The True Significance of GPT-6 Astra: Bringing the "AI Narrative" Back from "Demand Disputes" to "Physical Limitations"

The release of a model is changing the direction of market debate.

For the past few months, the market has been debating one question: “How much infrastructure is required to serve the known demand for AI?” The underlying implication of this question is whether AI infrastructure has already been over-invested.

However, according to news from BlockBeats Trading Desk, a report released by Morgan Stanley on September 7 points out that the importance of GPT-6 Astra should not be simply understood as a continuation of another scaling narrative. This is not just yet another model upgrade.

Astra represents a substantial leap in the breadth of capabilities and interoperability, specifically in four directions: reasoning capabilities, engineering capabilities, computer usage, and performing tasks in the physical world. This means that AI is starting to cross into executing complex reasoning, professional engineering tasks, direct computer operations, and even handling real-world tasks. The range of tasks it can execute has expanded, as has the range of profitable scenarios. OpenAI’s latest publication also states that better models open up new “job domains.”

AI Narrative Returns from “Demand Debate” to “Physical Bottleneck”

Astra’s emergence may shift the market’s core question from “How much infrastructure is needed for the known AI demand,” to: “As model intelligence improves, how many new workloads will become economically viable?”

This is a fundamentally different question. The former is skepticism on the demand side; the latter represents supply-side pressure.

• Elastic Effect: More Intelligence per Dollar Spent

Analysts proposed a core economic rationale in the report: Better models can create an elastic effect—as AI gets more intelligence and utility for every dollar spent, the number, duration, and complexity of inference workloads all increase.

This is similar to the classic “Jevons Paradox”: efficiency improvements do not reduce consumption, but instead increase total consumption. As the output of intelligence per unit cost increases → previously uneconomical application scenarios become feasible → new workloads surge → demand for computing power and infrastructure actually grows.

• The Accessible Market Expands Substantially

Analysts believe that if Astra's capabilities are translated into commercially useful applications, the total addressable AI revenue pool will expand dramatically, far beyond today’s usage scenarios focused mainly on chat and coding.

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

• The Physical Constraints on the Supply Side are the True Bottleneck

The analysts’ conclusion is: the emergence of Astra shifts the bottleneck narrative from “demand formation” back to “physical supply constraints”—that is, whether these intelligences can be delivered at scale.

What does that specifically refer to? Computing power. Electricity. Materials. Labor. Etc.

Where Specifically are the Physical Bottlenecks: ABF and HBM4E

The supply-side tension has two concrete anchors.

The first is ABF substrate.

Morgan Stanley expects ABF substrates to start being in short supply from 2027, and the shortfall will continue to widen through 2030. The key constraint: it takes at least two years for new capacity to come online.

The second is the complexity of HBM4E back-end-of-line (BEOL) processes.

The report points out that HBM4E's BEOL represents a major paradigm shift in semiconductor manufacturing—HBM has evolved from a dedicated 3D memory stack to a highly integrated, customized chiplet logic system.

The specific impact chain:

• Increased number of HBM4E interconnection layers (e.g., SK Hynix introduces dummy bump), a large amount of DRAM capex will be forced into BEOL capacity expansion
• FEOL process node migration and accelerated DRAM GB shipment volume; not expected until the second half of 2027
• DRAM manufacturers will prioritize equipment capacity, meaning NAND capacity expansion may be delayed

The common characteristics of these two bottlenecks: both are verifiable, traceable physical constraints—not driven by market sentiment.

Electricity: The Most Real Physical Bottleneck

With regulatory pressure on data centers rising, power supply has become another key physical bottleneck. Analysts expect total computing power of hyperscale cloud providers to rise from about 35GW in 2025 to about 145GW in 2028, a roughly fourfold increase.

Analysts point out that the US faces a gap of 38GW in electricity, and data centers will increasingly adopt behind-the-meter self-generation solutions.

The bank estimates that behind-the-meter solutions will add about $3 billion in capex per gigawatt. Using the example of Nvidia’s Rubin Ultra generation of chips, the all-in cost (including behind-the-meter power) is about $5 billion per gigawatt.

What is the Market Underestimating?

Analysts believe that the market is currently underestimating three things:

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

Second, the duration of supply tension may be longer. The constraints of ABF and HBM4E BEOL cannot be solved in the short term.

Third, some stocks that do not depend on AI are quietly strengthening. Analog chips (STM, NXP, Renesas), after an L-shaped bottom of over three years, are in the early phase of cyclical recovery—inventory correction, pricing stabilizing, industrial orders picking up.

The Conclusion is Not “Buy More AI”

Morgan Stanley’s investment priorities:

AI computing power (top priority) > Networking (next) > Memory (selectively) + Analog chips (early cycle hedge), including:

• AI Computing Power: GPU (Nvidia), ASIC (MediaTek, GUC), ABF substrate (Unimicron, Ibiden), MLCC (Murata, Samsung Electro-Mechanics), backend packaging and testing (Advantest, Tokyo Electron, ASE, King Yuan), power (Delta)
• Networking: GLW, LITE, COHR, KEYS, Furukawa Electric, Fujikura
• Memory: Prioritize structural share growth and localization (CXMT), SK Hynix, Samsung, Kioxia; tactical upside given persistent tight supply
• Non-AI: Analog chips (STMicroelectronics, NXP, Renesas)

The bank states: “We prefer positioning in companies at the intersection of a broader inference cycle, limited physical capacity, rising content intensity, and increased manufacturing complexity.”

Points Where Caution is Required

Morgan Stanley’s report is not one-sidedly optimistic, and clearly flags three points for attention:

First, AI expectations are already very high. The market’s tolerance for AI companies has shifted from “good performance” to “must deliver perfect performance”. Even if Astra brings substantial progress, unless performance far exceeds expectations, stock price reactions may still be limited.

Second, 2028 capex growth is expected to slow. Equity pricing depends on the direction of growth rate changes; a deceleration may continue to pressure valuations, even if the absolute value is still rising.

Third, macro headwinds remain. The report mentions oil prices, inflation, the Fed’s interest rate path, and the uncertainties of the 2028 US general election, especially potential political resistance to data center expansion if the Democrats win executive power.