AI News

Jiwei Consulting Releases "AI Industry Series Report: TPU Chips (2026)": TPU-type Custom Chips Stand at the Strongest Outlet

Jiwei.comUpdated Nov 16 20:58

AI summary

Jiwei Consulting released the "AI Industry Series Report: TPU Chips (2026)," reviewing TPU technology evolution, market landscape, and domestic substitution pathways. Citing data, the report states that in the first half of 2026, China's data center AI chip shipments totaled approximately 2 million units, with non-GPU accelerators shipping 1.016 million units, surpassing GPU accelerators for the first time; multiple institutions expect China's AI chip self-sufficiency rate to rise above 40% in 2026 and reach around 70% in 2027-2028.

Body · AI translation

The production model of AI computing power is being rewritten. The 2026 capital expenditure guidance of the four cloud vendors—Amazon, Microsoft, Google, and Meta—totals approximately $720 billion, nearly three times that of two years ago; massive investment is being deployed along the custom chip route, and multiple research institutions predict that global high-end cloud AI accelerator shipments will exceed 25 million units in 2027, with ASIC shipments surpassing GPU shipments for the first time—custom computing power represented by TPU is standing at an inflection point in the industrial structure.

TPU (Tensor Processing Unit) is a cloud-based custom ASIC centered on tensor operations and optimized for AI training and inference. According to Gartner's forecast, global semiconductor revenue will rise from $809 billion in 2025 to $1.6 trillion in 2026, and further to $1.9 trillion in 2027; it is estimated that global AI server shipments in 2026 will be about 3.7 million units, up 55.9% year-on-year, constituting the direct demand carrier for accelerator chips such as TPU.

Recently, JW Insights officially released the "AI Industry Series Report: TPU Chips (2026)," which is organized into four major chapters—industry overview, industry chain analysis, representative enterprises, and future outlook—and systematically reviews the technological evolution, market landscape, and localization path of TPU.

The following are selected contents of the Report:

Market size: computing power investment and domestic volume growth resonate

AI computing power demand is driving the semiconductor industry into a high-growth range. In overseas markets, Google Cloud revenue reached $30.7 billion in the third quarter of 2026, up 103% year-on-year, and management stated that demand for TPU from various customers is strong; Broadcom's AI semiconductor revenue in the same period was $16.7 billion, up 221% year-on-year, with custom chips becoming one of the fastest-growing segments of the global semiconductor market. In the Chinese market, according to the "2026 China Artificial Intelligence Computing Power Development Assessment Report" jointly released by IDC and Inspur, China's intelligent computing power scale is expected to reach 2,576.5 EFLOPS in 2026, up 87.9% year-on-year, and will reach 10,524.3 EFLOPS in 2030.

Figure 1 Global semiconductor market size and forecast (Source: Gartner, compiled by JW Insights)

Figure 2 China's intelligent computing power scale and forecast (Source: IDC, compiled by JW Insights)

Increment and cost: long-term agreements lock in demand, memory and packaging prices rise

Demand increments are being locked in by multi-year contracts: Anthropic disclosed that it will pay Google Cloud at least approximately $114.5 billion to obtain TPU computing power, and signed a five-year TPU equipment lease agreement with Broadcom totaling approximately $161.2 billion. On the supply side, HBM and advanced packaging constitute the main bottlenecks; TSMC's CoWoS monthly capacity is planned to increase from about 130,000 wafers at the end of 2026 to about 260,000 wafers in 2028; according to TrendForce estimates, the global HBM market will grow 58% to $54.6 billion in 2026, and the average selling price of HBM will rise about 121% year-on-year in 2027. Driven by costs, the quoted price of Huawei's Ascend 950DT accelerator card has reached more than 250,000 yuan, with a cumulative increase of about 60% within three months.

Figure 3 Forecast of changes in HBM's share of DRAM wafer input (Source: TrendForce, compiled by JW Insights)

Competitive landscape: Google supplies nearly 100 external customers, domestic echelon takes shape

Google's TPU external customers have approached 100, and it has landed for the first time on a third-party hyperscale cloud platform through Oracle Cloud, with OpenAI also planning to purchase; custom design services are highly concentrated, with Broadcom and Marvell together accounting for about 90% of the share. Domestically, according to IDC data, China's data center AI chip market shipped about 2 million units in the first half of 2026, with NVIDIA holding a 33.8% share and Huawei following closely at 28.5%, narrowing the gap to about 6 percentage points; non-GPU accelerator shipments were 1.016 million units, surpassing GPU accelerators for the first time. Cambricon achieved operating revenue of 9.526 billion yuan in the first three quarters of 2026, up 106.76% year-on-year.

(Source: compiled by JW Insights)

Figure 4 Shipments of major manufacturers in China's data center AI chip market in the first half of 2026 (Source: IDC, compiled by JW Insights)

Localization: self-sufficiency rate heading toward 70%

Domestic manufacturers are advancing along the path of "complete machines and clusters first, chip design following, and memory and packaging filling the chain": Huawei has deployed more than 1,000 Ascend 910C super nodes, and Ascend 950 super nodes have entered the stage of large-scale commercialization; Alibaba's T-Head Zhenwu series has shipped a cumulative 560,000 units and served more than 650 customers; Kunlunxin won approximately 70%, 70%, and all of the shares in the three bid packages of China Mobile's billion-level inference device centralized procurement. Multiple institutions predict that China's AI chip self-sufficiency rate will rise to more than 40% in 2026 and reach about 70% from 2027 to 2028.

Figure 5 TPU chip industry chain map (Source: TrendForce, IDC, compiled by JW Insights)

The Report judges that the focus of industry competition has shifted from single-chip performance to system-level capabilities encompassing chips, interconnection, packaging, and clusters, with HBM supply, advanced process foundry capacity, and ecosystem software constituting the main constraints; domestic forces have cost-effectiveness advantages in complete machine integration, cluster engineering, and inference scenarios, and the structural inflection point at which custom ASIC shipments exceed GPUs will first emerge in the Chinese market.

At present, the "AI Industry Series Report: TPU Chips (2026)" has been officially launched on the iJiwei official website and APP. Welcome to log in to the iJiwei official website or the Jiwei APP and click the "Jiwei Reports" section on the homepage to place an order.

Click to view more reports

Source

Jiwei.com