On November 10th Anthropic released Claude Sheets, Claude Whiteboards and Claude Meetings, and also made their dashboards product generally available under a new name - Claude Data. Two months before that it was Claude Docs and Claude Slides. And a day after those, Claude Code projects learned how to split work into parallel tasks, each one on its own branch, with a coordinator that hands out the work, keeps track of it and puts the result together. Call it whatever you want but that’s clearly a project management tool.

So, in about eight weeks, one AI company shipped a spreadsheet, a whiteboard, a document editor, a slide editor, a meeting assistant, a dashboard builder and a project manager. All of them living inside a chat window.

Around the same time, Bending Spoons bought Airtable for roughly 80% less than what it was valued at in 2021. Miro (yes, the whiteboard people) went for nearly 90% off. I am not saying Anthropic killed them - both deals happened before Claude Sheets or Whiteboards were a thing. But if you want a feel of what it means to be a startup right now – that’s a pretty good one.

So here’s my hot take: the survival ratio of new startups is getting abysmally low. And there are at least four reasons I can think of.

Everyone can build it

Stripe’s latest annual letter has a few numbers that are presented as “good news”: more new companies joined Stripe in 2025 than ever before, a quarter of all new Delaware corporations are now created through Stripe Atlas, and 20% of Atlas startups charged their first customer within 30 days (it was 8% in 2020).

Well, sure. Shipping something has never been cheaper. I spent a good chunk of this year having agents build tens of fully working useful projects: github.com/pavkam. It took a lot of babysitting, but I did not need a team or an investor for any of it.

The problem, as I’ve said numerous times before (and never tire of repeating it) is, that everyone else can do exactly the same thing. If you build it in a month, you probably went through the trouble of iterating over multiple versions until you got your product just right. Now, anyone can copy it in a week since they do not have to iterate - you did all the job.

In August an indie developer posted a video on X of a game where you have to find a single needle in 5,000,000 pieces of hay. It went viral and at least fourteen clones showed up on Steam, some of them overnight. Then, in October, the developers of one of those clones complained that another clone copied their code, with their own error messages still in it word for word. Steam users called it “karma”.

Other startups are not even the worst of it. Your potential customer’s internal team can build it too, and so can the AI company whose model your product is running on (see above). So the differentiation between startups is more or less zero and every category is crowded, but the number of customers on the other hand, did not grow at the same rate as the number of startups chasing them.

Who is going to bet on you?

Who will choose to take a chance on an unproven startup that can go under in 6 months?

As I’ve been looking for some solutions to problems that I needed for myself, I avoided anything unproven or unknown as the plague. I don’t need to become dependent on a “soon to be dead” product.

I am not the only one, either. A Madrona survey of 150 enterprise IT buyers found that 77% of them re-evaluate their AI vendors every six months or on a rolling basis, and that less than half of AI pilots even make it into production. With good reason, too: Carta counted 966 startup shutdowns in 2024, up from 769 in 2023, and those are only the companies using Carta for their cap tables.

Every shutdown leaves behind a bunch of customers who got burned, and they remember. So with every dead startup, the idea of betting on yet another one that will probably die soon becomes more unpalatable.

Now, to be fair, this is not much of a problem for the complete bullshit startups. If yet another chat bot dies, you lose a chat window and pick a different one tomorrow. It becomes a real mess when the startup is core tech - your database, your deployment platform, the stuff your own product is built on.

This exact thing just happened to me. A few months back we used InstaCloud for a hackathon - a Y Combinator company from the Spring 2026 batch that runs your whole backend (servers, database, auth, storage) for your coding agents. I made an account. Earlier this week they sent an email that they are shutting down: InstaCloud gets retired on January 4, 2027 and all remaining data is deleted, so the team can focus on their other product. That’s less than two months to move everything somewhere else - fine for a hackathon project, a lot less fine if your actual product runs on it.

Same story with Zeabur, a deployment platform, which is shutting down - its customers have until the end of December to move everything off it. Instant, a database, announced in August that its team is joining OpenAI and its cloud goes away at the end of August 2027. Instant says 400,000 apps were created on it. Every one of those now needs a new home, or someone willing to self-host it.

So if you are a young company, your customer looks at you again in six months, and by then somebody else does the same thing - another startup, their existing vendor, or just a chat window. Good luck getting a multi-year deal out of that, and without those deals I don’t see how you survive long enough to become the “proven” vendor.

Nobody is buying

Then there’s the old-school concept of “selling your startup to a larger company”. That’s dead too.

Why would a big company buy your startup when building the same thing in-house is cheaper than ever? Anthropic just showed everyone what a large company does with a category it likes - it builds it and adds it to the menu. Sherlocking at industrial scale!

And when big tech actually wants something from a startup, it just takes the people. Last year Google paid $2.4 billion for a non-exclusive license to Windsurf’s technology plus its CEO, a co-founder and the top researchers. Google got no stake and no control, and the rest of the 250-person company was left to carry on without them. It even has a name now – the “reverse acquihire”.

The exit numbers are not pretty either. PitchBook’s Q3 2026 Venture Monitor counted just 18 IPOs in the quarter and 992 active unicorns, worth a combined $5.7 trillion, with no obvious way to cash out. The ones that do sell, sell at the kind of discounts Airtable and Miro got. If billion-dollar companies can’t find a buyer at a sane price, what are the odds for your 6-month-old startup?

The money is not for you

According to the same PitchBook report, $515.8 billion went into US deals through Q3 2026. That’s already 44% more than all of 2021, which was the record, and 82.7% of it went to AI.

But that money goes to very few companies. Carta’s State of Private Markets for Q3 2026 shows valuations at record highs and, at the same time, the total number of rounds down 21% year over year. Carta themselves say that for a lot of founders the question now is “will a startup be able to raise capital at all?” The VC funds are no better: established firms took 88.2% of all capital raised this year, and Andreessen Horowitz alone closed $23.8 billion - more than 20% of all US venture fundraising.

So, unless you are one of the chosen few, money is in short supply. That means a seed round, if lucky, and then a Series A that never comes. Most new startups cannot get funded to survive the first rounds and that is where they die.

The obvious counter point

One counter point to all of this would be that AI makes migration easier. And that is very true. I have successfully migrated a complex backend from Azure and MSSQL to PostgreSQL on GCP, running on GKE, with the whole infrastructure-as-code thing. But that was easy because those systems were not actively used in production.

When you have some core dependency in your code that’s actually live, it’s a lot harder to replace. The code is one thing, the practices in the company that grew around those tools are another.

AAAANNNNDDD finally, the coup de grâce: if your solution is so hard to replace, then maybe it’s not a solution anyway.

Final words

I’ve seen people asking on X.com about what to work next in the current climate and I seriously have no idea. My take is that software without a physical aspect is dead, kaput, done-zo.

You either build a solution to a real-world problem (no Uber for X please!) or you are out.