COMPANY PROFILE
Hong Kong Zhongke Hangxing Technology Co., Limited
ZK-Storage WS5000 all-flash accelerated storage · Make every GPU earn its keep · Disaggregated · Self-controlled
300 GB/sAggregate bandwidth
50MRandom IOPS
90.9%7-metric median reduction
WS5000Finalized & in production
AT A GLANCE
Hong Kong Zhongke Hangxing Technology Co., Limited
| Item | Detail |
|---|---|
| Entity | Hong Kong Zhongke Hangxing Technology Co., Limited (香港中科航星科技有限公司) |
| R&D & mfg | Shenzhen Zhongke Hangxing Technology Co., Ltd. |
| Flagship | ZK-Storage WS5000 (WS-HBMM5000) all-flash storage |
| Stage | Mature — finalized, mass-produced, validated |
| Manufacturing | Luxshare Precision · ~1,000/mo |
~10 yrs
R&D heritage
Risk materially retired
~RMB 1.0B
Cumulative R&D
Self-funded
85.17×
Peak load speedup
Third-party measured
1,000/mo
Capacity
Within a month
01
COMPANY OVERVIEW
Company Overview
From precision electronics manufacturing to AI storage infrastructure.
WHO WE ARE
Hong Kong Zhongke Hangxing Technology Co., Limited
ZK-Storage is the brand of Hong Kong Zhongke Hangxing Technology Co., Limited (香港中科航星科技有限公司), the group's outward-facing entity for international business and partnerships. R&D and manufacturing are carried out together with its affiliate, Shenzhen Zhongke Hangxing Technology Co., Ltd., whose founding team has worked in electronics manufacturing since 1996 — nearly three decades — across semiconductor R&D, intelligent-terminal production and systems solutions.
Guided by "innovation-driven, quality-first," the group focuses on AI compute infrastructure, with disaggregated storage as its core route — delivering high-bandwidth, low-latency, self-controlled all-flash accelerated storage that frees GPUs from waiting on data.
Flagship line
ZK-Storage WS5000 (WS-HBMM5000) — developed over ~10 years with ~RMB 1.0B invested — is now finalized and in mass production.
1996
Heritage since
Nearly three decades
2025
Group focus
AI compute infrastructure
~10 yrs
Sustained R&D
Disaggregation & interconnect
~RMB 1.0B
Cumulative R&D
Self-funded
MATURITY
From concept to maturity: four certainties
~10 yrs
Sustained R&D
Tech/product/mfg risk materially retired
~RMB 1.0B
Cumulative R&D
Group's self-funded history
1,000/mo
Production capacity
Luxshare Precision, within a month
2 units
Demo stock
Ready for immediate PoC
Four certainties
Technology — Beijing Information Science and Technology University completed an independent test on Huawei Ascend Atlas 910B; Product — WS5000 finalized and in mass production; Manufacturing — pre-production agreement with Luxshare Precision, ~1,000/mo; Ecosystem — AMD / xFusion adaptation in progress (subject to final reports).
MILESTONES
Milestones and development stages
| Stage | Milestone |
|---|---|
| Technology | ~10 years of disaggregation and high-speed interconnect; ~RMB 1.0B invested |
| Productization | WS5000 (WS-HBMM5000) finalized; hardware / software stack mature |
| Validation | Independent test by Beijing Information Science and Technology University; leading the NFS baseline on all 7 metrics |
| Manufacturing | Pre-production agreement with Luxshare Precision; ~1,000/mo; 2 demo units in stock |
| Ecosystem | AMD / xFusion adaptation in testing; deep tuning for domestic accelerators |
Development stage
The group is in a mature stage — technology, product and manufacturing risk materially retired — entering scaled, repeatable delivery.
02
PRODUCT & TECHNOLOGY
Core Product & Technology
ZK-Storage WS5000 all-flash storage · Disaggregated architecture · Self-controlled.
OVERVIEW
ZK-Storage WS5000 (WS-HBMM5000)
ZK-Storage WS5000 is a high-performance all-flash accelerated storage appliance for AI training and inference. A disaggregated architecture and an end-to-end high-speed data path free GPU clusters from waiting on data — raising effective utilization and cutting total cost without changing the upper-layer framework.
All-flash EBOFNVMe-oF / RDMAGPUDirectKV-cache accelerationSelf-controlledTurnkey
300 GB/s
Bandwidth
Line-rate path
50M
Random IOPS
Small-file friendly
20 µs
Latency
Microsecond
48-72 h
Deployment
Turnkey
SPECS
Core specifications at a glance
300 GB/s
Aggregate bandwidth
Line-rate data path
50M
Random IOPS
High-concurrency small files
20 µs
Access latency
Microsecond response
90%+
GPU coverage
Broad accelerator support
48-72 h
Fast deployment
Live within a day
-40%
Total cost
vs. mainstream 3-yr TCO
-60%
Scale-out cost
Elastic, on-demand
2-3×
GPU utilization
High-switch / long-context
PORTFOLIO
Product portfolio: four delivery models
| Product / Service | Form | Customer | Core value |
|---|---|---|---|
| ZK-Storage WS5000 appliance | Hardware | New AI clusters | High-bandwidth all-flash, turnkey |
| ZK-Storage storage software | Subscription | Existing hardware | Disaggregation, continuous updates |
| Brownfield retrofit | Solution + service | Existing data centers | Speed-up without downtime |
| Accelerated storage service | Capacity / compute | SMB / cloud | On-demand, low barrier |
One platform, four ways to monetize
The same disaggregated capability — sell the box, sell software, retrofit, rent compute — covering the full AI build lifecycle.
WHY STORAGE
Why storage? Turn it into a compute amplifier
In the LLM era, stacking more GPUs yields rapidly diminishing returns; the real bottleneck is on the data-supply side — model loading, checkpoint I/O and KV-cache scheduling.
Measured
ZK-Storage upgrades storage from a supporting act into a compute amplifier — cutting KV-cache-related cost by about 74% in testing; the gains are largest for long-context, agentic and multi-tenant inference.
DISAGGREGATION
Disaggregation: compute pool ⟷ lossless fabric ⟷ all-flash pool
- Decouple: detach storage media from compute nodes into an independently scalable all-flash pool.
- Connect: link to the GPU compute pool over a high-speed lossless fabric — data goes direct.
- Elastic: compute and capacity scale independently, pooled and efficiently shared.
- Transparent: no changes to upper-layer training / inference frameworks.
NVMe-oF
over RDMA/RoCE
Near local-disk
GPUDirect
Direct path
Lower CPU/latency
EBOF
All-flash pool
Near-linear scaling
KV-cache
Scheduling
Higher utilization
FOUR PILLARS
Four core technologies
01NVMe-oF over RDMA/RoCE
Carry NVMe over remote direct memory access, bypassing redundant copies to approach local-disk performance.
02GPUDirect
Data moves directly between storage and GPU memory, shortening the path and cutting CPU and latency overhead.
03All-flash EBOF
Controller-less, high-density flash pool; bandwidth and IOPS scale near-linearly with capacity, at lower power.
04KV-cache scheduling
Offload and reuse KV cache for long-context / high-switch inference, lifting effective GPU utilization.
Self-controlled
Deeply tuned for Huawei Ascend and domestic accelerators, with 90%+ mainstream GPU coverage — aligned with self-controlled infrastructure.
03
INDEPENDENT VALIDATION
Independent Third-Party Validation
Beijing Information Science and Technology University · Huawei Ascend Atlas 910B · leading on all 7 metrics.
SETUP
A reproducible, independent third-party test
To objectively verify performance, the group commissioned Beijing Information Science and Technology University to run an independent test on the Huawei Ascend Atlas 910B platform, against an NFS (NFS over TCP, 10GbE, ~1.25 GB/s) baseline; the ZK-Storage side used a NVMe-oF over RDMA/RoCE (2×200GbE, ~50 GB/s) high-speed data path.
- Tester: Beijing Information Science and Technology University (national university, independent third party).
- Coverage: inference load/service, training I/O and token throughput — 7 metrics.
- Headline: median latency reduction 90.9% (lower is better), leading on every metric.
85.17×
Peak load speedup
DeepSeek-32B
7
Metrics covered
Inference/training/token
90.9%
Median reduction
Lower is better
NFS
Baseline
10GbE TCP
INFERENCE
Inference: model load and service speedup
| Model | ZK-Storage load | NFS load | Load speedup | Latency cut | Service speedup |
|---|---|---|---|---|---|
| DeepSeek-32B | 6.62 s | 563.85 s | 85.17× | 98.83% | 6.17× |
| DeepSeek-70B | 35.38 s | 1284.66 s | 36.31× | 97.25% | 9.33× |
How to read it
A 70B model load drops from 1285 s to 35 s — bring-up / switching moves from minutes to seconds.
TRAINING
Training: weights and checkpoint I/O
| Test | ZK-Storage | NFS baseline | Speedup | Latency cut |
|---|---|---|---|---|
| Model load | 12.72 s | 140.23 s | 11.02× | 90.93% |
| Model save | 31.16 s | 165.87 s | 5.32× | 81.21% |
| Checkpoint load | 10.55 s | 131.37 s | 12.45× | 91.97% |
| Checkpoint save | 81.94 s | 451.14 s | 5.51× | 81.84% |
How to read it
Training I/O speeds up 5–12×: the bigger the model and the more frequent the checkpoints, the more idle GPU time you save.
THROUGHPUT
Token throughput (= effective GPU utilization)
| Switch frequency | ZK-Storage util. | NFS util. | Relative gain |
|---|---|---|---|
| 10/day | 99.8% | 80.4% | +24.1% |
| 20/day | 99.5% | 60.8% | +63.6% |
| 40/day | 99.1% | 21.7% | +356.9% |
Key takeaway
The more frequent the switching, the wider the gap: at 40 switches/day, effective token output improves +356.9% — greatest value for multi-tenant / multi-model inference.
RESULTS
The conclusion on one page
85.2×
Peak model-load speedup
DeepSeek-32B
9.33×
Peak service speedup
DeepSeek-70B
+356.9%
Peak token efficiency
High-switch case
90.9%
7-metric median reduction
Lower is better
In one line
In Beijing Information Science and Technology University's independent test, ZK-Storage WS5000 reached ~85× peak model-load speedup, 5–12× training I/O and up to +357% token efficiency — reproducible, verifiable third-party endorsement.
04
ECOSYSTEM · VALUE · FUTURE
Ecosystem · Value · Future
Alliances + unit economics + global roadmap.
PARTNERS
Alliances and industrial ecosystem
MManufacturing · Luxshare Precision
A pre-production agreement leverages precision manufacturing and supply chain to deliver about 1,000 units within a month of order.
VValidation · Beijing Information Science and Technology University
An independent test on the Huawei Ascend platform provides national-university endorsement of performance.
TAdaptation · AMD / xFusion (in test)
Platform adaptation tests with AMD and xFusion are in progress (work in progress; subject to final reports).
SSelf-control · Ascend ecosystem
Deep tuning for Huawei Ascend and domestic accelerators, aligned with the self-controlled trend.
Honesty discipline
We separate delivered (validation / production) from in progress (AMD / xFusion in test) — truthfulness is our strongest trust asset.
TCO
Customer value: three-year total cost of ownership
Three-year TCO (US$M, lower is better)
$144.6M
ZK-Storage
All-flash accelerated
$241M
Industry baseline (high-end)
Representative high-end
How the case works
Against a mainstream high-end option, the ZK-Storage three-year TCO is about $144.6M vs. ~$241M baseline — about -40% total cost (~$96.4M saved over three years) — and a higher GPU utilization amplifies the customer's compute ROI.
UNIT ECONOMICS
Per-system economics (modeled)
| Revenue component (per system / annualized) | Amount | Note |
|---|---|---|
| Hardware (appliance) | RMB 2.80M | One-off |
| Software subscription (CPFS + acceleration + KV-cache) | RMB 0.50M/yr | Recurring |
| O&M service | RMB 0.34M/yr | Recurring |
| Compute service | RMB 0.30M/yr | Recurring |
| Blended revenue / gross margin | ~RMB 3.94M · ~49% | HW + SW + O&M + compute |
Amplification
Per-system blended revenue is about RMB 3.94M at ~49% gross margin; combined with a ~2.5× utilization uplift, the customer's effective compute ROI is further amplified. Modeled; quotations per formal proposal.
USE CASES
Applications and market network
- LLM training clusters: accelerate model loading and checkpoint I/O to shorten iterations.
- LLM inference serving: long-context and high-frequency switching — higher effective utilization.
- AI centers / domestic stack: disaggregation + domestic adaptation for sovereign infrastructure.
- Brownfield retrofit: no GPU swap, no downtime — revive idle compute in place.
| Market segment | Go-to-market |
|---|---|
| Domestic greenfield | Appliance + software |
| Domestic retrofit | Retrofit + token revenue share |
| Overseas greenfield | Appliance export (certified premium) |
| Overseas retrofit | Local integrators + share |
Sequencing
First prove scalable replication in domestic greenfield + retrofit, then extend overseas.
ROADMAP
Future blueprint and roadmap (modeled)
Revenue (RMB 100M, modeled)
0.6
2026
3.3
2027
8.9
2028
19.3
2029
37.0
2030
- 2026 · M1 Product & base — mass-production release; R&D / pilot base; first batch delivered.
- 2027 · M2 Commercial validation — benchmark customers; subscription ARR scaling.
- 2028 · M3 Profitable scale — EBITDA positive; 90%+ domestic-GPU coverage; KV-cache suite GA.
- 2029 · M4 National × overseas — hub-node coverage; enter SEA / Middle East.
- 2030 · M5 Global & exit window — revenue RMB 3.0B+, net margin 20%+.
Note on figures
2026–2030 revenue is modeled (see the Business Plan); the 2030 figure is about RMB 3.70B. Actuals per final disclosure.
MARKET
AI storage: a fast-growing core market
$133B
AI storage market 2030
Global
27.5%
CAGR
2025–2030
$67B
China AI-chip TAM 2030
Addressable
140T
China daily tokens
Demand anchor
Why now
AI storage is projected to reach about $133B by 2030 at a 27.5% CAGR; with domestic AI-center utilization below 60%, retrofit-for-efficiency is a policy and market imperative that ZK-Storage addresses head-on. (Public sources S2 / S6 / S7 / S11.)
CONTACT
Contact · let's talk
| Item | Detail |
|---|---|
| Entity | Hong Kong Zhongke Hangxing Technology Co., Limited (香港中科航星科技有限公司) |
| R&D & manufacturing | Shenzhen Zhongke Hangxing Technology Co., Ltd. |
| Registered office | Hong Kong SAR, China — registered office to be published upon completion of registration |
| R&D base | Room 302, Building 3, Ship Front Plaza, Sea World, Nanshan District, Shenzhen, China |
| Phone / Email | +86 138 2372 8880 · 13823728880@139.com |
| Contact | Lisa CHEN (CEO) |
| Focus | ZK-Storage all-flash accelerated storage · AI compute infrastructure |
PPoC units
2 demo units in stock for immediate customer PoC.
VJoint validation
National-university third-party testing, reproducible.
MVolume delivery
Luxshare Precision foundry, ~1,000/mo.
EEcosystem co-build
AI-center co-build and platform adaptation.
Let's talk
We welcome AI centers, model teams and industry partners to engage and scale ZK-Storage in AI compute infrastructure together.
THANK YOU
Make every GPU earn its keep
ZK-Storage WS5000 · All-flash accelerated storage · Hong Kong Zhongke Hangxing Technology Co., Limited
PoC units2 in stock, ready now
Joint validationNational-university test
EcosystemAI-center co-build