used: a handset SoC vendor adding persistent on-device assistant memory without designing the lifecycle machine themselves. An accelerator company that wants session semantics to interoperate with the same compiler target.
Sizing
Licence fees plus royalty per unit shipped. Costs us little to hold, since the subset is defined anyway to keep the compiler honest.
Terms are not set.
Long-form generative video
Media, entertainment and marketing, where scene and character state must persist across a sequence rather than be regenerated per shot. Scope is demand-gated and deliberately unscheduled.
ConceptSovereign and national-scale
National laboratories and sovereign AI programmes serving a very large single model. Structurally deferred, and listed for completeness of the architecture's reach rather than as a near-term plan.
DeferredAgent serving is the wedge
RM-1 is the part the architecture was designed for, and the only one we are sizing seriously today. Everything else reuses the same catalog and the same software.
Training reuses the same trick
RM-T needs no new idea. Hardware fork and determinism, built for serving, happen to be exactly what rollout trees want.
Licensing costs us nothing to hold
Profile-E is a subset we define anyway to keep the compiler honest. Offering it as IP is a distribution choice, not a separate programme.
How we size demand, and where we stop
We size from the bottom up, starting with a unit we can actually measure: how much session state a fleet has to keep resident. Every input below is stated so you can change it and see what happens. We stop before converting it to revenue, because we have no price to convert it with.
The inputs, and how confident we are in each
Concurrent live sessions. A round one million, chosen only to make the arithmetic legible. Scale it to whatever a given operator actually runs; the chain is linear.
Session footprint, about 0.9 GB. Derived from our own modelling of a latent-attention model at four-bit precision, at the context lengths measured in published production agent traces. This is the input we are most confident in and the one most likely to be argued about.
Managed session memory per node, about 12 TB. Our own design target: sixteen packages, each carrying an active tier and a capacity tier. This is a design figure, not a measurement.
What we deliberately do not do. We do not multiply the result by a rack price to produce a market size. We have no product and no price, so any such number would be invention rather than analysis. What the chain establishes is the unit of demand, which is the thing worth arguing about first.
How to read this page
- Figures attributed to third parties are reproduced as published on the dates stated and are not independently verified by us.
- The sizing chain uses our own design and modelling figures, labelled as such, and is illustrative rather than a forecast.
- We assert no independent market estimate, and claim no revenue, pipeline, customer or valuation.
- Nothing on this page is an offer of securities or a solicitation of investment.
Where we would sit
Agent-serving fleets
Operators running long-lived tool-using agents, where resident-session capacity inside a wake objective is the binding cost.
Reinforcement learning
Rollout-heavy training, where branching from a shared prefix dominates wall-clock and hardware fork applies directly.
Session semantics as IP
A licensable core for parties who will implement session behaviour in their own silicon rather than buy a part.
How to read this page
- Every figure above is a third-party reported number, attributed and dated. We assert no independent market estimate.
- No revenue, pipeline, customer or valuation is claimed, because none exists to claim.
- Memory pricing is volatile. Figures are a point in time and are not a forecast.
- Nothing on this page is an offer of securities or a solicitation of investment.