Samsung Electronics is developing a PC AI accelerator called "GAIA," the Korea Economic Daily reported on July 9. According to the report, GAIA is manufactured using a 4nm process, and Lenovo and HP are reportedly verifying the performance of prototypes. Mass production could begin as early as 2027, and the report also mentioned that cooperation with Processing-in-Memory (PIM) technology, which Samsung has been advancing, is under consideration. However, the computational performance and power consumption needed to judge its capability as an AI PC chip remain unclear. The connection method and software support are also unknown at this point.
Prototypes to Lenovo and HP, Mass Production Possibly by 2027
According to the Korea Economic Daily, GAIA is being developed by Samsung's System LSI division, part of its semiconductor business. The chip is reportedly designed around a Neural Processing Unit (NPU) core, aimed at efficiently executing generative AI processing on PCs. The report states that prototypes have already been delivered to Lenovo and HP, with both companies verifying performance.
The delivery of prototypes indicates that development has moved from the lab stage to evaluation by PC manufacturers. However, the start of evaluation does not necessarily mean product adoption. PC makers weigh speed and battery life against heat generation and board space as conditions for finished products. Component cost and driver stability also factor into selection. The "as early as 2027" mass production timeline mentioned in the report also presumes that customer adoption and verification proceed as planned.
Samsung has not officially announced GAIA at this time. While the 4nm manufacturing process has been confirmed, the TOPS figure representing NPU performance and the thermal design power remain unknown. The supported numerical precision, memory capacity, and bandwidth are also unclear. The report does not reveal which product lines Lenovo and HP are evaluating the chip for, or whether the prototypes reflect a design close to the final product.
A Path to Add AI Compute Separate from Integrated NPUs
Current Copilot+ PCs are centered around configurations that use NPUs built into CPUs or SoCs from Intel, AMD, and Qualcomm. Integrating the NPU with the CPU and GPU while sharing memory makes it easier to reduce component count and data transfer. For PC makers, choosing a processor also determines the AI compute capability together, simplifying design.
The reports on GAIA describe it as a dedicated AI accelerator separate from existing processors. However, it has not been disclosed whether it will be implemented as an independent chip on the motherboard, whether multiple dies will be packaged together, or whether connections will use something like PCI Express. Without knowing the implementation method, it cannot be assumed to simply be "an NPU that can be added to existing PCs."
If offered as an independent component, PC makers would not be constrained by the NPU performance built into the CPU and could add AI computing capability on a per-product basis. If local AI capability can be enhanced without choosing an expensive high-end CPU, this could expand adoption to mid-priced PCs as well. On the other hand, an additional chip requires board space along with power circuitry and cooling mechanisms, increasing component costs. If data must travel back and forth between the accelerator and system memory, the benefits in speed and power efficiency would also be diminished. Which price segment GAIA ends up in will depend more on overall system design than on peak performance alone.
A Mechanism to Reduce Data Movement, Ahead of Performance
If the reported PIM cooperation is realized, GAIA would bring memory-side AI processing—which Samsung has been developing—into the PC space. PIM places computing functions inside or near memory, reducing the number of times data must travel between the processor and memory. Because AI models frequently read and write weights and intermediate data, memory transfer can push up processing time and power consumption even when the computation itself is fast.
In 2021, Samsung announced HBM-PIM, which incorporates PIM, along with AXDIMM, which places an AI engine on a DRAM module, and LPDDR5-PIM for mobile use. While the applications and implementations differ, the goal is common: bringing computation closer to where data resides to reduce the time and power spent on data transfer.
That said, performance gains achieved by past PIM products cannot simply be applied to GAIA. How GAIA will connect to the low-power memory used in PCs, and how processing will be divided between the NPU and PIM, have not been disclosed. Which processes can leverage PIM also depends on the supported operations and data placement. To confirm GAIA's real value, both the chip's standalone TOPS figure and the speed and system power consumption when running actual AI models will be needed.
Beyond 40 TOPS Lies Windows Compatibility
Microsoft has set 40 TOPS or higher NPU performance as the benchmark for Copilot+ PCs. Whether GAIA will exceed this threshold remains unknown. Furthermore, TOPS represents a peak value of operations that can be processed per unit time at a specific numerical precision; actual speed can vary with the same figure depending on the supported AI models and memory bandwidth.
To use an NPU from applications on Windows, hardware-matched drivers and runtimes are also required. Microsoft's Windows ML combines ONNX Runtime with an Execution Provider (EP) optimized for each chip. Official documentation updated on April 24, 2026 lists Intel's OpenVINO, Qualcomm's QNN, and AMD's VitisAI as EPs for NPUs. An EP for Samsung-made NPUs is not on the current published list.
There is a path for Samsung to distribute its own EP, but gaining adoption by PC makers and application developers would require preparing drivers and model conversion tools. Supported operations and update methods would also need to be established. If it can be distributed via Windows Update as a Microsoft-certified EP, adoption would be easier. Without reaching that point, individual support would need to be added for each app using GAIA's performance, potentially slowing hardware adoption.
The next milestone for GAIA will be an official announcement and specification disclosure from Samsung. Whether it exceeds 40 TOPS and at what wattage it operates will be key judgment factors. Memory connection method and Windows ML compatibility are also essential. If Lenovo or HP then announces adoption for a 2027 model, GAIA can be judged to have progressed from a prototype-stage concept to an AI foundation for PCs.