Opinion
The AI conversation we're not having
The companies that will define this era of AI are not the ones in the headlines. The edge belongs to the founders and investors willing to look where building is still hard.
A dozen companies own the entire conversation about artificial intelligence: OpenAI, Anthropic, xAI, Cursor, Perplexity, and a handful of others. They take the headlines, and most of the money follows them. The most interesting companies being built right now are not on that list. Many are not in the news at all.
That is not an accident, nor is it a complaint about where capital goes. It is the most useful thing to understand about investing in this moment.
Consider SpaceX. For most of its first two decades, it was treated as a moonshot serious investors were right to avoid: too hard, too capital-intensive to be worth the wait. Those were the years that mattered. The company dismissed for so long is now among the most valuable ever built, and the quiet, uncovered years were the investor's years.
There is a reason the most durable opportunities sit outside the coverage. Consensus is priced. Where attention and capital concentrate, valuations inflate, the field crowds, and the reward for being right shrinks. Coverage also follows what is legible: the press covers a product you can open, a valuation you can put in a headline. It does not cover the systems software that makes thousands of chips behave as one machine, or the data engine inside a vertical software company, even though that is where lasting value forms. The loud layer is loud because it is easy to see and easy to build, which is exactly why a foundation model can absorb it with a single feature. The quiet layer is quiet because it is hard. And hard is the moat.
Venture returns live on scarcity. AI has collapsed the cost of building a software business. Taking a company from zero to a working product is now within reach of almost anyone. Good for founders, good for the economy. But a working product is no longer scarce. Very few will be the rare generational outcome venture exists to fund. What collapsed is the middle: the validated, decently growing company with no real defensibility, the kind that closed a round a year ago and cannot today. Two things still earn venture money: technology that is hard to build, or a market position that is hard to catch. Everything in between is a good business. It is just not a venture bet.
Knowing where value forms is only half of it. The other half is how it gets captured. Most enterprise AI disappoints not because the models fall short but because no one rebuilt the work around them. Capturing value is a founder act, not a model upgrade: it takes the agency to own a still-imperfect technology and push it into the real world. This is where I am most optimistic about Israel. Startup Nation produces that kind of founder in unusual volume, and on the harder end, it has long been one of the best places to build deep technology. The instincts that built the last generation of infrastructure and security companies are well suited to the layer where this revolution is being decided.
So where is that layer? Three directions, and the list grows every quarter.
The first: the AI buildout is breaking the infrastructure it runs on. The power grid is becoming the ceiling on adoption, and the compute stack built over forty years is hitting walls in how chips connect and how they reach memory. These are the bottlenecks everything else depends on, and few teams can build them. When we led Teramount's seed in 2021, chip-to-chip photonic connectivity was the piece the market had overlooked; it became an industry standard. On the energy side, we backed Lava in 2022 to turn the data center's largest waste, its own heat, back into power, well before the grid became the constraint everyone now talks about.
The second: AI is opening markets software never touched. For thirty years software stayed in the digital world. Reasoning can now reach the physical one, where it meets sensors and machines, and the human service economy, the work that was always too messy to automate. The last industrial revolution commoditized goods; this one will commoditize services. The physical world is the last unindexed data, the one place the models do not already have the answer. It is why we backed LimitlessCNC to model machining physics in real time and build proprietary data from the factory floor, and why NoTraffic is doing the same for how cities move.
The third: the durable winner owns data that compounds. While the market races to build the next model, the lasting advantage belongs to whoever holds data that grows richer with every customer. Proprietary data alone is only a head start, and a head start depreciates the day a better model ships. A real moat is a flywheel: each customer makes the product better and the dataset deeper, and the loop strengthens as the company scales. Alice, formerly ActiveFence, understood early that a proprietary record of harmful behavior, one that deepens with every threat it sees, is what keeps a trust-and-safety product defensible long after the underlying models commoditize.
One more change is worth naming, because I had not seen it before this year. Companies are doubling revenue without adding a single person. One closed a Fortune 50 customer within six months of starting to sell. Others have grown tenfold in two years running. Revenue and headcount used to move together; they have come apart. Speed has become its own form of defensibility, and the best of the hard-technology companies are now also the fastest.
This is the best time there has ever been to build a company, and to invest in one. The edge is no longer being in AI. Everyone is in AI. The edge is judgment about where building is still hard, and where a lead is still hard to catch. That is the work. And most of it is happening off the front page.
Lotan Levkowitz is a Co-Founder and Managing Partner at Grove Ventures, where he has spent over a decade investing at the earliest stages of company building, with a focus on AI, infrastructure software, and deep technology.
