The real question has never been whether AI demand will shrink, but rather where value will ultimately settle in the supply chain.

Author and source: 0x9999in1, ME News

TL;DR

• Morgan Stanley's report, "Open Weight Models and Three Future Scenarios," identifies the core judgment: the rise of open-weight models such as Kimi K3, Qwen, and Llama acts as a catalyst for accelerating AI adoption, not as a threat to market leaders.
• The argument is the Jevons paradox—after single-instance reasoning becomes cheaper, companies will apply AI to more tasks, ultimately increasing the total demand for tokens, computing power, electricity, and infrastructure.
• Morgan Stanley explicitly warns: Open weights do not mean free. Costs for GPUs, cloud services, operations, and security still exist—real economic viability depends on the use case.
• The cited MIT study shows that switching from closed-source to open models can reduce average costs by approximately 70%, corresponding to annual savings of about $25 billion for enterprises; however, the study was completed in December 2025 and did not fully account for deployment costs.
• Carnegie Mellon's research is more measured: the payback period is highly dependent on scale—small-scale deployments with fewer than 30 billion parameters can recoup costs quickly, but differences become significant as scale increases.
• Three possible outcomes: closed-source oligopoly, hybrid coexistence, or open-weight dominance. The real-world anchor is that 63% of companies are already using two types of models simultaneously.
• The only two constant beneficiaries around the clock are NVIDIA and the on-site power supply segment. The reason isn’t about taking sides—it’s that, regardless of whether computing power is centralized in massive data centers or distributed to the edge, the fundamental logic underlying GPUs and electricity remains unchanged.
• The real question has never been whether AI demand will shrink, but rather where value will ultimately settle in the supply chain.

I. What Did the Market Misread?

In mid-July 2026, Kimi K3 was released. The Philadelphia Semiconductor Index dropped in response, and NVIDIA, Micron, SK Hynix, and even EDA companies like Cadence were all affected.

Familiar? Very familiar.

More than a year ago, during the DeepSeek surge, the market’s reflex was the same: the stronger the open-source model, the less valuable compute becomes. The logic seemed sound—if an open-weight model can match the top tier, then the premium for closed-source models disappears; if efficiency improves, the need to pile up GPUs diminishes. As a result, the entire chain—selling chips, equipment, and electricity—was cut down.

In its report on August 4, Morgan Stanley directly referred to this feedback loop as a misinterpretation.

Where is the misunderstanding? It lies in the market equating "lower unit cost" directly with "lower total expenditure."

Between them lies a major variable: usage.

Morgan Stanley invoked Jevons Paradox. In the 19th century, economist Jevons observed an counterintuitive phenomenon: after steam engines became more efficient, allowing more work to be done with the same amount of coal, Britain’s coal consumption did not decrease—it surged. Because coal became cheaper, it became affordable; and because it became affordable, it was used everywhere.

Put it on AI: The cost of a single inference drops from one cent to one-tenth of a dime. Companies won’t pocket the savings—they’ll embed AI into processes that were previously too expensive to justify: full review of customer support tickets, continuous scanning of codebases, line-by-line contract comparison, and automatic email categorization. These tasks weren’t done before not because they weren’t wanted, but because they were too costly to make sense of.

Once the price drops below that line, demand doesn't increase linearly—it collapses inward in a surge.

II. What Exactly Is Opened by "Open Weights"?

We must first clarify the concepts, as half of the market’s confusion stems from terminology ambiguity.

Morgan Stanley made a distinction in its report that, while not new, is necessary: open weights ≠ open source.

Strictly defined open-weight models—such as Meta’s Llama, Google’s Gemma, and Qwen, DeepSeek, and Kimi—make the trained parameters publicly available. You can download them, fine-tune them, and deploy them yourself. However, the training data may not be disclosed, and licensing agreements may impose restrictions.

True open source means laying out the entire technology stack on the table: the code, training methods, and data details. The report cites AI2's OLMo as an example. This community is much smaller than most people realize.

Therefore, a more accurate statement is: the key difference between open-weight and closed-source models lies not in parameter size or rankings, but in how they are accessed and licensed.

Open weights: Typically available for free, supporting fine-tuning with flexible deployment options—on-premises, private cloud, public cloud, or API, your choice.

Closed-source: License + token-based billing; all users receive the same set of weights, which you cannot modify or view.

Morgan Stanley also made a practical point: today’s most advanced models are still mostly closed-source. The reason is simple—the training costs are astronomically high, and no one will give away the previous generation until the next one is ready.

This sentence is worth reading twice. It means that the pursuit of open weights is essentially a delayed race. The gap will narrow, but structural time lags are not easily eliminated.

Three: The Most Expensive Free Things—Costs Not Listed on the Price Tag

The most valuable section of the report is precisely the one that tempers expectations around open weights.

Morgan Stanley cited MIT research: switching from closed-source to open models can reduce average costs by approximately 70%, equivalent to annual enterprise savings of about $25 billion.

The numbers look great. But Morgan Stanley quickly added a caveat—the study was completed in December 2025 and did not fully account for deployment costs.

What is deployment cost? You must buy or rent GPUs yourself, have someone maintain the cluster, have someone fine-tune the model, and have someone ensure security and compliance. Open weights give you the blueprints and parts for a car—not a car you can drive away right away.

Carnegie Mellon University's research quantified this even more sharply: the payback period is highly dependent on deployment scale. For deployments under 30 billion parameters, payback can be rapid; once scale increases, the financial picture changes completely.

This is a rarely discussed but critically important reality: the economics of open weights is not a smooth curve, but a piecewise function.

Small-scale, high-frequency, clearly defined tasks—open weights are incredibly valuable.

Complex reasoning, long-range agents, and extreme reliability—closed-source money isn't spent for nothing.

Look further ahead at ultra-large MoE architectures like Kimi K3: even when only a small number of experts are activated, the massive weights still need to be distributed across clusters with high-capacity memory and high-speed interconnects.所谓"开放"，开放的是权重的所有权，不是对硬件门槛的豁免。

In other words, open weights have shifted the compute demand from a few closed-source labs to cloud providers, sovereign clouds, and enterprise-owned data centers. The total demand hasn’t decreased—it’s just changed hands.

This is the unsettled account from the market's decline in July.

Four: Three Possible Outcomes, Where Does Value Settle?

Morgan Stanley did not bet on a single outcome but presented three scenarios. This approach itself is a judgment: no one knows the final outcome, but the受益 structure of each possible outcome can be mapped out in advance.

Assuming the performance of frontier models is difficult to replicate, workloads will continue to maintain high compute density and consolidate further into hyperscale clouds. Token prices will still decline, but more slowly—due to the oligopolistic structure itself maintaining pricing discipline.

Beneficiaries: Cloud service providers, security software companies, optical networks, semiconductors, and power supply chains.

The report highlights a specific variable: Google. If Gemini 4 achieves a breakthrough and regains a leading position, Google’s return on investment (ROI) for running model APIs on its own infrastructure could reach approximately 45%. If it fails to lead at the forefront, Google would resemble more of an infrastructure provider, with a corresponding ROI of around 30%.

A fifteen-percentage-point gap rests on the shoulders of the first-generation model. That’s the odds in the frontier race.

No single architecture dominates all use cases. Proprietary frontier models lead in complex reasoning and AI agent tasks, while open-weight models serve high-frequency, cost-sensitive, and vertically specialized needs.

Workloads span public cloud, private cloud, on-premises, and edge environments—making "middle layer" software such as model orchestration, routing, evaluation, and governance critical to value migration upward. Demand for security software grows as workloads become more distributed.

This scenario is strongly supported by a real-world fact: currently, 63% of companies are using both open-weight and proprietary models.

This number speaks volumes. Businesses aren't choosing based on "which ideology to believe in," but rather "which option is cheaper and more reliable for each specific task." This is a procurement decision, not an ideological battle.

Open-weight performance gradually converges to state-of-the-art levels, making the intelligence of base models widely accessible. Model API prices have dropped significantly, triggering a wave of enterprise adoption.

AI infrastructure is moving toward decentralization: demand for private data centers, sovereign clouds, on-premises deployments, and edge devices has significantly increased. The focus of innovation is shifting from pre-training to fine-tuning, inference optimization, agents, and practical application deployment.

In this scenario, security software is one of the biggest winners. This is also easy to understand—in a world where weight can flow freely, the attack surface expands geometrically.

The divergence at the individual stock level is intriguing: AVGO could be the largest single beneficiary under a closed-source scenario; DELL and HPE, as local infrastructure providers, and Apple, as an edge device manufacturer, are more likely to unlock value under an open-source scenario. Cloud providers, however, depend on where the workload ultimately flows—under an open-source scenario, Microsoft’s advantage in hybrid and on-premises deployment becomes clearer, while Amazon and Google are reclassified as "other beneficiaries."

Five: All-Weather Beneficiaries: Why NVIDIA and Electricity

After analyzing the three scenarios, Morgan Stanley identified two threads that run through all of them.

The first one is NVIDIA.

The logic is not complicated, but solid: whether computing power is concentrated in hyperscale data centers or distributed to enterprise server rooms and edge locations, the underlying rationale for GPU demand remains unchanged.

Concentration means a few buyers purchasing large amounts; dispersion means tens of thousands of buyers purchasing small amounts. The total volume does not decrease—and may even increase due to broader adoption.

Notably, Morgan Stanley recently increased its cost estimate for AI clusters—the pricing for NVIDIA’s Vera Rubin system has been set at approximately $49 billion per gigawatt, nearly 20% higher than previous forecasts.

This number is somewhat nuanced here. On one hand, models are becoming cheaper; on the other, capital expenditure per unit of computing power is rising. These two trends are not contradictory: marginal costs for inference are falling, while fixed investment in training and deployment is increasing. The former drives demand expansion, while the latter determines who is qualified to participate in supply.

The second is on-site power.

The reason is that the electricity demand curve encompasses both centralized and distributed configurations. Regardless of how the load is distributed, AI expansion requires a simultaneous increase in power consumption and supply capacity.

Electricity is the most unforgiving constraint in the entire AI narrative. Models can be open-sourced, weights can be downloaded, and chips can be substituted—but electricity cannot be open-sourced. A gigawatt is a gigawatt, and no one can create more.

This might be the most certain line among the three scenarios—so certain it’s almost boring.

Six: The true value of this report lies not in its conclusions, but in its framework.

Share your own thoughts.

If one only reads the headline of Morgan Stanley’s report—“Open weights accelerate adoption rather than undermine leaders”—it’s easy to dismiss it as mere market reassurance. But the truly significant part is that it reframes the question.

The original market question was: Will AI demand shrink due to increased efficiency?

Morgan Stanley's question is: AI demand is certainly not declining, but the issue is where the value will ultimately settle—in compute, cloud, software orchestration and governance, or in local and edge distributed infrastructure?

The difference between these two questions lies in bear market thinking versus structural thinking.

I believe Morgan Stanley's direction is correct, but there are two areas worth remaining cautious about.

First, the Jevons paradox does not automatically take effect.

Jevons' Paradox holds under one condition: demand must have sufficient elasticity. Coal usage increased despite greater efficiency because, in 19th-century Britain, many industries had yet to be mechanized. AI presents a similar but not identical situation—there are indeed many scenarios where businesses can integrate AI, but whether those integrations yield measurable returns is another question entirely.

If a large number of companies use AI only in ways that appear busy but cannot clearly demonstrate returns, demand growth will hit a ceiling that may suddenly appear in a given quarter. Jevons' Paradox describes energy conversion efficiency, not guaranteed value creation.

Second, the conclusion that "NVIDIA benefits around the clock" is just a bit too comfortable to be correct.

It’s logically hard to fault—GPU demand doesn’t change with architectural shifts. But between sound judgment and good investment judgment lies price. Whether the market has already fully priced in this "all-weather" expectation is not addressed by the report, nor should it be.

A logically airtight bullish argument is often fertile ground for crowded trades. This isn’t contrarianism—it’s a reminder: certainty of conclusion doesn’t equal attractive odds.

Seven, for businesses, this is actually simpler.

Setting aside the valuation games of the secondary market, from the perspective of corporate procurement, the insights from this report are much more straightforward.

63% of companies use both—this number says it all: don’t pick a side, pick the use case.

Tasks that are high-frequency, predictable, data-sensitive, and have well-defined boundaries—self-hosting with open-weight models can save real money, especially at scales below 30 billion parameters, with payback periods so short they’re highly attractive.

Complex reasoning, long-range agents, tasks requiring continuous alignment with cutting-edge advancements—paying for closed-source APIs for reliability and iteration speed is nothing to be ashamed of.

What truly requires investment is the middle layer: orchestration, routing, evaluation, and governance. How do you determine which path a request should take? How do you ensure output quality doesn’t degrade after switching models? How do you enforce data boundaries in a world where weights can be freely downloaded?

Morgan Stanley has identified this layer as a primary beneficiary under a hybrid scenario; I find this assessment more practical than any individual stock recommendation, because this layer is a capability that companies must build themselves—it cannot be purchased.

By the way, security. Open weights decentralize model deployment—and also decentralize the attack surface. Previously, you only needed to manage a few API keys; now, you must secure a collection of locally running weight files, fine-tuning data, and inference endpoints. The fact that security software is called out in all three scenarios is no coincidence.

Eight, Finalization

On the day Kimi K3 was released, semiconductor stocks plummeted collectively. Looking back today, it felt more like market muscle memory than a thoughtful reaction.

Open weights aren't the ones knocking over the table—they're the ones extending it.

After the table is extended, will the person who originally sat in the main seat be pushed out? Possibly—but not because of the open weights themselves, but because they failed to maintain their position at the frontier. These are two separate issues.

The most honest part of Morgan Stanley's report is that it doesn't tell you what the endgame is—only who will remain in the same position in every possible endgame.

NVIDIA and electricity.

The former is consensus; the latter is physical.

As for the companies in the middle that follow price swings, switch sides, and reprice themselves—their fates depend heavily on the circumstances. This may sound like common sense, but in a market where everyone seeks definitive answers, admitting “it depends on the circumstances” is the most valuable form of honesty.

Models will become cheaper generation after generation—that’s the trend, and it’s unstoppable.

Computing power will gradually spread out layer by layer—that’s a trend that cannot be stopped.

But cheap and diversified never add up to decline.

Jevons understood this back in 1865. One hundred and sixty years later, humanity is still paying the price for the same intuitive misconception.

That's interesting.

Source citation

1. Morgan Stanley, *Open-Weight Models and Three Future Scenarios* (Research Report), August 2026
2. Gelonghui / Sina Finance: "Morgan Stanley: Open-weight models reduce AI costs; Jevons Paradox may lead to sustained growth in computing demand," August 4, 2026
3. BigGo Finance: *Open-Weight Models to Accelerate AI Adoption: Morgan Stanley Outlines Three Scenarios, Names Nvidia and Power Chain as Beneficiaries*, August 4, 2026
4. Huo Xing Cai Jing: Morgan Stanley: The Rise of Open-Weight AI Models Will Accelerate AI Adoption by Reducing Costs, August 4, 2026
5. Blockspace Media: *Morgan Stanley says open-weight AI could expand compute and power demand*, August 2026
6. MIT research on the price difference between open-source and closed-source models (completed in December 2025, cited in a Morgan Stanley report)
7. Carnegie Mellon University's research on the relationship between the payback period for open model deployment and scale (cited in the Morgan Stanley report)
8. Morgan Stanley Insights: *Opportunities in AI Infrastructure* (Including Upward Revision of Vera Rubin System Cost per Gigawatt Estimate)