Damo fights back against the “AI capital expenditure theory”. It is expected that AI semiconductor capital expenditure will continue to “overtake” in 2028, and 2.5D packaging production capacity will increase by another 50%

The Zhitong Finance App learned that the Morgan Stanley Asia Pacific Technology team recently published the latest research report “AI Supply Chain: Prelim 2028 AI Semis vs. CSP Capex Growth” to respond to market concerns raised after US internet analyst Brian Nowak predicted “the CSP capital expenditure growth rate will drop to 12% in 2028” in the “Morgan Stanley AI Guidebook.” Damo's core conclusion through cross-verification of the supply chain is that in 2028, AI semiconductor capital expenditure, especially computing chips, will continue to exceed overall AI capital expenditure; 2.5D packaging production capacity will still increase by 50%, infrastructure spending will stabilize, and memory prices will moderate.

This means that the story of the AI supply chain is not over, but rather a shift from a “sharp overall capital expenditure” to a “structural rivalry between computing chips and advanced packaging.”

CSP capital expenditure growth rate fell to 12%, but computing chips “outperformed”

The US team at Damo predicts that the data center capital expenditure of the four largest hyperscale cloud vendors will rise from about 917 billion US dollars in 2026 to about 1.47 trillion US dollars in 2027, an increase of about 60% over the previous year; however, by 2028, the growth rate will plummet to about 12%, and the scale is about 1.64 trillion US dollars.

This deceleration has raised questions from investors: Will the growth momentum of AI semiconductors stop at the same time?

The Damo Asia Pacific technology team gave a negative answer. Its latest supply chain inspection shows that although overall AI infrastructure spending growth is slowing, the growth rate of AI semiconductor capital expenditure, especially computing chips, is expected to exceed overall AI capital expenditure in 2028. The report introduced preliminary assumptions for CowOS, CoPOS, and EmiB-T production capacity in 2028. It is estimated that the total production capacity (average) of 2.5D packaging will reach 374 kwpm in 2028, an increase of about 50% over 250 kwpm in 2027.

Growth is driven by two aspects: larger chip sizes and more application scenarios. These 2.5D packaging technologies cover AI GPUs, XPUs, CPUs, and network chips, and are no longer exclusive to GPUs.

2.5D package three-wire parallel: CoOS, CoPOS, EMIB-T

Damo will disassemble the 2028 advanced packaging expansion into three main lines:

CoWoS (wafer level): TSM.US is still expanding, but the OS segment may require more partners. According to the survey, AMKR.US (AMKR.US) and Core Test will each be produced in the US and Taiwan for diversified OS. TSMC's CoWOS production capacity is expected to further increase from 130/200 kwpm in 2026/2027 to 260 kwpm in 2028. The expansion is mainly concentrated on AP9/10 in Arizona, and may also include AP7. The non-TSMC camp will expand to 110 kwpm by the end of 2028, with ASE/SPIL being 50-60 kwpm and AMKOR being 55-60 kwpm, focusing on Cowos-L and Cowos-R.

CoPOS (panel level): After conversion to the equivalent of 12-inch CoWoS wafers, Damo expects production to be 5-10 kwpm in 2028 and expand to 10-20 kwpm in 2029. The initial target project should be Nvidia's (NVDA.US) Feynman Ultra GPU. TSMC is currently testing production in Taoyuan. If it goes well, it may build production capacity at AP9/10 or AP7P3.

EMIB-T (substrate type): Although it is a substrate technology, Damo also converted it to the equivalent production capacity of a 12-inch CoWOS wafer. An eMiB-T substrate can generate 16 9x photomask size chips, which is about 4 times the number of CoWOS wafers. Damo expects to have an EMIB-T production capacity equivalent to 40k-45k wafers in 2028 to produce at least 2 million Humufish (9x mask) substrates, which is equivalent to 10% to 15% of the 2.5D package market share.

Intel EMIB-T amplification, Humufish the key

Intel (INTC.US) became an important variable in this report. According to Damo, Intel's current average EmiB-M production capacity is 110 kwpm, and is expected to drop slightly to 95 kwpm in 2027, due to part of the conversion of production capacity to EMIB-T. Emib-M's main customers are Trainium3, upcoming Trainium4, and internal server CPUs.

The average production capacity of EmiB-T is about 5 kwpm in 2026, expanded to 15-20 kwpm in 2027, and 40-45 kwpm in 2028. Assuming that each EMIB-T wafer can produce 4-5 chips, Damo maintains its judgment: Intel EmiB-T can still produce close to 3 million HumuFish from 2027 to the end of 2028. Although a few small projects are underway, MediaTek/Google Humufish is expected to consume most of the production capacity.

ASIC battle: Broadcom raises guidelines, MediaTek attacks second customer

The ASIC field is dynamic and intense.

According to Daimo analyst Joe Moore's report, Broadcom (AVGO.US) has steady earnings and raised its earnings forecast for the 2028 fiscal year. The specific revenue forecast for the 2027 fiscal year is now $115 billion, and is likely to double in the 2028 fiscal year. Cowos supply chain inspections show that Broadcom AI Lab customers have ordered about 15k-30k Cowos wafers in 2027.

Damo continues to see more opportunities for MediaTek to win a second AI CSP or AI Lab customer. Through its partnership with Nvidia's NVLink Fusion, relevant decisions may be made in the fourth quarter of 2026, which will potentially be an important catalyst for MediaTek. Damo gave MediaTek an “gain” rating.

Qualcomm (QCOM.US) announced a partnership with AWS's ASIC, Alchip's key ASIC customer. But the press release points to reasoning, unlike Alchip's Trainium 4, which covers training and reasoning. Therefore, Damo believes this will have limited impact on Alchip.

AMD and Microsoft: Venice downgraded, Maia upgraded

On the AMD (AMD.US) side, Damo expects its total CoWOS consumption to increase 165% year-on-year to 345,000 tablets in 2027. All AI GPU-related production will be carried out at TSMC Cowos. The MI455/450 is the focus in 2027, and the MI500 series (Arcadia) will be produced in small quantities at the end of 2027. Production of MI455 and MI450 may reach about 500,000 to 650,000 units, respectively. TSMC's Cowos-L pre-orders are expected to increase 200% year over year, reaching 210,000 units.

However, it is not the TSMC faction that climbed the slope more slowly than expected. The number of Cowos pre-orders for AMD Venice CPUs increased from 50,000 units in 2026 to 210,000 in 2027, but it was lower than the initial forecast, and production estimates were lowered from 5.6 million to 4.4 million units. TSMC also needs to support about 80,000 CoWS-L tablets for Venice production, partially squeezing AI GPUs to build production capacity.

Microsoft (MSFT.US) Maia200 has been upgraded. Damo has been watching the Microsoft Maia 200's positive correction demand for 2027. It believes that demand may reach about 5k CoOS in 2027, which means about 100,000 to 150,000 chip pre-orders.

HBM and wafer consumption: demand explosion in 2027

Damo predicts that the total demand for AI HBM will reach about 44.9 billion GB in 2027 and 29 billion GB in 2026. NVIDIA Rubin series and Google TPU v8i/v9 are the main drivers.

In terms of wafer consumption, total AI wafer consumption is expected to be at least US$55 billion in 2027 and US$27 billion in 2026. TSMC's AI-related revenue is likely to grow at a compound annual rate of 60% between 2024 and 2029. Damo still expects AI chip revenue to increase sequentially.

Global demand for CoOS is also growing rapidly. Total demand for Damo is expected to increase from 1.394 million tablets in 2026 to 2.59 million tablets in 2027, an increase of 80% over the previous year. Among them, Nvidia increased from 780,000 to 1,222,000, Broadcom from 300,000 to 484,000, AMD from 130,000 to 345,000, MediaTek from 40,000 to 180,000, and Maywell Technology (MRVL.US) from 26,000 to 90,000.

From “grabbing GPUs” to “competing for efficiency”: AI investment enters the ROIC test

The US team at Damo previously pointed out in a report that AI investment is not a capital black hole. For model companies that use their own computing power to provide API services, ROIC can be as high as 46%; model companies that rely on third-party infrastructure return about 25%; IaaS model ROIC is about 31%. The most appealing aspect is not pure computing power leasing, but companies with modeling capabilities, infrastructure, and commercialization capabilities at the same time.

Meanwhile, Damo expects the operating cash flow of the four major cloud vendors to rise from US$739 billion in 2026 to US$1.23 trillion in 2028, and demand for incremental debt financing to drop from US$238 billion to US$90 billion. By 2028, the giant's demand for incremental debt was only about 7% of operating cash flow, and financing pressure did not worsen at the same time as capital expenditure.

However, after a sharp jump in capital expenditure, 2028 may enter a “digestion period.” Realistic constraints such as chips, racks, land, electricity, and labor are limiting further expansion. Hyperscale manufacturers are already “pre-building” a large amount of data center production capacity from 2027 to 2029, and space for further forward capex is limited. The market focus will shift from “who can build more data centers” to “who can turn already built computing power into actual revenue and profit”.

Damo believes that as the growth rate of capital expenditure slows and the penetration rate of AI applications rises, capital may gradually rotate from hardware, semiconductors, memory to models, cloud platforms, and software applications. The “ROIC test” for AI investment is moving from the computing power construction stage to the commercialization stage. Whether the application layer can continue to contribute revenue and profit will be the core valuation driver in the next phase.