TPU v10 complexity expands the share of design services; MediaTek remains the core integrator; GUC and Alchip target prices raised

Overview

The report examines how Google TPU v10 design services will ultimately be allocated among multiple suppliers and further assesses the revenue and valuation implications for GUC, Alchip, and MediaTek amid the global expansion of AI ASICs. The key conclusion is that greater TPU v10 complexity will increase the number of design modules available for outsourcing, but will not naturally imply the replacement of MediaTek. Meanwhile, the scaling of GUC's and Alchip's projects with other cloud service providers represents a clearer source of earnings upgrades.

Core views

First, Morgan Stanley believes the technical complexity of TPU v10 allows the same project to accommodate more design service providers, but MediaTek remains the principal coordinator in the Asian supply chain. The report identifies three structural changes: first, the compute die adopts a KGD model, under which Google owns the compute die design and reticle and can outsource back-end design services to other suppliers; second, TPU v10 uses 3.5D packaging, requiring 3D stacking of the compute die with other chips and thereby increasing demand for 3D IC IP and design services; third, the chip's reticle size could expand from approximately 9 reticles currently to more than 12. Marvell, AMD, GUC, and Alchip may secure opportunities within the TPU ecosystem or in specific modules, but the report does not currently believe these participants will replace the incumbent core TPU designers. MediaTek is still expected to be responsible for TPU v10's I/O chip design, internally developed SerDes IP, and the packaging integration of the compute die KGD, I/O chips, and HBM, with responsibilities similar to those for TPU v9. The report believes TPU v10 may continue to use Intel EMIB-T, while MediaTek's continuously improving SerDes IP also strengthens customer stickiness. Even if Google supplies the compute die itself and MediaTek does not recognize revenue for that component, the larger I/O chip area could still result in higher recognized revenue per chip for MediaTek than for TPU v9. The report estimates approximately US$18k per TPU v10. MediaTek's TPU revenue forecasts for 2027 and 2028 remain at US$13-15bn and US$43-45bn, respectively: corresponding to 3mn TPU v8t chips in 2027, and 1mn TPU v8t chips plus 3mn TPU v9 chips in 2028. By 2029, the report forecasts 4mn TPU v9 chips and approximately 1mn TPU v10 chips, generating approximately US$70bn in revenue. On the report's estimates, MediaTek currently trades at 12 times 2028 EPS, and Morgan Stanley reiterates its Overweight view. GUC's principal growth drivers come from Google CPU, Microsoft, and Meta projects. Due to capacity constraints, the report lowers its forecast for 2026 Google Axion2 CPU shipments from 1.5mn-2mn chips to 1.2mn-1.3mn chips. At an ASP of approximately US$700, the expected 2026 revenue contribution to GUC is US$900-1000mn. Some demand is deferred to 2027, and together with the ramp-up of Axion3, the report expects Google CPU revenue to approach US$2.8bn in 2027, including 3mn Axion2 chips and 500k Axion3 chips. Because the number of 3nm compute dies increases from one to two, Axion3's ASP is expected to be approximately twice that of Axion2. The 2028 Axion3 forecast is 2.5mn chips, below Axion2's annual run rate of 3mn-4mn chips, also primarily due to the doubling of compute dies. This business follows a Turnkey 2 or 3 model, with GUC providing advanced packaging design support such as silicon interposers and ABF substrates, but not owning the ASIC reticle. Consequently, its gross margin may be only approximately 10%, and Axion3's gross margin could be slightly lower than Axion2's due to its higher ASP and larger project scale. In addition to Google, GUC participates in the Microsoft Maia 200 and Cobalt 200 projects. Supply chain checks indicate that Microsoft demand strengthened over the latest quarter, and management expects demand to more than double in 2027. The report forecasts a revenue contribution of at least US$250mn to GUC in 2027. Morgan Stanley also believes GUC may secure a Meta ASIC project led by the Rivos team, inferred to be MTIA 600, targeting tape-out in the first half of 2027 and potential CoWoS output in late 2027 or the first half of 2028. Increased TSMC 3nm wafer allocations and potential TPU v10 3D IC design services support the report's upgrades to GUC's earnings and target price. EPS forecasts for 2026, 2027, and 2028 are raised by 13%, 11%, and 11%, respectively. As low-margin turnkey revenue represents a larger share, the main text expects gross margin to decline from 20.8% in 2026 to 15.3% in 2028, while the financial summary table lists 15.8% for 2028. GUC's formal target price revision table and valuation section list NT$6,288, approximately 10% above the previous NT$5,688; elsewhere in the main text, the figure is written as NT$6,388. The formal target price of NT$6,288 represents 45 times 2027 EPS, below one standard deviation above the company's forward 12-month P/E since 2013. Key assumptions in the residual income model remain unchanged, including a 9.2% cost of equity, comprising a beta of 1.2, a 6% risk premium, and a 2% risk-free rate, as well as a 14.5% medium-term growth rate and a 5.0% perpetual growth rate. Bull- and bear-case values rise to NT$8,545 and NT$3,870, respectively. The report believes long-term AI ASIC growth will exceed that of AI GPUs, and expansion of the design services market and subsequent project wins could support further valuation rerating, although the recovery in non-AI businesses remains slow. Alchip's principal catalyst comes from AWS Trainium. Management believes Trainium4's potential market size could approach that of Google TPU. Morgan Stanley's industry checks indicate that Trainium4 shipments could reach at least 2.5mn chips in 2028. Assuming an ASP of approximately US$10k excluding HBM and an approximately 30% turnkey share for Alchip, the report raises its forecast for Trainium4's 2028 revenue contribution from US$4bn to US$8bn. For Trainium3, because the previous share assumption was overly optimistic, revenue contributions for 2026 and 2027 are modestly reduced to US$1.8bn and US$2.8bn, respectively. However, the larger Trainium ASIC scale brings better TSMC support, and gross margin is expected to improve from the previously anticipated low- to mid-teens percentage range to nearly the high-teens percentage range. These changes drive increases of 17%, 25%, and 67% in Alchip's earnings forecasts for 2026, 2027, and 2028, respectively, with 2028 revenue and earnings more than doubling from the previous forecasts. The target price rises from NT$5,088 to NT$5,888, which the report says implies more than 50% upside and incorporates the potential to win a second cloud service provider customer. The target price continues to use a residual income model: the cost of equity remains at 10.4%, comprising a beta of 1.4, a 6.0% equity risk premium, and a 2.0% risk-free rate. Considering competition from global peers and new entrants for AI ASIC projects, the medium-term growth rate is lowered from 16% to 12.5%, while the perpetual growth rate is reduced from 5.0% to 4%. Bull- and bear-case values are raised from NT$5,950 and NT$2,130 to NT$6,880 and NT$2,465, respectively. On the report's estimates, Alchip trades at 10 times 2028 EPS. The report also tests its AI supply chain forecasts using CoWoS capacity, chip volumes, and power demand. Its implied 2027 number of GPUs/ASICs from CoWoS is approximately 19mn chips. Assuming an average thermal design power of 2kW per chip, required installed power capacity would be approximately 38GW, indicating that power supply could remain a global bottleneck. As an order-of-magnitude example, if computing capacity expands from 2GW to nearly 10GW, the additional 8GW would be equivalent to approximately 4mn Rubin GPUs, representing approximately 20% of the AI accelerators TSMC could produce in 2027. Another model assumes only an additional 3GW, reaching 5GW by the end of 2027. The report also expects AI HBM demand to reach up to 50bn Gb and AI wafer consumption to reach at least US$59bn in 2027; the corresponding 2026 figures are up to 30bn Gb and at least US$26bn. Finally, Morgan Stanley calibrates GPU shipments against end demand using server configurations: the model assumes each HGX server uses 8 GPUs and that 9 HGX servers are equivalent to one NVL72 rack. The report forecasts 5.4mn Blackwell shipments in 2026, with supply sufficient to meet Grace Blackwell NVL72 demand in the second half of 2026. Blackwell chips previously regarded as inventory are actually primarily supply chain buffer inventory and are expected to be fully absorbed during 2026. Rubin is expected to begin ramping in the third quarter of 2026, with rack shipments commencing in the fourth quarter. Combined Rubin and Rubin Ultra shipments are expected to approach 7mn chips in 2027, with approximately 90k Rubin NVL72 racks. The report expects Rubin inventory to display a similar buffering and absorption pattern.

Analysis framework

The report first dissects TPU v10's compute die ownership, I/O design, 3.5D packaging, and reticle size to determine which modules different design service providers can undertake. It then combines supply chain checks, customer management commentary, and wafer and CoWoS allocations to model shipment volumes, ASPs, allocation shares, and gross margins for each project. It subsequently maps project revenue to GUC's and Alchip's earnings forecasts and derives target prices using residual income models and comparisons with historical P/E multiples. Finally, it checks whether supply chain forecasts align with infrastructure demand by converting among GPU/ASIC volumes, average power consumption, HBM, wafer consumption, and rack shipments.

Methodology notes

• Industry/Sector Analysis FrameworkUpstream, Midstream, and Downstream Industry Chain Transmission

  AI ASIC Supply Chain and Project Module Breakdown

  Starting with chip designs and reticles owned by cloud service providers, the report traces back-end design, I/O chips, 3D IC, advanced packaging, wafers, and HBM, and then determines how revenue and profit are allocated among MediaTek, GUC, and Alchip.

• Industry/Sector Analysis FrameworkVolume-price decomposition

  Shipment Volume, ASP, and Turnkey Share Estimates

  The report estimates revenue by multiplying project shipment volumes by chip ASPs and supplier allocation shares. For example, it derives US$8bn in 2028 revenue for Trainium4 based on at least 2.5mn chips, an ASP of approximately US$10k, and an approximately 30% share for Alchip.

• Industry/Sector Analysis FrameworkSupply-demand framework

  CoWoS, Wafer, HBM, and Power Constraint Testing

  The report maps forecast chip volumes to advanced packaging capacity, HBM demand, wafer consumption, and power capacity to identify whether capacity and power supply are sufficient to support forecast demand.

• Valuation MethodRIM Residual Income Model

  Residual Income Valuation of GUC and Alchip

  The report discounts future excess returns based on the cost of equity, medium-term growth rate, and perpetual growth rate to derive base-, bull-, and bear-case values for the two companies.

• Valuation MethodP/E and PEG Valuation

  Historical P/E and Forward EPS Comparison

  The report compares the 45 times 2027 EPS implied by GUC's target price with its forward 12-month P/E range since 2013 and lists the trading multiples of MediaTek, GUC, and Alchip based on 2028 EPS.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• MediaTek (2454.TW)

  Viewed by the report as the principal Asian coordinator for Google TPU v10, responsible for I/O chips, SerDes IP, and packaging integration.

  Strengths

  Established TPU design experience, continuously improving SerDes IP, EMIB-T packaging integration responsibilities, and high customer stickiness.

  Weaknesses

  Under the COT model, it does not recognize revenue from Google's internally owned compute die KGD, and some design modules can be allocated to other service providers.

  Comparison

  The report believes new participants are more likely to secure incremental modules than to replace MediaTek and the incumbent core US designers.

  Risks

  Changes in the division of TPU v10 modules could reduce the scope of content recognizable by MediaTek.

• Global Unichip Corp (3443.TW)

  Benefits from Google Axion CPU, Microsoft Maia/Cobalt, potential Meta MTIA 600, and TPU v10 3D IC design services.

  Strengths

  Rising exposure to cloud AI semiconductors, increased 3nm wafer allocations in 2027, and coverage of AI accelerators, AI server CPUs, and BMC chips.

  Weaknesses

  Gross margin on the Google CPU turnkey business may be only approximately 10%, recovery in non-AI businesses remains slow, and increasing turnkey revenue will reduce overall gross margin.

  Comparison

  The target price represents 45 times 2027 EPS, below one standard deviation above the company's forward 12-month P/E since 2013.

  Risks

  Capacity constraints, project timing, and a rising share of low-margin turnkey revenue could affect earnings realization.

• Alchip Technologies Ltd (3661.TW)

  Primarily benefits from AWS Trainium3 and Trainium4; the report also mentions the potential to win a second cloud service provider customer.

  Strengths

  Trainium's larger scale provides better TSMC support, and Trainium3's gross margin is expected to improve to nearly the high-teens percentage range.

  Weaknesses

  The Trainium3 allocation share is below the previously more optimistic forecast, and the business is highly dependent on major cloud service provider projects.

  Comparison

  Trades at 10 times 2028 EPS on the report's estimates; valuation assumptions have been lowered due to competition from peers such as GUC, QCOM, and AMD, as well as new entrants.

  Risks

  Competition from global peers and new entrants for AI ASIC projects could reduce project share, growth rates, or valuation.

Key data

• TPU v10 Reticle Size>12 reticlesCurrently approximately 9; greater complexity increases the number of separable design modules
• MediaTek Recognized Revenue per TPU v10approximately US$18kCould still exceed TPU v9 even without recognizing revenue from the compute die KGD supplied by Google
• MediaTek TPU Revenue Forecast2027 US$13-15bn; 2028 US$43-45bn; 2029 approximately US$70bnAssumes 4mn TPU v9 chips and approximately 1mn TPU v10 chips in 2029
• GUC 2026 Google Axion2 Forecast1.2mn-1.3mn chipsLowered from the previous 1.5mn-2mn chips due to capacity constraints
• GUC 2027 Google CPU Revenueclose to US$2.8bnIncludes 3mn Axion2 chips and 500k Axion3 chips
• GUC 2027 Microsoft Revenueat least US$250mnManagement expects related demand to more than double in 2027
• GUC EPS Forecast Upgrades2026 +13%; 2027 +11%; 2028 +11%Primarily reflects revenue growth from cloud service provider projects and Chinese ADAS customers
• GUC Target PriceNT$6,288Raised from NT$5,688 in the formal revision table and valuation section; elsewhere in the main text, it is written as NT$6,388
• Alchip 2028 Trainium4 RevenueUS$8bnRaised from US$4bn, assuming at least 2.5mn chips, an ASP of approximately US$10k, and an approximately 30% share
• Alchip Earnings Forecast Upgrades2026 +17%; 2027 +25%; 2028 +67%Improved Trainium3 gross margin and expanded Trainium4 revenue
• Alchip Target PriceNT$5,888Raised from NT$5,088, which the report says implies >50% upside
• 2027 GPU/ASIC Volume and Power Demandapproximately 19mn chips; approximately 38GWEstimated based on an average of 2kW per chip
• AI HBM Demandup to 30bn Gb in 2026; up to 50bn Gb in 2027Based on AI chip supply chain forecasts
• AI Wafer Consumptionat least US$26bn in 2026; at least US$59bn in 2027The report's estimate of AI-related wafer demand
• Blackwell and Rubin Shipment Forecasts5.4mn Blackwell chips in 2026; close to 7mn Rubin and Rubin Ultra chips in 2027Approximately 90k Rubin NVL72 racks forecast for 2027

Impact & implications

The report believes TPU v10's modularization will not create a single winner but will expand the overall ASIC design services market: MediaTek retains high-value integration and I/O responsibilities, while other suppliers can participate in modules such as compute die back-end design and 3D IC. GUC's and Alchip's earnings leverage comes primarily from the expansion of Google CPU, Microsoft, Meta, and AWS projects, although a rising share of turnkey business will reduce gross margins for certain projects. At the industry level, whether CoWoS, 3nm wafer, HBM, and power supply can expand in tandem will determine whether current chip shipment forecasts can be realized.

Risks

• Global power infrastructure may not be able to support the GPU/ASIC installation volumes forecast in the report in a timely manner, and power supply remains a bottleneck for AI expansion.
• Wafer, 3nm, and CoWoS capacity constraints have already deferred some Google CPU demand, and insufficient future allocations could still affect the timing of project revenue.
• A rising share of turnkey revenue could reduce the overall gross margins of GUC and Alchip.
• Intensifying competition for AI ASIC projects from GUC, QCOM, AMD, and other new entrants could affect Alchip's share and long-term growth assumptions.
• The slow recovery of GUC's non-AI businesses could partially offset growth from cloud AI projects.

What to watch

• Monitor the final division of TPU v10 modules and the design content secured by MediaTek, GUC, Alchip, Marvell, and AMD.
• Monitor whether TPU v10 continues to use Intel EMIB-T and whether the 3.5D packaging and design involving more than 12 reticles can be implemented as planned.
• Monitor GUC's TSMC 3nm wafer allocations in 2027 and the ramp-up of Google Axion3.
• Monitor whether Meta MTIA 600 can tape out in the first half of 2027 and generate CoWoS output in late 2027 or the first half of 2028.
• Monitor Alchip's actual Trainium4 turnkey share, the assumption of at least 2.5mn shipments in 2028, and the project with a second cloud service provider.
• Monitor whether infrastructure can support approximately 38GW of power demand in 2027 and whether CoWoS, HBM, and wafer supply can expand in tandem.
• Monitor the execution progress of Rubin's ramp-up in the third quarter of 2026, rack shipments in the fourth quarter, and shipments approaching 7mn chips in 2027.