Morgan Stanley raises focus on the AI supply chain: 2027 demand for CoWoS, Rubin, and ASIC remains strong
AI summary card
Morgan Stanley raises focus on the AI supply chain: 2027 demand for CoWoS, Rubin, and ASIC remains strong
The report maintains its view of AMD's 2027 CoWoS allocation of 240k and expects Rubin/Rubin Ultra in 2027 to be close to 7mn units, with Rubin NVL72 racks around 90k; Blackwell inventory concerns are interpreted as supply-chain buffering.
- AMD's 2027 CoWoS allocation is held at 240k for now, with MI455 expected to ship 1mn units and Meta custom MI450 expected 500k units.
- Blackwell chip "inventory" is assessed as supply-chain buffering, expected to be fully drawn down by 2026, with Rubin potentially following a similar pattern.
- Google TPU Sunfish shipments are more concentrated in 4Q26, while Zebrafish 4Q26 ramp is unchanged.
- Meta cancelled the original 2nm ASIC Olympus and replaced it with Apollo; Broadcom continues to provide design services, with mass production expected in 1Q28.
- AI HBM consumption in 2027 may reach up to 50bn Gb, and AI wafer consumption at least US$47bn.
Report interpretation
Overview
This report is a Morgan Stanley event update on the Asia-Pacific technology AI supply chain, centered on the 2027 global CoWoS allocation, AMD MI400 series, Google TPU pace, NVIDIA Blackwell/Rubin chips and rack matching, and changes to Meta ASIC projects. The report views AI chips, HBM, advanced packaging, and cloud provider capex as the main drivers of the supply chain.
Core views
The report maintains AMD's 2027 CoWoS allocation outlook of 240k, but notes that AMD reduced CoWoS bookings in 2026, so execution risk should not be ignored. For NVIDIA, so-called Blackwell inventory looks more like supply-chain buffering rather than a demand issue, and is expected to be fully absorbed by 2026; Rubin/Rubin Ultra chip shipments in 2027 are expected near 7mn units, with Rubin NVL72 racks around 90k. TPU progress from Google is delayed somewhat but does not indicate order cuts; Meta's new ASIC Apollo is pushed out to volume production in 2028.
Analysis framework
The report evaluates AI semiconductor supply-chain demand strength in 2026-2027 through supply-chain research, CoWoS capacity allocation, chip-to-rack consumption modeling, HBM and wafer usage calculation, and tracking of key customer and ASIC project progress.
Methodology notes
Chip shipments and upstream resource consumption are estimated using CoWoS capacity, chips per wafer, HBM configuration, and customer demand.
This approach compares demand pull intensity for advanced packaging, HBM, and wafers across different AI GPU, CPU, and ASIC projects.
The chip consumption model incorporates HGX servers, treating nine HGX servers as equivalent to one NVL72 rack.
This model assesses whether Blackwell and Rubin chip shipments are sufficient to meet AI server rack demand and whether so-called inventory reflects real surplus.
ASIC supply-chain pace is judged by tape-out timing, volume production timing, supplier service relations, and changes in customer customized variants.
The report uses this framework to assess the demand impact of Google TPU, Meta Apollo, potential GUC projects, and Broadcom design services.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- AMDAI GPU and CPU demand driver
- Strengths
- MI455, MI450, and Venice CPU drive CoWoS and HBM demand in 2027; key customers include Microsoft, AWS, Oracle, and Meta.
- Weaknesses
- The report notes AMD reduced CoWoS bookings in 2026, and that the 2027 allocation of 240k carries execution risk.
- Comparison
- Compared with NVIDIA, AMD's CoWoS scale is smaller but the incremental trend is clear, with the MI400 series split into standard and Meta customized versions.
- Risks
- Changes in customer orders, CoWoS booking adjustments, and MI400 execution pace below expectations.
- NVIDIA Blackwell/RubinCore variable for AI GPU platform and rack demand
- Strengths
- Blackwell chip inventory is seen as supply-chain buffering, while Rubin/Rubin Ultra are expected to be near 7mn units in 2027.
- Weaknesses
- Rubin ramp and rack shipments still need to materialize according to the 3Q26 and 4Q26 timelines.
- Comparison
- The Blackwell 2026 chip consumption model is extended to HGX-equivalent racks, and Rubin is expected to follow a similar supply-chain buffering pattern.
- Risks
- Rubin mass-production delay, rack-delivery bottlenecks, and lower-than-expected cloud capex.
- Google TPUASIC and CoWoS/HBM demand source
- Strengths
- Sunfish full-year volume remains 960k, and the view of a Zebrafish ramp in 4Q26 is unchanged.
- Weaknesses
- KYEC 3Q26 revenue growth guidance was reduced from 15% q/q to near 10% q/q; Sunfish and Rubin are somewhat delayed.
- Comparison
- TPU revenue recognition is more weighted to 4Q26 or 1Q27, differing from the timing of some GPU chains.
- Risks
- Mass-production pace pushback, customer order adjustments, delayed supply-chain revenue recognition.
- Meta ASICCustom AI ASIC demand driver
- Strengths
- The new 2nm Apollo succeeds Olympus, Broadcom continues design services, and GUC may receive relevant projects from the Rivos team.
- Weaknesses
- Olympus was canceled, and the new project’s mass-production timing has been delayed to 1Q28.
- Comparison
- Meta appears in both AMD MI450 custom demand and its own ASIC project, reflecting a multi-path AI compute strategy.
- Risks
- Tape-out delay, delayed CoWoS onboarding, and project scope changes.
- TSMC and CoWoS supply chainCore beneficiary of advanced packaging and AI wafer consumption
- Strengths
- The report estimates TSMC's AI-related revenue CAGR for 2024-2029e could reach 60%, with CoWoS demand driven by NVIDIA, AMD, Google, AWS, and other major customers.
- Weaknesses
- Capacity expansion needs to match customer timing; delayed projects could cause quarterly revenue volatility.
- Comparison
- Compared with a single-chip company, TSMC with the CoWoS chain covers a broader range including GPU, ASIC, CPU, and wafer test.
- Risks
- Customer concentration, advanced packaging capacity mismatch, and capex cycle volatility.
- OSAT and assembly test housesCoW and late-stage packaging/test beneficiary
- Strengths
- ASE/SPIL, Amkor, and Powertech are cited as concentrated sites for AMD Venice CPU CoW production.
- Weaknesses
- Revenue timing is affected by upstream chip ramp and customer orders.
- Comparison
- Compared with front-end wafer fabs, OSATs are more directly impacted by CoW mass-production timing.
- Risks
- Project delays, utilization volatility, and customer pricing pressure.
Key data
- AMD 2027 CoWoS allocation240kThe report maintains this view, but flags execution risk.
- AMD MI455 2027 shipment1mn unitsStandard MI400 series, paired with Helios rack, with key customers including Microsoft, AWS, and Oracle.
- AMD MI450 2027 shipment500k unitsMeta customized split-size chip, with 1 compute die and 6 HBM4 12hi.
- AMD Venice CPU 2027 shipment5.7-6mn unitsVenice is AMD's first CoWoS-based CPU; CoW production is concentrated at OSATs such as ASE/SPIL, Amkor, and Powertech.
- Sunfish full-year shipment960k unitsShipments are expected to be concentrated in 4Q26, with revenue potentially recognized in 4Q26 or 1Q27.
- Blackwell 2026 shipment5.4mn unitsChip volume is expected to meet Grace Blackwell NVL72 demand in 2H26.
- Rubin/Rubin Ultra 2027 shipmentNearly 7mn unitsRubin is expected to ramp from 3Q26, with rack shipments starting in 4Q26.
- Rubin NVL72 2027 racks90kEstimated using the chip-rack matching model in the report.
- 2027 AI HBM consumptionup to about 50bn GbFrom Exhibit 1's AI HBM consumption calculation.
- 2027 AI wafer consumptionat least US$47bnFrom Exhibit 2's AI wafer consumption calculation.
- 2026 AI HBM consumptionup to about 30bn GbAI HBM demand disclosed in the report charts.
- 2026 AI wafer consumptionat least US$26bnAI wafer consumption disclosed in the report charts.
Impact & implications
The report’s implications for the AI semiconductor chain are broadly positive: advanced packaging, HBM, AI GPUs, AI ASIC, and OSAT continue to benefit from cloud AI capex and chip platform upgrades. For investors, the key issue is not a single inventory number but whether chips, HBM, CoWoS capacity, and rack deliveries are in sync. In the near term, TPU and some Rubin revenue may shift to 4Q26 or 1Q27, but the report does not show core orders being cut.
Risks
- AMD's 2027 CoWoS allocation has execution risk, as AMD reduced CoWoS bookings in 2026.
- Revenue recognition for Google Sunfish and Rubin may shift to 4Q26 or 1Q27.
- Meta ASIC has shifted from Olympus to Apollo, with volume production expected by 1Q28, and project timing carries delay risk.
- If AI server rack delivery, HBM supply, and CoWoS capacity are not matched, chip shipment conversion could be affected.
- If cloud provider AI capex comes in below expectations, AI GPU, ASIC, and advanced packaging demand may weaken.
- The report includes disclosures on export controls, U.S. executive orders, and restricted entities; related securities and trading may face compliance constraints.
What to watch
- Whether AMD's 2027 240k CoWoS allocation is maintained, and whether MI455/MI450 customer orders are delivered.
- Whether Rubin ramping in 3Q26 and rack shipments in 4Q26 proceed as planned.
- Whether Blackwell chip buffering inventory is fully absorbed in 2026.
- Google Sunfish 4Q26 shipments, Zebrafish 4Q26 ramp, and TPU revenue recognition pace.
- Progress of Meta Apollo tape-out, Broadcom design services, and potential GUC projects.
- TSMC CoWoS capacity expansion, HBM supplier share, and changes in AI wafer consumption.
- Whether China AI inference demand and potential rebound in NVIDIA 5090 GPU prices continue.