AI memory needscompute that grows
Extracting, organizing, recalling and evolving memory all run on compute. MemoryLake Compute delivers it in standard MCU units with an SLA, so you buy measurable AI compute, not hardware.
Multiple regions · Backed by an SLA
Every step of memory runs on compute
Long-term memory is not static storage but a pipeline that never stops. Its four stages load the hardware in different ways; MemoryLake Compute gives them one predictable, measurable base.
Extract
Pull facts, preferences and experience out of conversations, documents and business data. Large-scale inference, running continuously.
Organize
Embed, cluster, deduplicate and index into a knowledge network. Heavy on vector compute and data processing.
Recall
Millisecond recall and context injection so every generation carries full memory. Low-latency online inference, around the clock.
Evolve
Compress, distill and replay-train memory so models understand the business better over time. Training-grade compute.
Pick by workload, not by hardware
Every package is just a number of MCUs plus a scheduling domain. Buy in multiples of 4; distributed training needs a tightly coupled domain so its MCUs share the fast interconnect.
Inference & development
Online recall and inference, debugging memory-enabled agents, light jobs.
Fine-tuning
LoRA / SFT fine-tuning and mid-sized inference clusters.
Training
Distributed training and memory distillation inside one coupled domain.
Cluster
Large-scale training or inference platforms across several domains.
Let memory grow. Leave the hard parts to us.
No infrastructure
No hardware to buy, no data center to build, no ops team to hire.
Measurable
Compute is counted in MCUs: how much you used, what is left and what ran.
Backed by an SLA
Platform-level SLA; failed nodes migrate automatically and memory services keep running.
Easy to manage
One console for resources, usage and bills, scaling with your memory.
Service level
Exact figures are set in the signed SLA.
Four steps from browsing to running
Size the plan with our sales team, then order in the web console. Contracts are signed offline and paid by corporate bank transfer; once payment is confirmed, the MCUs appear in your account.
- 1
Browse inventory
Available MCUs, price and load by region, in real time.
- 2
Order
One order, quantities per region, priced automatically.
- 3
Sign & pay
The contract is generated online; sign it and pay by bank transfer.
- 4
Run
Delivered to your account; dashboards go live right away.
Transparent, metered in MCUs
Locked for 12 months, billed annually. Each region is priced on its own line (currently the same everywhere); the order total is the sum of the lines.
Workload coefficients
Different stages draw on resources differently. On pool plans the monthly fee is fixed; coefficients only decide how many MCU·hours a job uses inside your pool.
Invoiced monthly as "information technology service" (6% VAT).
Guide
MemoryLake Compute product guide
Product definition, MCU specs, buying process, SLA, metering and billing, and a procurement checklist. Written for business and procurement review.
Frequently asked questions
The MCU is defined by guaranteed capability (TFLOPS, GPU memory, system memory, NVMe, vCPU, interconnect), not by a GPU model or card count. The platform can upgrade what is underneath as long as the committed spec never goes down.
Ready to let your AI memory grow?
Tell our sales team about your memory workloads. We will size the MCUs, regions and plan with you and send a quote.