The release of a single model is now redirecting the entire market conversation.
Over the past several months, the market has been locked in a debate: "How much infrastructure is truly needed to serve existing AI demand?" The underlying implication is whether AI capital spending has already gone too far.
According to trading desk sources, a September 7 report from Morgan Stanley argues that the importance of GPT-6 Astra should not be dismissed as merely another chapter in the scaling narrative. This is not just another routine model upgrade.
Astra represents a substantial leap in capability breadth and interoperability, manifesting across four key areas: reasoning, engineering proficiency, computer usage, and execution of physical-world tasks. This signals AI's transition into handling complex logic, specialized engineering, direct computer manipulation, and even real-world operations. With a wider range of achievable tasks comes a wider range of monetizable opportunities. OpenAI's latest communications also indicate that improved models unlock new "work domains."
AI narrative shifts from demand debates back to physical bottlenecks
The arrival of Astra may pivot the market's central query from "how much infrastructure is required to serve known AI demand" to "as model intelligence rises, how many new workloads will become economically viable?"
These are fundamentally opposite questions. The former is a demand-side challenge, while the latter is a supply-side constraint.
The elasticity effect: more intelligence per dollar
Analysts outline a core economic logic in the report: better models create an elasticity effect. As the intelligence and utility gained per dollar spent on AI increases, the volume, duration, and complexity of inference workloads all expand accordingly.
This mirrors the classic Jevons paradox: improved efficiency does not curb consumption but instead amplifies overall usage. Higher intelligence output per unit cost makes previously unviable applications feasible. New workloads pour in, and demand for computing power and infrastructure actually grows larger.
Total addressable market expands significantly
Analysts believe that if Astra's capabilities translate into commercially useful applications, the accessible AI revenue pool will expand well beyond today's dominant use cases of chat and coding.
Morgan Stanley estimates the global knowledge-work TAM at roughly $22.5 trillion, based on approximately 900 million knowledge workers with an average annual salary of about $25,200. Consumer spending TAM stands at approximately $30 trillion, covering retail plus travel (around $16 trillion), autonomous driving and mobility (around $4 trillion), food delivery (around $4 trillion), and advertising (around $3 trillion).
Physical constraints on the supply side are the true bottleneck
The analysts conclude that Astra pulls the bottleneck narrative back from "demand formation" to "physical supply constraints," specifically whether this intelligence can be delivered at scale.
What exactly does this mean? Computing power. Electricity. Materials. Labor. And so on.
Where physical bottlenecks are located: ABF and HBM4E
Supply-side tightness has two concrete focal points.
The first is ABF substrates. Morgan Stanley projects that ABF substrates will see supply shortages starting in 2027, with the gap continuing to widen through 2030. The key constraint is that new capacity takes at least two years to come online.
The second is the back-end-of-line (BEOL) complexity of HBM4E. The report highlights that HBM4E's BEOL processes represent a major paradigm shift in semiconductor manufacturing, as HBM evolves from dedicated 3D memory stacking into highly integrated, customized chiplet logic systems.
The specific chain of impact includes:
Increased interconnect layers in HBM4E, such as SK Hynix's introduction of dummy bumps, forces substantial DRAM capital expenditure toward BEOL capacity expansion. Front-end-of-line (FEOL) process migration capex and DRAM GB shipment acceleration are not expected to materialize until the second half of 2027. DRAM manufacturers are prioritizing equipment capacity, which may delay NAND capacity expansion.
Both bottlenecks share a common trait: they are verifiable, traceable physical constraints, not market sentiment.
Electricity: the most real physical bottleneck
With rising regulatory pressure on data centers, power supply emerges as another critical physical constraint. Analysts anticipate that total computing capacity for hyperscale cloud providers will grow from approximately 35 GW in 2025 to around 145 GW by 2028, a roughly fourfold increase.
Analysts point out that the United States faces a 38 GW electricity shortfall, and data centers will increasingly adopt behind-the-meter self-generation solutions.
The firm calculates that behind-the-meter power solutions add approximately $3 billion in capital expenditure per gigawatt. Taking Nvidia's Rubin Ultra generation as an example, the all-in cost including behind-the-meter power is roughly $50 billion per gigawatt.
What the market is underestimating
Analysts believe the market is currently undervaluing three key factors.
First, the global tech beneficiaries of GPT-6 Astra are not yet fully priced in.
Second, supply tightness may persist longer than expected. The constraints on ABF and HBM4E BEOL are not short-term issues that can be quickly resolved.
Third, some stocks that can rise independently of AI are quietly strengthening. Analog semiconductor names such as STMicroelectronics, NXP, and Renesas, after more than three years of an L-shaped bottom, are in the early stages of a cyclical recovery, marked by inventory destocking, stabilizing pricing, and improving industrial orders.
The conclusion is not simply "buy more AI"
Morgan Stanley lays out the following investment priorities: AI computing power (highest priority), followed by networking, then memory (selective), plus analog semiconductors (early-cycle hedge).
AI computing: GPUs (Nvidia), ASICs (MediaTek, GUC), ABF substrates (Unimicron, Ibiden), MLCCs (Murata, Samsung Electro-Mechanics), back-end testing and packaging (Advantest, Tokyo Electron, Walton Advanced Engineering, ASE, KYEC), and power supplies (Delta Electronics).
Networking: GLW, LITE, COHR, KEYS, Furukawa Electric, and Fujikura.
Memory: prioritizing structural share growth and localization (CXMT), with SK Hynix, Samsung, and Kioxia having tactical upside under persistently tight supply.
Non-AI: analog chips (STMicroelectronics, NXP, Renesas).
The firm states: "We prefer positioning at the intersection of a broader inference cycle, limited physical capacity, rising content intensity, and increasing manufacturing complexity."
Where caution remains necessary
The Morgan Stanley report is not one-sidedly optimistic and explicitly highlights three areas of concern.
First, AI expectations are already elevated. Market tolerance for AI companies has shifted from "good results" to "flawless results." Even if Astra delivers tangible progress, unless earnings significantly exceed expectations, stock price reactions may remain muted.
Second, capital expenditure growth is expected to slow by 2028. Stock pricing reflects the direction of growth changes, and deceleration itself may continue to pressure valuations even if absolute numbers still rise.
Third, macroeconomic headwinds persist. The report cites uncertainties including oil prices, inflation, the Federal Reserve's rate path, and the 2028 U.S. election, with particular attention to potential political resistance to data center expansion should Democrats win executive power.
Comments