a16z's latest in-depth analysis of the AI ​​market: Is your company still operating at a loss?

Author: Deep Thinking Circle

Have you ever considered that the software industry might be undergoing a transformation even more dramatic than the shift from command-line to graphical interfaces? I recently attended an in-depth analysis of the AI ​​market by David George from a16z, and I was struck by a set of data: the fastest-growing AI companies are expanding at an annual growth rate of 693%, while their sales and marketing spending is far lower than that of traditional software companies. This isn't an isolated case; the entire AI company group is growing at more than 2.5 times the rate of non-AI companies. What I find even more incredible is that these companies' ARR per FTE (Annual Recurring Revenue per Employee) has reached $500,000 to $1 million, while the standard for the previous generation of software companies was $400,000.

What does this mean? It means we are witnessing the birth of a completely new business model, an era of creating greater value with fewer people and less cost.

In his presentation, Avid George mentioned that this is not a minor adjustment, but a complete paradigm shift. Core concepts—version control, templates, documentation, and even the concept of a user—are being redefined by AI agent-driven workflows. I firmly believe that within the next five years, companies unable to adapt to this change will be completely eliminated.

The surprising truth behind the growth of AI companies

David George's presentation, which included data that made me rethink what true growth really means, highlighted 2025 as a year of accelerated growth for AI companies. After the slowdowns in 2022, 2023, and 2024 caused by rising interest rates and a contraction in the tech industry, 2025 completely reversed this trend. Most strikingly, among companies ranked in different tiers, the truly outlier companies experienced unbelievable growth rates.

My first reaction upon seeing this data was: Is there something wrong with these numbers? The top-performing AI company group saw a year-on-year growth of 693%. David said his team also checked the numbers three times before believing them. But this perfectly aligns with the actual situation and cases they've seen from their portfolio companies. This isn't an isolated phenomenon, but rather a systemic change happening across the entire AI field.

More importantly, it's about the quality of growth. Traditional software companies typically take a long time to reach $100 million in annual revenue, while the fastest-growing AI companies reach this milestone much faster. David emphasizes a crucial point: this isn't because they spend more on sales and marketing; quite the opposite, the fastest-growing AI companies actually spend less on sales and marketing than traditional SaaS (Software as a Service) companies. They grow faster, yet spend less. Why? Because end-customer demand is extremely strong, and the product itself is incredibly attractive.

I think this reveals a profound shift in business logic. In the past, during the software era, growth often relied on strong sales teams and huge marketing budgets. You needed to educate the market, persuade customers, and overcome adoption barriers. But in the AI ​​era, truly excellent products speak for themselves. When a product can immediately create value for users, allowing them to experience increased efficiency from the first use, market demand will automatically arise. This product-driven growth model is far healthier and more sustainable than the traditional sales-driven model.

Another set of data David presented was also quite interesting. AI companies actually have slightly lower gross margins than traditional software companies. Their team's perspective is unique: for AI companies, a low gross margin is, in a way, a badge of honor. Because if the low gross margin is due to high inference costs, it indicates two things: first, people are genuinely using the AI ​​functionality; second, these inference costs will decrease over time. So, to some extent, if they see an AI company with an exceptionally high gross margin, they might be a little skeptical, as it could mean that the AI ​​functionality isn't something customers are actually buying or using.

Why are AI companies able to achieve greater efficiency?

I've been pondering a question: why are AI companies, also software companies, able to generate more revenue with fewer people? David focused on the metric ARR per FTE in his presentation, which is the annual recurring revenue generated by each full-time employee. This metric is actually a comprehensive indicator of a company's overall operational efficiency, encompassing not only sales and marketing efficiency but also administrative and R&D costs.

The top AI companies have an ARR per FTE of $500,000 to $1 million, while the standard for the previous generation of software companies was around $400,000. This may seem like just a numerical difference, but it reflects completely different business models and operating methods. David believes the main reason for this difference is the extremely strong market demand for these products, so they need fewer resources to bring them to market.

But I think that's just the surface reason. The deeper reason is that AI companies were forced from the outset to think differently about how to operate. They had no choice but to use AI to redesign their internal processes, product development methods, and customer support systems. This forced innovation, in turn, led them to a more efficient business model.

David shared a particularly vivid example. He said he recently chatted with the founder of a company who was dissatisfied with the progress of one of their products. So, he directly assigned two engineers with deep expertise in AI to rebuild the product from scratch using the latest programming tools like Claude Code and Cursor, giving them an unlimited budget for these tools. The result? The founder said he felt the progress was 10 to 20 times faster than before. Moreover, the billing from these tools was so high that it made him start rethinking what the entire organization should look like.

What struck me most about this example was that this wasn't incremental improvement, but an order-of-magnitude leap. What does a 10 to 20-fold speed increase mean? It means a project that would have taken a year to complete might now only take one or two months. This difference in speed can have a decisive impact on competition. The founder concluded: I need to get the entire product and engineering teams working this way, and I believe this will happen within the next 12 months. But this also means a fundamental change in the team's organizational structure. Where are the boundaries between product, engineering, and design? These questions need to be redefined.

I believe December 2024 will be a turning point for programming. David feels the same way. He says it feels like there will be a qualitative leap in programming tools at that point. In the next 12 months, this change will either truly take root in companies, or those that don't adopt it will lag far behind their peers. This isn't alarmist; it's a reality.

Adapt to AI or be eliminated?

David raised a very serious point in his presentation: for companies founded before the AI ​​era, it's either adapt to the AI ​​era or die. This sounds extreme, but I completely agree. Moreover, this adaptation needs to happen simultaneously on two levels: front-end and back-end.

On the front end, companies need to think about how to natively integrate AI into their products, rather than simply adding a chatbot to existing workflows. This requires reimagining what products can do with AI and radically disrupting themselves to make changes. David shared several interesting examples. One pre-AI era software company, whose CEO has been completely transformed by the AI ​​philosophy, said: "We want to become an AI product. We want the product to be able to say, 'Your employees have now become your AI agents. How many agents do you have?'" These are the topics he discusses now.

Here's an even more extreme example. One CEO said, "For every task we need to accomplish now, I ask myself one question: Can I do this with electricity, or must I do it with blood?" This is an extreme shift in mindset. "Electricity" refers to using AI and automation, while "blood" refers to using human labor. This shift in thinking is profound; it forces you to re-examine every process and every task in the company.

On the backend, the company needs to fully adopt the latest programming models and tools. All developers should use the latest programming aids, and every functional area should use the latest tools. So far, the adoption rate in the programming field is the highest, and this is where we've seen the biggest leaps. But this change is spreading to other functional areas.

David noted that the good news for pre-AI companies is that the evolution of their business models is still in its early stages. The most disruptive scenario is when technology and products change simultaneously with business models. While technology and products are indeed undergoing dramatic changes, the transformation of business models has not yet fully unfolded.

He views business models as a spectrum. On the far left is the licenses model, the licensing and maintenance model of the pre-SaaS era. Then comes SaaS and the subscription model, typically based on seat fees—a major and disruptive innovation. You can look at what happened to Adobe during this transition. Next is the consumption-based model, which is how cloud services are charged; many task-based businesses have shifted from seat-based to consumption-based.

The next phase will be an outcome-based model. You'll be charged based on the successful completion of a task. Currently, the only area that truly supports this model is customer support and customer success, because you can objectively measure problem-solving. However, as model capabilities improve, if other functions besides customer support can also measure these kinds of outcomes, it will be incredibly disruptive to existing companies.

I find this evolutionary path very insightful. From licenses to subscriptions, from subscriptions to consumption, and from consumption to outcomes, each shift has disrupted the previous generation of business models. We are currently on the eve of this shift from consumption to outcomes. Once AI agents can reliably perform tasks and can be objectively evaluated, outcome-based pricing will become mainstream. At that point, companies still charging per seat will find themselves completely uncompetitive.

The AI ​​adoption dilemma of large companies

David's observations on the adoption of AI by Fortune 500 companies are quite interesting. He says there's a huge gap between what he hears from these large company CEOs and what's actually happening. The CEOs are all saying: We have to adapt, we desperately want to know what AI tools we need, we're ready to change, our business will fully roll out these tools, we want to become an AI company.

But the reality is quite different. The biggest disconnect between this mindset and actual business changes lies in the fact that change management is incredibly difficult. Even simply getting people to use AI assistants to help them do their jobs better is challenging enough. Actual business management, changing business processes, and change management are all extremely difficult.

David said he wasn't surprised by rumors circulating that things were progressing slower than expected. But for the best companies that have truly embraced AI and know what to do, the impact has been enormous. He gave several specific examples: Chime said they reduced support costs by 60%; Rocket Mortgage said they saved 1.1 million hours on underwriting, a six-fold year-over-year increase, equivalent to $40 million in annual operating cost savings.

I believe this reveals a crucial issue: the gap between willingness and capability. CEOs of large companies are willing to embrace AI, but their ability to implement it is another matter entirely. The difficulty of change management is often underestimated. It's not just about buying some tools or hiring some AI engineers; it requires fundamentally changing the company's processes, culture, and organizational structure.

Moreover, many large companies need to first adjust their operations to prepare them for AI. Using chatbots is one thing, and the resulting productivity gains may be limited. But if you have to completely overhaul your systems, information, and backends to adapt to AI, much of the work is potential, accumulating, and the results are yet to be seen.

David predicts the next 12 months will be very interesting. He believes we'll see more cases, but some companies will succeed and some won't. Those that succeed will gain a huge productivity advantage, while those that don't will be at a huge disadvantage. I think this differentiation will come faster and more dramatically than people imagine.

Model Busters and the Future of the Market

David mentioned a concept in his presentation that I found particularly insightful: Model Busters. This refers to companies whose growth rate and duration far exceed what anyone could predict in any given situation. The iPhone is a classic example of this concept. If you look at the consensus predictions before the iPhone's release and its actual performance 4-5 years later, the consensus predictions deviated by a factor of three. And this was one of the most watched companies in the world.

David believes that AI will be the biggest model buster he's ever seen in his career. Many AI companies will significantly exceed any spreadsheet's predictions. I completely agree. When a technology platform delivers orders-of-magnitude leaps instead of incremental improvements, traditional predictive models become ineffective.

He noted that technology itself is a model buster. But since 2010, technology has delivered high-margin revenues at an unprecedented speed and scale. So it always seems expensive in its early stages, but it repeatedly exceeds expectations, creating value far exceeding the required capital. He has no reason to believe this time will be any different.

David's data on capital expenditures is also quite interesting. Compared to the dot-com bubble era, current capital expenditures are actually supported by cash flow, and the percentage of capital expenditures to revenue is much lower. The largest burden of capital expenditures is borne by hyperscale cloud service providers, and these are some of the best-performing businesses of all time.

David specifically mentioned that, as a portfolio company, they welcome this kind of capital expenditure. He said: "Building as much capacity as possible to provide as much supply as possible for training and inference is a very good thing. And the majority of the burden is borne by the best commercial companies in history."

One phenomenon they started to notice was the influx of debt into the equation. You can't finance all projected future capital expenditures with just cash flow, and the market started to see some debt. But overall, they were quite comfortable with companies that financed with cash flow, continued to generate cash flow, and used debt, as long as their counterparties were companies like Meta, Microsoft, AWS, and Nvidia.

David cited a case worth noting: Oracle. Oracle has consistently been profitable and has been buying back shares, but their committed capital expenditures are enormous—a gamble. They will experience negative cash flow for many years to come. The market has begun to take notice; Oracle's credit default swap (CDS) costs have risen to approximately 2% in the past three months. This is a signal that warrants attention.

I believe this capital-intensive construction phase is necessary, but not without risk. The key is to ensure these investments ultimately yield a return. Currently, demand far exceeds supply. All hyperscale cloud service providers report demand far exceeding supply. Gavin Baker, in an interview with David, gave a good analogy: the internet age laid a lot of fiber optic cables, and then those cables lay idle, unused—that's called dark fiber. But in the AI ​​era, there's no such thing as dark GPUs. If you install a GPU in a data center, it will be fully utilized immediately.

The astonishing rate of income growth

David presented a particularly striking set of data. He compared cloud services, publicly traded software companies, and the net revenue generated in 2025. Publicly traded software companies collectively generated $46 billion in new revenue in 2025. If you only look at OpenAI and Anthropic, their new revenue, calculated by operating income, is almost half that figure.

Furthermore, David believes that if the same comparison is made for 2026, AI companies (model companies) could account for 75% to 80% of the new revenue generated by the entire publicly traded software industry (including SAP and established software companies, not just SaaS). This speed is simply incredible. This means that in just a few years, the new value created by AI companies will exceed that of the entire traditional software industry.

Goldman Sachs estimates that AI infrastructure will generate $9 trillion in revenue. Assuming a 20% profit margin and a P/E ratio of 22, this translates to $35 trillion in new market capitalization. Approximately $24 trillion is already priced in. While we can debate whether this is entirely attributable to AI or the performance of large tech companies, there is still significant room for growth, and if these assumptions are correct, there is substantial upside potential.

David also did a simple calculation. According to current estimates, by 2030, the cumulative capital expenditure of hyperscale cloud service providers will be slightly less than $5 trillion. To achieve a 10% threshold return on this $4.8 trillion, or close to $5 trillion, investment, AI annual revenue would need to reach approximately $1 trillion by 2030. In context, $1 trillion is roughly 1% of global GDP to generate a 10% return.

Is this even possible? It might be slightly less than ideal. But David believes that looking only at 2030 is limiting. The returns on these investments could materialize over a longer period, such as between 2030 and 2040. And if we currently have around $50 billion in AI revenue (his rough estimate), and that's largely been generated in the last year and a half, then growing from $50 billion to $1 trillion isn't inconceivable.

My thoughts on the future

After listening to David's presentation, my biggest takeaway is that we are at the beginning of a historic turning point, not in the middle or end. This is a product cycle that could last 10 to 15 years, and we've only just begun. This makes me both excited and anxious.

What's exciting is the immense opportunity this transformation presents. Companies that can quickly adapt and fully embrace AI will not only gain a competitive edge but also have the potential to define the next era. We will see new unicorns emerge, new business models appear, and entirely different ways of organizing companies.

The worrying thing is that this change may be happening much faster than most people expect. The statistic David mentioned is particularly telling: the average time an S&P 500 company stays in the index has decreased by 40% over the past 50 years. This means that companies are being disrupted at an accelerating pace. In the age of AI, this pace may accelerate even further.

I believe a clear divergence will emerge. Some companies will truly understand the potential of AI and fundamentally rethink their products, processes, and organizational structures. These companies will achieve orders-of-magnitude efficiency gains and competitive advantages. Others, even if willing to change, will progress slowly due to difficulties in change management, organizational inertia, and technological debt. This divergence will become increasingly apparent in the coming years.

For entrepreneurs, now may be the best of times. Market demand is extremely strong, technological capabilities are advancing rapidly, and the capital market remains willing to support companies with real potential. Moreover, compared to the previous generation of software companies, it's now possible to achieve the same scale with fewer resources and at a faster pace. This lowers the barrier to entry for starting a business, but it also raises the bar for product quality and market fit.

For investors, the key is to identify the true model busters. These companies grow at rates and for durations far exceeding the predictions of any traditional model. But this also requires investors to have sufficient foresight and patience, and to be willing to believe in growth curves that seem unreasonable.

For professionals, whether you're an engineer, product manager, designer, or in any other role, the ability to quickly learn and adapt to new tools and ways of working is crucial. David's example—two engineers using the latest programming tools being 10 to 20 times faster than before—is not an isolated case, but a trend. Those who master these new tools and methods will gain a significant career advantage.

Finally, I want to say that this transformation is not just a technological one, but a shift in mindset. It's a shift from "how we should do it" to "what results we want to achieve," from "adding more manpower" to "how to use AI to solve this problem," from "following established procedures" to "reimagining possibilities." The question of "electricity or blood," though seemingly extreme, captures the essence of this transformation.

We are witnessing the rewriting of the software world. This is not an incremental upgrade, but a complete reconstruction. And those individuals and companies who understand and embrace this will define the next era.