The Great AI Capital Misallocation

There has never been more capital chasing so few names

The post-Covid boom had a problem: too much capital funded too many companies that shouldn’t have raised. Today we have the opposite problem. The flood of capital into AI is leaving many companies that should get funded without it.

We see it in our own portfolio. In the past year, every company that has raised a subsequent round did so less than three months after its pre-seed, and at eye-popping valuations. Nearly all were tied to the AI infrastructure trade or sell tokens or services. There is increasingly nothing in between.

Three things are driving this.

1. Venture has become a game of capital absorption

There’s an old joke in venture: come for the carry, stay for the fees. And the fees have never been better! But when you manage tens of billions, the question shifts from what multiple a company can return to how much capital it can absorb. Data centers, chips, inference sold below cost, data suppliers to the frontier labs. Each can consume billions, which is exactly what a giant fund needs.

That logic trickles down the venture ecosystem. Series B investors chase the companies they think can raise mega-rounds quickly. Series A investors chase the companies they think Series B investors will chase. You’re an obvious winner almost immediately, or you’re invisible. This is all exacerbated by GPs fear of becoming irrelevant if they’re not in the right names or if the portfolio has not exited when the music stops.

2. AI has broken the growth benchmarks

The concentration at the top creates another problem further down the stack: AI has completely reset investors expectations around fast growth. The infrastructure buildout has produced companies going from zero to hundreds of millions in revenue faster than anything we’ve seen since the telecom buildout of the late 1990s.

Not all revenue is equal, though. Some of these businesses sell to the same five frontier labs. Many have negative margins once inference is counted. So many are services businesses valued like software, or count pilot budgets as recurring revenue. Many would not survive public-market scrutiny, the same way e-commerce brands valued at 10x revenue in venture were repriced near 1x once they went public. Still their growth becomes the benchmark for everyone else.

As a result, an AI-native software company going $0 → $2M → $10MM, with passionate customers and solid margins, gets tossed around. By any historical software standard that’s a top-decile trajectory for a Series A. VCs would have killed for it three years ago.

3. The winner-take-all fallacy

But there is a deeper assumption underneath all of this: that AI will be winner-take-all. If a handful of companies capture most of the value, overpaying for the likely winners is a smaller mistake than missing them.

We’ve heard this before—with railroads, telecom, the internet, mobile and cloud. But AI comes with more creative ideas for why this time it’s different. First it was AGI, yet whether or not AGI has arrived, model competition is fiercer than ever. So today’s term is RSI, recursive self-improvement: once a model can improve itself, the leader’s advantage “compounds” (a favorite word in the AI lexicon) until everyone else is hopelessly behind.

But recursive doesn’t mean exponential. You can recursively improve your way toward an asymptote, and that’s a fair description of what we see today: more compute and capital producing better models with diminishing gains, while competitors are catching up at an ever faster pace. For RSI to crown a permanent winner, each generation would have to get better at improving the next one while making it harder for rivals to follow. There’s little evidence of either.

To be clear, there are real winner-take-most businesses: Network effects, distribution, economies of scale, and market places do create winner-take-most businesses. But AI is a technology, and technologies don’t have market structures. Businesses do. And it’s not driven by capital spend. The railroad builders spent fortunes laying track, and many of them lost money. The businesses that used the track did quite well.

This is the part I think we’re missing in AI. Hundreds of billions of dollars are being spent to make intelligence better, cheaper and more abundant. If that works, the cost of software’s most important new input keeps falling, and that should create more companies, not fewer. Venture may be financing the commoditization of the very thing it treats as scarce.

So what do you do about it

It depends on where you’re sitting.

If you’re building AI infrastructure, take the money.

This is one of the greatest capex cycles we’ve ever seen. Enjoy it but don’t confuse a great market with a permanent one. Eventually capacity catches up and prices fall. Cisco is probably the most useful cautionary tale. During the internet buildout, Cisco briefly became the most valuable company in the world. The internet thesis was completely right. The assumption that those infrastructure economics would persist was not.

So raise while capital is available, take secondaries (a lot of secondaries) when you can, and entertain any offer that prices today’s economics as permanent. Learn from Cursor who saw their market share dropping rapidly but were quick to land a spectacular deal.

If you’re building applications, stay alive.

If you have diversified customers, good retention, solid margins and 3x yoy growth, nothing is wrong with your business. The market today is pricing a different game, but venture capital appetite turns on a whim. Keep growth healthy and burn under control. Make sure you own the system of record or a critical workflow for your customers. Don’t contort the company into whatever raised $200 million this month, whether that’s selling data to the labs or deploying an army of forward-deployed engineers to generate services revenue that hides a lack of product-market fit. Treat margin as an asset, because eventually investors will care again.

Your defense against the labs is what they don’t have: deep workflow integration, proprietary customer data, and distribution into buyers who don’t take calls from model companies. And when the over-building is done (we always overbuild), your most important input will get cheaper.

At the height of the dot-com and telecom boom, Netflix offered to sell itself to Blockbuster for $50 million and was turned down. A few years later, it built a streaming business on the cheap bandwidth the boom left behind, and today it’s worth $300 billion.

You don’t need to win the infrastructure race. You need to survive it.

If you’re allocating capital, don’t make the same bet twice.

This may be the biggest implication. Most institutional portfolios already have enormous exposure to the AI infrastructure trade: Nvidia and the hyperscalers in public equities, frontier labs through growth managers, and the same consensus names through venture managers. A new fund chasing those companies is the same bet expressed through a different vehicle, now with mounting debt and stretched valuations attached.

There is another way to underwrite the AI opportunity. You can believe AI becomes one of the most important technologies of our lifetime without knowing which model wins. In fact, if intelligence becomes abundant and cheap, the number of businesses that can profit from it should explode. You can also be extremely bullish on AI and simultaneously believe that the AI buildout is mis-pricing many of the companies best positioned to take advantage of AI.

The infrastructure bet requires you to be right about who wins. The application bet requires you to be right that AI works. Those are very different bets.

The paradox

That’s what makes this moment so strange.

We are spending unprecedented sums to make intelligence cheap and abundant while underfunding the companies best positioned to use it. When an important input gets radically cheaper, history says the biggest returns go to whoever figures out what to do with it. For application (and dare I say SaaS) startups, this isn’t the end. It’s a whole new beginning.