Intel and Tencent Cloud Release New-Generation Cloud Instances Based on Xeon 6+, Showcasing Full Agentic AI Infrastructure Suite

Zhidongxi, November 3 report: Recently, at the 2026 Tencent Global Digital Ecosystem Summit, Intel and Tencent Cloud released a new-generation Tencent Cloud cloud instance based on Intel Xeon 6+ processors, and comprehensively showcased Intel's chip products and cloud practices.

The all-new Xeon 6+ processor has up to 288 efficiency cores, and a single processor can support the stable deployment of more than 1,000 agents, which also provides Tencent Cloud servers with cloud computing services featuring higher concurrency, lower latency, and better energy efficiency.

Supported by Xeon 6+, the comprehensive performance of Tencent Cloud's new-generation cloud instance is more than 30% higher than the previous generation, making it suitable for scenarios such as cloud-native microservices, high-concurrency Web, and agent task orchestration, further expanding Tencent Cloud's product capabilities for agentic AI workloads.

In addition, the new-generation cloud instance introduces Intel TDX (Trust Domain Extensions), which, through hardware-level isolation and attestation, builds a full-lifecycle trusted defense line of "data usable but not visible" for agents.

For the next stage of higher-density task scheduling and concurrent computing, Intel will launch a Xeon processor (codenamed Diamond Rapids) designed specifically for next-generation agentic AI infrastructure, achieving a leap in throughput for multi-task high-load scenarios.

In response to the high-concurrency, low-latency requirements that agentic AI places on infrastructure, Intel and Tencent Cloud also shared the latest practices around topics such as CPU performance, AI storage solution innovation, sandbox startup, and software ecosystem, jointly exploring the evolution path of cloud-based agentic AI infrastructure.

"Looking to the future, Intel looks forward to further deepening 'chip, cloud, and device' integrated collaboration with Tencent, jointly building a full-stack computing foundation for agentic AI, so that computing power is allocated where it is needed, intelligence is within reach, and together we promote agentic AI from technological innovation to large-scale application," said Guo Wei, Vice President of Intel Marketing Group and General Manager of Intel China.

▲Guo Wei, Vice President of Intel Marketing Group and General Manager of Intel China

1. Leveraging Intel IAA to innovate memory solutions

For example, limited memory must carry more tasks, KV Cache must flow, compress, and be scheduled efficiently across different tiers, sandboxes must be started securely and quickly, and multimodal data must also be processed efficiently.

▲Intel booth at the Tencent Global Digital Ecosystem Summit

The deeper agentic AI goes into real business, the more precious memory becomes. How to simultaneously meet the dual requirements of density and latency in high-density, high-concurrency business scenarios is a problem that software solutions cannot solve, and it is also the starting point for Tencent Cloud and Intel to advance memory application innovation.

According to the introduction, the Tencent team initially tested software solutions such as LZ4 and ZSTD. LZ4's latency was still acceptable during single-channel decompression, but after multi-channel concurrency, the maximum latency once reached 5 milliseconds, which could not meet the requirements of this business scenario. In addition, increased concurrency also brought memory bandwidth contention.

With the help of Intel IAA (In-Memory Analytics Accelerator) built into Intel Xeon processors, Tencent Cloud innovated its memory solution, introducing hardware acceleration for memory compression and decompression to meet customers' requirements for extremely latency-sensitive decompression that must be strictly controllable.

IAA can provide hardware acceleration for compression and decompression, with advantages such as low latency, low power consumption, and lower CPU usage. The Tencent team made lightweight adaptations to the kernel driver, simplified the relevant call paths, and tested different batch parallel configurations, ultimately choosing IAA-batch-8.

Compared with software solutions, after 2M-page decompression is accelerated by IAA, decompression latency has been brought down to the microsecond level, with average time accelerated by 16 times and maximum time accelerated by 9 times. While improving memory utilization, it achieves more stable and predictable responses, so that memory is no longer the performance bottleneck for high-density, high-concurrency agentic AI. This solution has been verified by Tencent Cloud and has completed business rollout.

2. Building a stable and reliable sandbox foundation, with a single machine supporting concurrent creation of 350 sandboxes

For agentic AI to operate efficiently, the execution environment must also be available on demand. As agentic AI workloads begin to be deployed at scale, another important challenge facing infrastructure is how to create and start massive numbers of sandboxes in batches within an extremely short time, and efficiently complete the orchestration and isolation of execution environments.

Based on deep collaboration between Intel Xeon processors and Tencent Cloud CubeSandbox, the two parties jointly built a sandbox foundation with high concurrency, extremely fast startup, and stability and reliability.

CubeSandbox is an elastic agent sandbox computing platform open-sourced by Tencent Cloud, carrying elastic workloads such as agent sandbox services and cloud functions. Relying on the computing performance, flexible scalability, and reliable platform capabilities of Xeon 6+ processors, CubeSandbox can efficiently support high-density sandbox workloads, achieving lower startup latency and higher resource utilization efficiency.

According to Tencent's actual measurements, while meeting the 200-millisecond SLA requirement, CubeSandbox equipped with Xeon 6+ processors can support 350 concurrent sandbox creation requests per machine, an increase of about 46% compared with other mainstream test platforms.

Facing the need to start a large number of environments simultaneously, Tencent Cloud has made considerable preparations at the software level. For example, some resources required to create sandboxes are pooled in advance, so they can be used directly when requests arrive, reducing the overhead of temporary creation; memory and disk content that have not been modified are shared as much as possible to reduce duplicate occupation.

The Intel and Tencent teams also jointly investigated the kernel execution path and merged relevant optimizations into the OpenCloudOS kernel to further reduce concurrent startup latency.

For agent services, such optimization is of very practical significance. If the environment is prepared faster, tasks can start earlier; the same machine can handle more concurrent creation requests, further improving resource utilization efficiency; and remaining stable under high load brings these capabilities closer to the requirements of a production environment.

3. Fully leveraging CPU resources to improve multimodal task efficiency

Execution environments must be ready quickly, and agents entering real business scenarios also face another type of challenge: for example, rapidly processing embodied intelligence data in batches. Agent interaction with the real world has long gone beyond text; video, images, audio, and other data are entering workflows in large volumes, placing higher demands on data processing and inference efficiency.

In response to this change, Tencent Cloud's multimodal computing engine Xpark applies the AI acceleration engine built into Xeon processors, Intel AMX (Advanced Matrix Extensions), to use CPU resources for small-model inference and lightweight multimodal tasks, thereby improving AI computing throughput and overall resource utilization efficiency.

Xpark has several key capabilities, including a very rich set of high-performance multimodal operators, heterogeneous resource collaborative acceleration, Intel AMX instruction set acceleration, and distributed implementation of single-machine operators to achieve performance breakthroughs.

In the Xpark video processing workflow, with the support of iVTAL (Intel Video Transcode Acceleration Library), video transcoding performance is improved by 1.5 times compared with open-source PyAV, and overall video processing efficiency is significantly improved.

At present, Xpark has been officially launched and is being used in actual business scenarios.

For storage requirements featuring large capacity, high performance, high energy efficiency, and easy deployment, Intel addresses issues such as the flow and reuse efficiency of KV Cache through high-speed I/O, built-in accelerators such as QAT and DSA, dedicated instructions such as UINTR+TPAUSE, and other software-hardware collaborative optimizations, enabling storage to move from fitting data in to flowing data smoothly and achieving high efficiency.

4. Building a system-level product portfolio to accelerate the large-scale deployment of agents

Intel has built a system-level solution for customers covering CPU, GPU, storage, and networking, with a product portfolio including Xeon, Ethernet controllers, and the GPU D series.

Among them, the Intel GPU D series (codenamed Crescent Island) is purpose-built for continuously running memory-intensive AI inference workloads: based on the Xe3P architecture, within a 350-watt power range, it can provide users with up to 480GB of LPDDR5X memory, 1.5TB/s memory bandwidth, and 1.3PFLOPS FP8 computing power, making large-capacity inference easier to deploy.

To truly translate hardware performance into developer productivity, an open and easy-to-use software ecosystem is equally critical. Intel insists on open-sourcing its AI software stack and prioritizing contributions to upstream communities, covering mainstream frameworks such as PyTorch, vLLM, and SGLang, as well as computing and communication libraries such as Triton, oneDNN, and oneCCL, helping mainstream models achieve out-of-the-box usability.

In addition, for cross-platform migration of existing code, Intel has introduced an AI Agent based on large language models, combined with professional Skill capabilities, to natively refactor code on the basis of understanding algorithm intent and complete the full closed loop of compilation, verification, and performance tuning, enabling cross-architecture migration to move from "running" to "running well."

At the same time, Intel extends computing power and agent capabilities from the cloud to terminals and the edge. End-side applications such as AI PCs equipped with WorkBuddy and Marvis and in-vehicle AI Boxes based on Intel Core Ultra processors were showcased together, vividly presenting the end-side intelligent experience of end-cloud collaboration and bringing agentic AI into thousands of households.

▲Intel booth at the Tencent Global Digital Ecosystem Summit

Guo Wei, Vice President of Intel's Marketing Group and General Manager of Intel China, said that from data centers to intelligent terminals, Intel and Tencent continue to expand the boundaries of cooperation, using Xeon 6/6+ to strengthen the cloud computing foundation, using agentic PCs to expand end-side AI application space, releasing full-stack intelligence potential through deep collaborative optimization of models and Agents, and relying on an open ecosystem to gather broader innovation forces, jointly connecting the key links of agentic AI from infrastructure and models to applications.

From encountering bottlenecks in software testing, to introducing hardware acceleration, and then to driver adaptation and business launch, the case of Intel's cooperation with Tencent Cloud demonstrates a complete optimization path. Facing the trend of agents moving from technical exploration to all-scenario productivity, relying on system-level innovation covering computing power, storage power, software optimization, and more, Intel is using Xeon as the computing hub and open, inclusive full-stack platform capabilities to build a high-performance intelligent computing foundation for the agentic AI era.