Vibe AI was an AI friend. From March 2025 to February 2026, I built it, ran it, watched it work in week one, watched it stop working by week three, and then watched the unit economics confirm what week three had hinted at. I shut it down in February 2026, the third shutdown in three years. This is the postmortem.

The category was the deadliest AI vertical of 2025 (TechStartups, December 2025). I wasn't surprised by the company; I was surprised by how cleanly the failure mode showed up in the numbers. This post is about the failure mode, not the product. The product is gone. The lesson is portable.

What Vibe AI was

Vibe AI was a chat surface designed for ongoing relationship,not for tasks, not for productivity, not for queries. The user named the AI, gave it a personality outline, and started a conversation. The AI remembered. It checked in. It asked the right follow-ups. It was warm where ChatGPT was professional, and patient where a real friend was busy.

The thesis: people are lonelier than they admit, and a thoughtful AI companion is more useful than another productivity tool. That thesis is partly right. Some part of it is durably right. The product is dead because being right about the demand side does not save you from being wrong about the cost side.

What worked in week one

Three things were unambiguously real in the early data.

First, the right user found the product. People who downloaded Vibe AI in week one used it heavily. Long messages. Multiple sessions per day. Real disclosure. The kind of usage that says the surface fits the demand.

Second, the personality system worked. Users named their AI, calibrated the tone, returned to the same conversation thread the next morning. The remembering felt like remembering, not like context-stuffing. The illusion held.

Third, week-one retention was strong. Users came back day two, day three, day four. The novelty curve was intact for at least the first week,and that is a stronger signal than most consumer apps get.

Where the curves crossed

Then week three happened. The novelty curve crashed into the cost curve.

Two things were happening at once. The first was retention: the kind of conversation that produces strong week-one engagement (long, emotionally invested) is exactly the kind of conversation that decays the fastest. Users who came in for "I need someone to talk to today" did not necessarily need that the same way three weeks later. Some converted to lighter, less compute-heavy use; many stopped entirely.

The second was cost. Long contextual conversations are not free. Each turn pulls more context, the model does more reasoning, the per-conversation cost is multiples of a one-shot ChatGPT prompt. At scale and at price points the audience would tolerate, the per-user economics did not work.

Retention drops, cost climbs, revenue lags,the standard companion-app shape Per-user $0 Wk 1 Wk 2 Wk 3 Wk 4+ Cost stays high Retention drops Revenue never catches up Retention (engagement) Compute cost / user Revenue / user Stylised. Vibe AI operator data, March 2025-Feb 2026. Pattern is general across companion apps.
The companion-app shape: cost stays high, retention drops, revenue lags. Growth deepens losses.

The math: if the per-active-user cost-of-goods exceeds the price the audience will pay, more users equals more loss. Subsidising acquisition only delays the same outcome. There is no scale-to-margin path from negative unit economics; the curve does not bend on its own.

Why companion apps lose money structurally

Vibe AI was not unique. The AI companion category was the deadliest AI vertical in 2025 (TechStartups, December 2025). Every product in this space runs into the same mismatch: emotional engagement is compute-heavy, and emotional-engagement audiences are not subscription-rich.

The macro version of this problem is visible at the largest scale of generative AI. Industry coverage of generative-AI economics has consistently documented compute costs running multiples of revenue at consumer price points across the largest products. Different category, different scale, same fundamental shape: compute cost outpaces willingness-to-pay for an entire class of products.

The aggregate pattern shows up in the structural failure data. CB Insights' 2026 corpus of 431 failed startups: 70% cite "ran out of capital" as the surface cause; the underlying drivers were poor product-market fit (43%), bad timing (29%), and unsustainable unit economics (19%) (CB Insights, "Why Startups Fail", 2026). For Vibe AI, the underlying driver was unit economics specifically. The audience was real. The cost-vs-price gap was structural.

The shutdown decision

The decision moment was a spreadsheet. By January 2026 I had run every reasonable price experiment,higher subscription, lower subscription, freemium with a paid tier, paid tier with a usage cap. None of them produced a margin-positive cohort across a 90-day window. The retention curve was the same shape; the cost curve was the same shape; the revenue curve never crossed cost.

I sat with the spreadsheet for a week and tried to find a model I had not tried. There were variations,pivot to "AI coach" with explicit goals, pivot to per-conversation pricing, pivot to a B2B "employee wellness" play. Each pivot was a different startup, not a path through to Vibe AI's success. The honest read was that Vibe AI as designed was not going to be margin-positive at any plausible scale. I shut it down in February 2026.

The hindsight read: I should have set the cost-per-user kill threshold at month three, not month nine. The signal was clear by week six. The "maybe scale fixes it" reasoning is the standard founder error in negative-unit-economics businesses, and I made it.

The lesson,and how Gravity uses it

The lesson tag for Vibe AI's failure: sustainable-margins test failed. The product was loved. It was not a business. Two distinct things.

For Gravity, the Vibe AI lesson surfaces in the pricing model. Capability-based pricing,charging for the agent's ongoing job, not per task,keeps the per-active-agent margin computable in advance. How Gravity works: the agent runs a recurring task; the cost is bounded by the task; the price reflects the value of the outcome. That structure makes margin a design constraint, not a hope.

The full synthesis across all three failures is in three startups, three shutdowns. MindWave failed the 10x test (postmortem). Super AI failed the scaling test (postmortem). Vibe AI failed the margin test. Three distinct check failures; the absence of any one was sufficient.

Frequently asked questions

What was Vibe AI?

Vibe AI was an AI companion product I built from March 2025 to February 2026,an "AI friend" designed for daily conversation, emotional support, and check-ins. Engagement in week one was real. The product worked. The unit economics did not, and I shut it down in February 2026.

Why do AI companion apps fail?

AI companion apps face a structural mismatch between cost and price. Long, contextual conversations are compute-expensive at scale, and the audience that wants an AI friend is largely unwilling to pay subscription prices that match the cost. The retention curve and the cost-per-active-user curve cross before the paid model converges. The category lost more startups in 2025 than any other AI vertical.

How long did Vibe AI run before shutting down?

Vibe AI ran from March 2025 to February 2026,almost exactly 12 months. The product was loved in week one. It was forgotten by week three for most users. The shutdown decision came after unit economics confirmed there was no price the audience would accept that covered compute cost.

Why are AI compute costs so high?

Generative-AI inference is computationally expensive at scale. Long-context conversations multiply the cost per turn, and consumer products at low subscription prices cannot recover that cost. Industry coverage of generative-AI economics has consistently documented compute costs running multiples of revenue across many large products at consumer price points,same shape of negative per-user economics that killed Vibe AI, just at much larger scale.

What is the lesson from Vibe AI's shutdown?

Engagement is not a business. A product can be loved and still be a money-losing operation if the per-user margin is negative. When per-user economics are negative, growth deepens losses; the math does not fix itself with scale. The lesson maps to one of three checks: sustainable-margins test failed.

Three takeaways before you close this tab

Sources