微软发布Maia 300并传出由联想承接AI机架组装，ASIC服务器供应链2027年放量预期升温

Microsoft unveils Maia 300 AI ASIC, boosting optimism for ASIC server shipments. Credit: DIGITIMES

Microsoft has unveiled its latest in-house AI application-specific integrated circuit (ASIC), Maia 300, and accompanying server rack, with its compute chiplets manufactured using TSMC's advanced N2P process (an enhanced, extended version of the company's baseline 2nm, or N2). Supply-chain sources indicate that Lenovo, rather than Taiwanese manufacturers, will assemble the Maia 300 server racks.

The launch adds to the in-house ASIC server offerings of the four major US cloud service providers (CSPs), positioning ASIC servers as potentially the fastest-growing segment of the AI server market in 2027, according to industry sources.
Lenovo secures Maia 300 assembly, Inventec eyes FPGA orders
Microsoft has traditionally relied on Taiwanese manufacturers, including Wiwynn, Inventec, Foxconn, and Quanta, for server production. Lenovo also maintains close ties with Microsoft, with their collaboration extending from PCs to servers and storage equipment. At a recent earnings briefing, Lenovo disclosed an AI server sales pipeline of approximately US$21 billion as of the end of March 2026 and said it was negotiating additional large contracts with global cloud providers and major enterprises.
Beyond Lenovo's reported assembly role, Maia 300 servers are expected to adopt liquid cooling, with Asia Vital Components (AVC) and Cooler Master identified as principal cooling suppliers. King Slide and Nan Juen International are reportedly supplying server rails. The companies declined to confirm individual customer relationships but expressed optimism about broader ASIC server demand.
Microsoft also plans to introduce next-generation field-programmable gate array (FPGA) servers in 2027, according to supply-chain sources. Inventec is reportedly positioned to undertake rack-level (L11) assembly, marking a potential milestone in its expansion from L6 manufacturing into higher-level system integration. The contract could provide a significant revenue opportunity. Inventec declined to comment.
Maia 300 targets inference efficiency and Azure deployment
Microsoft has spent years developing proprietary ASICs but has progressed more slowly and at a smaller scale than Google and Amazon Web Services (AWS), according to supply-chain sources. Maia 300 will initially serve Microsoft's own data centers, although growing customer demand could eventually open the possibility of wider external availability, following its cloud rivals.
Microsoft describes Maia 300 as an "inference-first" accelerator designed for long-context reasoning, latency-sensitive conversational AI, and multistep agentic tasks. The company expects it to deliver 40% more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration, rack-level token throughput can reach seven times that of Maia 200.
Initial deployment is scheduled for US data centers in 2027, with other regions to follow once capacity expands to several hundred megawatts and eventually gigawatt scale. Early Maia 300 deployment is expected to remain focused on Microsoft's internal workloads.
Despite its comparatively late expansion, Microsoft is projected to record the fastest ASIC shipment growth among major cloud providers in 2027. A DIGITIMES Intelligence report forecasts that high-end cloud AI accelerator shipments will exceed 25 million units that year, with ASICs overtaking GPUs for the first time and high-bandwidth memory (HBM) and substrates emerging as supply bottlenecks.
DIGITIMES analyst Stella Weng estimates Microsoft's ASIC shipments will reach 419,000 units in 2027, surging 272% year-on-year. The increase reflects a low comparison base in 2026 and the gradual deployment of Maia 300 and Maia 200 across Azure data centers.
ASIC servers reshape the 2027 supply-chain outlook
With Microsoft's Maia 300 joining the proprietary AI accelerators developed by the other major US cloud providers, industry sources expect ASIC servers to become a principal growth engine for AI server manufacturers in 2027.
The DIGITIMES Intelligence report projects a structural shift in the high-end cloud AI accelerator market, from GPU-led demand towards parallel growth in GPUs and ASICs. GPU accelerator shipments are forecast to rise 25.2% year-on-year in 2027, while ASIC shipments are expected to surge 110.9%, exceeding GPU shipments in total volume for the first time.
Supply-chain companies are optimistic not only about shipment growth but also about profitability. Unlike Nvidia-based GPU servers, proprietary ASIC platforms vary across cloud providers, giving original design manufacturers (ODMs) and component suppliers greater scope to differentiate their engineering and system designs.
Although individual ASIC server orders remain smaller than those for Nvidia GPU systems, suppliers report more attractive gross margins. The shift towards customized accelerators could therefore create opportunities for manufacturers to capture greater design value even without matching the volumes of established GPU platforms.