In response to market anxiety triggered by its US internet analyst Brian Nowak's projection of a slowdown in hyperscaler capital expenditure growth to 12% by 2028, Morgan Stanley's Asia-Pacific technology team has released a fresh research report offering a more nuanced perspective on the AI investment landscape.

The report, titled "AI Supply Chain: Prelim 2028 AI semis vs. CSP capex growth," employs supply chain cross-validation to reach a pivotal conclusion: AI semiconductor capital expenditure, particularly for compute chips, is set to outpace overall AI infrastructure spending by 2028. Additionally, 2.5D packaging capacity is projected to climb another 50%, while infrastructure spending stabilizes and memory prices rise modestly. This indicates the AI supply chain narrative is far from concluded—rather, it is shifting from a broad-based spending surge to a phase where compute chips and advanced packaging emerge as structural outperformers.

Morgan Stanley's US team had forecast that data center capital expenditure from the four major hyperscalers would escalate from approximately $917 billion in 2026 to $1.47 trillion in 2027, marking a 60% year-over-year increase. However, by 2028, that growth rate is expected to plummet to roughly 12%, reaching a scale of around $1.64 trillion. This anticipated deceleration has fueled investor skepticism about whether AI semiconductor momentum would similarly wane, a notion the Asia-Pacific team explicitly rejects.

Based on initial capacity assumptions for CoWoS, CoPoS, and EMIB-T technologies in 2028, Morgan Stanley projects total 2.5D packaging capacity will average 374 thousand wafers per month (kwpm), a 50% increase from the anticipated 250kwpm in 2027. This expansion is driven by larger chip sizes and a broadening range of applications, as these 2.5D packaging solutions now serve AI GPUs, XPUs, CPUs, and networking chips—extending well beyond their GPU-centric origins.

The 2028 advanced packaging expansion is segmented into three primary trajectories. For CoWoS at the wafer level, Taiwan Semiconductor Manufacturing Co (TSM.US) continues its expansion, though the oS segment may require additional partners. Investigations indicate that Amkor Technology (AMKR.US) and a semiconductor testing firm are diversifying oS production across the US and Taiwan. TSMC's CoWoS capacity is forecast to grow from 130/200kwpm in 2026/2027 to 260kwpm by 2028, concentrating on Arizona's AP9/10 and potentially AP7. Non-TSMC players are expected to collectively reach 110kwpm by end-2028, with ASE/SPIL contributing 50-60kwpm and Amkor adding 55-60kwpm, focusing on CoWoS-L and CoWoS-R variants.

In the CoPoS (panel-level) segment, after converting to 12-inch CoWoS wafer equivalents, Morgan Stanley predicts production of 5-10kwpm in 2028, scaling to 10-20kwpm by 2029. The initial target project is expected to be Nvidia Corp's (NVDA.US) Feynman Ultra GPU. TSMC is currently piloting production in Taoyuan, with potential capacity built at AP9/10 or AP7P3 contingent on progress. Meanwhile, EMIB-T (substrate-based) technology is also converted to 12-inch CoWoS wafer equivalents by the bank. A single EMIB-T substrate can generate 16 chips of 9x reticle size—approximately four times the chip count from a CoWoS wafer. Morgan Stanley projects roughly 40k-45k wafer-equivalent EMIB-T capacity in 2028, producing at least 2 million Humufish (9x reticle) substrates, capturing a 10-15% share of the 2.5D packaging market.

Intel Corp (INTC.US) emerges as a critical variable in this report. Morgan Stanley observes Intel's current average EMIB-M capacity at 110kwpm, which is expected to slightly decline to 95kwpm in 2027 as some capacity transitions to EMIB-T. Key EMIB-M clients include Trainium3, the forthcoming Trainium4, and internal server CPUs. EMIB-T average capacity is projected to rise from roughly 5kwpm in 2026 to 15-20kwpm in 2027, reaching 40-45kwpm by 2028. Assuming each EMIB-T wafer yields 4-5 chips, Morgan Stanley maintains that Intel's EMIB-T can still produce nearly 3 million Humufish units by end-2027 to 2028. While several smaller projects are underway, the MediaTek/Google Humufish venture is anticipated to consume the bulk of this capacity.

In the ASIC arena, significant developments are unfolding. Per analyst Joe Moore's assessment, Broadcom (AVGO.US) delivered robust earnings and raised its fiscal 2028 earnings per share outlook, with fiscal 2027 revenue expectations now pegged at $115 billion and potential doubling by fiscal 2028. CoWoS supply chain checks indicate that Broadcom's AI lab customers have reserved roughly 15k-30k CoWoS wafers for 2027. Morgan Stanley continues to see increasing prospects for MediaTek securing a second AI CSP or AI lab customer. Leveraging its Nvidia NVLink Fusion collaboration, a related decision could materialize in Q4 2026, serving as a significant catalyst for MediaTek, which holds an "Overweight" rating from the bank.

Qualcomm Corp (QCOM.US) has announced an ASIC partnership with AWS, a key ASIC client for Alchip. However, the press release specifically references inference workloads, distinguishing it from Alchip's Trainium4, which covers both training and inference. Consequently, Morgan Stanley views this as having limited impact on Alchip.

For Advanced Micro Devices Inc (AMD.US), Morgan Stanley projects total CoWoS consumption in 2027 to grow 165% year-over-year to 345,000 wafers. AI GPU-related production will be exclusively housed in TSMC CoWoS, with MI455/450 as the 2027 focus and the MI500 series (Arcadia) entering low-volume production in late 2027. MI455 and MI450 output could reach approximately 500,000-650,000 units each. TSMC CoWoS-L reservations are expected to surge 200% year-over-year to 210,000 wafers, though non-TSMC capacity ramp-up is slower than anticipated. AMD's Venice CPU CoWoS reservations increase from 50,000 wafers in 2026 to 210,000 in 2027, yet fall short of initial forecasts, with production estimates revised down from 5.6 million to 4.4 million units. TSMC must also allocate roughly 80,000 wafers of CoWoS-L for Venice production, partially squeezing AI GPU build capacity. Conversely, Microsoft Corp's (MSFT.US) Maia200 is seeing upward revisions, with Morgan Stanley observing positive demand adjustments for 2027, potentially reaching approximately 5k CoWoS, implying 100,000-150,000 chip reservations.

On the HBM and wafer consumption front, Morgan Stanley projects total AI HBM demand in 2027 will reach approximately 44.9 billion GB, up from 29 billion GB in 2026, driven primarily by Nvidia's Rubin series and Google's TPU v8i/v9. Wafer consumption for AI is expected to total at least $55 billion in 2027, compared to $27 billion in 2026. TSMC's AI-related revenue could achieve a compound annual growth rate of 60% from 2024 to 2029, with Morgan Stanley still anticipating sequential growth in AI chip revenue. Global CoWoS demand is expanding rapidly, with total demand projected to rise from 1.394 million wafers in 2026 to 2.509 million in 2027, an 80% year-over-year increase. This includes Nvidia growing from 780,000 to 1.222 million wafers, Broadcom from 300,000 to 484,000, AMD from 130,000 to 345,000, MediaTek from 40,000 to 180,000, and Marvell Technology Inc (MRVL.US) from 26,000 to 90,000.

As the market transitions from a GPU-grabbing frenzy to an efficiency-driven era, AI investment is entering its "ROIC examination" phase. Morgan Stanley's US team previously highlighted that AI investment is not a capital black hole. Model companies providing API services via proprietary compute can achieve ROIC as high as 46%, while those relying on third-party infrastructure see returns around 25%, and IaaS models yield approximately 31%. The most compelling segment is not pure compute leasing but companies possessing integrated model capabilities, infrastructure, and commercialization prowess. Concurrently, Morgan Stanley forecasts operating cash flow for the four major cloud providers will rise from $739 billion in 2026 to $1.23 trillion by 2028, with incremental debt financing needs declining from $238 billion to $90 billion. By 2028, incremental debt requirements will represent only about 7% of operating cash flow, indicating that financing pressure is not deteriorating in tandem with capital expenditure.

However, following the capital expenditure surge, 2028 may usher in a period of digestion. Real-world constraints—chips, racks, land, power, and labor—are limiting further expansion. Hyperscalers have already "pre-built" substantial data center capacity for 2027-2029 demand, leaving limited room for additional front-loaded capex. Market focus will pivot from "who can build more data centers" to "who can convert built compute into actual revenue and profit." Morgan Stanley suggests that as capex growth moderates and AI application penetration rises, capital may gradually rotate from hardware, semiconductors, and memory toward models, cloud platforms, and software applications. The "ROIC examination" for AI investment is transitioning from the compute construction phase to the commercialization stage, where the sustained revenue and profit contribution of the application layer becomes the primary valuation driver for the next phase.