At MWC Shanghai 2026, CypressTel made its first domestic appearance at MWC — and used the stage to formally announce a major brand upgrade: the company is now positioned as a Global AI Computing and Network Infrastructure MSP.
After 18 years of deep experience in enterprise networking and SD-WAN, serving tens of thousands of mid- to large-sized enterprises and deploying more than 140 global PoPs and 30+ data centres worldwide, CypressTel is now fully focused on the AI computing and network track.
In an exclusive interview with Phoenix Tech, CypressTel CEO Connee Zhang shared her perspective on the real pain points behind AI compute adoption, the future of distributed compute networks, how to solve the structural mismatch between compute supply and demand, and why new applications such as AI Agents are reshaping network requirements from the ground up.
“The industry already agrees on one thing: buying GPU hardware is only the first step. The real challenge is making compute truly usable and manageable end-to-end. That’s where a lot of enterprises end up with idle GPUs and sunk costs,” Connee began.
Based on CypressTel’s experience serving thousands of mid- to large-sized enterprises, she summarised four major pain points enterprises face when using AI compute.
Four Core Pain Points in Enterprise AI Compute
1. Fragmented multi-model management
Today, there are dozens of mainstream large language models (LLMs) on the market. Most enterprises access multiple models in parallel, often with separate accounts, permissions, and billing systems. Context cannot be carried over when switching models, employee experience is poor, and management complexity increases sharply.
2. Network becomes the compute bottleneck
Many enterprises have sufficient GPU hardware, but when doing cross-regional training and inference, or remote model calls, public internet latency and packet loss become significant issues. This directly drags down training and inference efficiency and leads to the common phenomenon of “compute idle, network holding things back”.
3. High operational overhead
From hardware deployment and network tuning to model operations and maintenance, the full stack of AI infrastructure requires specialised teams. For small and mid-sized teams, the manpower and time costs are difficult to sustain.
4. Imbalanced utilisation of compute resources
Enterprises that build their own GPU clusters typically make one-off, heavy capital investments. But AI workloads are highly volatile: peak periods suffer from insufficient compute, while off-peak periods leave large amounts of hardware idle. This mismatch leads to serious resource waste.
To address these pain points, CypressTel has launched three complementary products — AI Gateway, AI-Ready High-Bandwidth SD-WAN, and GPU Computing Services — forming a “three-layer” solution across the model, network, and compute layers.
CypressTel’s AI Gateway is designed to solve the fragmentation of multi-model access, operations, and billing.
It supports unified access to more than 50 mainstream LLMs, including ChatGPT, DeepSeek, Qwen, Kimi, and others.
It provides a single, one-stop interface for calling different models and centralised token-based billing.
It offers cross-model memory, enabling context continuity even when users switch between models, significantly improving AI usage experience.
It supports private deployment, keeping data on-premises to meet strict information security requirements.
At a leading online professional education platform, the AI Gateway has already gone live. After deployment, the enterprise saw AI operations workload reduced by around 70%, and end-user satisfaction with AI-assisted Q&A services increased by more than 20%.
CypressTel’s high-bandwidth SD-WAN solution is tailor-made for large model training and inference scenarios, tackling the widely observed issue of “enough compute, insufficient network”.
It is built on CypressTel’s 10T-class global backbone, supporting single-link bandwidth beyond 100G.
Cross-regional transmission latency can be controlled within approximately 20ms.
Intelligent traffic scheduling and dynamic bandwidth allocation ensure the network can meet high-throughput, low-latency demands for AI workloads.
A leading new energy vehicle manufacturer’s autonomous driving team has commercially adopted this solution. Previously, its cross-city public internet links experienced latency over 200ms and packet loss above 5%, leading to unstable training. After moving to CypressTel’s dedicated network:
Latency dropped to below 35ms.
Packet loss was reduced to near zero.
Model training efficiency improved by around 40%.
The enterprise has since migrated all cross-data-centre training traffic onto this dedicated network.
To complete the compute layer, CypressTel offers GPU Computing Services that provide a full range of GPU resources:
Bare metal, virtual machines, and container-based environments.
Support for mainstream GPU models such as H100, H200, and others.
Hourly billing, elastic scaling, and “tidal” capacity management, well suited to highly volatile workloads.
A leading short-video platform has adopted CypressTel’s GPU Computing Services for AI-based content moderation. Through elastic scaling:
Compute capacity flexes with nighttime peaks and daytime troughs.
Overall compute cost has dropped by around 30%.
All business data is transmitted over CypressTel’s private backbone, avoiding public internet risks and significantly improving data security.
“Simply put, we help enterprises not only get compute, but use compute well, use it cost-effectively, and use it securely,” Connee summarised.
When asked whether enterprise AI compute will eventually be consumed “on-demand” like today’s cloud services, Connee’s answer was unequivocal:
“AI compute moving fully towards cloud-like, on-demand consumption is inevitable. The model where enterprises buy and operate their own large GPU clusters will gradually weaken — just as most enterprises today no longer build their own servers and instead rely on public cloud.”
She highlighted two key reasons:
Rapid hardware iteration: AI evolves extremely fast, GPU generations turn over quickly, and individual hardware models become outdated in short cycles. For non-infrastructure-core enterprises, building their own GPU clusters means heavy upfront investment (hardware purchase, data centre build-out) and ongoing pressure from hardware refresh, operations, and power consumption. The cost-performance of this heavy-asset model is increasingly unfavourable.
Highly volatile workloads: AI workloads such as large model training, AI moderation, and interactive AI experiences have strong peaks and troughs. Fixed, self-built compute capacity struggles to match dynamic demand, inevitably causing idle resources and wasted investment.
Leveraging 18 years of global networking technology and node deployment, CypressTel is now fully focused on building a global distributed AI computing network. The core objective is to pool and service compute resources, not just sell hardware.
By integrating more than 140 PoPs and 30+ data centres worldwide, CypressTel aggregates distributed GPU compute, network links, and operations capabilities into standardised services. Enterprises can then consume compute resources:
Without heavy, upfront hardware purchases.
On-demand, similar to water, electricity, or public cloud.
With elastic expansion and pay-per-use billing aligned to their business needs.
“This doesn’t mean self-built GPU clusters will disappear entirely. For a small number of very large organisations with extreme data security requirements and dedicated compute needs, private clusters will still exist,” Connee noted. “But for around 90% of government, enterprise, internet, manufacturing, and education customers, on-demand compute consumption will become the mainstream choice. Our distributed compute network is designed to be the bridge between compute supply and demand — making compute flow freely and be available as and when needed.”
Today, a common phenomenon in the AI compute market is that operators are rapidly building intelligent compute centres, yet many enterprises still report: “Compute exists, but can’t find the right customers — and customers can’t find suitable compute.”
Connee defines this as a structural mismatch between supply and demand.
Operators and data centres are deploying large-scale intelligent compute centres, but resources are fragmented, information is siloed, and cross-regional scheduling is difficult.
Enterprises with strong AI compute demand struggle to find capacity that matches their workload specifications, network conditions, and security requirements — especially when connecting across regions or operators.
CypressTel addresses this mismatch from three key dimensions: matching, network, and full-stack services.
Unified compute matchmaking and scheduling platform CypressTel works closely with operators and third-party data centres to integrate idle and dedicated compute resources across intelligent compute centres into a unified resource pool. At the same time, it systematically collects enterprise demands for different scenarios — including large model training, AI inference, autonomous driving simulation, and AI content moderation — and intelligently matches these demands to appropriate compute resources. This ensures idle compute finds real-world workloads, while high-demand customers quickly obtain the right capacity.
Global backbone network for cross-region compute flow Many compute centres have strong hardware but limited cross-regional connectivity. Their compute can only serve local customers. CypressTel’s mature cross-border and cross-city backbone network connects different regions and operators, breaking geographical barriers. Idle compute in western regions can be allocated to high-demand enterprises in the east, while edge compute can serve nearby AI terminals, activating resources at the national and global scale.
Integrated “Compute + Network + Operations + Security” service A significant number of compute centres only provide hardware, lacking network, operations, and data security capabilities. Even when they connect to customers, they struggle to support complex AI workloads. CypressTel fills this gap by providing network transport, intelligent operations, and security as an integrated package. Compute providers gain robust network and operations support, while customers receive end-to-end solutions, ensuring that compute is not only “found” but also “used reliably”.
“Operators and data centres focus on building compute. CypressTel focuses on connecting, scheduling, and operating that compute. Together, we address structural gaps along the industry chain and deal with the core pain points behind supply–demand mismatch,” Connee summarised.
AI Agents, smart robots, and autonomous driving are moving from concept to large-scale deployment. Hui Zhang emphasised that these real-time, interactive, and mobile applications fundamentally change what is required of the network.
Traditional enterprise and internet workloads can no longer define network standards. These new applications bring four core requirements:
Ultra-low latency
Use cases such as autonomous driving, remote robot control, and real-time AI Agents require millisecond-level response. Excessive latency directly impacts safety and user experience.
Extremely high reliability and ultra-low packet loss
These applications often involve safety-critical environments and business continuity. Network interruptions and frequent packet loss are unacceptable and must be designed out.
Very high bandwidth with strong uplink capacity
Robots and autonomous driving systems send back HD video, radar signals, and sensor data in real time. Uplink traffic bursts are the norm, and traditional network architectures designed mainly for downlink traffic are no longer sufficient.
Cloud–edge–endpoint collaboration
These applications run across endpoints, edge nodes, and cloud compute. Networks must support seamless flow of compute, data, and commands across cloud–edge–endpoint architectures.
In response, CypressTel has built clear technical strengths and scenario-focused deployments:
Its AI-Ready High-Bandwidth SD-WAN natively supports high-bandwidth, low-latency workloads, keeping cross-regional link latency in the 20–35ms range, with packet loss approaching zero.
Leveraging global edge PoP nodes, CypressTel has created an edge-centric architecture, pushing compute and network closer to endpoints for local processing and transmission.
The OneSight AI Intelligent Operations Platform enables automated fault detection and rapid resolution across the entire network, improving end-to-end stability and resilience.cypresstel+1
Scenario-wise, CypressTel is already:
Providing dedicated network services for autonomous driving training for leading domestic new energy vehicle manufacturers.
Supporting AI Agent scenarios by using the AI Gateway for unified access and management of multiple intelligent agents.
“Looking ahead, we will continue to deepen cloud–edge–endpoint collaboration, and optimise specialised network strategies for different AI applications, helping these frontier use cases scale confidently,” said Connee.
At the end of the interview, Connee reiterated CypressTel’s industry role:
“We are a Global AI Computing and Network Infrastructure MSP. We don’t build large AI models ourselves, and we are not competing with cloud providers for public cloud market share. Our mission is to be the foundation that makes compute flow — working with operators, data centres, model vendors, and cloud providers to build an open, collaborative ecosystem. That has always been our direction.”
She highlighted three major commercial opportunities CypressTel is most optimistic about:
Compute network scheduling and operations services
Whether for operator-intelligent compute centres or enterprise-built clusters, professional compute–network scheduling and operations are needed to fully unlock value. This is the largest, most foundational opportunity in the ecosystem — and CypressTel’s core arena.
Edge “compute + network” integrated services
As AI Agents, robots, and autonomous driving become more common, edge nodes near endpoints will need integrated solutions for connectivity, local compute, and data security. These scenarios are rich and offer broad room for adoption.
End-to-end AI security and compliance services
AI workloads involve large volumes of enterprise data, user privacy, and training data. Multi-model access, cross-network data transfer, and private deployments all carry security and compliance risks. Solutions that balance high AI performance with robust security and compliance will become baseline requirements for enterprise decision-making.
In addition, vertical, industry-specific AI computing and network solutions hold significant potential:
“Financial services, education, manufacturing, and internet companies all have different requirements for compute specifications, networking strategies, deployment models, and compliance. Generic products can’t meet these needs. Industry-specific solutions will become a major growth engine.”
Looking ahead, CypressTel plans to continue deepening ecosystem cooperation with operators and data centres:
On one hand, aligning network, compute, and node resources to launch standardised, joint solutions.
On the other, co-developing industry-specific offerings for more government and enterprise customers.
At the same time, CypressTel will keep refining its three core products — AI Gateway, AI-Ready High-Bandwidth SD-WAN, and GPU Computing Services:
AI Gateway is scheduled to be fully open for commercial use in Q3 2026.
By the end of 2026, CypressTel aims to serve more than 100 enterprise customers with its AI infrastructure portfolio.
“Together with our ecosystem partners, we want to solidify the AI computing and network infrastructure layer, unlock new opportunities across the industry, and push digital and intelligent transformation forward in a steady, sustainable way,” Connee concluded.
To explore the complete Chinese article as published on Phoenix Tech, please visit: GPU像水电一样按需调用？专访赛柏特CEO张慧：90%的企业不用自己买
CypressTel uses cookies to operate this website and understand how it's used so we can improve our services. By clicking 'Accept', you consent to this use. Learn more in our Privacy & Cookie.
