⚡️ Should you build your own model, partner to enhance an existing LLM, or procure it off-the-shelf? Here is what it may take.

🗽 Boston Consulting Group (BCG) projects an incredible growth of 66% CAGR for the Generative AI, forecasting it as a 120B+ market in just four short years.

With potential for automating 20-30% of tasks across every job category (Accenture), AI is what everyone is talking about.

Here are just a few examples of AI-driven wins:

88% of developers reported higher productivity with GitHub Co-Pilot.

Insurtech COVU managed to cut customer service costs by 30%.

AI-generated fashion images increased retailer conversion rates by 1.5X.

So how do you tap into this potential? There are three main routes:

🛠️ Building a new foundation model in-house from scratch

Estimated cost: $50 - $90M+

This option is costly due to hardware required (GPUs/TPUs) ~$30M.

In addition to multiple training runs at $10M+ (GPT-3 training run estimated at ~$12M).

Not to mention the need for rare, expensive and highly specialized talent for R&D.

🤝 Partnering with an LLM provider

Use case: Significantly enhance an existing model by feeding in complex, proprietary company data.

Estimated cost: $1 - $10M

Costs here primarily stem from less intensive training runs and partnership costs.

🛍️ Using an off-the-shelf foundation model and fine-tuning it for related tasks

Use case: fine-tune ChatGPT for legal memo writing.

Estimated cost: $10K - $100K

Costs in this option are from data labeling, etc..

Unless you have abundant resources, time, data, and talent, the choice likely boils down to partnering or using an off-the-shelf LLM.

And to deploy AI successfully at scale and customize the model with proprietary data, companies will likely need a robust partner ecosystem.

The typical Generative AI journey starts by brainstorming customer outcomes and use cases, followed by considering the required technology & data, and ends by finding the talent and partners to make it all happen.

BCG research