HP Cuts Local AI Costs by 45%

HP says it moved 40–50% of internal AI workloads from the cloud to workstations and AI stations, reducing related costs by 45% and recouping its equipment investment in 12 months. As agentic AI increased token costs, HP adopted local AI, which runs AI close to where data is produced. The company is using this experience to pursue South Korea’s semiconductor, medical and broadcast markets.

Jim Nottingham, senior vice president of HP’s Advanced Compute Solutions (ACS) division, said AI assistants had greatly reduced work time, while token consumption continued to rise. HP’s analysis found that smaller models could produce the same quality as frontier models. Running workstation-class small models on company equipment increased inference speed and removed token-use limits. A company development team that had switched to per-token private cloud pricing also adopted local equipment, allowing 30 to 80 developers to use AI agents without restrictions.

Nottingham cited data privacy and sovereignty, low latency, data gravity (the tendency for data to attract applications and services), and cost as reasons companies operate AI outside the cloud. Demand is particularly strong in manufacturing. Global automaker A wanted to install AI inference equipment on its production line, but large workstations with many GPUs were difficult to place because of heat and noise. HP proposed ZBoost, which lets multiple devices share workstation GPUs. The large equipment stays in the back office while smaller devices are placed on the line; automaker A is moving toward company-wide deployment after a proof of concept.

Automaker B used ZBoost to avoid reloading a large rendering model each time, raising design productivity by 570% and cutting total cost of ownership (TCO) by 23%. Another customer uses about 300 workstations that would otherwise sit idle after work to fine-tune AI models about 1,000 times each night, reducing cloud costs. Nottingham said the key is a hybrid approach that combines local and cloud computing where each is appropriate.

HP expects demand for local AI to grow in South Korea, particularly in industries such as semiconductors, automotive and shipbuilding, where design and process data are a source of competitive strength. It cited semiconductor inspection equipment, semiconductor factory automation and logistics systems, and dental and medical imaging equipment as local use cases. Broadcasters are adopting more augmented reality (AR) solutions, and HP Z workstations were used for a live broadcast solution that ran for more than 12 hours during this year’s nationwide local elections. HP also plans to strengthen support for professionals doing demanding work, including architects and engineers.

Nottingham said his goal during the South Korea visit was to listen to and learn from customers and partners, and that he would support HP Korea’s customer success globally. He previously led HP Printing Korea and oversaw integration after HP acquired Samsung Electronics’ printing business.

HP says it moved 40–50% of internal AI workloads from the cloud to workstations and AI stations, reducing related costs by 45% and recouping its equipment investment in 12 months. As agentic AI increased token costs, HP adopted local AI, which runs AI close to where data is produced. The company is using this experience to pursue South Korea’s semiconductor, medical and broadcast markets.

Jim Nottingham, senior vice president of HP’s Advanced Compute Solutions (ACS) division, said AI assistants had greatly reduced work time, while token consumption continued to rise. HP’s analysis found that smaller models could produce the same quality as frontier models. Running workstation-class small models on company equipment increased inference speed and removed token-use limits. A company development team that had switched to per-token private cloud pricing also adopted local equipment, allowing 30 to 80 developers to use AI agents without restrictions.

Nottingham cited data privacy and sovereignty, low latency, data gravity (the tendency for data to attract applications and services), and cost as reasons companies operate AI outside the cloud. Demand is particularly strong in manufacturing. Global automaker A wanted to install AI inference equipment on its production line, but large workstations with many GPUs were difficult to place because of heat and noise. HP proposed ZBoost, which lets multiple devices share workstation GPUs. The large equipment stays in the back office while smaller devices are placed on the line; automaker A is moving toward company-wide deployment after a proof of concept.

Automaker B used ZBoost to avoid reloading a large rendering model each time, raising design productivity by 570% and cutting total cost of ownership (TCO) by 23%. Another customer uses about 300 workstations that would otherwise sit idle after work to fine-tune AI models about 1,000 times each night, reducing cloud costs. Nottingham said the key is a hybrid approach that combines local and cloud computing where each is appropriate.

HP expects demand for local AI to grow in South Korea, particularly in industries such as semiconductors, automotive and shipbuilding, where design and process data are a source of competitive strength. It cited semiconductor inspection equipment, semiconductor factory automation and logistics systems, and dental and medical imaging equipment as local use cases. Broadcasters are adopting more augmented reality (AR) solutions, and HP Z workstations were used for a live broadcast solution that ran for more than 12 hours during this year’s nationwide local elections. HP also plans to strengthen support for professionals doing demanding work, including architects and engineers.

Nottingham said his goal during the South Korea visit was to listen to and learn from customers and partners, and that he would support HP Korea’s customer success globally. He previously led HP Printing Korea and oversaw integration after HP acquired Samsung Electronics’ printing business.