Morgan Stanley updates 2027 CoWoS allocation and AI ASIC dynamics, maintaining a constructive demand read

Overview

This report, focused on the Asia-Pacific Technology AI supply chain, updates 2027 global CoWoS capacity allocation, AMD MI400-series, Google TPU, Meta ASIC, and NVIDIA Blackwell/Rubin chip and rack shipment matching. The overall conclusion is that AI semiconductor demand remains strong, with continued benefits for CoWoS, HBM, AI GPUs/ASICs, and related OSAT segments, although there is execution risk in some customer and product ramp timing.

Core views

The report maintains its 2027 AMD CoWoS allocation call at 240k wafers, with MI455 shipments expected to be about 1 million units, Meta-custom MI450 about 500k units, and, after Venice CPU adopts CoWoS, related CPU chips likely to ship 5.7 to 6.0 million units in 2027. On Google TPU, some Sunfish and Rubin-related revenue may shift from 3Q26 into 4Q26 or 1Q27, but the report views this as a timing shift rather than demand reduction. For NVIDIA, Blackwell chip inventory is seen as supply-chain buffering and is expected to be fully digested in 2026; in 2027, Rubin plus Rubin Ultra is expected to be close to 7 million units, with Rubin NVL72 racks about 90,000 units. On Meta ASIC, Olympus has been cancelled and Apollo will follow, with Broadcom continuing design-service support and mass production expected in 1Q28.

Analysis framework

The report primarily uses supply-chain research, CoWoS allocation modeling, chip-to-rack conversion, HBM consumption calculations, and customer-order cadence tracking to assess AI semiconductor demand. The analysis includes HGX servers in the chip-consumption model and equates nine 8-GPU HGX servers to one NVL72 rack to assess the matching between Blackwell and Rubin shipments and rack demand.

Methodology notes

• Supply-chain estimationCoWoS capacity-allocation model

  Estimate AI chip shipments and advanced packaging demand by customer, chip type, and CoWoS wafer allocation.

  The report derives 2027 demand for AMD, NVIDIA, Google TPU, AWS, Microsoft, Meta and others by using CoWoS allocation across AI chips, chips per CoWoS wafer, and customer mix.

• Shipment matchingChip-to-rack conversion

  Align GPU chip shipments with NVL72 rack and HGX-equivalent rack demand.

  The report treats nine HGX servers as one NVL72 equivalent rack to test whether Blackwell and Rubin chip supply can support AI server rack shipments.

• Capacity demandHBM consumption estimation

  Estimate HBM demand by chip counts, HBM capacity, stack specification, and supplier mix.

  The report expects 2027 AI HBM consumption to rise to as much as about 50bn Gb and 2026 around 30bn Gb, indicating AI accelerators continue to significantly pull the HBM supply chain.

Asset mapping & comparison

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

• AMD

  Core beneficiary with execution risk

  Strengths

  MI455, MI450 and Venice CPU provide CoWoS and chip shipment increments in 2027, with key customers including Microsoft, AWS, Oracle and Meta.

  Weaknesses

  AMD had cut CoWoS bookings in 2026, and the report explicitly flags execution risk for 2027 as well.

  Comparison

  Compared with NVIDIA Rubin, AMD's CoWoS allocation scale is smaller, but MI400-series and CPU adoption of CoWoS broaden product coverage.

  Risks

  Customer demand realization, CoWoS booking revisions, MI400 production ramp timing, and supply-chain execution.

• NVIDIA Blackwell/Rubin

  Core anchor for AI GPU demand and rack shipments

  Strengths

  Blackwell inventory is viewed as supply-chain buffering and expected to be digested in 2026; Rubin/Rubin Ultra is expected to near 7 million units in 2027.

  Weaknesses

  The supply chain needs to stay synchronized across chips, HGX servers, and NVL72 racks.

  Comparison

  The report suggests Rubin may replicate Blackwell's supply-chain buffering and demand-digestion pattern.

  Risks

  Rack delivery pace, advanced packaging capacity, and AI capex volatility.

• Google TPU supply chain

  ASIC and CoWoS demand source

  Strengths

  Sunfish full-year shipments are still expected around 960k units, while the 4Q26 ramp view for Zebrafish remains unchanged.

  Weaknesses

  KYEC 3Q26 revenue growth expectation is lower than previously assumed, and there is slight timing delay for Sunfish and Rubin.

  Comparison

  Compared with the GPU chain, TPU revenue is more exposed to production-timing shifts at a single customer.

  Risks

  TPU ramp delays, smartphone SoC order reductions, and delayed revenue recognition.

• Meta ASIC

  Shift in self-developed ASIC strategy

  Strengths

  Apollo succeeds Olympus, Broadcom continues to provide design services, and GUC may win projects tied to the Rivos team.

  Weaknesses

  Original Olympus was cancelled, Apollo ramp is moved to 1Q28, and CoWoS lift may not release until end-2027 or 1H28.

  Comparison

  Meta-custom MI450 and self-developed ASIC initiatives coexist, showing a diversified AI compute supply strategy.

  Risks

  tape-out progress, 2nm production start, CoWoS release timing, and project ownership uncertainty.

• HBM and CoWoS supply chain

  Upstream constraint and beneficiary path for AI chip ramping

  Strengths

  AI HBM consumption in 2027 is expected to peak at about 50bn Gb, while AI wafer consumption is at least US$47bn.

  Weaknesses

  Capacity expansion and customer allocation remain highly concentrated in a few large customers and advanced packaging nodes.

  Comparison

  CoWoS and HBM are shared bottleneck resources for GPUs, ASICs, TPUs, and AI server CPUs.

  Risks

  Capacity mismatch, customer order changes, advanced packaging yield, and HBM supplier delivery.

Key data

• AMD 2027 CoWoS allocation240k wafersThe report maintains this forecast but flags execution risk.
• MI455 2027 shipments1mn unitsStandard version, 2 compute dies, 12 HBM4 12hi, paired with Helios rack.
• MI450 2027 shipments500k unitsMeta custom half-size version, 1 compute die, 6 HBM4 12hi.
• AMD Venice CPU 2027 shipments5.7-6mn unitsCompared with 2026 at about 1mn units, CoWoS production is concentrated at ASE/SPIL, Amkor, and Powertech OSAT.
• Sunfish full-year shipments960k unitsShipments are expected to be concentrated in 4Q26, while revenue may shift to 4Q26 or 1Q27.
• Blackwell 2026 shipments5.4mn unitsThe report believes chip volumes can meet Grace Blackwell NVL72 demand by 2H26.
• Rubin/Rubin Ultra 2027 shipmentsclose to 7mn unitsRubin is expected to ramp starting 3Q26, with rack shipments starting in 4Q26.
• Rubin NVL72 2027 racks90k racksUsed as a cross-check between Rubin chip supply and rack demand.
• 2027 AI HBM consumptionup to 50bn GbFrom Exhibit 1 supply-chain estimates in the report.
• 2027 AI wafer consumptionat least US$47bnFrom Exhibit 2 in the report.

Impact & implications

For investment implications, the report reinforces the durability of AI semiconductor demand, benefiting CoWoS advanced packaging, HBM, OSAT, wafer foundries, AI ASIC design services, and AI server CPU-related chains. NVIDIA Blackwell inventory risk is de-emphasized, while Rubin demand expectations remain strong. AMD has incremental opportunities in MI400 and Venice CPU, but execution stability is the key variable. Timing shifts in Google TPU and Meta ASIC suggest quarter-level revenue mismatch may occur in the supply chain rather than demand disappearance.

Risks

• AMD's 2027 CoWoS allocation carries execution risk, especially considering its CoWoS bookings were cut in 2026.
• Timing delays for products such as Sunfish and Rubin may shift supply-chain revenue from 3Q26 to 4Q26 or 1Q27.
• The Meta ASIC shift from Olympus to Apollo, with ramp pushed to 1Q28, adds timing uncertainty for project milestones.
• If AI capex slows, it may impair realization of chip, rack, CoWoS, and HBM demand.
• Export controls, executive orders, and compliance restrictions may affect securities or supply-chain transactions.

What to watch

• Actual 2027 AMD MI455, MI450, and Venice CPU CoWoS bookings and shipment realization.
• The ramp pace of Google TPU Sunfish, Zebrafish, and later TPU products in 4Q26 to 1Q27.
• Whether Rubin 3Q26 ramp and 4Q26 rack shipment launch proceed smoothly.
• Whether Blackwell chip inventory is fully digested in 2026 as the report expects.
• Tape-out progress, Broadcom design-service execution, and GUC-linked project rollout for Meta Apollo ASIC.
• TSMC, OSAT, and HBM supplier CoWoS and HBM expansion speed and changes in customer allocations.