AMD Finalizes MK1 Acquisition to Boost High-Speed AI Inference

• The MK1 team has joined the AMD Artificial Intelligence Group to enhance performance across AMD’s AI hardware and software stack.
• MK1’s engine, Flywheel, processes over 1 trillion tokens per day and is deeply optimized for AMD Instinct GPUs.
• The deal follows earlier AI-related acquisitions (including Nod.ai and Silo AI) as AMD builds a full-stack alternative to NVIDIA’s ecosystem.

On November 10, 2025, AMD officially announced the closing of the deal and welcomed the MK1 team into its Artificial Intelligence Group. According to AMD, the goal is to improve  performance in enterprise AI especially as more customers look for efficient ways to run LLMs on Instinct GPUs.

AMD did not disclose the purchase price, but the company positioned this acquisition as another strategic software investment, similar to its earlier deals with Silo AI and Nod.ai. All of these acquisitions are meant to strengthen AMD’s growing AI ecosystem and help it better compete with rivals that already have integrated hardware-software platforms.

AMD has been steadily building a full-stack AI platform that spans both hardware and software. Its Instinct accelerators, including the MI300X and upcoming MI400 series, serve the hardware layer, while acquisitions like Nod.ai, Silo AI, and now MK1 strengthen the software foundation.

Bringing MK1 on board gives AMD three major benefits:

• It boosts the company’s ability to deliver high-speed inference and reasoning at scale, an area AMD explicitly highlighted as critical for its roadmap.
• It makes Instinct GPUs more competitive by pairing them with software that can fully unlock their bandwidth and memory advantages, improving performance per dollar against NVIDIA for LLM serving.
• It allows AMD to offer customers a more integrated “chips + software” platform, reducing deployment friction for enterprises that want a ready-made stack rather than building one themselves.

Who Is MK1?

MK1 is a specialized AI company focused on high-speed inference and advanced reasoning. Its main product is an inference engine called Flywheel that processes over 1 trillion tokens per day in live deployments, along with additional “comprehension” engines designed for more complex reasoning tasks.

MK1 was founded by AI veterans including Paul Merolla, who was part of the founding team at Neuralink before shifting his focus to high-performance inference software. Even before the acquisition, MK1’s technology was already used on large AMD Instinct MI300X cloud clusters through partners such as TensorWave, where it supported extremely high-volume inference workloads.

What MK1’s Technology Does

MK1 builds software that makes GPUs more efficient at the serving side of AI, a quick and cost-effective system that can answer user prompts once a large model has been trained.

Flywheel is heavily optimized for AMD’s Instinct GPUs. It takes advantage of the MI300X memory architecture to push out more tokens per second and improve overall throughput making it well-suited for applications like chatbots, copilots, and large-scale LLM APIs that require high-speed & low-latency responses.

The company also emphasizes traceable reasoning, offering tools that record how models responds, a growing requirement in enterprise environments with strict governance or compliance needs. Although the software is most deeply tuned for AMD hardware, MK1 also supports multiple GPUs and mixed environments, giving enterprises flexibility as they scale.