It has become a completely different company.
Bae Tae-won, president of Intel Korea, described Intel this way after Lip-Bu Tan took office as chief executive officer (CEO) in Mar. last year. Meeting at Intel Korea's headquarters in Yeouido, Seoul, on the 28th, Bae stressed that the organizational structure, workforce, and business direction have all changed significantly, saying, "In the era of agentic AI, physical AI, and hybrid AI, the biggest question for Intel now is what role the central processing unit (CPU) should play."
Bae said the growing importance of CPUs in artificial intelligence (AI) infrastructure was a fully anticipated trend. He said, "As the bottleneck in Generative AI began to emerge from the graphics processing unit (GPU), it was a natural progression that the next would be memory, followed by the CPU's orchestration role tailored for the agentic AI era."
As AI spreads from the cloud to the edge and further into Robotics, the CPU's role is growing even larger. Bae explained that for a Humanoid Robot, installing a discrete GPU is virtually impossible due to power consumption and expense constraints, making an architecture where the CPU orchestrates the integrated GPU and the Neural Processing Unit (NPU) the key. In fact, Intel's laptops equipped with the latest-process Panther Lake products have begun to be used to run local large language models (LLMs) and as robotics reference boards, and lightweight LLM inference is possible without a GPU by using the AI engine (AMX) of Intel Xeon 6.
Hybrid AI, cited as an alternative to the massive expense incurred in running Generative AI, is also being led by Intel. Bae said, "Corporations and consumers using Generative AI services are beginning to suffer such a heavy expense burden from paid services that phrases like 'the tokens are melting' are coming up," adding, "There is growing demand to balance things by handling general workloads with local LLMs and using online services for heavy tasks, and Intel's 'Superclaw,' to be introduced at Computex 2026, could be an alternative."
Another key challenge in this era of hybrid AI is security. Bae introduced Intel's confidential computing technology, TDX (Trust Domain Extensions), saying, "As AI becomes routine, the exposure risk of personal information and corporate confidential data also grows." TDX manages data directly at the CPU hardware level, not at the software virtual machine level, and is currently undergoing PoCs (proofs of concept) with several corporations, including domestic telecom companies. Bae emphasized, "If we do not properly design AI infrastructure now, it may be too late to change the frame later."
Bae also mentioned Intel's organizational changes. Since Lip-Bu Tan became CEO, the decision-making process has been compressed to five stages and the number of executives has been drastically reduced, while a strong push is underway to shift to a customer-centric culture. The following is a Q&A with Bae.
— Intel is said to have changed a lot since Lip-Bu Tan took office as CEO. What changes are you sensing?
It feels like a completely different company. There have been many changes organizationally and personnel-wise, and I assess those changes positively. The crux is that the company's direction has moved away from the traditional PC- and server-centric CPU toward the question of what role the CPU will play in the era of agent AI, physical AI, and hybrid AI. On the organizational side, the decision-making steps have been compressed to five, and the number of executives has been drastically reduced. If, in the past, Intel was in a position to give guidelines to customers or partners, now there is even feedback that we have become a company that more actively listens to the voices of customers and partners.
— Was the AI agent beneficiary effect, cited as the biggest driver of Intel's recent stock surge, an expected tailwind or a surprise?
We naturally expected it. As the bottleneck of Generative AI began to emerge from the GPU, it was a natural sequence that the next would be memory, followed by the CPU's orchestration role. Two years ago at the AI Summit, when Intel highlighted the linkage between high-bandwidth memory (HBM) and CPUs by presenting SK hynix as a premium partner, even internal employees asked, "What do SK hynix memory and we have to do with each other?" Now, of course, the relationship is recognized. The hybrid AI and physical AI markets, in fact, have not even truly opened yet.
— If you were to explain the role differences among the GPU, NPU, and CPU in simple terms?
In the end, it's about computation. The GPU is specialized in handling simple calculations with a vast number of cores. Its downside is high power consumption. The NPU is a dedicated chip that processes specific computations with power efficiency. The CPU takes a comprehensive view of all these resources, determining where the bottleneck is and which parts it will handle, and orchestrates them. A system can run without a GPU or without an NPU, but nothing runs without a CPU. That is why, as the inference market grows, the ratio in data centers—once one CPU per eight GPUs—is narrowing to as tight as one-to-one.
— How will roles be divided between the CPU and GPU in the agentic AI era?
Looking at AI's development stages, it has evolved from traditional AI to Generative AI to agentic AI. Traditional AI could be handled by CPUs, and in the Generative AI era, GPUs took on a larger role due to large-scale model training. Now, moving into agentic AI, the CPU's role is growing again. Agentic AI is autonomous AI in which AI judges and acts on its own without human commands, and in this stage, role-sharing between the GPU and CPU is essential. The GPU performs large-scale parallel computation and inference tasks, while the CPU is responsible for orchestration that manages the actions of numerous agents. Task scheduling, memory management, and API calls are all the CPU's domain.
In practice, many corporations' AI inference tasks can be sufficiently handled by CPUs without GPUs. Intel Xeon 6, thanks to its built-in AI accelerator AMX, delivers performance comparable to entry-level GPUs like Nvidia's L4. In MLPerf tests, dual-socket Xeon 6 recorded more than double the throughput of the L4 in image recognition and more than four times in speech recognition. Ultimately, in an agentic AI environment, system performance is determined less by the specs of a single accelerator and more by balanced use of CPU and GPU resources.
— The expense of running GPU-based Generative AI models is also growing.
That's right. In response, Intel is making changes not only in hardware but also in software. The 'Super Claw' introduced at this Computex is a representative example. If you try paid, cloud-based services like ChatGPT, Gemini, or Claude, the expense burden is heavy—hence the phrase "the tokens are melting." Thus, there is emerging demand to balance by processing general workloads with local LLMs and using online services for heavy tasks, and Intel is providing such software solutions as well.
— AI PCs have not shown the explosive momentum expected. Where will consumers first experience agentic AI in everyday life?
I think the first everyday touchpoints for agentic AI will be smartphones and PCs. On the way to work, scenarios will become reality where a smartphone proactively informs you before you ask, such as "Fine dust is bad today; bring a mask," or "If you leave now, you'll arrive 10 minutes earlier; shall I start navigation?" On PCs as well, agentic functions will be integrated at the operating system level, preparing email replies, document organization, and schedule management before the user instructs them. The current AI PC concept is still in an early stage, but as agents in PCs and smartphones increasingly learn users' work and life patterns and begin to support them proactively, we expect this will lead to a new paradigm in which AI works and thinks alongside humans.
— How can security issues be resolved?
As AI becomes routine, the exposure risk of personal information and corporate confidential data also grows. Intel has a confidential computing technology called TDX (Trust Domain Extensions), which manages data directly at the CPU hardware level rather than at the software virtual machine level. We are currently conducting PoCs with several corporations, including domestic telecom companies. If we do not properly design AI infrastructure now, it may be too late to change the frame later. That is because there are not many corporations that have the full lineup spanning CPUs, manufacturing capability, ecosystems, and security.