Wuray's WQS: 24 GPUs + one storage node match 48-GPU token output, raises $14M+
创业邦 · wechat · 2026-11-16
Can AI inference scale without just adding GPUs? A customer validation disclosed by storage startup Huari Index Cloud (Wuray) shows 24 GPUs plus one WQS storage node delivering token throughput on par with a 48-GPU setup. The company also announced a new funding round of over 100 million RMB.
Core idea: KVCache context storage
- Inference repeatedly reuses historical context (multi-turn chats, agent tool calls). Storing computed KV results avoids recomputation, but limited VRAM causes two losses: recompute after cache eviction, and GPU idle time waiting for data. Adding GPUs alone doesn't fix either.
- NVIDIA launched a推理 context storage platform in Jan 2026 and detailed its CMX architecture in March — evidence that context storage is becoming a new layer of AI infrastructure.
Why rebuild storage from scratch
- General-purpose storage adds unnecessary overhead via file interfaces. Inference context has clear I/O traits (write-once, content-addressable, framework knows which blocks are reusable). WQS, built on the in-house WiDE engine, uses a native-KV path with user-space RDMA, bypassing filesystem overhead.
- Full-stack co-design: frameworks pre-announce needed cache blocks so reads overlap compute; cross-node cache sharing widens scheduling. Ecosystem work with NVIDIA/AMD/Ascend and vLLM/SGLang integration is underway.
Measured results
- One token-factory validation: 24 GPUs + 1 WQS node ≈ 48 GPUs of token output.
- Software-only deployment at a large enterprise: no new hardware, TTFT down 80%, throughput up 3x+.
Background: founded 2020 with alumni from Huawei, Tencent, IBM, Microsoft, EMC; WDS block storage set an SPC-1 record in 2023 with 27.2M IOPS. Products run in core production at 200+ customers; H1 2026 contract value up 150% YoY. Goal: 50+ production deployments within 12 months.
More from Venture
- Indie dev cuts $400-600/month by ditching Vercel, Ahrefs and Sanity — yihui_indie · 2026-11-16
- Who absorbs failed AI datacenter deals? Microsoft took Abilene's 900MW in three weeks — tengyanAI · 2026-11-16
- Nvidia bylaws make the 'shareholder vote to reinstate Jensen' narrative nearly impossible — firstadopter · 2026-11-16
- Founders slip clause into term sheet forcing VCs to attend Cowboys game in full gear at $100M ARR — pxd · 2026-11-16
- Apple tops $5.18T market cap while OpenAI and Anthropic still trail Saudi Aramco — peterwildeford · 2026-11-16
- WSJ: Nvidia's days as the AI economy's open checkbook are numbered — Gregory_C_Allen · 2026-11-16