Liqid announces managed memory pooling service for AI inference servers

Executive Briefing
“At SC26, Liqid announced a managed memory pooling service and outlined an approach to...”
Key Takeaways
- 1.The article says many AI servers retain a fixed GPU-to-memory ratio throughout their lifetime. Liqid co-founder and chief technology officer Sumit Puri argues that this is a poor fit when future model deployments are unknown.
- 2.Puri described Liqid’s approach as pooling GPUs and memory within a rack, then assigning those resources to servers according to the model or workload. This is the company’s proposed approach, not a measured performance result.
- 3.Liqid evaluates inference economics using tokens per second, per watt and per dollar, with cost per token receiving the greatest emphasis.
- 4.At SC26, Liqid announced Liqid as a Service as a fully managed, single-tenant offering. The article also says the company is moving toward preconfigured appliances.
- 5.Liqid sponsored the theCUBE interview on which the article draws. The broadcaster says sponsors have no editorial control over its content.
At SC26, Liqid announced a managed memory pooling service and outlined an approach to assigning GPUs and memory according to each AI model’s needs. The company wants to move beyond fixed server configurations.
Product Features & Key Highlights
The article says many AI servers retain a fixed GPU-to-memory ratio throughout their lifetime. Liqid co-founder and chief technology officer Sumit Puri argues that this is a poor fit when future model deployments are unknown.
Puri described Liqid’s approach as pooling GPUs and memory within a rack, then assigning those resources to servers according to the model or workload. This is the company’s proposed approach, not a measured performance result.
Liqid evaluates inference economics using tokens per second, per watt and per dollar, with cost per token receiving the greatest emphasis.
At SC26, Liqid announced Liqid as a Service as a fully managed, single-tenant offering. The article also says the company is moving toward preconfigured appliances.
Liqid sponsored the theCUBE interview on which the article draws. The broadcaster says sponsors have no editorial control over its content.
Practical Assessment & Evaluation
If memory can be assigned according to application needs as Puri proposes, a server cluster could avoid reserving the same amount for every configuration. Actual savings would still depend on how many servers need memory at once.
For AI inference, any economic benefit needs to be weighed against the effect of resource sharing on token generation speed and power use. Assigning less memory to each server alone does not establish a lower cost per token.
Pricing, Availability & Next Steps
The source provides no independent measurements or comparative figures for cost per token, latency or power use with shared resources.
The source does not specify pricing, general availability or the scale of real-world Liqid as a Service deployments.







