Nothing captures the excitement about AI as the $10 trillion rise in the market capitalization of the Magnificent Seven since January 2023, just after OpenAI released its generative AI model called ChatGPT. Nvidia has benefited the most; its market capitalization has risen more than ten times, or almost $3 trillion because of its near monopoly on the high-performance chips that are used to process AI models. Microsoft, Google, Amazon, and Meta are also profiting from AI, mostly through their cloud services that use those chips.

One surprising part of this story is that there are few revenues for OpenAI (less than $10 billion), the ten-year old startup that started the boom. Other startups that build these types of foundational models, such as Anthropic, xAI, and Mistral, also have few revenues, much less profits. Even members of the Magnificent Seven such as Microsoft and Google receive few revenues from licensing or selling subscriptions to their generative AI models.

Who benefits from AI?

The second most surprising part of this story is that few users seem to benefit from AI. Most of their spending is on experiments that use AI cloud services. There are lots of products, but few success stories or even product revenues to tie the success stories and the products together. Call centers were once thought to be a big application but an analysis by Forbes says the Philippines isn’t worried. Anthropic wants you to use agentic software to do various applications, “just not to apply for its jobs.” Microsoft lists many happy customers in its 2024 annual report, from aqua farmers to street vendors, but just how big are those applications? Even the 10-year old voice assistants such as Alexa, Siri, and Google Assistant have accumulated tens of billions in losses. Attempts to document and categorize other business cases demonstrate the incredible narrowness of the benefits.

Where are the successful applications?

What types of workers use AI the most? According to an analysis of queries to one popular chatbot, 37.5% of the queries were for coding and 10.3% for creative work. The first is consistent with reports of big tech companies using AI for coding, and the second with analyses of Meta’s higher profits and revenues in the fourth quarter of 2024: the increased profits were “aided by artificial-intelligence improvements to its ads business.”

Will other applications emerge quickly? The CEO of Amazon Web Services recently said: “Almost every CEO or CIO that I talk to, they’re basically saying, ‘Look, my organization did 100, 200 proof of concepts,” referring to companies’ AI experiments. Then they say, ‘How do I go find the one, two, five of those proof of concepts that are valuable, and that are delivering real ROI?” Similarly, a survey of almost 3,000 executive concluded that “proof of concept projects is stuck in pilot phase as investors get itchy feet.”

This is partly what Sequoia’s David Cahn, Goldman Sachs’ Jim Covello, and Citadel’s Ken Griffen meant when they said that the small revenues for AI was a problem last summer. They suggested that there is a big bubble in AI. David Cahn said in mid-2024 that $600 billion is needed in revenues to justify the current stock prices of AI companies, a number that has probably grown a lot over the last six months as investments have also significantly grown.

Most of the current revenues fall into two categories: 1) subscriptions to AI services such as OpenAI’s ChatGPT; and 2) cloud services that enable users to access the AI models. The first category is still worth less than $10 billion while the second is worth much more. For 2024, Microsoft reports a little over $100 billion for its intelligent cloud services while Google reports about $43 billion for its services. Neither Amazon nor Meta break out the revenues for their AI services, and those revenues are believed to be much smaller than those for Microsoft or Google.

Optimistically speaking, annualized AI revenues for subscriptions and cloud service probably don’t exceed $200 billion. Yet Jim Covello said $600 billion was required last summer — a figure that has probably increased by $100 to $200 billion since then. For those who claim the dotcom bubble was similar, Gary Smith and I found $1.5 trillion in revenues for internet subscriptions, e-commerce, and purchases of personal computers, needed to access the internet services, at the peak of the bubble in 1999.

Cracks are beginning to form

In the short run this system is being propped up by the big tech players. Microsoft has invested more than $13 billion in OpenAI over the last six years, thus giving OpenAI the funds necessary to develop ChatGPT while providing Microsoft with a basis for its cloud services. Google, Meta, Amazon, Apple, and Tesla have developed their own models, alone or through similar investments in startups. Some of these companies deliver the models to users through their cloud services while others deliver them through their devices (Apple).

However, cracks are beginning to form in the system. OpenAI doesn’t receive revenues from Microsoft’s AI cloud services, yet needs money to develop new models, including agentic ones, and Microsoft hasn’t offered new funding. Thus, Softbank is leading a funding round that is purported to almost double OpenAI’s valuation to $300 billion while Musk offered to buy it for $97.4 billion; very nice for a company with about $5 billion in annual revenue. The $500 billion promised in the Stargate project, by Oracle, Softbank, and Abu Dhabi’s Sovereign Wealth Fund — which was announced at the White House — also raises similar questions. How will OpenAI obtain revenues?

Other AI startups such as Anthropic, Mistral, and xAI are in the same boat. They must create services or partner with a cloud service provider because that is the main way of reaching customers. Thus, despite the rumors about better models and AGI, there appears to be no money at the end of the tunnel for many of these startups, while some insiders claim that AGI is a pipe dream, something that is at least a decade away.

Enter DeepSeek

DeepSeek has much lower training costs, but higher inference costs. Some investors claim that DeepSeek has rewritten the economics of AI, which resulted in a big one-day loss on January 27. But the market subsequently recovered and since then more questions have emerged than answers. Did DeepSeek really use fewer of Nvidia’s high priced processors or is DeepSeek hiding something? Although very cheap cloud services are available, particularly from Chinese companies, many countries are starting to ban them for security reasons and reports of the Magnificent Seven retraining their models can’t be found. Even if America’s startups begin offering similar services, revenues are hard to obtain, and happy customers are even harder to find.