Cloud revenue and earnings for the three major cloud service providers (CSPs)—Microsoft Azure, Amazon Web Services (AWS), and Google Cloud Platform (GCP)—surpassed high expectations in the first quarter of 2024, driven by strong demand for generative AI (GenAI). Collectively, they generated $53.7 billion in revenue, with a quarterly sequential increase of $2.6 billion. Azure led the pack with 31.1% year-over-year revenue growth. GCP also performed exceptionally well, with revenue increasing 28.4% year-over-year. According to Gartner, overall cloud spending is expected to account for 58% of IT spend by 2027, growing at 18% CAGR.

As cloud spending grows, BCG is keeping an eye on recent pricing changes using our Nimbus Pricing Index (NPI). The NPI, which reflects average pricing in US dollars converted at constant currency across the three main CSPs, saw a small decrease of 0.4% since the last installment of our Cloud Cover report in November of 2023. The decrease would have been greater, but GCP actually increased prices by 0.9% on average, driven by new high-cost locations. AWS was able to reduce prices, in part, by omitting some compute services in the high-cost city of Cape Town.

Azure was the only CSP that changed rates in existing locations in the first half of 2024, driven by adjustments in compute rates. (See Exhibit 1.) Storage and network rates remained consistent. Globally, Azure’s pricing adjustments resulted in mostly small changes in each region. The most dramatic swings occurred in Osaka (+$128), Abu Dhabi (+$31), Cape Town (+$30), Mumbai (–$50), and São Paulo (–$105). Even with the substantial Osaka rate increase, Azure still offers the lowest rates in that city. Moreover, its competitive pricing in Tokyo means Azure is the most affordable CSP in Japan.

Providers Add Locations

While AWS, Azure, and GCP all added locations in the first half of 2024, there were notable differences in the expansion strategies. (See Exhibit 2.) Azure added two locations (Tel Aviv, Warsaw); AWS also added two (Tel Aviv, Calgary); while GCP added four locations (Berlin, Dammam in Saudi Arabia, Doha, and South Carolina).

Serverless Pricing

The downside to serverless is that, in contrast to the traditional virtual machines pricing model, pricing for a given set of parameters is often inconsistent across regions. (See Appendix 2 for pricing differences and Appendix 3 for city-level pricing.) Azure is the only player in the market which shows consistent per-unit pricing across all regions, something that engineering teams appreciate since they can run their applications on Azure functions without worrying too much about the most cost-effective region. (See Exhibit 4.)

Fine-Tuning Your GenAI Model

In our last Cloud Cover report, we outlined a framework to estimate the cost of “building versus buying” GenAI models. (See Appendix 4.) The “buy” models we looked at are ready-made, without the option to tailor them. In this report, we dive deeper into the “buy” question and explore the cost factors to consider when deciding whether or not to fine-tune a GenAI model.

Learn More: Cloud Cover Series

Learn More: Cloud Cover Series

Cloud Computing

Our cloud computing consultants understand the importance of linking a cloud migration to a business strategy.

Article

Cloud Cover: Introducing a Standardized Pricing Index

This inaugural snapshot of the cloud industry introduces a rigorously standardized pricing index and looks at four trends shaping the industry.

Fine-tuning usually involves pre-trained models and a custom dataset. Pre-trained models are broad-purpose models that have been initially trained on large, unlabeled datasets. These models are further trained, or “fine-tuned,” on a smaller, more specific dataset relevant to the task at hand. Through fine-tuning, models can learn the nuances of a particular field or dataset, resulting in better predictions, translations, or other outputs more aligned with the company's requirements.

Customized models come with some drawbacks. They are resource intensive and consume more computational power than generic models. The data used for fine-tuning can at times introduce or perpetuate biases, requiring careful selection and monitoring to ensure fairness and relevance.

Costs of Fine-Tuning

The cost of fine-tuning GenAI models depends primarily on the size of the training dataset in tokens and the number of iterations (or epochs) necessary for training. Typically, the more epochs, and thus more exposure to the data, the better the model becomes. For simpler and smaller use-cases, a general recommendation is to train using 5 to 10 epochs.

As model complexity escalates, so does the cost of fine-tuning—from a mere 0.032 million USD for a simple model, to 0.48 million USD for an average model, to 0.64 million USD for a complex model. (See Appendix 5 for model complexity scenarios.) Once built, there’s an exponential increase in the resources needed to run the models across all usage tiers. Interestingly, while usage costs escalate from pre-trained to fine-tuned, the scale of the increase for complex models is less drastic (4.5x) compared to simple models (8x) and average models (12x). (See Exhibits 6-8 slideshow).

Also worth noting is that while the simple model offers better efficiency at lower usage levels, at higher volume the complex model might be more economical. Complex models can become less expensive to run than average models at high usage rates. (See Exhibit 9.) These insights demonstrate that aligning model complexity with expected usage is extremely important to optimize performance and costs.

Please keep an eye out for the next issue of Cloud Cover, where we’ll continue to develop the NPI and address key questions on cloud services and pricing—including insights on specialist offerings, other CSPs, and price evolution over time. We will also be conducting a more detailed analysis comparing the economics of deploying VM machines versus serverless functions.

Appendices 1-6