Will frontier labs survive the generative AI arms race?

This article is part of our series that explores the business of artificial intelligence

Advances in generative models have been nothing short of stunning in the past year. Unfortunately, these advances come at a huge cost. Frontier AI labs and big tech companies are pouring billions of dollars into building GPU clusters and training new models. 

This pace can only be maintained while there is excitement and appetite for investing in genAI startups. But once the dust settles and the funding sources dry up, companies without a profitable business model will need an exit strategy. And quite a few might die in the process.

Funding still hot

Less than two years since the release of ChatGPT, there is still a lot of interest in pouring cash into AI startups. In June, Paris-based Mistral raised €600 million at a €5.8-billion valuation. Also in June, Canadian AI startup Cohere raised $450 million and is projected to reach a $5-billion valuation. Stability AI also reported raised millions of dollars this month.

Even robotics companies are cashing in on the excitement around generative AI and raising hundreds of millions of dollars as they bring together the power of foundation models and humanoid robots.

What is the money spent on?

A considerable part of the money raised is being spent on either building GPU clusters or paying for the cloud computing costs of training frontier models. Models such as GPT-4, Claude 3, and Llama 3 require tens of thousands of GPUs to be trained. This is why Nvidia has managed to add more than $2 trillion to its market cap in less than one year and become the world’s most valuable company for a short while.

Although not as expensive as training, running the models and inference costs also account for a considerable portion of the expenses of AI companies. As they are all commercial companies that have customer-facing platforms, these AI labs are processing hundreds of millions of queries per day. And while customers pay to use these services, a lot of them are offered for free.

For example, OpenAI’s recent deal with Apple will add ChatGPT capabilities to hundreds of millions of iPhones and MacBooks. However, the feature will be offered for free and OpenAI—not Apple—will be bearing the costs. 

As competition shifts toward having access to better data, companies are spending more on curating better data, striking licensing deals, and hiring experts to create training data for their models. In 2024 alone, OpenAI entered licensing deals with several media companies, worth hundreds of millions of dollars, to train its models on their data.

Is the business model sustainable?

Nothing lasts forever, and the current craze around AI will eventually die down and adjust itself. Therefore, AI startups can’t burn investor cash forever and need a path to profitability. The main question is, will all this spending eventually enable the AI startups to break even and become profitable? 

All AI companies are making money through paid access to their models, either through a flat-fee monthly subscription or token-based pricing on API platforms offered to developers and enterprises. There are a few reasons to believe that the current business model doesn’t work. 

First, the switching costs are very low. Prompts that you run on OpenAI models are likely to work on Anthropic and Cohere with little adjustments. For most daily uses, any of the models work decently. And for specialized applications, you don’t need the most advanced model. You need one that has a long enough context window to support in-context learning with your application-specific examples or can be fine-tuned on your proprietary data. The commoditization of generative models will drive competition to the level of price, preventing AI companies from raising their profit margins.

And open models are another factor that are disrupting the business strategy of companies that serve private models. The most money is to be made in the enterprise sector. But as the market matures, more companies will gravitate toward running open models that are hosted on their own servers and are considerably cheaper to run.

At the same time, the model update cycles are shrinking as AI labs compete to release better models every few months. This makes it difficult for AI companies to recoup the costs of training the models. With small profit margins and tens of millions of dollars going into training each model, the companies will need mass scale to make sure the payback window matches the model release cycle. 

What’s the end game?

For the moment, OpenAI is the most successful of the pack, with its annualized revenue jumping from $1.6 billion in December 2023 to $3.4 billion in June 2024. But how will that stand against the costs of training and running its models, legal fees, licensing fees? It is not clear. And OpenAI has the first-mover advantage and has branded itself as the go-to model for newcomers and developers who want to get started on their generative AI journey. OpenAI also enjoys massive distribution through its partnership with Microsoft. 

It is not clear how the other companies will lay out their business strategy as they move forward. Anthropic is testing a dual partnership with Google and Amazon. Mistral is testing a hybrid approach, releasing both open models and private models that will be distributed through cloud partners. 
Some of these strategies might succeed. Some might evolve into full acquisitions or acqui-hires such as Inflection AI. Some might pivot toward techniques that require less capital. And some will fail. But what’s for sure is that not everyone will make it to the finish line.