Have you noticed how every second pitch now seems to include the letters A and I, as if those two characters could carry a whole business on their own? I have. Sitting through founder calls over the past year has felt a bit like watching a fashion cycle speed up and then snag on a zipper. The clothes still look expensive. The stitching, though, is starting to show. Capital is still plentiful in pockets of the market, yet the people writing checks are asking a harder question: does this thing actually make money, save money, or just sound clever in a slide deck?
Why An Early-Stage AI Shakeout Is Coming Into View
The boom did not appear out of thin air. Chipmakers beyond the usual household name have been bid up because data centers need power, racks, networking gear, and a lot of patient construction. That buildout is real. Trucks still roll to sites. Transformers still take months to arrive. None of that vanishes because a few seed rounds get marked down. What can vanish, and rather quickly, is the benefit of the doubt given to early-stage software shops that sprinkled a model on top of an old workflow and called it a category.
In my experience, markets do not punish ambition first. They punish sloppy unit economics dressed up as destiny. That is the mood shift I keep hearing from people who allocate venture money. They are not anti-technology. They are anti-fog. They want to know where the tech creates genuine value and where it is simply a feature wearing a business costume.
We are likely to see a shakeout as investors become much more demanding about where the technology creates genuine value and where it is simply a feature dressed up as a business.
– Founding partner at a European venture firm
That line stuck with me because it is blunt without being theatrical. Capital, this person argued, will get far more selective over the next six to twelve months. Not frozen. Selective. There is a difference, and founders who miss it will burn months chasing rooms that have already moved on.
The Buildout Is Not The Same Thing As The Business
Here is the split that keeps getting blurred in casual conversation. On one side you have infrastructure: silicon, cooling, power purchase agreements, networking, and the unglamorous work of keeping clusters online. On the other side you have applications that are supposed to sit on top of that stack and produce productivity, revenue, and eventually cash flow. The first side can look busy for years. The second side has to prove it is more than a demo.
Wealth managers I speak with keep returning to the same checkpoint. After the factories and the server halls, does the end user actually generate enough output to justify the bill? If the answer is fuzzy, the valuation should be fuzzy too. If the answer is crisp, the multiple can stay rich a while longer. Simple in theory. Messy in practice, because growth figures have been so large that they hide the quality of the growth.
The next question is whether applications and end users generate enough productivity, revenues and cash flow to justify that investment.
– Chief executive at a wealth management firm
I like that framing because it refuses the usual binary of “bubble” versus “revolution.” You can believe the technology will rearrange large parts of the economy and still think plenty of pitch decks do not deserve an extraordinary price. Those two thoughts can live in the same head. They should.
Where Valuations Start To Look Stretched
Valuation is not a morality play. It is a claim about future cash, discounted with some guess about risk. When that claim leans on narrative more than invoices, the room gets nervous. Perhaps the most interesting aspect is how quickly “AI-native” became a default label rather than a description of architecture. Labels are cheap. Switching costs are not.
Investors who still want exposure are being told, quietly, to watch productivity growth inside the companies themselves. Not vanity metrics. Not weekly active this-or-that. Can a mid-market customer fire a painful process, keep the lights on, and still show a line of savings that survives an audit? If yes, the story has a spine. If not, you are underwriting a feature request.
Does that sound harsh? Maybe. Markets get harsh when money has options. And right now money has options: later-stage infrastructure names, public semiconductor suppliers, power-related industrials, and a thinner set of application businesses with actual contracts. Early-stage software that cannot show a path to cash will feel that competition for attention.
Features Versus Expensive Problems
One venture partner put it in a way I wish more founders would tape above their laptops. The winners, in this view, will be companies using models to solve expensive and highly complicated problems. Not mildly annoying ones. Expensive ones. The kind that already have a budget owner, a procurement cycle, and a reason to renew.
A concrete example that circulated in recent conversations involves store analytics. Cameras plus on-device models watch how shoppers move through a physical space. Retailers then get data they can act on: staffing, layout, shrink, dwell time. You can argue about privacy and implementation. Fine. The point is narrower. The buyer already spends money on the problem. The product is not asking a grocer to invent a new line item from scratch. That is a different sales motion from a chatbot bolted onto a help desk that already had macros.
I’ve found that the “expensive problem” test saves time. If the pain is cheap, customers will tolerate a clunky workaround forever. If the pain is dear, they will sit through a security review. Early-stage teams that cannot name the budget they are attacking tend to raise on vibes. Vibes have a half-life.
What Pickier Capital Actually Looks Like In Practice
Selective does not mean cruel. It means longer diligence, smaller clubs, more milestone-based closings, and fewer “we loved the demo” emails at midnight. Founders should expect questions that used to arrive in Series B to show up at seed. Who pays? How often? What happens if the model vendor raises prices? What is the human-in-the-loop cost that never made it onto the slide?
- Proof that a customer would still buy if the novelty faded next quarter
- Gross margin that survives real inference costs, not slide-deck costs
- A sales cycle that does not depend on a single champion leaving the company
- Data rights that do not collapse the moment an enterprise lawyer wakes up
- A roadmap that is not just “wait for the next foundation model”
None of those bullets are romantic. That is the point. Romance is what the last cycle sold. This cycle, at least at the early stage, is drifting back toward accounting.
Will some funds keep spraying seed checks to keep their logos on announcement posts? Of course. Brand is a strategy for some partnerships. But the center of gravity is shifting toward underwriting. If you have sat in both kinds of meetings, you can feel the temperature change in the first ten minutes.
Capex Without A Ceiling Makes People Twitchy
Public markets have their own version of this argument. Firms keep lifting capital expenditure with no tidy end date. Growth prints look spectacular. Sustainability questions follow like a shadow. That shadow is useful. It keeps analysts from treating every upward revision as a permanent law of physics.
Worries of a bubble persist for a reason. They persist because the spend is front-loaded and the payback is distributed across customers who have not all signed up yet. That is not automatically a disaster. Railroads looked insane before they looked obvious. It is, however, a reason to separate the shovel-makers from the souvenir stands.
Even if froth pops, some of the metal in the ground will still be used. An asset manager I read recently made that case without the usual chest-thumping. Damage, in that telling, lands hardest on people who bought the fear-of-missing-out layer. The halls, the power contracts, the trained operators: those can outlive a bad vintage of software startups.
If a bubble does pop it will mostly damage those that invested in the FOMO-led froth. But the infrastructure that has been created will still be used and prove game-changing for many companies.
– Founding partner at an Asia-focused asset manager
That is the grown-up version of optimism. Not “numbers only go up.” More like “the wiring will still be there when the slogans quiet down.”
How Public Markets And Private Rounds Talk Past Each Other
Listed chip suppliers live on order books, utilization, and guidance. Seed-stage application companies live on narrative velocity and a handful of design partners. Those two languages share vocabulary and almost no grammar. When commentators mash them together, you get headlines that treat every private markdown as a verdict on the whole stack. That is sloppy.
A shakeout in early-stage software can run in parallel with a multi-year infrastructure cycle. Both can be true on the same Tuesday. I keep repeating that because it is the distinction most dinner conversations erase after the second glass of wine.
| Layer | What Investors Now Want | Where Risk Hides |
| Infrastructure | Visible demand, power access, delivery timelines | Capex overshoot, delayed sites |
| Models and platforms | Usage that turns into paid seats | Price wars, commodity feel |
| Applications | Cash flow path, sticky workflows | Feature-only products |
| Services wrappers | Repeatable delivery, not heroics | Labor costs that never fall |
Use the table as a map, not a commandment. Real companies straddle rows. The useful habit is asking which row is doing the work in the valuation.
A Founder’s Checklist Before The Window Narrows
If you are raising in the next two quarters, skip the poetry and bring a short, slightly boring packet. Boring travels well right now.
- Write the expensive problem in one sentence a CFO would recognize.
- Show inference cost per completed job, not per token fantasy.
- Name the buyer, the budget, and the renewal trigger.
- Explain what still works if the underlying model gets 30 percent cheaper for everyone.
- Admit the part of the product that is still a feature and the part that is a system.
That last item is the one teams dodge. Honesty here is strangely attractive. Investors are tired of being told the roadmap will invent moats later. Later is a crowded neighborhood.
And if you cannot complete the list without squinting? That is information. Raise less, extend runway, or pair with a customer who will pay for a pilot that looks like work rather than a press release.
What Allocators Should Actually Watch
For people sitting on the other side of the table, the job is not to swear off the theme. The job is to stop paying novelty prices for incremental tools. Watch cohort retention after the pilot team leaves. Watch whether support tickets shrink or explode. Watch whether the product needs a new model drop every six weeks to stay interesting. Products that require constant fireworks are entertainment. Entertainment is a tough enterprise category.
I also look at hiring patterns. Teams that staff mostly prompt wranglers and marketers, with a thin layer of domain people, often discover too late that the customer’s world is weirder than the demo. Domain density is not a vibe. It shows up in implementation notes.
A rough filter I keep on a sticky note: Can it survive a cheaper model? Can it survive a skeptical finance lead? Can it survive the champion’s vacation?
Three yeses and I lean in. Two yeses and I slow down. One yes and I smile, take notes, and leave the term sheet in the bag.
Productivity Is The Unfashionable Word That Matters
Productivity does not trend on social feeds. It shows up in hours, error rates, and the unsexy line called cost of goods. That is why some commentators keep dragging the conversation back to output per worker and cash conversion. They are not trying to kill the party. They are trying to measure whether the party paid the caterer.
When productivity is real, valuations can look high and still be defensible. When productivity is a hope, even a modest multiple is a stretch. The coming period of pickier funding is, at heart, an argument about that gap.
Will every company with a model in the stack fail? Obviously not. Some will become the quiet plumbing of industries that do not issue flashy keynotes. Those are the ones I want in a portfolio when the headlines cool. Plumbing does not go viral. Plumbing gets paid.
The Human Texture Of A Cooling Market
Shakeouts are not only spreadsheets. They are delayed offer letters, co-founders who suddenly remember they like product design more than fundraising, and board meetings that last twenty minutes because there is nothing new to celebrate. That part rarely makes it into market notes. It should. Culture inside startups changes when the default answer stops being “raise more.”
I’ve sat with teams that treated a down round as a moral failure. It is not. It is a price. Prices move. The failure is pretending the old price was a law. If that sounds cold, consider the alternative: burning eighteen months to defend a number that only existed because last year’s comparables were drunk.
There is a kinder version of selectivity too. Funds that slow down often spend more time on the companies they already own. That can be a gift if the advice is practical. It can be a trap if the advice is just anxiety with a calendar invite. Founders should learn the difference fast.
Six To Twelve Months Is Not A Prophecy
Timelines in this business are guesses wearing a tie. Still, the six-to-twelve-month window keeps coming up, and it matches what I see on calendars. Partnership meetings that used to rubber-stamp “AI” now split the agenda into infrastructure, applied tools, and everything else. Everything else is getting shorter slots.
Does that mean seed volume collapses to zero? Unlikely. It means the distribution of checks gets lumpier. A few teams with painful, well-owned problems will still raise quickly. A long tail of “we added a copilot” products will recycle slides until the language changes again.
Language always changes. Remember when every company was a platform? Then every company was a marketplace. Then every company was an API. The pattern is old. The compute is new. Do not confuse the two.
A Note On Bubbles That Is Not A Pep Talk
People love to ask whether this is a bubble as if the answer were a light switch. History is sloppier. You can have wasted equity in one layer and indispensable assets in another. You can have brilliant research and terrible go-to-market. You can have a transformative tool that still should not trade at a fantasy multiple in year two.
If froth unwinds, the people who bought only the slogan will feel it first. The racks will still hum. The better application teams will hire some of the displaced talent and ship quieter products. That is not a fairy tale. It is how prior computing waves digested their excess.
My own bias, since we are being adults: I would rather own a dull workflow that prints invoices than a dazzling demo that prints screenshots. Dull compounds. Dazzling needs an audience.
Putting The Pieces On One Desk
So where does that leave a reader who is neither a partner nor a founder, just someone trying to make sense of the noise? Start with the split. Infrastructure spend can stay loud while early-stage software gets a haircut. Treat “AI company” as a weak description. Ask what problem is expensive, who already pays for it, and whether cash shows up before the next model drop.
Then watch behavior, not slogans. Longer diligence. Fewer spray rounds. More talk of margins. More skepticism toward feature-only stories. That behavior is the shakeout, even before the first ugly cap table hits a blog.
And keep a little humility. Some of the teams that look ornamental today will stumble into a wedge that becomes a system. Some of the teams that look rigorous will still miss the market. Selectivity reduces waste. It does not grant omniscience.
If you remember only one thing, make it this. The letters on the slide are not the business. The business is the expensive problem, the buyer, the cost to run the model, and the cash that remains when the applause stops. Capital is beginning to act like it knows that. About time, if you ask me.
The next year will sort a lot of costumes from companies. That process will feel uneven and occasionally unfair. It usually does. The consolation is ordinary and sturdy: useful infrastructure tends to stay useful, and products that earn their keep tend to find a price. Everything else can take a number and wait in the hall.
