Dell takes AI server lead as memory costs push Huawei buyers to the black market

Dell takes the lead as AI server revenue jumps

The firm puts market growth close to 20% year on year.

Unit sales grew just 4.4% over the same period. Buyers are paying more per machine, not buying many more of them, and that changes the arithmetic for anyone budgeting an AI cluster this year. Hyperion attributes the gap to growing GPU content in systems and higher memory costs. Dell took the market-share lead with $4.2 billion in server revenue in the first half of 2026, ending a run of more than 25 years at the top for HPE. Hyperion credited large procurements and a fast-growing segment of mid-range AI-focused machines, bought to add AI to scientific and engineering workflows. HPE had recognised significant revenue in 2025 from multiple exascale-class machines, and Hyperion noted that scale of procurement is not realistic every year.

Earl Joseph, chief executive of Hyperion Research, said replacement demand and AI adoption are both driving spending. "Organizations and institutions are continuing to invest in their advanced computing infrastructure due to both replacing aging machines and the desire to adopt AI into their workflows to support their businesses and missions," he said.

Why AI systems cost more

Memory is where the pressure concentrates. Huawei is telling some customers to source memory for its AI chips themselves, according to The Information, as reported by The Decoder.

Prices for smuggled HBM3E memory have jumped from a few hundred dollars to around $4,000 per unit in a year, according to a Shenzhen broker. Each AI chip needs six to eight units, so the memory alone can cost up to $32,000. That shifts risk onto buyers. A customer sourcing HBM independently has to account for where it came from.

Black market fills the HBM gap

Its Chinese suppliers cannot produce enough high-bandwidth memory, and U.S. export rules block direct purchases from Samsung and SK Hynix. Those two companies control more than 80% of the global market.

Smugglers bundle HBM with cheap server processors whose performance falls below the export threshold, then ship the packages to China through Hong Kong. Workshops there separate the memory chips from the processors. The process is fragile and the error rate is high, The Decoder reported, but the margins make it worth it. The black market does not run on smuggling alone. Older Nvidia GPUs stripped for parts and excess inventory from scrapped projects also feed supply.

South Korea's customs authority had uncovered trade violations worth $560 million by May, already above the previous full-year record. A third of that total involved strategic goods. Those included chips exported without proper authorisation or with false declarations.

South Korea told to defend chip profits

Samil PwC, the South Korean member firm of PwC, published a report on Monday. It said Chinese competition would likely heat up first in conventional memory, mature-node manufacturing, and packaging and testing. The firm urged South Korean chipmakers to sharpen cost competitiveness in commodity memory while doubling down on advanced packaging and high-bandwidth memory. The warning is about mix as much as volume. Chinese rivals are scaling up production and market share, while the report says Korean firms should double down on advanced packaging and HBM.

Longer term, the industry expects AI to reshape how chips are designed. Cambricon co-founder Chen Yunji made a prediction at a science forum in Shanghai. Over the next 10 years, he said, AI models would improve AI models and AI chips would improve AI chips. The South China Morning Post reported the remarks.

What happens next

The near-term test is the Ascend 960DT. Huawei says it will double the performance of its current chips when it launches early next year. Whether customers can source memory for it at a workable price is a separate question, and one the company has pushed onto them.