Global AI server shipments are forecast to exceed 2.5 million units in 2026, as Nvidia’s GB300 platform accelerates demand and cloud customers shift towards branded systems.
DIGITIMES’ latest research points to another year of rapid expansion in the AI server market, with global shipments forecast to exceed 2.5 million units in 2026. Of those, 2.37 million are expected to be high-end systems built around high-bandwidth memory, a segment the research firm says will rise 56% from 2025 and account for more than 15% of all server shipments for the first time. The shift underlines how quickly data centre operators are moving from general-purpose infrastructure towards heavier, more specialised AI hardware.
The report also suggests that the supplier map is continuing to evolve as demand broadens beyond the biggest US cloud groups. Google, working through its manufacturing alliance with Celestica, is expected to remain the largest supplier, with Quanta and Foxconn holding second and third place. Dell is projected to stay in fourth, but with high-end AI server volumes approaching 300,000 units this year, up more than 162% year on year. Wistron is forecast to climb to fifth, overtaking Supermicro, as cloud customers increase orders.
That ranking change reflects a broader structural shift in procurement. DIGITIMES analyst Hsiao Sheng-lun said that as second-tier cloud companies expand their purchases and increasingly prefer to buy through branded vendors, OEM shipments will make up more than a fifth of the high-end market in 2026. By contrast, direct buying from contract manufacturers is expected to fall to 79%, signalling a market in which the balance between branded systems and direct supply relationships is changing.
At the product level, Nvidia’s GB300 NVL72 is expected to be a major driver of growth. DIGITIMES says shipments of GPU-based high-end AI servers will account for nearly 60% of the market in 2026, helped by smoother deliveries of the GB300 NVL72 compared with its predecessor. The NVL72 platform is forecast to become the most common rack configuration, taking 42% of shipments next year, up from 24% in 2025, while NVL8 and PCIe-based designs continue to lose share. The report also expects TPU-based systems to weaken as supply constraints limit volumes.
Customer demand is diversifying as well. According to the report, the share of purchases from North American hyperscalers is set to fall from 78% to 68%, while Neo Cloud operators are expected to rise to 17% and leading AI laboratories to contribute close to 6% for the first time. Google and Amazon remain the two largest customers, but Oracle and SpaceXAI are forecast to move ahead of Microsoft. Anthropic is expected to enter the ranking in eighth place with 135,000 units, while Nebius is projected to post more than fourfold shipment growth.
Even so, the near-term outlook is not without risk. Hsiao said concerns about AI model safety could cause some operators to review their computing plans and slow training activity in the short term. However, he argued that the associated computing capacity is likely to shift towards longer inference workloads and agentic AI applications, including so-called harness engineering. That would delay some shipments, but not necessarily alter the wider trajectory. S&P Global has separately warned that the rush into high-bandwidth memory is tightening supplies of conventional DRAM and pushing up prices across servers, PCs and consumer electronics, reinforcing the view that AI infrastructure demand is now affecting the broader semiconductor market.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
