The Bumpy Road Ahead for AI Startups

Ever since ChatGPT burst onto the scene late last year, the hype and excitement around artificial intelligence has reached a fever pitch. This has fueled a veritable gold rush, with over $35 billion pouring into AI startups just in the first half of 2024 - accounting for a staggering 21% of global venture capital funding.

But as the dust settles, doubts are starting to emerge about the long-term viability of these AI-powered business models. Sequoia's recent analysis posed the "$600 billion question" - where's the additional revenue that justifies the staggering investments in GPUs and data centers? Even the typically bullish Goldman Sachs issued a more measured report, highlighting both the promise and pitfalls of the technology.

Of course, the bigger existential questions about AI's impact on jobs and society as a whole continue to loom large. These are certainly fascinating debates to follow. But in this article, I want to take a closer look at a more specific area that has me feeling skeptical - the wave of VC-backed startups trying to sell generative AI products like virtual agents, workflow automation tools, and custom chatbots to enterprise customers.

I’ll start with the short version and then explain each factor in more detail below:

The TL;DR Version of My Skepticism Toward (Enterprise) AI Startups

Most of the tangible benefits of generative AI technologies will ultimately flow to two key groups - individual consumers and small businesses that use the end products, as well as the tech giants and consulting firms that are responsible for training the underlying models or providing the critical infrastructure.

The reality is that most enterprise-focused AI startups are structured more akin to service businesses rather than traditional software or SaaS companies. This means their margins and scalability potential is inherently limited, often falling short of the lofty "Rule of 40" benchmarks. And let's be honest, the majority of startups simply struggle to effectively deliver complex services to large corporate customers.

Another key concern is that most AI startups are not actually doing anything truly defensible or proprietary. The training data is really the key value driver for these AI models, products, and features - and that data is something the major tech giants have in spades. Without unique proprietary data or models, the long-term competitive moats for these startups start to look rather thin.

In fact, when you step back, AI is more accurately characterized as an additional feature or capability to be incorporated into existing products and services, rather than a wholly new "product category" in its own right. This is a crucial distinction, especially when you consider how quickly the tech behemoths have adapted and seized the initiative in previous platform shifts, like the rise of the internet or mobile devices.

And speaking of exits, a troubling trend has emerged where many AI startup acquisitions are structured more as licensing deals or acquihires, where the acquirer doesn't even technically "buy" the startup. This practice could significantly reduce the exit values and investment returns for VC backers, while also disadvantaging the rank-and-file employees of these AI firms.

Do AI Startups Benefit from AI?

Most people's immediate, gut reaction to the arguments I've laid out so far is likely something like this: "But ChatGPT or that new AI feature is so darn useful! It saves me so much time and helps me get way more done."

And you know what, I completely agree. For individual users and small businesses that are perpetually strapped for time and resources, these generative AI tools can be a game-changer. They provide a way to boost productivity and efficiency without having to bring on additional employees or contractors.

However, the key question here is not about which users benefit the most from these AI capabilities. The real issue is who ultimately profits the most from them. And at this stage, the answer appears to be tilted heavily towards the tech giants and consulting firms, for a few key reasons:

1. They're the ones selling the underlying hardware and infrastructure powering these AI models, like Nvidia.
2. The tech titans like Microsoft, Google, and Facebook possess the massive troves of data and computing resources required to train advanced AI systems, and can seamlessly integrate these features into their existing product ecosystems.
3. Consulting firms like BCG, McKinsey and Accenture are cashing in by offering expensive implementation and advisory services to help enterprises adopt these AI technologies.

In contrast, I believe the VC-backed AI startups find themselves stuck in a sort of no-man's land. Sure, they can leverage tools like CoPilot and GPT-4 to develop products more quickly. But they simply lack the inherent advantages in training data and sales/marketing muscle that the tech behemoths possess. Those are critical ingredients for building truly valuable B2B products and services.

At the end of the day, simple ChatGPT-powered wrappers may have their niche uses. But I don't foresee any large, sustainable businesses being built solely on these types of generic, non-defensible AI capabilities.

Most AI Startups Are Service Businesses, Not Software Businesses

You might be surprised to learn that a consulting behemoth like Accenture has already generated a staggering $3.6 billion in annualized bookings for its generative AI services, compared to OpenAI's own annualized recurring revenue of $3.4 billion.

But if you have a deeper understanding of how these large language models (LLMs) actually work, this dynamic starts to make a lot more sense.

The core issue is that it's incredibly difficult to build a truly "one-size-fits-all" AI product that can seamlessly serve the needs of diverse enterprise customers. Each organization has its own unique workflows, proprietary data, and specialized requirements. As a result, customizing and training the AI models to meet their needs requires a significant amount of computing power and human labor.

And the work doesn't stop there. Once the initial version is deployed, it must be continually tweaked, updated, and retrained over time to maintain optimal performance.

Now, there's nothing inherently "wrong" with this services-driven business model. But it's a far cry from the classic enterprise software or SaaS startup playbook that many venture capitalists have grown accustomed to. The margin and valuation profiles simply aren't going to be the same.

For reference, consulting giants like Accenture, IBM, and Cognizant typically trade at revenue multiples of 2-4x - a far cry from the 5-10x multiples that VCs may be hoping for from their AI portfolio companies. These are solid, profitable businesses, to be sure. But they operate much more like specialized services firms than traditional high-growth software plays.

Most Startups Do Not Know How to Be Service Businesses

There's one specific problem with being a services-oriented business beyond just the lower margins and valuation multiples - most startups and growth-stage companies simply aren't very good at it.

Sure, there are some notable success stories out there, like Palantir and Red Hat, but those tend to be the exceptions that prove the rule. The reality is that selling complex products and services to large enterprises is an immensely challenging endeavor.

These big corporate customers have a laundry list of stringent compliance, safety, and privacy requirements that any third-party solution must be certified against. And as this insightful Hacker News thread illustrates, navigating that minefield is no easy feat. It requires significant resources and expertise that many scrappy startups simply don't possess.

In contrast, the tech titans inherently have a huge advantage in this arena. They already deeply understand the unique needs of large enterprises and have the organizational scale to efficiently deliver and support their offerings.

AI is Mostly a New Feature, Not a Product

And we've seen this dynamic play out time and again, with AI-powered "features" getting quickly bolted onto mainstream products like Word, Gmail, Photoshop, and Google Search. Now, we can debate the actual usefulness of these AI integrations all day. But the core point is that these big tech firms can easily inject these capabilities into the products already used by millions or billions of people.

This is a meaningful difference compared to previous technology waves, like smartphones and mobile apps. In those cases, the new platforms genuinely enabled entirely new categories of products - think Uber, which couldn't have simply been "added" to Windows or Google Search. But so far, most enterprise AI startups have been focused on services that could theoretically be folded into existing software offerings.

So for any AI-driven startup targeting the corporate market, the real challenge is making a convincing case that their solution offers something truly unique and differentiated - something that won't just get absorbed into the next version of Slack, Acrobat, or Microsoft Office.

That's why, in my view, the more promising opportunities may lie in hardware-centric AI startups doing things that simply aren't possible with today's existing physical-world capabilities.

Big Tech Companies Are?On the Ball?This Time Around

After the 2008 financial crisis, the tech giants were relatively cautious, focused on smaller deals and IP acquisitions rather than major business model overhauls. This allowed startups like Stripe, Uber, and Airbnb to gain traction before the big players took notice.

But the landscape has shifted dramatically since then. Companies like Google, Facebook, and Microsoft now dominate the S&P 500 and have experienced astounding revenue growth - Google over 10x, Facebook over 100x. These tech titans have far greater resources to invest in developing their own AI capabilities.

They are unlikely to "sleep on" promising AI startups this time around. The one potential exception is OpenAI, given Microsoft's 49% stake - but even then, it's essentially a strategic extension of Redmond's AI ambitions.

The big tech firms' ability to quickly integrate new AI features into their existing products poses a major challenge for standalone AI startups targeting the enterprise market. The days of young upstarts catching the giants off-guard may be behind us, at least when it comes to AI.

Startup AI Deals Could Hurt VC Investors

When examining the structure of many AI startup exits so far, a common pattern has emerged:

• The big tech acquirer pays a "licensing fee" to the startup, rather than executing a full acquisition. This may or may not end up recouping the original VC investment.
• Some of the startup's employees will then join the big tech company as part of the deal.
• However, the AI startup itself continues to operate, but without the involvement of the co-founders or other key personnel.

We've seen this play out with acquisitions like Google's deal for Character.AI, Microsoft's acquisition of Inflection, and Amazon's purchase of Adept.

The tech giants are likely using these structures to avoid heightened antitrust scrutiny. But the unintended consequences could hurt both the venture capitalists and the rank-and-file employees.

VCs are less likely to achieve the high exit valuations they covet if these are essentially just "licensing" transactions rather than outright sales. And many startup employees could potentially lose their equity stakes in these deals - a dynamic that could erode the incentives for top talent to join these AI ventures in the first place.

Ultimately, the tech behemoths seem to recognize that most AI startups lack viable standalone business models. But they still see value in acquiring the talent and technology to seamlessly integrate as features within their mass-market products. Hence, these structured deals become an efficient way to achieve that goal.

And Why I Could Be Wrong About AI Startups

I'm confident that additional AI startups will continue to find success. However, I suspect the scale of that success may be more modest than what many VCs are expecting, and will likely happen more outside the enterprise sector.

Certain startups addressing underserved niche markets could do quite well, even if they don't reach the towering heights of tech giants like Google or Facebook. And I could see some promising spin-offs emerge from large companies or institutions that possess troves of valuable, proprietary data.

Moreover, there's no shame in building a successful service or agency-style business, even if they lack the financial profiles of traditional software or biotech plays. These types of companies can absolutely thrive.

And of course, AI also presents abundant opportunities for small, cash-cow businesses to succeed - think of all the consumers who will happily pay modest monthly fees for simple AI-powered apps for photo generation, outlines, summaries, and the like.

The Biggest Misunderstanding About AI Startups

The key issue is that usefulness doesn't necessarily equate to profitability as an investment. Take Facebook - it remains hugely "useful" to hundreds of millions of users, but the real money comes from its lucrative ad business, not basic user functionality.

Similarly, much of the output from enterprise AI startups so far falls into the "potentially useful but hard to monetize independently" camp. That may shift over time, but for now, I'm more bullish on the prospects for Big Tech and AI-focused consulting firms.

Personally, after the grind of my current business, I don't have much appetite to start something new. But if I were launching a new online venture today, I'd certainly look to leverage generative AI to boost efficiency and productivity. That may not benefit VCs, but that's their concern, not mine.

The bottom line is that usefulness and profitability don't always go hand-in-hand. And for many AI startups, cracking the code on sustainable, scalable business models remains an elusive challenge.