Generative AI is changing the world as we know it. ChatGPT has already become one of the most widely adopted consumer apps in history – the app has well above 800m weekly active users, which represents 10% of the world’s population. Anthropic, another model winner, now has over $2b in revenue. Both model providers are becoming household (or enterprise) names, and countless reports have detailed the multi-trillion dollar opportunity ahead.

As we previously wrote, it is definitely living up to its 2023 hype. Generative AI has the promise to transform nearly every industry. Today, routine tasks are being rapidly automated, paving the way for a wave of new businesses. Already, tens of startups in the space have surpassed $100 million in ARR faster than ever (total startup ARR in the space is north of $6.5b already), and we’ve seen our first generative AI IPO with CoreWeave going public at $23b – a whopping 17x revenue multiple despite volatile market conditions.

Despite the excitement and capital inflows, a critical question remains unanswered: Will there be a payback on the hundreds of billions invested into Generative AI?

Most of these investments by dollar amount have been in the infrastructure layer where they are “wholesaling” Generative AI capability. In the traditional rules of trade, there’s no wholesale market without a retail market. Put simply, in order for the massive spend on GPUs, data centers, and AI talent to be justified, end-user applications need to be deployed at scale, driving revenue and clear returns.

Let’s consider Microsoft’s case, arguably the most advanced example of AI ROI at scale. How much Generative AI revenue will Microsoft need to generate to pay back their Capex spend?

At the end of 2024, Microsoft generated $10 billion in Generative AI revenue run-rate, more than doubling from the previous year. At the same time, the company dramatically ramped up Capex spend from $28 billion in 2023 to $44 billion in 2024. Per Microsoft’s earnings calls, 50% of that spend, c. $23 billion, is attributable to AI.

Microsoft has a large suite of AI products that are in production in the enterprise today. Microsoft Copilot Pro is increasingly rolled out to customers starting at $20 per user per month (although average price is $8 per user per month given enterprise discounts). Github Copilot was one of the first AI businesses to reach $100m ARR, and is at ~$400m ARR today, driving 40% of Github growth. On top of that, Microsoft has been heavily commercialising Azure services around AI (both compute as well as large teams of deployment engineers to assist businesses to get AI into production).

Satya Nadella famously said that he is closely scrutinising Microsoft’s return on AI investment, likely using the below formula:

Return on Invested Capital (ROIC) = Net Operating Profit After Tax / Invested Capital 

If we assume that Microsoft is targeting its historic ~20% ROIC, that they get to 45% EBIT margins on Generative AI (in line with their current margin), and spend >$200b total in AI Capex ($80b is planned for this year alone, so this assumes Microsoft doubles AI investment to date over time), the company would need roughly $130b of annual Generative AI Revenue to clear their ROIC hurdle. This means that Microsoft needs to earn $130b of ARR for five years to pay back their Capex investments.

Practically, this means that Microsoft needs to grow their Generative AI revenues by 12x from today (over multiple years). It’s an ambitious target, but given the fact that their Generative AI revenue grew 5x last year, and they have clear productisation across multiple pre- and post-production product lines, this equation might work out for them.

Does Microsoft Maths work for the rest of Generative AI?

What happens if we expand this logic to the rest of the AI ecosystem? Following the ChatGPT “aha moment,” $140 billion has been poured into generative AI startups, per Pitchbook. A large portion of this capital flowed to the foundational model players (e.g., OpenAI, which recently announced a $40 billion round led by SoftBank), but many application and infrastructure companies also raised substantial funding.

Now, let’s see if VC’s AI Maths holds up in the same way as Microsoft’s. Unlike businesses, VCs think of their “money-on-money” return on investments. 

Money on Money (MoM) = Total Cash Proceeds at Exit / Total Cash Invested 

To justify the $140 billion investment into Generative AI, assuming VCs are underwriting a 5x payback, we estimate that Generative AI needs to create $3 trillion of enterprise value for VCs if investment continues at the current pace. If we then assume that VCs own c. 40% of companies they invest in, VC-backed Generative AI companies need to be generating $1 trillion of revenue. That is 1.3x larger than the entire enterprise software market today, which generates $785b.

$3 trillion of enterprise value is an astronomical figure and $1 trillion of revenue is no small feat. This new revenue will have to come from somewhere.

The obvious source that everyone points to is replacing existing software with AI-native solutions. However, the B2B software incumbents have not been asleep at the wheel. As we mentioned, Microsoft has built a real case for being a Generative AI winner, and others like ServiceNow and Salesforce are likely getting many things right when it comes to AI. If we assume 50% of the software market, or $400 billion, could be vulnerable to new VC-backed Generative AI companies, that still leaves a $600 billion gap in revenue to fill from somewhere other than current software spend.

The additional revenue for AI therefore needs to come from Generative AI companies helping enterprises either a) grow their revenue or b) improve their productivity. Productivity gains have to come from reducing headcount drastically (a brave endeavour for a CEO), or more likely in the near to medium-term, supporting enterprise growth without adding new headcount.

In a world where AI touches every P&L line, businesses could see dramatic improvements in operating income. We’ve decided to take efficiency gains to the max in the illustrative example below to see if sufficient value can be created with a 5x uplift to profits. This could be enough to get us to the additional $600 billion of revenue, if we assume that Generative AI companies are able to capture at least 10% of the value they create. It does imply though that Generative AI simultaneously grows total enterprise revenues by 5% while cutting costs by 30%.

The unspoken reality is that there will also likely be a lot of venture dollars lost in Generative AI. Generative AI is very likely to follow Carlota Perez’s framework on disruption, with an “installation phase” – where new technology emerges and its underlying infrastructure is built – followed by a collapse, ahead of a “deployment phase” where technology is broadly adopted. In the deployment phase, net new products and services that weren’t conceivable prior to LLMs will be born. Think about how the birth of computers led to the creation of Microsoft, or how the birth of the internet led us to Google. These are the kinds of exciting LLM startups we are excited to back – companies that are redefining the future as we know it.

The potential upside from Generative AI is immense, but following the math is critical. VC-backed startups in Generative AI need to create up to $3 trillion in value (depending on investment losses), and replacing the existing software market is not enough. Companies will need to help enterprises grow their topline and drive radical productivity improvement to have a shot at reaching this value. This type of innovation will not come from incumbents but, excitingly for VCs, from a whole new breed of startups that will emerge, creating products and services that were not even conceivable prior to LLMs.

Look out for Part II of our article, where we highlight some of the European B2B companies that could become winners. If you’re building in the space (or just want our AI Maths cheat sheet!), please reach out to shamillah@dawncapital.com and norman@dawncapital.com.