Morgan Stanley's report states that the release of GPT-6 Astra will shift the focus of market discussions. Astra represents a substantial leap in reasoning, engineering, computer usage, and execution of physical-world tasks, signaling AI's expansion from simple interactions to complex task execution.

Author: Long Yue

Source: Wall Street Journal

The release of a model is shifting the direction of market discussions.

For the past several months, the market has been debating a question: “How much infrastructure is truly needed to serve known AI demand?” The underlying implication is: Has AI infrastructure been over-invested in?

But according to Zhui Feng Trading Desk, Morgan Stanley's report released on September 7 stated that the significance of GPT-6 Astra should not be simply understood as a continuation of another scaling narrative. This is not another model upgrade.

Astra represents a substantial leap in breadth of capabilities and interoperability, manifested in four key areas: reasoning, engineering, computer use, and execution of physical-world tasks. This means AI is now advancing to the point where it can perform complex reasoning, professional engineering tasks, directly operate computers, and handle real-world tasks. With more things now possible, more opportunities to generate revenue emerge. OpenAI’s latest post also notes that better models are opening up new “domains of work.”

AI narratives are once again shifting from "demand debates" back to "physical bottlenecks"

The emergence of Astra may shift the core market question from "How much infrastructure is needed to serve known AI demands?" to "As model intelligence improves, how many new workloads will become economically viable?"

This is a completely opposite issue. The former questions demand-side factors, while the latter reflects supply-side pressure.

Analysts present a core economic logic in their report: better models create an elasticity effect— as AI gains more intelligence and utility per dollar spent, the volume, duration, and complexity of reasoning workloads increase accordingly.

This is similar to the classic “Jevons Paradox”: increased efficiency does not reduce consumption, but instead expands total consumption. Higher intelligent output per unit cost → previously uneconomical use cases become viable → new workloads flood in → demand for computing power and infrastructure grows even larger.

Analysts believe that if Astra's capabilities are translated into commercially useful applications, the addressable AI revenue pool could expand significantly beyond today's use cases dominated by chat and coding.

Morgan Stanley estimates the global TAM for knowledge work at approximately $22.5 trillion (based on 900 million knowledge workers globally with an average annual salary of about $25,200); the consumer spending TAM is approximately $30 trillion, covering retail and travel (about $16 trillion), autonomous driving/mobility (about $4 trillion), food delivery (about $4 trillion), and advertising (about $3 trillion).

The analysts' conclusion is that Astra's emergence has shifted the bottleneck narrative from "demand formation" back to "physical supply constraints"—whether these intelligences can be delivered at scale.

What exactly is being referred to? Hashrate. Electricity. Materials. Labor. Etc.

Where are the physical bottlenecks specifically: ABF and HBM4E

The supply-side tightness has two specific leverage points.

The first is an ABF substrate.

Morgan Stanley expects ABF substrates to experience supply shortages starting in 2027, with the gap continuing to widen through 2030. The key constraint is that new capacity requires at least two years to come online.

The second is the backend-of-line (BEOL) complexity of HBM4E.

The report indicates that the backend processes of HBM4E represent a major paradigm shift in semiconductor manufacturing—HBM is evolving from a dedicated 3D memory stack into a highly integrated custom chiplet logic system.

Specific impact chain:

• The increase in HBM4E interconnect layers (such as SK Hynix’s introduction of dummy bumps) will force significant DRAM capital expenditures toward BEOL capacity expansion.
• Capital expenditures for FEOL process migration and DRAM GB shipments are expected to accelerate only by the second half of 2027.
• DRAM manufacturers prioritize access to production capacity, so NAND capacity expansion may be delayed.

The common characteristic of these two bottlenecks: both are verifiable and traceable physical constraints, not market sentiment.

Electricity: The most real physical bottleneck

As regulatory pressure on data centers increases, power supply has become another critical physical bottleneck. Analysts expect the total computing power of hyperscale cloud providers to grow from approximately 35 GW in 2025 to about 145 GW in 2028, an increase of roughly four times.

Analysts note that the U.S. faces a 38-gigawatt power shortfall, and data centers will increasingly adopt behind-the-meter self-generation solutions.

This line estimates that the post-meter power solution will add approximately $3 billion in capital expenditure per gigawatt. For example, with NVIDIA’s Rubin Ultra first-generation chips, the all-inclusive cost including post-meter power is approximately $50 billion per gigawatt.

What is the market undervaluing?

Analysts believe the market is currently undervaluing three things:

First, the global tech beneficiaries of GPT-6 Astra have not been fully priced in.

Second, the supply constraints may last longer. The bottlenecks in ABF and HBM4E BEOL cannot be resolved in the short term.

Third, some stocks that don’t rely on AI are quietly gaining strength. Analog chips (STM, NXP, Renesas) have emerged from a three-plus-year L-shaped bottom and are now in the early stages of a cyclical recovery—inventory levels have been reduced, pricing has stabilized, and industrial orders are improving.

The conclusion is not "buy more AI"

Morgan Stanley's investment priorities:

AI computing power (highest priority) > Network (secondary) > Memory (optional) + Analog chips (early-cycle hedge), including:

• AI computing power: GPU (NVIDIA), ASIC (MediaTek, GUC), ABF substrates (Unimicron, Ibiden), MLCCs (Murata, Samsung Electro-Mechanics), back-end packaging and testing (Advantest, Tokyo Electron, Hua Feng, ASE, Kyocera) power supplies (Delta)
• Network: GLW, LITE, COHR, KEYS, Furukawa Electric, Fujikura
• Memory: Prioritize structural growth in memory and localization (CXMT); SK Hynix, Samsung, and Kioxia have tactical upside potential amid continued supply tightness.
• Non-AI: Simulation chips (STMicroelectronics, NXP, Renesas)

This line states: “We favor companies positioned at the intersection of longer reasoning cycles, limited physical capacity, rising content intensity, and increased manufacturing complexity.”

Areas where you need to stay alert

Morgan Stanley's report is not one-sidedly optimistic and clearly highlights three areas to watch:

First, expectations for AI are already very high. The market's tolerance for AI companies has shifted from "strong performance" to "must deliver perfect results"; even if Astra makes substantial progress, if earnings do not significantly exceed expectations, the stock reaction may still be limited.

Second, capital expenditure growth is expected to slow in 2028. Stock valuations focus on the direction of growth; a slowdown may continue to suppress valuations, even if absolute values are still rising.

Third, macroeconomic headwinds persist. The report highlights uncertainties such as oil prices, inflation, the Federal Reserve’s interest rate path, and the 2028 U.S. election, with particular attention to potential political resistance to data center expansion if the Democratic Party gains executive power.