AI’s Unsustainable Burn Rate: The Harsh Realities of the Industry

AI’s Reckoning: The Mounting Losses and Uncertain Future of the Industry

The AI Gold Rush: Navigating the Path to Profitability

In less than three years since the launch of ChatGPT, I’ve witnessed OpenAI become one of the world’s most valuable tech startups, reportedly achieving an astounding $80 billion valuation in a recent share sale.

Within this short period, AI has become a booming business, with OpenAI’s revenue reaching a run rate of $2 billion by the end of 2023 and targeting up to $5 billion in 2024.

However, as AI companies have seen their revenue skyrocket over the past two years, so too have the massive computational costs of running increasingly complicated AI models.

It is widely believed that most AI companies, including OpenAI, are currently losing money.

We strongly recommend that you check out our guide on how to take advantage of AI in today’s passive income economy.

The Cost of AI Success

In 2022, Microsoft launched GitHub Copilot, a tool powered by OpenAI technology that helps software engineers write code.

Despite its popularity and affordable price of just $10 per month, it is estimated that Microsoft is losing $20 per user per month due to the massive data center costs.

Recently, Google parent Alphabet’s chairman indicated to Reuters that generative AI chatbots cost approximately ten times more per query than a traditional Google search.

While investors are currently caught up in the AI hype train, it’s crucial to take a step back and ask: what is the path to profitability?

The Opaque Financials of AI Companies

While there are numerous AI companies, it’s challenging to get a sense of how much money they’re making or losing.

Tech giants like Microsoft and Google have generative AI offerings, but their economics are difficult to determine as they’re consolidated within their other businesses.

There are pure-play AI companies like OpenAI and Anthropic, but they are privately held, so our only information about their profitability comes from sporadic media reports.

According to a report by The Information, OpenAI’s 2022 operating expenses were estimated to be $540 million, including $420 million in computing costs, $90 million in employee costs, and $30 million in other costs.

However, ChatGPT was only released in November 2022, so they incurred very little cost in operating the service, with most of the computing cost related to training and testing ChatGPT.

Unfortunately, we don’t have any solid information about how much money they’re spending to power ChatGPT now that it’s being used by tens of millions of people.

The Staggering Costs of AI Development

Another AI company, Anthropic, which is a competitor to OpenAI, has raised over $7 billion over the past two years, which they’re burning through very quickly.

According to reporting by The New York Times, it is expected that Anthropic will seek additional equity financing in the near future.

By the end of 2023, Anthropic’s Claude chatbot reached $8 million in monthly revenue, or almost $100 million annualized, which is impressive for a new startup but pales in comparison to the $7 billion they’ve raised.

The biggest cost for AI companies is the massive amount of computing power they use to train and operate their increasingly complex models.

Thus, it should be expected that they burn a lot of money upfront with the hope of recouping this cost once they commercialize their product.

The True Cost of Running AI Models

The more important question is: how much does it cost to operate an AI model once it is complete? Stated differently, what prices will they have to charge to make a gross profit?

One proxy for computing cost is electricity usage. Operating data centers requires electricity, and a lot of it.

According to estimates from Livemint, one query on ChatGPT-4 uses between 0.001 and 0.01 kilowatt-hours to process.

While the accuracy of this estimate is questionable, if we take the midpoint of 0.005 kilowatt-hours per query, this is about 17 times greater than the amount of electricity used to power one Google search.

As a sanity check, these estimations are at least within the same ballpark as statements made by Alphabet Chairman John Hennessy to Reuters, indicating that AI chatbots like ChatGPT or Google’s Gemini cost about ten times more per query than traditional Google searches.

The Hidden Costs of AI Infrastructure

To be clear, the high cost of AI is not due to electricity itself. The average cost of electricity for industrial customers in the US is 8 cents per kilowatt-hour, so 8 cents of electricity can get you roughly 200 ChatGPT queries.

Electricity only represents a small minority of the cost associated with building and operating a data center.

The largest cost of AI data centers is the upfront cost of the GPUs.

In the calendar year 2023, Nvidia generated roughly $61 billion in revenue, more than double the amount it made in the previous year, with substantially all of this growth attributed to artificial intelligence applications.

Most of the GPUs are sold to cloud service providers such as Microsoft, Amazon, and Google, which then lease their computing power to AI companies such as OpenAI, Anthropic, etc.

Nvidia disclosed that 33% of their revenue was attributed to a single customer, which they call “Customer A,” and 19% of revenue was attributed to a single indirect customer, totaling $8 billion and $1.6 billion, respectively.

While we don’t know for certain, the most likely candidate for this anonymous customer is Microsoft.

Analysts at the brokerage firm D.A. Davidson estimate that nearly 40% of Microsoft’s total capital expenditures are now spent on Nvidia GPUs.

In the calendar year 2023, Microsoft spent about $35 billion on capital expenditures. If Microsoft was indeed Nvidia’s largest customer, they would have spent about 33% of their capex on Nvidia GPUs, which is at least within the ballpark of D.A. Davidson’s estimate.

The AI Value Chain: From GPUs to End Users

With companies like Microsoft spending billions upon billions of dollars to build out new AI computing capacity, will the end demand for AI applications be enough to justify this investment?

If we look at the AI value chain, the cloud service providers are collectively spending tens of billions of dollars on Nvidia GPUs and other equipment.

Once they buy the equipment, they incur ongoing costs to operate their data centers.

AI startups like OpenAI and Anthropic rent computing power from the cloud service providers.

The cloud service providers themselves need to make a profit, so they charge a markup over their own cost.

Ultimately, the AI startups need to charge their own customers a high enough price to pay for all of this.

The Profit Puzzle: Microsoft’s GitHub Copilot

Earlier, we talked about how Microsoft is losing $20 per user per month on their GitHub Copilot.

Microsoft is presumably using its own data centers to power GitHub Copilot, so it’s not clear how these costs are calculated.

It’s possible that Microsoft’s cloud computing business is making profits at the expense of GitHub Copilot.

It can be hard to decipher the real economic situation when intersegment transfers are involved.

Anthropic’s AI Economics

Another example we can look at is Anthropic.

A publication called Contexto claims to have gotten its hands on some of Anthropic’s financial data.

According to them, Anthropic’s gross profit margins are around 50%, with the cost of goods sold only accounting for the ongoing cost of operating the AI models once they are complete.

Additionally, they claim that Anthropic incurs up to $100 million in server costs to train each model.

It’s unclear what is referred to by a “model.”

Anthropic recently released its Claude 3 model family, which includes three models: Haiku, Sonnet, and Opus.

If it cost $100 million to generate each one, this would be $300 million in total.

Given that the company recently raised $7 billion and is reportedly burning through it very quickly, $300 million in server costs to develop the Claude 3 family is certainly possible, although they don’t cost the same amount to develop.

The largest model, Opus, certainly costs much more than the smallest model, Haiku.

Anthropic’s Pricing Strategy

Anthropic offers a customer chatbot called Claude Pro, which costs $20 per month, but it is believed that most of their revenue comes from their enterprise offering.

Their enterprise offering has a consumption-based pricing model based on tokens, where each token represents approximately 3.5 English characters.

Let’s say the average input prompt is 50 words, which would be about 350 characters, costing about 100 tokens.

The output is generally longer than the input. Let’s assume 250 words, which would be roughly 500 tokens.

If you use their cheapest model, Haiku, one prompt and response will cost you about 0.065 cents.

If you use their most expensive model, Opus, one prompt and response will cost you about 4 cents.

While 4 cents doesn’t sound like a lot, it can add up very quickly.

The reason Anthropic is already able to make a gross profit is because of their consumption-based pricing model. Power users can rack up bills in the hundreds of dollars per month.

The Loss Leader Strategy

We can see how GitHub Copilot was losing so much money.

The people who use it are mostly software engineers who code all day for a living. They could easily give it 100 prompts per day as they experiment and try to refine the outputs.

Even if it only costs Microsoft 1 cent per prompt, that could be $1 per day, or close to $30 per month, and they were only charging $10 per month for it.

The $10 per month was likely an intentional loss leader to get software engineers used to the experience.

Microsoft subsequently released more powerful and versatile versions targeting businesses and large enterprises, costing $19 per month and $39 per month, respectively.

OpenAI appears to be using the same strategy.

Their free version of ChatGPT is obviously a loss leader. Even ChatGPT Premium at $20 per month probably isn’t generating much profit.

But by gaining a user base tens of millions strong, many companies are willing to pay OpenAI to make ChatGPT plugins.

ChatGPT’s consumer offering also creates market share and market awareness, helping them drive sales for their API, which follows a similar consumption-based pricing model as Anthropic.

This is likely where the bulk of their revenue comes from.

The Profitability Paradox

It’s certainly possible for AI companies to be profitable, so long as they charge a high enough price for their services.

But the more you charge, the fewer people will be able to afford it.

The question is: will the market be big enough to justify the investment?

In the fiscal year 2024 (the 12 months ended January 28, 2024), Nvidia generated $47 billion in revenue in their data center segment, $32 billion greater than what this segment generated in the previous year.

Substantially all of this $32 billion in growth is related to AI end markets.

So how big does the end AI market need to be to pay for this investment?

The Breakeven Point

If you amortize the investment in Nvidia GPUs over ten years, the cloud service providers will incur $3.2 billion in depreciation per year.

Collectively, they also have to buy other related equipment and pay for the ongoing operational cost of the data centers.

To make things simple, let’s say these other expenses double the total costs.

So the cloud service providers need to generate $6.4 billion from their AI customers per year to break even.

The AI startups themselves need to make enough gross profit to cover their massive R&D expense and corporate overhead.

For them to break even at scale, they will likely need to charge their end customers double whatever they pay to the cloud service providers.

This lines up with Anthropic’s reported 50% gross margin.

So the AI startups will need to generate $13 billion in annual revenue to achieve a 50% gross margin.

They’ll have to charge the end customers a high monthly cost, so let’s say $40 per month, which is the same cost as GitHub Copilot Enterprise.

To generate $13 billion per year, you need 27 million enterprise seats, each paying $40 per month.

These are all very rough estimates because of the limited information that we have.

The point of this exercise is to show just how massive end AI adoption has to be to cover the cost of the Nvidia GPUs purchased in 2023 alone.

The AI Investment Conundrum

The current growth in AI is currently funded mostly by venture capitalists and cloud service providers.

Companies like Anthropic and OpenAI may be generating gross profits, but they’re still burning through billions of dollars on research and development.

Cloud service providers and other tech companies are spending tens of billions of dollars on Nvidia GPUs in hopes that the AI market will grow enough to justify this investment.

It was recently reported that OpenAI CEO Sam Altman is talking to sovereign wealth funds in the Middle East, trying to raise $1 trillion to build new AI chips.

This is obviously absurd and is never going to happen.

Regardless, it portrays a mindset shared by many tech bros and investors alike: the idea that the opportunity in AI is so big that an almost unlimited amount of investment can be justified.

But there’s no definitive proof that enough end users will be willing to pay high enough prices to make the industry viable.

The AI Revolution: Hype vs. Reality

Many people say that generative AI represents a paradigm shift in technology, with some saying its impact will be even bigger than the internet.

But it’s important to remember that in the early days, the internet was expensive and inefficient.

It nevertheless created a huge amount of hype, and investors were willing to fund thousands of dotcom companies whose businesses were not yet viable due to the primitive nature of the internet at the time.

This malinvestment ultimately created the dotcom bubble.

Eventually, networking technology became more efficient and cheaper.

Once the majority of the population gained access to affordable high-speed internet, online business models finally became viable at a large scale.

But this didn’t happen until many years after the original hype cycle, and most of the early dotcom companies had already gone bankrupt.

For AI to have the same revolutionary impact as the internet, costs need to come down a lot.

This will happen eventually, but it could be many years into the future.

Conclusion

The AI industry is currently in a gold rush phase, with investors and tech companies pouring billions of dollars into the development and deployment of AI models.

However, the path to profitability remains unclear, as the costs of running these models are still very high.

While some AI companies are generating revenue, they are still burning through cash at an alarming rate.

For the AI revolution to truly take hold, costs need to come down significantly, and end users need to be willing to pay high enough prices to justify the massive investments being made.

Only time will tell if the current level of investment in AI is justified or if we’re in the midst of another tech bubble.

As the industry matures and the hype settles, we’ll have a clearer picture of the true potential and limitations of AI technology.

We strongly recommend that you check out our guide on how to take advantage of AI in today’s passive income economy.