Morgan Stanley Analysis: Why Are Amazon and Google Still Optimistic Amid AI Price Wars?

TL;DR
· Morgan Stanley estimates that when 1GW GB300 computing power is used for model APIs, the model layer ROIC is approximately 20%-60%.
· If GPUs are directly rented out by cloud platforms, with a utilization rate of 75%, the IaaS ROIC is estimated to be around 23%-39%.
· The report continues to be optimistic about Amazon and Google; Meta remains a Top Pick, but its AI returns come more from advertising, proprietary applications, and an open model ecosystem.

The Brian Nowak team at Morgan Stanley presents a counterintuitive judgment in their latest report: open-weight models are driving down AI token prices, but platforms that possess computing power, scheduling capabilities, and enterprise access may still achieve substantial returns from AI capital expenditures.

The report calculates two sets of accounts.

At the model API layer, Morgan Stanley bases its analysis on a 1GW GB300 data center, assuming a price of $1 to $2.5 per million tokens, a single GPU throughput of 2000 to 3500 tokens per second, and that 50% to 80% of computing power is used for inference. Under different scenario combinations, the model layer ROIC is estimated to be around 20% to 60%.

At the cloud platform layer, the report assumes that a 1GW data center is configured with approximately 410,000 GB300 GPUs, with a utilization rate of 75% and rental prices ranging from $7 to $10 per hour. Based on this, the IaaS ROIC for renting GPU computing power is estimated to be around 23% to 39%.

These two sets of figures jointly explain Morgan Stanley's core judgment: the decline in AI model prices primarily impacts the unit economics of model providers; however, for cloud platforms like Amazon and Google, enterprises still require GPUs, storage, databases, security, and model deployment tools to adopt AI. While models may become cheaper, the infrastructure that supports them does not lose its value.

Open weights drive down prices and encourage more AI workloads to materialize

The direct impact of open-weight models is to lower the barriers for enterprises to use and deploy AI.

Unlike closed models, open-weight models allow users to download model weights, fine-tune or retrain using private data, and choose to deploy locally, in the cloud, or via API. Enterprises do not have to rely entirely on a single model provider and can more flexibly control data, costs, and deployment methods.

Differences between open-weight and closed-weight models include whether weights can be downloaded, support for private data fine-tuning, and whether they can run on proprietary hardware.

The downside is that it is becoming increasingly difficult for model APIs to maintain high prices. As low-cost models approaching frontier capabilities proliferate, model labs must compete for customers through lower prices, faster responses, and more complete toolchains. The report mentions that low-cost near-frontier models like Meta's latest Muse Spark 1.2 are further intensifying price competition at the model layer.

This is also why the market is concerned about the returns on AI capital expenditures. Tech companies are investing billions of dollars in purchasing GPUs, building data centers, and securing power; if inference prices continue to decline, the ultimate returns will increasingly depend on three variables: GPU utilization, per-card token throughput, and whether AI workloads can drive other cloud service revenues.

Morgan Stanley believes that while price declines will compress model layer profits, it does not necessarily mean that the returns on the entire AI infrastructure will deteriorate simultaneously. Cheaper and more flexible models may also encourage more enterprises to adopt AI, thereby expanding demand for underlying computing power and cloud services.

20%-60%: Return Estimates for Model API Layer

The estimates for the model API layer revolve around three core variables: token price, GPU throughput, and the proportion of computing power used for inference.

Morgan Stanley sets a baseline price of approximately $1.75 per million tokens and conducts scenario analyses within the range of $1 to $2.5. The report references public inference benchmarks of open-weight models like DeepSeek-V4-Pro 1.6T on GB300 hardware, assuming a single GPU throughput of 2000 to 3500 tokens per second.

In the baseline scenario, the report assumes that 65% of computing power is used for inference, with a single GPU processing 2750 tokens per second and a price of $1.75 per million tokens, corresponding to approximately $40 billion in revenue/GW. The complete scenario range corresponds to revenues of approximately $23 billion to $52 billion/GW.

Based on this, the report estimates that the model layer ROIC for 1GW GB300 computing power is approximately 20% to 60%.

Revenue and cost framework for model APIs. Different token prices, throughput, and inference computing power proportions correspond to revenues of approximately $23 billion to $52 billion per GW.

It is important to note that the 20% to 60% return is not automatically achieved at a fixed price of $1.75, but rather forms a scenario range based on different combinations of prices, throughput, and computing power allocations.

Whether this set of accounts can be realized hinges on whether efficiency improvements can outpace price declines. If token prices continue to drop rapidly while GPU utilization and throughput do not increase in tandem, the returns at the model layer will narrow significantly.

23%-39%: Another Set of Accounts for Cloud Platforms Renting GPUs

For cloud platforms like Amazon and Google, a more direct source of revenue is renting out computing power rather than bearing the price risk of model APIs themselves.

Morgan Stanley assumes that a 1GW GB300 data center can be configured with approximately 410,000 GB300 GPUs. At a 75% utilization rate, if the rental price for each GPU is $7 to $10 per hour, the corresponding annual revenue is estimated to be around $18.9 billion to $27 billion/GW.

After deducting costs such as IT equipment depreciation, non-IT asset depreciation, electricity, labor, and maintenance, the report estimates that the after-tax operating profit for cloud platforms is approximately $8.9 billion to $15.3 billion per GW, corresponding to a ROIC of about 23% to 39%.

Cloud platform GB300 GPU rental model. Rental prices of $7 to $10 per hour correspond to a ROIC of approximately 23% to 39%.

This set of calculations directly supports Morgan Stanley's optimistic judgment regarding AI capital expenditures for cloud platforms.

Even if model API prices decline, enterprises still need to rent GPUs to run open-weight models. For cloud platforms, the key is not how much a single token can be sold for, but whether scarce computing power can maintain high utilization rates and rental prices.

Morgan Stanley also emphasizes that low-cost models may lead to increased usage. Even if unit prices and profit margins decline, as long as the overall scale of AI workloads expands, the absolute EBIT of cloud platforms may still grow.

Cloud platforms sell more than just GPUs and tokens

Morgan Stanley believes that cloud platforms can withstand price wars for four main reasons.

First, computing power remains scarce. Even if enterprises choose open-weight models, they still need GPUs, inference clusters, storage, networking, and security environments. Large enterprises can build their own infrastructure, but more customers will still rent computing power and host models through AWS, Google Cloud, or Azure.

Second, cloud platforms can improve system efficiency through scale. Larger memory capacities, faster interconnections, MoE architectures, and request batching can all enhance the token throughput per GPU. Large cloud platforms have a greater flow of enterprise customers, making it easier to merge and schedule requests, thereby improving GPU utilization.

Third, self-developed chips can reduce service costs. Amazon has Trainium, and Google has TPU. Self-developed chips and software optimizations give cloud platforms greater flexibility in price competition and reduce dependence on a single GPU supplier.

Fourth, AI workloads can drive additional cloud service revenues. When enterprises deploy AI, they typically also need to purchase object storage, databases, vector retrieval, RAG, security governance, identity verification, and monitoring services. These additional services can expand the cloud revenue generated by each AI workload and enhance overall profitability.

Therefore, while open-weight models may turn API token revenue into a "loss leader," they may also lead to increased GPU rentals, data storage, and database demand.

Amazon, Google, and Meta Have Different Return Logics

Based on the above judgments, Morgan Stanley maintains an "Overweight" rating for Amazon and Alphabet, with target prices of $335 and $400, respectively.

Amazon's advantages lie in its AWS enterprise customer base, Trainium self-developed chips, and complete cloud service ecosystem. Google, on the other hand, possesses TPU, Gemini models, search and YouTube application scenarios, as well as the enterprise distribution capabilities of Google Cloud. For both companies, the returns on AI capital expenditures primarily come from computing power rentals, model hosting, and cloud service bundling.

Meta's logic is different.

Morgan Stanley still lists Meta as a Top Pick, giving it an "Overweight" rating and a target price of $775, but Meta is not an enterprise cloud platform in the sense of AWS or Google Cloud. Its AI investments are primarily realized through an open model ecosystem, advertising recommendation efficiency, proprietary applications, and data center capabilities.

Meta's target price and risk-return assumptions.

The report suggests that Meta's AI investments may improve user engagement, Reels monetization, and advertising measurement capabilities, providing growth opportunities for new products. However, if data center construction is poorly executed, capital intensity continues to rise, and the new computing power fails to generate sufficient returns, it could drag down free cash flow and valuations.

The biggest risk is that prices fall faster than efficiency improvements

This report does not deny the AI price war.

As low-cost open-weight models approach frontier capabilities, enterprises will pay more attention to costs, latency, deployment flexibility, and data control. It will become increasingly difficult for model labs to achieve high profits solely through high-priced APIs, and the importance of toolchains, enterprise distribution, application scenarios, and ecosystems will continue to rise.

Cloud platforms also face pressure from heavy assets. A 1GW GB300 data center involves chips, machine rooms, electricity, networking, and depreciation; any increase in costs in any of these areas will affect the final returns. If the spread of enterprise AI applications is slower than expected and GPU utilization is insufficient, it will be even harder to convert capital expenditures into profits.

Throughput improvements will not automatically materialize either. Hardware upgrades, MoE architectures, and request batching can enhance efficiency, but they require coordination among models, chips, networks, software stacks, and real customer requests.

Therefore, Morgan Stanley's conclusion is not that "the AI price war has no impact," but rather: when token prices decline, the model layer is the first to feel the pressure; platforms that possess scarce computing power, system optimization capabilities, and enterprise cloud access are better positioned to absorb price competition through GPU rentals, scale effects, and additional cloud services.

What truly needs to be observed next is whether cloud platforms can continue to outpace the speed of AI price declines with improvements in throughput and utilization rates.