TPU v10 complexity expands the division of design services work; MediaTek remains dominant, while earnings potential for GUC and Alchip is revised upward

Overview

The report focuses on the division of design services work for Google TPU v10 and the revenue opportunities for MediaTek, GUC, and Alchip in the supply chains for U.S. cloud service providers' in-house chips. The core view is that the increased complexity of TPU v10 will create more outsourced design modules, but MediaTek will remain the primary coordinator due to its capabilities in I/O dies, SerDes, and packaging integration. Meanwhile, GUC and Alchip could derive greater revenue contributions from Google, AWS, Microsoft, and Meta projects.

Core views

Google TPU v10 is not a single-winner scenario in which a “new supplier replaces an incumbent supplier.” The report believes MediaTek will remain the primary integrator in the Asian supply chain, while participants such as Marvell are more likely to secure adjacent or newly added modules within the TPU ecosystem rather than replace Broadcom in the U.S. or MediaTek in Taiwan. TPU v10 features three structural changes: first, the compute die adopts a KGD model, under which Google owns the compute die design and mask set and can assign back-end design services to other suppliers; second, it adopts 3.5D packaging, requiring 3D IC IP and stacking design services; third, the reticle size could expand from approximately 9 reticles currently to more than 12. The report expects TPU v10 to continue using Intel EMIB-T, while MediaTek's continuously improving SerDes IP strengthens the stickiness of its design services. MediaTek's responsibilities for TPU v10 are expected to be similar to those for TPU v9, including designing the I/O die using its proprietary SerDes IP and integrating the compute die KGD, I/O die, and HBM. Because TPU v10's I/O die is larger, even if MediaTek does not recognize revenue from Google's proprietary compute die, its recognized revenue per chip could still exceed that of TPU v9; the report estimates approximately US$18k. MediaTek's TPU revenue forecasts for 2027 and 2028 remain at US$13-15bn and US$43-45bn: corresponding to approximately 3mn TPU v8t chips in 2027, and approximately 1mn TPU v8t chips and 3mn TPU v9 chips in 2028. By 2029, the report expects shipments of 4mn TPU v9 chips and approximately 1mn TPU v10 chips, driving TPU revenue to approximately US$70bn. The report reiterates OW on MediaTek and notes that it trades at approximately 12 times 2028 EPS. GUC's primary growth driver comes first from Google Axion CPU. Due to capacity constraints, the report lowers its 2026 Axion2 shipment forecast from 1.5mn-2mn chips to 1.2mn-1.3mn chips; based on a chip ASP of approximately US$700, this would contribute US$900-1000mn in revenue to GUC. Demand for 2027 is revised significantly upward, including some Axion2 shipments deferred from 2026 and the start of Axion3 volume production. Google CPU revenue is expected to approach US$2.8bn, comprising 3mn Axion2 chips and 500k Axion3 chips. Because Axion3 increases the number of 3nm compute dies from one to two, its 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. For these Turnkey 2 or 3 projects, GUC provides advanced packaging design support such as silicon interposers and ABF substrates but does not own the ASIC mask set; therefore, gross margin may be only approximately 10%. Axion3's gross margin could be slightly lower than Axion2's due to its larger project scale and higher ASP. GUC also benefits from Microsoft and Meta projects. The report expects demand for Microsoft Maia 200 and Cobalt 200 to more than double year over year in 2027, contributing at least US$250mn in revenue to GUC. Supply-chain research also indicates that GUC may secure the Meta MTIA 600 project led by the Rivos team. The chip is targeted for tape-out in the first half of 2027 and could enter the CoWoS stage by the end of 2027 or in the first half of 2028. Additional TSMC 3nm wafer allocations also support the upward revision to Google CPU revenue and create the possibility that GUC will participate in 3D IC design services for Google TPU v10 in 2028. The report raises GUC's 2026, 2027, and 2028 EPS forecasts by 13%, 11%, and 11%, respectively. As low-margin turnkey revenue rises as a share of the mix, the main text expects gross margin to decline from 20.8% in 2026 to 15.3% in 2028, while the 2028 forecast in the financial tables is 15.8%. GUC's target price is raised from NT$5,688 to NT$6,288, primarily reflecting the 2027—2028 earnings forecasts, additional 3nm capacity, and opportunities from Meta and Microsoft projects. The target price equals 45 times the report's 2027 EPS forecast, below one standard deviation above the company's historical average next-twelve-month P/E since 2013. The bull and bear cases are NT$8,545 and NT$3,870, respectively, corresponding to 61 times and 28 times 2027 EPS. The residual income model assumptions remain unchanged: a cost of equity of 9.2%, including a beta of 1.2, a risk premium of 6%, and a risk-free rate of 2%; a medium-term growth rate of 14.5%; and a perpetual growth rate of 5%. The report expects GUC's 2027 and 2028 EPS to be NT$142.53 and NT$212.89, respectively, and states that it currently trades at approximately 25 times 2028 EPS. Elsewhere in the report, the main text gives the target price as NT$6,388, but the cover page, revision table, and valuation section all show NT$6,288. Alchip's primary incremental growth comes from AWS Trainium. Based on management's statement that Trainium4's potential market size could approach that of Google TPU and on supply-chain research, the report assumes Trainium4 shipments of at least 2.5mn chips in 2028, an ASP excluding HBM of approximately US$10k, and an approximately 30% turnkey share for Alchip, similar to its roughly one-third share of Trainium3. On this basis, it raises Trainium4's 2028 revenue contribution from US$4bn to US$8bn. For Trainium3, because the previous share assumption was too high, the report slightly lowers its estimated revenue contribution to Alchip in 2026 and 2027 to US$1.8bn and US$2.8bn. However, large-scale Trainium ASICs are expected to receive strong support from TSMC, and gross margin is projected to improve from the previously expected low-to-mid teens to near the high teens. These changes lead the report to raise Alchip's 2026, 2027, and 2028 earnings forecasts by 17%, 25%, and 67%, respectively, with 2028 revenue and earnings more than doubling relative to the previous forecasts. The financial forecasts show 2027 and 2028 EPS of NT$225.82 and NT$363.29, respectively, with 2028 EPS growth of approximately 60% year over year. The target price is raised from NT$5,088 to NT$5,888, implying >50% upside, and the report also mentions the potential to secure a second cloud service provider customer. Valuation continues to use a residual income model, with the cost of equity maintained at 10.4%, corresponding to a beta of 1.4, a risk premium of 6.0%, and a risk-free rate of 2.0%. Considering competition from GUC, QCOM, AMD, and other new entrants, the medium-term growth rate is lowered from 16% to 12.5%, and the perpetual growth rate is lowered from 5.0% to 4%. Bull, base, and bear case values are NT$6,880, NT$5,888, and NT$2,465, respectively, corresponding to 30 times, 26 times, and 11 times 2027 EPS. The report states that Alchip trades at approximately 10 times 2028 EPS. The report also cross-validates its AI supply-chain forecasts using CoWoS capacity, chip shipments, and power demand. If approximately 19mn GPUs/ASICs in 2027 are estimated at an average TDP of 2kW per chip, they would require approximately 38GW of installed power capacity, making power supply a potential global bottleneck. As an order-of-magnitude example, an additional 8GW of computing power is equivalent to approximately 4mn Rubin GPUs, or roughly 20% of the AI accelerators TSMC could produce in 2027. The report also estimates that AI HBM demand could reach up to 50bn Gb in 2027, while AI wafer consumption would be at least US$59bn; the corresponding 2026 figures are up to approximately 30bn Gb and at least US$26bn. In its GPU supply-demand validation, the report incorporates HGX systems into its chip-consumption model, assuming that each HGX contains 8 GPUs and that 9 HGX systems are approximately equivalent to one NVL72 rack. It expects Blackwell shipments of 5.4mn chips in 2026, with supply sufficient to meet Grace Blackwell NVL72 demand in the second half of 2026. Blackwell chips that previously appeared to be inventory are viewed as supply-chain buffer inventory and are expected to be absorbed during 2026. Rubin is expected to begin ramping in the third quarter of 2026, with rack shipments starting in the fourth quarter. Shipments of Rubin and Rubin Ultra are expected to approach 7mn chips in 2027, with approximately 90k Rubin NVL72 racks. The report therefore believes chip and system shipments can be aligned and that Rubin may follow a similar buffer-inventory pattern.

Analysis framework

The report first dissects TPU v10's die ownership, I/O design, advanced packaging, and reticle structure to determine which design modules each supplier can undertake. It then uses chip volumes, ASPs, turnkey shares, and capacity allocations derived from supply-chain research to estimate project revenue for MediaTek, GUC, and Alchip, before mapping changes in revenue and gross margin to earnings forecasts. Finally, it determines target prices using residual income models and historical P/E ratios and conducts supply-chain cross-validation using CoWoS capacity, HBM and wafer demand, rack shipments, and power consumption.

Methodology notes

• Industry/Sector Analysis FrameworkVolume-price decomposition

  Decomposition by chip volume, ASP, and supplier share

  The report estimates project revenue by multiplying expected shipment volume by ASP per chip and then applying the supplier's turnkey share. For example, it derives approximately US$8bn in revenue for Alchip using at least 2.5mn Trainium4 chips, an ASP of approximately US$10k, and an approximately 30% share.

• Industry/Sector Analysis FrameworkUpstream-midstream-downstream industry chain transmission

  Transmission from cloud service provider chip demand to design services, wafers, advanced packaging, and HBM

  The report sequentially maps demand for in-house chips from Google, AWS, Microsoft, and Meta to ASIC design services revenue, TSMC 3nm and CoWoS allocations, HBM consumption, and final rack shipments.

• Valuation MethodRIM Residual Income Model

  Residual income valuation

  The base-case target prices for GUC and Alchip are both derived using residual income models that explicitly incorporate the cost of equity, medium-term growth rate, and perpetual growth rate, with revaluations under bull and bear scenarios.

• Valuation MethodPE/PEG valuation

  Historical and scenario P/E comparison

  The report compares the three companies using P/E ratios based on 2027 or 2028 EPS and benchmarks GUC's target valuation against its range of next-twelve-month P/E ratios since 2013.

• (Method outside the vocabulary)

  Chip-system-power consumption cross-validation

  The report converts GPU/ASIC volumes into power demand using an average TDP of 2kW and matches chip shipments with HGX and NVL72 rack volumes to assess whether supply-chain forecasts are constrained by system deliveries or power infrastructure.

Asset mapping & comparison

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

• MediaTek (2454.TW)

  The primary integrator for Google TPU v10, responsible for the I/O die, SerDes, and final packaging integration.

  Strengths

  Its proprietary SerDes IP continues to improve, and it has experience with existing TPU projects and sticky system-integration capabilities; 2029 TPU revenue is expected to reach approximately US$70bn.

  Weaknesses

  Under the COT model, it does not recognize revenue from Google's proprietary compute die KGD, and some back-end design modules may be undertaken by other suppliers.

  Comparison

  The report states that it trades at approximately 12 times 2028 EPS, below GUC at approximately 25 times but slightly above Alchip at approximately 10 times.

  Risks

  Further disaggregation of TPU v10 modules, as well as weaker-than-expected project ramp-up and packaging capacity.

• Global Unichip Corp (3443.TW)

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

  Strengths

  Increased 3nm wafer allocations in 2027 and expanded cloud AI semiconductor exposure across AI accelerators, server CPUs, and BMC chips.

  Weaknesses

  Gross margin on Google CPU turnkey projects may be only approximately 10%; changes in the business mix pressure overall gross margin, while the recovery in non-AI businesses is relatively slow.

  Comparison

  The target price corresponds to 45 times 2027 EPS, while the shares currently trade at approximately 25 times 2028 EPS, above MediaTek and Alchip.

  Risks

  Insufficient advanced-process allocations, project delays, and a greater-than-expected increase in the share of low-margin turnkey revenue.

• Alchip Technologies Ltd (3661.TW)

  AWS Trainium design service provider, with its forecast 2028 Trainium4 revenue contribution raised to US$8bn.

  Strengths

  Large-scale Trainium ASIC projects can receive TSMC support; Trainium3 gross margin is expected to improve to the high teens, and there is an opportunity to secure a second cloud service provider customer.

  Weaknesses

  The Trainium3 share forecast has been lowered, revenue is relatively concentrated in cloud service provider projects, and the 2028 business ramp-up will be accompanied by declining gross margin.

  Comparison

  The report states that it trades at approximately 10 times 2028 EPS, below MediaTek at approximately 12 times and GUC at approximately 25 times.

  Risks

  Competition from GUC, QCOM, AMD, and other new entrants for AI ASIC projects could reduce its share and long-term growth rate.

Key data

• MediaTek 2029 TPU revenueUS$70bnBased on 4mn TPU v9 chips and approximately 1mn TPU v10 chips
• MediaTek recognized revenue per TPU v10 chipApproximately US$18kStill expected to exceed TPU v9 even without recognizing revenue from Google's proprietary compute die KGD
• GUC 2027 Google CPU revenueClose to US$2.8bnIncluding 3mn Axion2 chips and 500k Axion3 chips
• GUC earnings forecast revisions2026 +13%, 2027 +11%, 2028 +11%Driven by cloud service provider projects and Chinese ADAS customers
• GUC target priceNT$6,288Raised from NT$5,688, approximately 45 times 2027 EPS
• Alchip 2028 Trainium4 revenueUS$8bnPrevious forecast was US$4bn
• Alchip earnings forecast revisions2026 +17%, 2027 +25%, 2028 +67%Improved Trainium3 gross margin and expanded Trainium4 contribution
• Alchip target priceNT$5,888Raised from NT$5,088, implying >50% upside
• Implied 2027 GPU/ASIC power demandApproximately 38GWEstimated based on approximately 19mn chips and an average TDP of 2kW per chip
• 2027 AI HBM demandUp to 50bn GbThe report's supply-chain demand forecast
• 2027 AI wafer consumptionAt least US$59bnDriven by production demand for AI GPUs and ASICs
• 2026 Blackwell shipments5.4mn chipsSupply is expected to meet Grace Blackwell NVL72 demand in the second half of 2026
• 2027 Rubin-series shipmentsClose to 7mn chipsCorresponding to approximately 90k Rubin NVL72 racks

Impact & implications

The report believes the increasing complexity of TPU v10 will expand the ASIC design services market rather than result in a single supplier completely replacing incumbent designers. MediaTek can retain the high-value roles of I/O and packaging integration; GUC could expand its cloud AI revenue through Google CPU, Microsoft, Meta, and potential TPU 3D IC projects; and Alchip benefits from the upward revision to Trainium4 scale. However, a rising share of turnkey business will depress gross margins, while competition, power supply, and advanced-process and packaging capacity will continue to determine the pace of revenue realization.

Risks

• Global power infrastructure may be unable to support in a timely manner the approximately 38GW of power demand corresponding to roughly 19mn GPUs/ASICs in 2027.
• Constraints on 3nm wafer and CoWoS capacity have already caused the 2026 Google Axion2 shipment forecast to fall below previous expectations.
• The rising share of GUC's low-margin turnkey revenue could cause gross margin to decline from 20.8% in 2026 to approximately 15% in 2028.
• Alchip faces competition from GUC, QCOM, AMD, and other new entrants for AI ASIC projects.
• The slow recovery in GUC's non-AI businesses could weaken the contribution of cloud AI project growth to overall performance.

What to watch

• How Google ultimately divides TPU v10 design modules such as compute-die back-end work, 3D IC, the I/O die, and packaging integration.
• Whether MediaTek's TPU v9 and TPU v10 can ramp according to the 2028—2029 shipment trajectory and achieve approximately US$18k in recognized revenue per TPU v10 chip.
• The TSMC 3nm wafer allocations secured by GUC in 2027 and the actual shipment cadence of Google Axion2 and Axion3.
• Whether Meta MTIA 600 can tape out in the first half of 2027 and enter the CoWoS stage by the end of 2027 or in the first half of 2028.
• Whether Alchip can secure an approximately 30% turnkey share of Trainium4 and whether Trainium3 gross margin can rise to the high teens.
• Whether CoWoS, HBM, wafer, and power supply can match demand for AI accelerators in 2027.
• Whether Rubin can begin ramping in the third quarter of 2026 and enter rack shipments in the fourth quarter.