The AI Gold Rush Has a Dependency Problem
AI has been the new startup gold rush for the past few years. Billions of dollars have poured into the industry, thousands of startups have emerged, and nearly every existing software company has tried to reposition itself as an AI company. But when you look at where most of the money has actually gone, it’s hard not to notice that the biggest winners have been the largest companies. Frontier AI labs have attracted enormous funding while also collecting inference fees from startups and businesses building on their models. Cloud providers and chipmakers have benefited from the infrastructure spending. Meanwhile, many startups that expected to benefit from this AI revolution have struggled to build sustainable businesses.
It’s strange that so many startups are effectively paying their potential future competitors for the technology that makes their products possible. The more successful these startups become, the more money they send upstream to the companies that control the underlying models. Unlike traditional infrastructure providers, these companies aren’t necessarily content to stay in infrastructure. They’re increasingly building applications themselves, often targeting the same customers and use cases as the startups that depend on them.
This creates a problem beyond inference costs. When you build your product around someone else’s model, you don’t really control your roadmap. You’re dependent on what that model can do, how much it costs, how quickly it improves, and whether its provider decides to compete with you. A model update can suddenly make your product significantly better, but it can also make your product largely unnecessary. Something that required a specialized application six months ago might become a built-in feature of the next model release.
For years, one of the laziest questions VCs would ask founders was, “What if Google built this?” In some cases, it was a reasonable concern, but you could ask it about almost any software startup. Today, that question has simply become, “What if OpenAI or Anthropic built this?”
The difference is that the threat is arguably much more immediate. Historically, Google had to decide to enter a market, allocate resources, build a product, and distribute it. Frontier AI labs can sometimes eliminate entire categories of startups simply by improving their general-purpose models or adding a relatively small feature to their existing products. They don’t even need to target a startup’s market specifically. A general improvement in reasoning, memory, tool use, or multimodal capabilities can make dozens of specialized products less valuable overnight.
Of course, this doesn’t mean every AI application is doomed. The same argument could apply to startups building on operating systems, cloud platforms, or the internet itself. Plenty of successful businesses have been built on infrastructure they don’t own. But there’s a meaningful difference between depending on infrastructure with relatively stable capabilities and depending on technology whose capabilities are expanding so quickly that it can absorb much of the value your application provides.
Many AI startups have essentially bet they can build a business faster than the underlying models improve. They identify something the models aren’t particularly good at, build an application that compensates for those limitations, and hope to establish enough distribution and customer loyalty before the next generation of models makes their work obsolete. Sometimes this works. But it’s a difficult way to build a lasting company when you can't control how fast the underlying technology advances.
Open-weight models have helped change this dynamic. The availability of increasingly capable models that companies can run themselves has reduced some dependency on frontier labs. Startups no longer necessarily need to send every request to OpenAI, Anthropic, or another proprietary model provider. They can host their own models, fine-tune them, optimize them for specific workloads, and potentially achieve much better economics at scale. They also gain more control over their infrastructure, data, and product behavior.
But this introduces another problem. Instead of being dependent on a proprietary model provider, startups can become dependent on the constant evolution of the open-weight model ecosystem.
New models are released at an extraordinary pace. A model that looks like the best choice today might be surpassed a few weeks later. Each release brings different capabilities, architectures, hardware requirements, inference characteristics, and tradeoffs. Switching models isn’t necessarily as simple as changing an API endpoint. Applications may need retesting, prompt adjustments, tool-calling validation, inference infrastructure reconfiguration, and performance optimization. If you’ve fine-tuned a model, you may need to repeat that work with a newer one.
For a startup with limited resources, this can become an enormous distraction. Instead of building a better product or acquiring customers, the engineering team ends up spending more time evaluating models, updating infrastructure, and trying to keep up with the latest developments. The company might have eliminated its dependency on a particular AI provider, only to replace it with a dependency on an ecosystem that never stops changing.
And the irony is that many of these improvements don’t necessarily create more value for customers. A new model might score better on benchmarks, reason more effectively, or generate responses faster, but that doesn’t automatically translate into a meaningfully better product. Yet the competitive pressure to adopt the latest models remains. If competitors advertise that they’re using the newest, most capable models, it’s hard to justify sticking with something older, even when the actual customer benefit is marginal.
This raises a more fundamental question: where will the value in AI applications ultimately reside?
For much of the software industry’s history, the application itself was the product. Companies built software to solve specific problems, and the complexity of building it created some defensibility. It took time to develop the features, infrastructure, and expertise needed to compete. With AI, general-purpose models are absorbing more of that complexity. Features that once required substantial engineering effort can now be implemented relatively quickly, sometimes with little more than a prompt and access to an API.
This dramatically lowers the barrier to entry, boosting innovation but making it harder for startups to build lasting competitive advantages. If a feature can be built in a weekend, it probably isn’t much of a moat. If customers use your product mainly because it makes a particular AI model easier to access, expect that advantage to disappear as the models and their interfaces improve.
I suspect many AI startups confuse access to powerful technology with a defensible business. Being able to do something impressive with AI doesn’t necessarily mean customers will pay for it, and even if they do, it doesn’t mean they’ll continue paying once that capability becomes widely available.
The companies most likely to survive are those where AI is an important component of the product, but not the primary reason the business exists. They solve problems customers already have, fit into workflows customers already depend on, and create value that doesn’t disappear as the underlying models improve. Their advantages come from distribution, customer relationships, proprietary data, integrations, domain expertise, or the difficulty of replacing an established workflow.
In other words, the more interesting question isn’t whether OpenAI or Anthropic could build your product. Given enough resources, they probably could. The question is whether they would have a compelling reason to do so, whether they could reach your customers effectively, and whether their version would actually be good enough to replace yours.
There’s also an important distinction between using AI to build a product and making AI the product itself. A company that uses AI to dramatically reduce the cost of delivering an existing service might have a much more sustainable business than one that simply wraps a frontier model in a specialized interface. The former can benefit from AI improvements without necessarily undermining its value proposition. The latter risks becoming less relevant with every major model release.
The same applies to open-weight models. Their real advantage isn’t necessarily letting startups compete with frontier labs on model capabilities. Most startups shouldn’t be trying to do that in the first place. Their advantage is that they give companies more control over their costs, infrastructure, and dependencies. If a smaller, older model can reliably perform a specific task at a fraction of the cost, there may be no reason to replace it every time something newer becomes available.
In fact, I think one of AI's more overlooked opportunities is figuring out how to do more with less. The industry has become obsessed with the most capable models, the largest context windows, and the newest releases. But most business applications don’t require frontier-level intelligence for every task. They require reliability, predictable costs, reasonable performance, and integration into existing processes. A carefully designed system using smaller models might deliver better business outcomes than one that constantly chases the latest frontier capabilities.
This is where I think AI startups' economics will become particularly interesting. As models become more capable and less expensive, the cost of intelligence itself should continue to decline for many workloads. That sounds like an enormous opportunity for application developers, and in many ways it is. But it also means that intelligence becomes less scarce, and therefore less valuable as a standalone differentiator.
When everyone has access to increasingly capable models, AI capabilities stop being special. The value shifts to knowing what to do with those capabilities, reaching the right customers, accessing the right data, and delivering a complete solution people will pay for.
We’ve seen versions of this before. The internet made distribution dramatically cheaper, cloud computing made infrastructure more accessible, and open-source software reduced the cost of building applications. Each wave created enormous opportunities, but it also commoditized capabilities that previously gave companies an advantage. AI is accelerating this process to a whole new level.
What’s unusual about this gold rush is how much economic value the companies supplying the picks and shovels capture, while the companies prospecting for gold also help finance the development of tools that might eventually replace them.
I don’t think startups should avoid frontier models or insist on running everything themselves. That would be replacing one form of dependency with another kind of constraint. A better approach is to treat models as interchangeable components wherever possible, choose them based on actual product requirements rather than hype, and avoid building a business whose entire value depends on capabilities likely to become standard features.
The startups that survive this period won’t necessarily be the ones using the most advanced models or adopting every new release first. They’ll be the ones that figure out how to turn increasingly commoditized intelligence into something customers actually need, can depend on, and are willing to keep paying for.
Ultimately, I think that’s the central contradiction of the AI gold rush. The technology is becoming more powerful and accessible at an extraordinary pace, but that same progress is making it harder for many companies built around it to stay differentiated.
AI might make it easier than ever to build a product. It doesn’t necessarily make it any easier to build a business.