AMD: "South Korea, Though It Has No Oil, Can Earn 100 Times More Than Oil-Producing Nations by Running a 'Token Factory'"

AMD, Moreh Form a "Token Factory" Alliance
"100 Times More Added Value Than Oil"
Break Nvidia's Lock-In and Target the Inference Market

AMD and South Korean AI software company Moreh have proposed a business model called a “token factory.” The strategy shifts the focus away from hardware specifications and toward lowering the cost of producing each token when running actual services. The goal is to use this approach to crack Nvidia’s dominant ecosystem.

Ramin Roane, AMD vice president of AI product application engineering, and Kangwon Jo, CEO of Moreh, jointly presented the next-generation data center business model and software optimization results at “AMD AI Developer Day Seoul 2026,” held at the Westin Seoul Parnas Hotel in Gangnam-gu on November 4.

Ramin Roane, Vice President of AI Product Application Engineering at AMD (left), and Jo Kangwon, CEO of Moreh, give a presentation at “AMD AI Developer Day Seoul 2026” at the Westin Seoul Parnas Hotel in Gangnam-gu on the 4th. Photo by Kim Jinyoung

The AI market’s center of gravity is rapidly shifting from model training to real-time inference. As a result, data centers around the world are being transformed from facilities that simply rent out computing capacity into factories for mass-producing tokens.

Roane compared the profitability of crude oil extraction in oil-producing countries with that of AI inference. “When a barrel of crude oil costs $100, an oil-producing country makes about $50 in profit. But if you use the equivalent amount of electrical energy to run inference on an open-source AI model, Kimi K3, you are left with $5,000 in net profit after cloud rental and licensing fees,” Roane said. “That is 100 times the profit made by an oil-producing country.” He added, “South Korea does not have oil as a natural resource, but it can generate 100 times more added value than an oil-producing country through data center infrastructure and software.”

Jo pointed to changing demand among corporate customers. “In the past, the core competitive advantage of data centers was the ease of GPU programming with technologies such as Nvidia’s CUDA. But customers no longer buy entire blocks of GPU computing resources; they simply request the token-generated output,” Jo said. “They do not care whether the hardware is an Nvidia GPU, an AMD GPU, or a locally developed neural processing unit (NPU). They treat it all as a single ‘black box.’” He added, “The only two things customers care about are whether the latest open-source models are supported immediately and the cost of generating tokens.”

A token is the basic unit of data that an AI language model recognizes and processes, whether a character or a word. It is used to measure the output of generative AI services and determine their usage fees.

The economic benefits of software optimization were also highlighted. According to Roane, integrating AI coding agents with AMD’s ROCm platform led to a 3.3-fold performance improvement in just six months. Meanwhile, according to semiconductor research firm SemiAnalysis, margins for Nvidia-based infrastructure in open-source inference environments were just 3% to 21%. By contrast, AMD stack-based infrastructure was found to achieve margins of around 44%.

“Our simulations show that companies selling tokens can consistently secure margins of more than 40% using AMD infrastructure when Moreh’s optimization technology is applied,” Jo said. “Companies looking to build their own AI data centers can also cut operating costs by about half compared with using external cloud-based token services.”

The two companies emphasized the importance of building open data centers that are not tied to specific hardware. “Because of the recent GPU shortage, no data center wants to rely on a single hardware vendor,” Jo said. “No single chip can optimize every workload, so Nvidia GPUs, AMD GPUs, and locally developed NPUs that may emerge in the future must be able to work together seamlessly in a single data center.”

This content was produced with the assistance of AI translation services.