Morgan Stanley says OpenAI's GPT-6 Astra launch reframes the AI infrastructure debate from demand skepticism to physical supply constraints, projecting hyperscaler compute will grow roughly fourfold to 145GW by 2028.
Retour
GPT-6 Astra shifts AI debate to supply as compute grows 4x
A single model release is resetting the question investors have argued over for months. Rather than asking whether AI infrastructure has been overbuilt, the debate now centers on whether physical supply — chips, materials, and power — can keep pace with what GPT-6 Astra makes possible.
Morgan Stanley's Sept. 7 report argues Astra's capability breadth — spanning reasoning, engineering, computer use, and physical-world task execution — expands the addressable AI revenue pool far beyond today's chat-and-coding workloads. The bank's analysts describe an elasticity effect: as each dollar of AI spend buys more intelligence, inference workloads multiply in number, duration, and complexity. Hyperscaler compute is projected to grow roughly fourfold from about 35GW in 2025 to 145GW by 2028.
The report estimates the global knowledge-work TAM at roughly $22.5 trillion, based on 900 million knowledge workers earning an average $25,200 annually. Consumer spending across retail, travel, autonomous mobility, food delivery, and advertising adds another $30 trillion in potential markets. The question shifts from how much infrastructure known AI demand requires to how many new workloads become economically viable as model intelligence rises — and whether physical supply can keep pace.
ABF substrates and HBM4E bind first
Morgan Stanley expects ABF substrates — the advanced laminate material used in high-performance chip packaging — to face a supply shortage starting in 2027 that widens through 2030. New capacity requires at least two years to come online, meaning near-term demand growth will outstrip what suppliers can add.
HBM4E's back-end-of-line (BEOL) complexity represents a separate structural shift. High-bandwidth memory is evolving from dedicated 3D stacking into highly integrated custom chiplet logic systems, with SK Hynix introducing dummy bump interconnects that add layers. The report says a large share of DRAM capital expenditure will be redirected toward BEOL capacity expansion, while front-end-of-line (FEOL) process migration and DRAM gigabyte shipment acceleration are not expected until the second half of 2027. NAND capacity expansion may be deferred as DRAM makers prioritize equipment.
Power is the most binding physical limit
The US faces a 38GW electricity gap, and data centers are increasingly turning to behind-the-meter self-generation. Morgan Stanley estimates this adds roughly $3 billion in capex per gigawatt; for Nvidia's Rubin Ultra generation, the all-in cost including behind-the-meter power reaches about $50 billion per gigawatt.
The bank's investment priority ranks AI compute first — GPU leader Nvidia, ASIC designers MediaTek and Global Unichip, ABF substrate makers Unimicron and Ibiden, MLCC suppliers Murata and Samsung Electro-Mechanics, and back-end test and assembly firms including Advantest, Tokyo Electron, and ASE. Networking follows with Corning, Lumentum, Coherent, and Furukawa Electric among the picks. Memory is selective, favoring structural share gainers such as ChangXin Memory Technologies, while SK Hynix, Samsung, and Kioxia offer tactical upside if supply stays tight. Analog chipmakers STMicroelectronics, NXP, and Renesas serve as an early-cycle hedge after more than three years of an L-shaped bottom.
What the market may be underpricing
Morgan Stanley flags three areas where investors may be underestimating the setup. First, the global technology beneficiaries of GPT-6 Astra are not yet fully priced. Second, supply tightness in ABF and HBM4E BEOL will persist longer than near-term fixes can address. Third, analog chip stocks that do not depend on AI are showing early signs of cyclical recovery — inventory has been drawn down, pricing is stabilizing, and industrial orders are improving.
The report also cautions that AI expectations have reached a level where tolerance has shifted from good earnings to must-be-perfect earnings. Even with Astra delivering substantive progress, stock reactions may remain muted unless results substantially exceed forecasts. Decelerating 2028 capex growth could also weigh on valuations even as absolute spending continues to rise. Macro headwinds — oil prices, inflation, the Federal Reserve's rate path, and potential political resistance to data center expansion if Democrats win the 2028 US election — add further uncertainty.
Morgan Stanley said it prefers companies positioned at the intersection of a broader inference cycle, limited physical capacity, rising content intensity, and increased manufacturing complexity. The bank's ranking: AI compute above networking, with memory selective and analog chips as a portfolio hedge.
This article is for informational purposes only and does not constitute investment advice.
Fonctionnalités
© 2026 Edgen fonctionné par