Author: Fang Yuan
If an enterprise prepares to deploy 1,000 agents, what else does it need besides purchasing GPUs?
The answer may be more complex than imagined. Each agent requires an execution environment, memory to save running states, storage to record context, and a CPU to coordinate tasks and invoke tools. As the number of agents increases, resource consumption that was previously negligible can become a major expense for data centers. This reflects a shift occurring in the AI industry. Over the past two years, competitive focus has centered on large models and GPU computing power. However, as AI transitions from answering questions to autonomously executing tasks, infrastructure pressure has expanded from simple model computation to include execution environments, memory management, data storage, and transmission.
On October 29, during the Intel session at the 2026 Tencent Global Digital Ecosystem Conference, Intel and Tencent Cloud showcased multiple collaborations involving Xeon processors, agent sandboxes, memory compression, and KV Cache storage. On the same day Intel unveiled these technologies, it also released its Q3 2026 financial results. Company revenue reached CNY 16.5 billion, up 21% year-over-year; specifically, Data Center and AI (DCAI) revenue hit CNY 7.3 billion, surging 78% year-over-year, making it the fastest-growing business segment.
01、Xeon6+ Lands on Tencent Cloud, Sandbox Concurrency Becomes Key Breakthrough
At this conference, Intel Xeon 6+ processors were deployed for the first time in Tencent Cloud's new generation cloud instances.
Xeon 6+ belongs to Intel's Clearwater Forest product line, based on the 18A process node, featuring up to 288 efficient cores (E-cores), with a focus on high-density computing and energy efficiency. Tencent Cloud launched new CVM cloud instances based on this processor. According to data disclosed on-site, overall performance improved by over 30% compared to the previous generation, targeting scenarios such as cloud-native applications, high concurrency, and agent task orchestration.
Beyond processor performance, the collaboration between Intel and Tencent Cloud regarding CubeSandbox is particularly noteworthy. Unlike traditional chatbots, agents not only generate content but also execute code, invoke tools, and access files. These operations pose potential security risks, necessitating isolation via independent execution environments. Tencent Cloud's open-source CubeSandbox is an elastic agent sandbox computing platform designed for this need. It utilizes MicroVMs combined with snapshot and resource reuse technologies, allowing multiple sandboxes to share base images, unmodified memory pages, and file caches, consuming additional resources only for changed parts.
For example, if 3,000 virtual machines are created from the same image, and each modifies only 2MB of memory, the incremental memory usage would be approximately 6GB given sufficient sharing of base resources. This approach reduces resource redundancy in traditional execution environments and improves single-machine deployment density. Furthermore, both parties jointly optimized Cgroups, network card resource pools, and the OpenCloudOS kernel execution path. Test results published on-site showed that while meeting a 200ms SLA requirement, a single machine equipped with Xeon 6+ running CubeSandbox could support 350 concurrent sandbox creation requests, an improvement of about 46% over the control platform. Meanwhile, even under high load conditions, the P99 latency for sandbox creation remained stable, with no significant deterioration as concurrency density increased. Thus, the value lies not just in increasing CPU core counts, but in improving actual business execution efficiency through joint tuning of processors, operating systems, and sandbox software. For cloud service providers needing large-scale agent deployments, the speed and stability of execution environment creation directly impact server utilization and deployment costs.
02、IAA and QAT Play Distinct Roles, Memory and KV Cache Become New Bottlenecks
If sandboxes address where agents work, memory and storage address where the data generated during their work is placed. The longer an agent runs, the more context and task state it retains. Especially when multiple agents execute tasks simultaneously, memory capacity easily becomes a limiting factor for server deployment density. To address this, Intel and Tencent Cloud demonstrated a memory compression solution based on IAA.
IAA, or Intel Advanced Matrix Extensions Accelerator (note: typically referred to as In-Memory Analytics Accelerator in this context, translating strictly as Intel Memory Analysis Accelerator per source implication of function), is a dedicated hardware acceleration unit integrated within Xeon processors capable of handling specific compression and decompression tasks, reducing the computational burden on general-purpose CPUs. Tencent adopted the OS Zswap mechanism to compress inactive memory pages, testing with 2MB huge pages. Previously, using software solutions like LZ4 or ZSTD often led to memory bandwidth contention during multi-path concurrent decompression, causing increased latency. Ultimately, both parties selected the IAA-batch-8 configuration and performed lightweight adaptations for drivers and kernel paths. According to published data, compared to software compression schemes, IAA improved average processing speed by 16x and maximum processing speed by 9x. In a comprehensive stress test with 256 concurrent paths, the IAA solution achieved an average latency of 56 microseconds and a maximum latency of 564 microseconds. This indicates it can improve memory utilization while minimizing the impact of decompression on business response times. For cloud platforms running large numbers of agents simultaneously, memory compression is not just about saving space, but also about maintaining business latency.
Another area worth noting is KV Cache. KV Cache is a critical mechanism for storing context calculation results during large model inference. In traditional Q&A scenarios, once the model answers, the interaction ends. However, agents may continuously execute multiple steps, repeatedly searching, reading files, and invoking tools, causing context to accumulate and KV Cache to grow larger.
If all cache resides in GPU VRAM, capacity and cost pressures arise. Therefore, tiered management of KV Cache across GPU VRAM, CPU memory, and SSDs is becoming an important direction for inference system optimization. The solution demonstrated by Intel employs a compute-storage separation architecture, connecting GPUs, CPUs, and backend storage via RDMA. Combined with self-developed KV Shrink hierarchical KV software and the QAT accelerator built into Xeon processors, it achieves tiered storage, transmission, and compression of cache data. It is necessary to distinguish between IAA and QAT. IAA primarily targets memory page compression, while QAT mainly handles data compression and decompression acceleration tasks; they are not the same technology.
In tests using the Qwen2.5-32B-Instruct model, with 8,000 input tokens, 128 output tokens, and a 64GB DDR memory limit, when KV Cache achieved 100% hits on backend high-performance storage, Time to First Token (TTFT) decreased by 90% compared to the native vLLM scheme without cache reuse. With concurrency levels of 8, 16, and 32, throughput increased by more than 5x, 6x, and 8x respectively. Additionally, enabling QAT compression resulted in no obvious change in overall throughput, but available capacity increased by nearly 20%.
03、GPUs, Software Ecosystem, and Edge Intelligence: What Is Intel's Next Card?
Beyond Xeon processors and built-in accelerators, Intel also introduced its GPU product layout for agent inference. Of particular note is the GPU D-series, codenamed Crescent Island. This product is based on the Xe3P architecture, supporting up to 480GB LPDDR5X memory, with memory bandwidth reaching 1.5TB/s, FP8 compute power up to 1.3PFLOPS, and power consumption around 350W. Unlike traditional GPUs emphasizing peak compute power, Crescent Island focuses more on VRAM capacity and energy efficiency. Large-capacity memory can accommodate more model parameters and KV Cache, suitable for long contexts, multi-user concurrency, and continuously running inference workloads. Intel stated the product positioning supports running models with up to 300 billion parameters on a single card, and models exceeding 1 trillion parameters on an eight-card system. Currently, the product has entered the customer sampling stage.
This reflects another strategy for Intel in the GPU market: seeking differentiation in the fiercely competitive compute market through large VRAM, relatively lower power consumption, and existing air-cooled data center deployment conditions. However, competition in the GPU market depends not only on hardware specifications but also significantly on the software ecosystem. Intel is lowering the cost of migrating CUDA projects to Intel platforms through open software technologies like SYCL and SYCL-TLA, as well as AI Agent-assisted code migration tools. In experiments displayed on-site, 10 different projects completed initial migration using large models and specialized Skills, though further operator optimization was still required.
In the terminal market, Intel is also advancing hybrid edge-cloud intelligence. The conference showcased WorkBuddy and Marvis local agent solutions in collaboration with Tencent. Marvis provides nearly 20 local skills, while WorkBuddy integrates over 30 local or hybrid experts, with some tasks accelerated by local models and OpenVINO. Thus, Intel's AI layout is not limited to data centers but aims to extend model inference, task execution, and privacy protection capabilities to AI PCs. With the cloud handling complex reasoning and terminals managing low-latency, high-privacy tasks, this hybrid model opens new application spaces for PC processors, integrated GPUs, and edge-side AI software.
04、NVIDIA and AMD Are Also Doubling Down
Intel is not the only vendor recognizing this market shift. NVIDIA and AMD have recently been restructuring their product portfolios around agents, expanding beyond single GPU products to include CPUs, storage, and system architectures.
In addition to Intel, NVIDIA also shared insights specifically focused on AI infrastructure at this conference. NVIDIA introduced five key network infrastructures for AI factories: Scale-Up, Scale-Out, Scale-Across, Context Memory, and Scale-In, corresponding to high-speed GPU interconnects, server cluster expansion, cross-data-center connectivity, context storage, and infrastructure management, respectively. Among these, Context Memory is particularly noteworthy. As agent tasks accumulate context and KV Cache scales grow, NVIDIA aims to enable different compute nodes to share and access cache data via network and storage architectures, reducing redundant computation. This aligns with Intel's direction of optimizing KV Cache through QAT and KV Shrink, although NVIDIA places greater emphasis on network interconnection and cluster-level resource management. The introduction of technologies like BlueField-4 and DOCA indicates that DPUs are evolving from traditional data transfer and offload tools to further participate in data access, security, and management within AI infrastructure.
AMD brought another notable direction: synergy between edge AI and storage. Together with partners like Longsys and Liulian Intelligent, it showcased an agent host solution based on Ryzen AI Max+ 395, combining Longsys' AIDIMM high-bandwidth memory, iSA storage intelligent agent, and AISSD to address memory capacity and data scheduling issues when running large models locally. Specifically, AIDIMM uses LPDDR5X chips with a 256-bit width design, while iSA coordinates data flow between DRAM and SSD through dynamic MoE expert offloading, cache tiering, and data prefetching. Joint tests published by Longsys showed that a 64GB AIDIMM configuration supports local running of 70B, 80B, and 122B level models. Related solutions extend to the AMD Ryzen AI Halo developer platform equipped with AISSD and iSA. The significance of this design is that it no longer requires all model data to reside in DRAM, but rather allocates data to different storage tiers based on runtime needs. For edge devices with limited memory capacity, this offers new ideas for running larger scale models.
From cloud-side KV Cache management to edge-side memory and SSD synergy, it is evident that storage is transforming from traditional peripheral hardware to a critical component influencing AI inference efficiency. NVIDIA focuses on high-speed connections and cache sharing within data centers, while AMD and its ecosystem partners explore breaking through edge device memory limits via storage tiering. These two distinct paths actually point to the same trend: competition in AI infrastructure is expanding from compute chips themselves to aspects like memory, storage, and data flow.
05、Conclusion
From the displays at this Tencent Global Digital Ecosystem Conference to Intel's latest Q3 financial results, a noteworthy change is visible: data centers are becoming a key growth source for Intel again, and agents may further expand demand for infrastructure such as server CPUs, memory, and storage.
For the entire semiconductor industry, the changes brought by agents may be more profound than a simple round of GPU upgrades. When AI begins to continuously execute tasks, determining infrastructure costs is no longer just about how many tokens a model generates per second, but also includes execution environment creation speed, context storage costs, data transmission efficiency, and overall system energy utilization.
In the past, the semiconductor industry competed on who could provide stronger computing power; in the era of agents, competition is shifting to who can convert computing power into more real-world work outcomes at lower system costs.