How Small AI Startups Can Outmaneuver Incumbents Without Outspending Them

Every founder competing against a better-funded company hears the same advice: You cannot win on resources, so win on speed. It is true and also incomplete. Speed without a clear sense of where your advantages actually live just means shipping the wrong things faster.

Speed needs a bit of unpacking, because big companies can also move fast. Anthropic is a massive company and they are moving fast in one direction: model improvements, coding capabilities, agentic infrastructure. You are not going to outrun them there.

But that kind of speed comes with momentum, and momentum makes it hard to change course. A startup can pivot in a week. A company with 500 engineers and a product roadmap committed to three enterprise customers cannot. The real advantage is nimbleness: being quick on your feet, and willing to change direction when new opportunities show up.

After spending time on the Fundamental AI Research team at AWS AI Labs and two stints in a Caltech lab before co-founding OpenLens (formerly Bread), where we track what AI says about you and your brand, I think the useful question for a small team is narrower – which of the incumbent’s advantages did frontier AI just erase, and which ones did it quietly make stronger?

Which layer is yours

Frontier AI dissolved a specific set of incumbent advantages. Custom interfaces, workflow lock-in, proprietary parsers, the integration teams that stitched it all together. Those existed because moving and structuring data used to be hard. When an agent can read, reason over, and act across that data directly, the bundle stops being something a customer has to buy as a unit.

But the same commoditization that lets you stand up a product quickly lets the incumbent bolt identical capability onto a base of customers, distribution and years of data you do not have. If what you’re building is genuinely just a feature on top of an already-entrenched product, and the incumbent already has the distribution, you are going to get crushed.

So the question becomes which layer you fight to own.

We got this wrong the first time. My co-founder Aman Bhargava and I started the company to build technology that would fine-tune a model’s weights as it operated. We wrote “Prompt Baking,” and the thesis was that models would learn at inference time, the way humans do. Continuous learning. It was a genuinely beautiful idea and it was a little bit insane for a seed-stage startup. Really quite insane. We were trying to solve one of the hardest open problems in machine learning with a $5 million seed round.

When we pivoted to OpenLens, that was our bet on the product layer. We stopped competing at the model level and started building for the people who actually need AI visibility: SEO professionals, agencies and marketers. And AI is becoming an entirely new channel, so the market is growing underneath us while we build.

Five people, 500 agents

The cost and headcount required to build, test and rebuild a product have collapsed at the same time. Harvard Business Review recently called this a second great compression of entrepreneurship, where small teams enter and disrupt markets at a cost structure that forces incumbents to rethink how they operate.

An example from my world was when we shipped AI site indexing in under a week, which shows which pages an AI has indexed. Ideation to customer feedback to delivered product. In Google Search Console you can see which of your pages are indexed by Google, but you cannot tell which pages have been indexed by AI.

Each major AI platform maintains a separate index (Anthropic appears to use Brave), each crawls pages independently, and the crawl is entirely decoupled from user queries. A platform can index your page without ever surfacing it, and surface your content without a fresh crawl.

Existing tools were designed for a world where one search engine controlled the index. So in under a week, we went from idea to live product. No procurement cycle, no security review committee, no change-management process, no architectural debt to route around. Just the idea and the customer, and hundreds of agents in between.

The literal math: A team of five people can command 500 agents. You can go so, so fast if you know how to orchestrate them properly.

None of this lets a small team beat an incumbent at the incumbent’s own game, and it doesn’t need to. Frontier AI quietly redrew the board. It dissolved a specific set of advantages that used to make well-funded companies hard to touch, and left a different set fully intact. The teams that win without outspending anyone read that map honestly, then spend their limited attention only on the squares where the giant can no longer reach them.

Everything is a wrapper

One thing that surprised me going from research to product: Everyone is building on top of everyone else, and that’s fine.

People used to complain about “GPT wrappers.” Perplexity co-founder and CEO Aravind Srinivas had a good line about this. Perplexity started as a wrapper around OpenAI’s models and got dismissed for it constantly. But if you squint hard enough, OpenAI is an Nvidia wrapper. Nvidia is a TSMC wrapper, because the fabrication doesn’t happen inside Nvidia.

As Aravind put it, there are wrappers at every level of the stack. You just don’t notice because they’ve delivered enough value that you stop thinking about the infrastructure underneath. The question is whether what you build on top actually matters to someone.

For us that someone is marketing agencies. They need to know what AI is saying about their clients. Which pages are being indexed by which AI platforms. Whether their brands show up when a consumer asks Claude or ChatGPT for a recommendation, and when they do, exactly what the AI says. That is a specific, messy, real problem, and it’s the one we think about every day.