China’s AI price war intensifies as Korea flags tight HBM supply into 2027

Asia’s tech agenda this week centered on two linked pressures, cheaper Chinese AI services and tighter Korean high bandwidth memory supply, that will shape weekend planning for global engineering and procurement teams. Chinese model providers cut headline prices and pushed bundle discounts for enterprise inference, while SK Hynix and Samsung indicated that HBM3E capacity is largely committed through next year, keeping GPU total cost of ownership elevated.

The combination matters because lower model pricing can pull demand forward, but compute bottlenecks still gate deployment. That gap keeps leverage with suppliers tied to Nvidia’s Blackwell platforms and CoWoS packaging timelines.

Chinese AI platforms escalated price competition around large language models used in search, coding, and customer service. Alibaba Cloud continued to promote aggressive Qwen API tiers on its mainland console, and Baidu kept enterprise discounts for Ernie Bot across usage bundles, moves flagged in Chinese developer forums and company rate cards. ByteDance has expanded Doubao model access inside Feishu and Toutiao with low entry thresholds, aiming to convert in-house usage into paid external workloads. Local procurement teams cite compliance with China’s algorithm filing regime under the Cyberspace Administration as a reason to consolidate on domestic providers for regulated industries.

Investors track how far unit economics can stretch. Alibaba has said publicly that Qwen will be embedded across its commerce and logistics stack to drive utilization, while Baidu ties Ernie to ad tech and cloud migration deals. Research firms such as IDC and local brokerages have noted that cut‑rate APIs can be loss leaders if inference runs on subsidized domestic accelerators rather than imported GPUs restricted by US export controls.

On the hardware side, Seoul’s memory champions signaled a slower easing of constraints. SK Hynix, the market leader in HBM for AI accelerators, has told customers and investors in recent updates that HBM3E lines are running at high utilization and that meaningful HBM4 output depends on substrate and TSV yield improvements. Samsung, which has worked to broaden qualifications at hyperscalers, guided that it is ramping HBM3E and early HBM4 in step with demand visibility. Industry trackers including TrendForce still place SK Hynix ahead on HBM share, with Samsung gaining as it aligns with US chipmakers beyond Nvidia.

For global CTOs the message is clear. Expect Chinese AI software to get cheaper and more localized, and plan for component lead times on premium memory and advanced packaging to remain the long pole that constrains large model deployments through 2027.