The 'Network Effect' Persists Even in the Age of Electronic Banking: Why Databricks Bought Mosaic and The Rapidly Decreasing Costs of AI

The 'Network Effect' Persists Even in the Age of Electronic Banking: Why Databricks Bought Mosaic and The Rapidly Decreasing Costs of AI

In the world of banking, the concept of the network effect has long been recognized as a significant factor in determining a financial institution's success. The network effect refers to the phenomenon in which large branch networks capture a disproportionate share of market deposits. This means that as a financial institution adds more branches to its network, the average deposit size per branch increases. Each new branch provides a lift to all preexisting branches, creating a mutually beneficial relationship.

Historically, financial institutions have pursued concentrated deployment strategies to leverage the network effect. For example, having 10 branches in one market rather than five branches in each of two markets. This broader representation allows institutions to maximize the benefits of the network effect. Studies conducted in 2004 and 2010 found significant positive correlations between network size and average branch size. The more recent study revealed a positive relationship between network size and average branch size in 44 of the top 50 metros, with an extremely strong correlation in 25 of those markets.

However, there are exceptions to this trend. Some financial institutions in the middle tier, known as community banks, have maintained lending levels and deposit funding needs even as larger banks focused more on problem loan resolution than asset growth. These community banks have managed to balance their deposits-per-branch ratio and maintain their rankings in the industry.

Additionally, during the financial crisis, large banks suffered reputational damage, which led some consumers to migrate to smaller banks. This migration rewarded fiscal responsibility over location convenience, resulting in smaller banks experiencing an increase in deposits per branch. By lowering the denominator in the deposits-per-branch calculation, these banks improved their ranking on that measure.

For smaller banks operating in a single market, it is essential to understand that the network effect holds not only at the metropolitan area level but also at the corridor level. Instead of scattering branches across an entire metro area, smaller banks should strive to build a cohesive network in one or two corridors while ceding others entirely. This strategic approach will yield increased per-branch deposits and maximize the benefits of the network effect.

While the network effect holds true in markets across the country, the strength of the effect varies. Markets with strong legacies of thrifts and mutual savings banks show lower correlations as these institutions tend to operate with a rate-based, branch-light model.

In a different realm, the field of artificial intelligence (AI) is experiencing its own set of changes. The decreasing costs of AI have been driven by two primary factors: algorithmic improvements and decreasing GPU costs. Companies like MosaicML have made significant algorithmic advancements that have made it more cost-effective for businesses to train and fine-tune their own AI models. This aligns with the vision of companies like Databricks, which aim to help businesses rapidly adopt machine learning to outpace the competition.

The cost of training AI models has decreased by 10 times in less than a year. For example, it now costs $50,000 to train stable diffusion and $200,000 to train a high-quality LLM. MosaicML has shown through several papers that they can train stable diffusion for as low as $50,000. This decrease in training costs can be attributed to the improvement in algorithms and a three-fold decrease in GPU costs over the past three years.

Furthermore, the cost of inference, which refers to the process of using a trained AI model to make predictions or decisions, has also significantly reduced. In just 16 months, inference costs dropped by 10 times, from $0.006 per 1,000 tokens to $0.0005 per 1,000 tokens. This decrease in costs has important implications for companies that heavily rely on AI technology. While most software companies spend around 50% of their COGS (cost of goods sold) on cloud costs, the decreasing costs of AI models mean that companies are now spending a similar amount on AI features as they are on regular cloud costs.

The decreasing costs of AI models and the increasing efficiency of algorithms will likely lead to more competition in the model layer of AI. This will put pricing pressure on closed-source model providers and may encourage more companies to utilize open-source models. The landscape of AI is rapidly changing, and businesses must adapt to these changes to stay competitive.

In conclusion, both the network effect in the banking industry and the decreasing costs of AI models are significant factors that businesses should consider. Financial institutions, particularly smaller banks, can benefit from strategically building cohesive branch networks to maximize the network effect and increase per-branch deposits. Meanwhile, companies exploring AI technology can take advantage of the decreasing costs of training and inference to adopt machine learning and stay ahead of the competition.

Three actionable advice:

1. For smaller banks, focus on building a cohesive branch network in one or two corridors rather than scattering branches across an entire metro area to maximize the network effect.
2. Businesses looking to adopt AI technology should explore the decreasing costs of training and inference, leverage open-source models, and stay updated on algorithmic advancements to stay competitive in the rapidly changing AI landscape.
3. Keep an eye on the strength of the network effect in different markets, as some markets may show lower correlations due to specific industry sectors or operating models. Adapt your strategies accordingly to optimize the benefits of the network effect.