GenstorAIge ProductsVP - Xu Wei at2026 WAIC Delivers Keynote

 

7Month18Day,2026World AI Conference (WAIC)(WAIC)Expo Center,GenstorAIge VP of Products Xu Weidelivered a keynote speech and officially launched AI90 ——FaaS (Flash as a Service) for AI architecture, drawing significant attention.

The 'memory wall' has become a critical bottleneck constrainingAIinference efficiency

and cost

GenstorAIge VP of Products Xu Wei Delivers Keynote at WAIC 2026

Xu Wei pointed out:

Over the past two decades, GPU computing power has increasedapproximately60,000times, while memory capacityhas only grown16times. The severe compute-storage imbalance results in GPU effective computing power utilizationof only30%-60%。

 

. More critically, as large models evolve from short-context to long-context and multi-turn dialogues,KV Cache KV Cache is rapidly expanding. The 'memory wall' has become a AI key bottleneck restricting inference efficiency and cost。

 

AI90 Reconstructing the Inference Path with Compute-Storage Convergence

Xu Wei detailed the breakthrough logic of theAI90 solution: by offloading KV Cache from HBM to high-performance SSD, a HBM + DRAM + SSD three-tier storage hierarchy is built.Intelligent P2P interconnect technologyoptimizes the data path betweenGPU. Cross-card data transfer no longer requires CPU and host memory copying,enabling GPU efficient direct connection between。

The PT200Z AI SSD powering this architecture utilizes PCIe Gen5 and pSLC media, achieving up to 100 DWPD, specifically optimized for high-frequency KV Cache write scenarios.

50Times Performance Leap Running Large Models on Consumer-Grade GPUs

Xu Wei announced live demonstration data:

• Llama 3 70Bmodel inference,4card5090cluster throughput increased from120 tk/sto610 tk/s，, a5.1times improvement；

 

• 64Kcontext first Token latency dropped from27.99secondsto0.564seconds, an improvement ofnearly50times；

 

• memory utilization increased from30%-40%to85%-95%，, a2-3times improvement。

 

• AI90 supported context expanded from approximately 8K to 128K+, and the longer the input sequence, the more pronounced the acceleration effect.

 

From 'Compute-Centric' to 'Compute-Storage Convergence'

 

GenstorAIge VP of Products Xu Wei Receives In-Depth Media Interview

Xu Wei stated thatAI inference is shifting from'compute-centric'to a new phase centered on'efficiency as the core'.Inference cost optimization has moved from simply stacking computing power to system-level efficiency gains through data and storage'compute-storage convergence'.

 

The essence of GenstorAIge's AI90 is to reconstruct the storage path, turning storage intoan 'amplifier for computing power'.——This is not just a technical choice, but an inevitable path for AI infrastructureto become universally accessible.

GenstorAIge VP of Products Xu Wei Receives Exclusive Media Interview

 

After the event, tech media including TMTPost and DOIT conducted in-depth exclusive interviews with Xu Wei and the GenstorAIge team regarding the deployment scenarios, compatibility, and implementation roadmap of AI90.

 

The full interview will be published soon on the respective media's official accounts and the GenstorAIge official account. Stay tuned.