Last week two stories landed within 48 hours of each other, and almost nobody put them side by side.
On November 13, LangChain raised $300 million at a $3 billion valuation, led by Alphabet's CapitalG. That more than doubles the company's valuation, and it comes with a number that should make every AI founder nervous: 7,000 organizations now use LangSmith, its agent observability and management platform.
On November 14, TechCrunch published a running list they are calling the AI graveyard. Character.ai filed for bankruptcy in October. Ollie, a family AI assistant that launched in June, shut down on November 6 saying it could not keep up with Muse and Instinct. Relay, the Zapier alternative built on AI agents, closed in August. ChatGPT Atlas never made it to a birthday. Sora shut down in March. According to S&P Global Market Intelligence, roughly 42% of AI initiatives get abandoned by the companies that started them.
So the infrastructure layer is compounding in value while the product layer is being buried. That is not a coincidence, and it is not bad news for everyone.
The thing LangChain is actually selling
Read the coverage carefully and you notice LangChain is not being funded for the framework. The framework was never a business. It was a distribution channel.
What CapitalG is paying for is the boring layer: tracing, evaluation, prompt versioning, cost tracking, and the audit trail that an enterprise needs before it lets an autonomous system touch anything that matters. LangSmith is where a company goes when its agent does something strange on a Tuesday and someone in compliance asks why.
That is the real product in 2026. Not intelligence. Accountability.
Look at the graveyard entries and the pattern is brutal and consistent. Almost every one of them sold a behavior: chat with a character, automate your inbox, talk to your family calendar, browse the web for you.
In my own work building agents, the failure mode I keep hitting is never "the model was too dumb." It is that I could not see what happened. A tool call returned something slightly wrong, the agent routed around it, and by the time the output looked off I had forty minutes of reasoning to reconstruct by hand.
Every engineer I know who ships agents has built some version of this trace viewer themselves. Some of them built it three times because they switched frameworks. That is not an observability problem, that is a tax on everybody who insists on rolling their own.
Why the products die and the plumbing survives
The graveyard list reads like a masterclass in the same mistake.
Character.ai had enormous engagement and died under legal and regulatory weight. That is not a model problem, that is a governance problem: nobody could prove what the system would and would not do.
Relay got crushed because OpenAI and Google folded workflow automation into their own surfaces. When your product is a thin agent wrapped around someone else's model, and that someone else ships your feature as a checkbox in a chat window, you have no moat. You have a head start.
Ollie's co-founder put it plainly: "We got close, but ultimately couldn't keep up." They launched in June and were gone in November. Five months.
Meanwhile Bharat Semi, an Indian chip startup you have probably never heard of, quietly closed a $24 million seed round this week with design wins including General Atomics. It is an AI sensor company. Nobody is going to write a thinkpiece about it. It also has a physical product that cannot be cloned by a feature flag.
The pattern is not "consumer AI is dead" or "agents were hype." The pattern is that the value is migrating to the layer where switching costs are real: hardware, data, and the unglamorous middleware that makes autonomous behavior auditable.
What this means if you are building an agent right now
I hate posts that end in a list of predictions, so here is the operational version instead. Three things I would actually do differently this quarter.
First, instrument before you add capability. Before I add a second tool or a second agent, I want a trace of the first one. Full prompt, full tool I/O, latency, token spend, and the decision that followed. If I cannot answer "why did it do that" in under a minute, I am not building a product, I am building a slot machine.
Second, assume your feature will be absorbed. Whatever your agent does, one of the big labs will eventually bundle a worse version of it into a chat box that a billion people already have open. So build the part they cannot bundle: your data, your integration depth, your compliance posture, your customers' trust. Relay did not lose to a better agent. It lost to distribution.
Third, treat evals as the product surface, not the internals. The teams I have seen survive scrutiny are the ones who can hand a skeptical security reviewer a report that says: here are the 400 cases we test, here is our failure rate, here is what happens when a tool times out. That artifact closes deals. A demo does not.
The uncomfortable part
There is a version of this trend that is genuinely bleak. If the money all flows to observability vendors and chip makers, and the application layer keeps dying, then a lot of very smart people are going to spend two years building beautiful things that get deprecated by a model release.
But I do not think that is the lesson. Look at what Nvidia is doing right now: privately telling Congress that chip-level tracking to fight smuggling is technically feasible for the Chip Security Act, while publicly the industry questions whether it is practical. Even the biggest company in the space is discovering that capability is the easy half. Verification is the hard half.
That is the whole story of this year, really. Everyone got very good at making models do things. Almost nobody got good at proving what they did, catching them when they are wrong, or shutting them off cleanly when they misbehave.
The graveyard is full of teams that solved the first problem and treated the second one as a stretch goal.
The $300 million is going to whoever sells the second one.
Sources: TechCrunch on LangChain's Series C and the AI graveyard; MIT Technology Review on Nvidia's chip tracking; TechCrunch on Bharat Semi.
Top comments (0)