For the past three years, the story of AI computing power has almost entirely revolved around GPUs.
From NVIDIA's H100, H200, to GB200, GB300, and the hundred-thousand-card clusters cloud providers are scrambling to build—the entire industrial narrative tells one thing: the bottleneck of computing power is in the GPU. In this story, the CPU has long been defaulted to a less important "supporting" role, following behind the GPU, responsible for tasks the GPU doesn't want to do.
But starting in 2026, cracks are appearing in this narrative.
On June 1, Intel launched the Xeon 6+ processor in Beijing, specifically designed for cloud-native, agent AI, and network-intensive workloads. This is the first data center CPU based on Intel's 18A process.
In Intel's own description, Xeon 6+ is not a "supporting actor" for the GPU, but the "control plane" of AI infrastructure, responsible for orchestration, concurrency, and data flow.
"The path to scaling AI isn't about adding more components, but about the coordinated operation of the system," said Kevork Kechichian, Intel Executive Vice President and General Manager of the Data Center and AI Group, during a briefing. "As AI moves into the agent era, orchestration, concurrency, and data flow have become the new limiting factors.
This reinforces a core truth: the CPU remains the control plane of modern AI infrastructure."
This isn't just Intel's judgment. In February, independent semiconductor research firm SemiAnalysis released a 2026 data center CPU landscape report titled "The CPU Comeback," offering a similarly direct assessment. In the current widespread deployment of AI training and inference, CPUs are being needed again in a way fundamentally different from the past three years.
However, this "comeback" needs to be examined closely. It's not about the CPU reclaiming the leading role, but about the CPU being redefined in a new position.
I. The Cracks in the GPU-Centric Narrative
To understand why the CPU is "coming back," we must first look at the changes happening within AI workloads themselves.
Over the past two years, the mainstream narrative of AI computing power has been about training. The scale of large model training has been increasing by four to ten times annually. Training requires massive parallel computing, and for this, the GPU is the absolute star. But training is not the entirety of AI workloads.
According to Intel's assessment during the briefing, the entire landscape of AI computing workloads can be roughly divided into three categories:
The first category is foundational workloads. Storage, databases, web services, microservices, CDN—these are not AI, but they are the underlying services needed for AI to run. This remains the traditional stronghold of CPUs.
The second category is training. The training of cutting-edge large models relies almost entirely on GPUs and dedicated accelerators. This has been the battleground for the past three years.
The third category is inference and agents. This part is growing rapidly, and is significantly different from training.
The key difference in the third category lies in the nature of the workload itself. Training is the process of "calculating" a model from scratch, with extremely high parallelism and demand for peak single-point computing power. But inference and agents are not—they involve deploying already-trained models to run in real business environments.
This means a significant portion of the work isn't "calculation," but orchestration: scheduling multiple models to collaborate, managing context, coordinating data flow between different agents, handling concurrent user requests, ensuring predictable latency.
These are things GPUs are not good at.
"In this scenario, we see workloads that incorporate GPU-level acceleration, but the main body remains centered around traditional CPUs," Kevork Kechichian said during the briefing.
Behind this lies a more concrete industrial fact. In its "CPU Comeback" report, SemiAnalysis cited an example: In Microsoft's "Fairwater" data center built for OpenAI, a 48-megawatt CPU and storage building supports a 295-megawatt GPU cluster.
In other words, to make that 295-megawatt GPU cluster actually run, thousands of CPUs are needed alongside to handle the petabyte-scale data streams generated by the GPUs, schedule tasks, and manage storage.
The higher the computing power of the GPU is pushed, the greater the "peripheral computing power demand" it generates. And this peripheral computing demand ultimately falls on the CPU.
Therefore, the CPU's comeback is not about "the CPU becoming faster than the GPU again." It's about the form of AI computing power expanding from "training one large model" to "running thousands of agents," and thus orchestration and data flow have re-emerged as bottlenecks. GPUs cannot solve this; CPUs can.
This is the other side of the AI narrative that has been overlooked for the past three years.
II. What Path is Xeon 6+ Betting On?
Intel's bet is reflected in the product definition of Xeon 6+.
The most striking number is up to 288 cores, all of which are Efficient-cores (E-cores).
E-cores and P-cores (Performance-cores) represent a fork in Intel's CPU architecture strategy in recent years. P-cores pursue ultimate single-core performance, the traditional design goal for server CPUs. E-cores are efficiency-focused, with somewhat weaker single-core performance but smaller size and lower power consumption, allowing more cores to be packed onto the same die area.
Xeon 6+ pushes this fork to the extreme. 288 Efficient-cores means Intel is betting not on "how fast each core is," but on "how many cores can be packed onto one CPU."
The logic behind this product definition is: Agent AI workloads are not about how fast a single core can run, but about whether thousands of lightweight tasks can run simultaneously. When a single server needs to orchestrate hundreds of agents, handle thousands of inference requests, and maintain tens of thousands of concurrent connections, the throughput capacity of 288 E-cores is far more important than the single-core performance of 64 P-cores.
This is a product definition that goes against the mainstream. For decades, the mainstream narrative for server CPUs has been about competing on single-core performance—higher clock speeds, stronger IPC, larger caches. The E-core path essentially acknowledges: that narrative may be coming to an end.
But several things must be considered together.
First, the E-core path is not unique to Intel. AMD launched Bergamo in 2023, based on density-optimized Zen 4c cores. AWS's Graviton series and Ampere's AmpereOne series have long followed the "high-density cores + efficiency first" path. In Ampere's 2024 Aurora roadmap for AmpereOne, the core count has already reached 512.
In other words, Xeon 6+ represents Intel catching up to an existing industry trend—Intel is not the leader here, but a player rejoining the mainstream direction.
Second, Xeon 6+ being the first data center CPU on the Intel 18A process might be more important within Intel's own context than the "288 E-cores."
Intel 18A is Intel's biggest bet in recent years. Its significance extends beyond a single CPU to whether Intel Foundry, Intel's contract manufacturing business, can establish itself. If the 18A process cannot deliver a competitive product, the Intel Foundry story falls apart.
Xeon 6+, built on 18A, pushing E-core count to 288, and publicly claiming "industry-leading performance density," is one of Intel's report cards to the market. Whether it will be recognized by the market, and whether it can hold its ground against TSMC's N2 and Samsung's 2nm in the same-generation competition, is another question.
Third, several significant names appear on Xeon 6+'s customer list—Ericsson is testing 5G core networks with Xeon 6+, and T-Systems, under Deutsche Telekom, is building private agent AI infrastructure with Xeon 6+. Both are traditional, steady-state purchasers of data center CPUs, and their procurement choices are a market signal in themselves.
Putting these three things together, Xeon 6+ is betting on this path: Leverage the 18A process for power efficiency advantage, use 288 E-cores for core density, and target the "high-density, high-efficiency, high-throughput" type of workloads in AI inference and agent scenarios.
This is not a story about the CPU returning to the main stage of computing power, but about the CPU finding a new position.
III. Is This Narrative Valid?
Is the "CPU comeback" story Intel is telling actually valid? It depends on several other variables in the industry.
The first variable is the reaction from GPU vendors.
NVIDIA has also been working on "orchestration"-related things in the past two years. The Grace CPU + Hopper GPU combination itself is NVIDIA filling the CPU role. If GPU vendors mainstream their own integrated "CPU + GPU" solutions, the position of independent CPU vendors could be squeezed. This is the biggest opponent to Intel's narrative of "CPU as the control plane"—not AMD, but NVIDIA itself.
The second variable is the trend of cloud providers developing their own CPUs.
AWS Graviton is already deployed at scale within AWS's own data centers, handling a significant portion of AWS's general-purpose compute workloads. Microsoft is developing Cobalt, Google is developing Axion, Alibaba is developing Yitian—almost all major cloud providers are developing their own ARM-based server CPUs.
These custom CPUs also follow the "high-density, efficiency-first" path—placing them in direct competition with Xeon 6+ in terms of product definition.
This means that the market Xeon 6+ aims to capture, cloud providers are making for themselves. Intel needs to prove there is still a large enough market outside of cloud providers' custom CPUs—for example, with telecom operators, private clouds, and vertical industry data centers.
The third variable is the 18A process itself.
Xeon 6+ being the first data center CPU on Intel 18A means this chip carries industrial significance far beyond the product itself. If the 18A process encounters issues in mass production yield, performance stability, or customer validation, Xeon 6+'s market performance will suffer. Conversely, if 18A performs well, Xeon 6+ could bring some breathing room for Intel Foundry.
But 18A doesn't operate in a vacuum—TSMC's N2 process will begin volume production in the second half of 2026, and Samsung's 2nm is also on the way. What Intel 18A aims for is not just "making it work," but "being competitive after making it work", which is a higher standard.
Combining these three variables, the ultimate success of Xeon 6+ depends not only on itself, but also on whether NVIDIA will absorb the CPU role, whether cloud providers will continue with custom CPUs, and whether Intel 18A can stand its ground in its same-generation competition with TSMC and Samsung.
This is why, while the "CPU comeback" is valid as an industry-level judgment, it remains uncertain whether Intel itself can capture the benefits of this resurgence.
The battle for the CPU's position on the AI computing power stage has been ongoing for three years.
The script for the past three years has been "GPU is the center, CPU is the support." This script began to loosen in 2026—not because CPUs are becoming faster than GPUs again, but because AI computing itself is changing. As AI expands from "training one model" to "running thousands of agents," orchestration, concurrency, and data flow have re-emerged as systemic bottlenecks, and the CPU becomes indispensable in this position.
Intel has bet on this, and Xeon 6+ is its answer. But whether this will hold true, and whether Intel itself can reap the benefits, will ultimately be answered in customer data centers in 2027 and 2028. AMD, the ARM camp, cloud providers' custom CPUs, and NVIDIA making its own CPUs—each variable could change the direction of the script.
The CPU's comeback is real, but who will lead it is yet to be determined.
