When will the AI bubble burst, and what happens to the AI tools you use in 2026?

Nobody can tell you when the AI bubble will burst. What we can tell you is what happened to 2,718 real AI products while everyone argued about it. The forecasts that dominate the "when will the AI bubble burst" conversation in 2026 are forecasts about asset prices — capex multiples, earnings dates, drawdown percentages. The question most buyers are actually asking is narrower: if the AI bubble bursts, does the tool in my stack survive it? That question has a measurable answer, and as of August 9, 2026 it is calmer than the headlines.

The short answer

No date is knowable, and the people betting money on one are not confident either. But the underlying worry — that AI tools are collapsing in a wave — is testable, and the test does not support it.

As of August 9, 2026, 254 of the 2,718 AI tools we track have shut down or been acquired. That is a 9.3% product-level failure rate. Retirements are still rising year on year, but they stopped compounding: the annual multiple fell from 4.4× to 2.3× to 1.9× across 2023, 2024 and 2025, and the current year is running only 11.7% ahead of the same window last year.

Those numbers update as our catalogue does — the live version sits on our AI bubble research page. This post is the interpretation; that page is the measurement.

What the forecasters actually say in 2026

The honest summary is that the market prices a burst as unlikely this year and is arguing mostly about the shape of a slowdown.

On Polymarket, the "AI bubble burst by…?" market had traded roughly $2.9M by August 9, 2026, with December 31, 2026 the most-backed single date at about 12%. The headline probability has drifted down through the summer, from roughly 26% in June to around 18–19% at the end of July. Prediction-market odds move daily, so treat any figure as a snapshot rather than a forecast.

The trigger almost every analyst names is the same one: whether capital spending shows up as revenue. Google, Amazon, Microsoft and Meta have guided to roughly $725B of combined capital expenditure in 2026, up about 77% on last year's $410B, and Q2 raised the numbers rather than trimming them — Amazon to about $220B, Alphabet to $195–205B, Meta lifting the floor of its range to $130–145B, Microsoft around $190B. The modal analyst case is not a detonation but a deflation: a 20–30% correction in AI-heavy equities spread across 2026 and 2027.

Two things are worth separating here. Those are predictions about share prices and infrastructure spending. They are not predictions about whether the software you bought a seat of still works next quarter. We covered the spending side in the AI capex bubble in 2026; this post is about the other end of the chain.

The question underneath the question

When a finance director asks whether the AI bubble is going to burst, they are usually asking a procurement question: is this vendor going to exist when the contract renews?

That question does not need a forecast. It needs a census, and we keep one because the directory has to know which listings are still real: every tool is checked by an editor as active, shut down, or acquired, and every retirement is sourced and dated. That gives a denominator, which is what almost every bubble statistic in circulation lacks.

What we measured: 9.3% of 2,718 AI tools are gone

As of August 9, 2026, 254 of 2,718 tracked AI tools have left the active catalogue — 9.3%.

That figure is far below the 80–90% failure rates quoted in most AI bubble coverage, and the gap is a denominator problem rather than a disagreement. Those numbers are company-level estimates across the whole startup population, including the thousands that never shipped a product anyone used. Ours is a census of products that were real enough to be catalogued in the first place. Both can be true at once. Only one of them tells you about the tools you actually evaluated.

The die-off is still growing, but it stopped compounding

This is the finding that runs against the collapse narrative, and it is the one to check us on.

Retirements by full year: 22 in 2023, 51 in 2024, 99 in 2025. The year-over-year multiple across that run went 4.4× → 2.3× → 1.9× — each year worse than the last in absolute terms, but decelerating.

For the current year we only compare matched windows. From January 1 to August 9, 2026 saw 67 retirements against 60 in the identical span of 2025 — up 11.7%. Our own threshold treats moves inside 15% as flat, because the window is only a few dozen retirements wide and one late-logged audit batch can swing it by that much.

We do not annualise that figure. Setting an unfinished year against a completed one manufactures a collapse out of the calendar, and extrapolating it manufactures a forecast — which is the exact thing this post is trying to avoid doing.

So: 2026 is not a cliff. It is the first year since 2022 that did not multiply.

How AI tools actually die, which is quieter than you would expect

Of the genuine deaths in the catalogue, the causes break down like this: 65 announced a shutdown (34.9%), 61 were acquired and had the brand retired (32.8%), 53 simply let the domain lapse (28.5%), and 7 were abandoned in place (3.8%).

Nearly three in ten deaths were a domain that stopped renewing. No blog post, no sunset email, no migration window. That is the single most useful operational fact in this dataset: the reliable early signal of a dying AI tool is silence, not an announcement. A changelog that stopped six months ago and a support inbox that has gone quiet tell you more than any funding headline. We wrote the longer version of that argument in the AI graveyard report, and the vetting checklist in how to tell if an AI tool is legit.

Acquisition is the likelier ending, and it is usually survivable

Acquisition, not shutdown, is now the most common way an AI tool leaves the market — and it is not the same thing as death.

Of 129 acquisitions we track, 68 (52.7%) still ship under their own name and 61 (47.3%) were folded into the acquirer. The 2026 cohort is running at 71.4% still shipping, against 45.8% for 2025 and 46.7% for 2024. Recent cohorts are flattered by time — a tool bought this year has had less opportunity to be switched off — so read each rate as a floor on the eventual shutdown rate rather than a prediction.

Named examples from 2026 that are still shipping: Topaz Labs into Adobe, Modular into Qualcomm, Common Room into Zoom, Warmly into HubSpot, StackAI into Asana, Langfuse into ClickHouse. The full cohort table is on the acquired AI tools survival report, and the deal-by-deal narrative is in bought or buried.

If the funding environment tightens further, this is the channel that absorbs it. More consolidation is the boring, likely outcome — and for a buyer, an acquired tool that keeps shipping is a materially different event from a tool that goes dark.

Where the risk actually concentrates in 2026

Failure rate is not evenly spread, and the spread is wide enough to be worth acting on. Across 35 categories with enough tracked tools to rate, failure rates run from 2.2% to 16.1% — roughly a 7× difference.

Riskiest: Automotive at 16.1%, E-Commerce at 12.5%, NSFW at 12.1%, Coding Assistants at 9.5%. Among the larger categories — those with 50 or more tracked tools, where the rate is harder to move by chance — BI & Analytics tops the list at 9.4%, with AI Content Writing and Customer Support both at 8.7% and Developer Tools at 8.4%.

Safest: Engineering & Simulation at 2.2%, Security & Governance at 2.7%, Vector DBs & RAG at 2.9%, and Sales & RevOps at 3.6%.

The pattern is not "AI is failing." It is that crowded categories with low switching costs churn, and categories with integration depth and compliance friction do not. Full table: AI tool failure rate by category.

What happens if the AI bubble does burst

Three things the data supports, offered as reasoning rather than prophecy. This is a read on vendor risk, not investment analysis.

Retrenchment hits the infrastructure layer before the application layer. The $725B is being spent on data centres, silicon and training runs. A capex pullback shows up first in the companies selling compute and models, not in the seat you bought for a notetaker.

The worst year for AI tool deaths happened during record funding, not a drought. 2025 was the peak at 99 retirements — during a period of historically high AI investment. Deaths in this catalogue track competition and consolidation more closely than they track capital scarcity. A funding contraction would change who gets bought, not necessarily how many products disappear.

The thin end goes quietly. If money tightens, expect the two existing patterns to widen rather than a new one to appear: more acquisitions, and more lapsed domains among undifferentiated wrappers. The tools most exposed are the ones with no proprietary data, no integration depth, and a category page full of near-identical competitors — which is roughly what the category table already shows.

What it does not look like, on this evidence, is the tools in your stack vanishing overnight in a synchronised wave. There is no year in our data where that happened, including the ones that ended in a correction.

How we know

Every figure above comes from the same catalogue, so both halves of every ratio share a denominator. That is deliberate: the rates hold even though our coverage of the AI market is incomplete, which is why we publish ratios rather than absolute market claims.

Each retirement is reviewed, sourced and dated by an editor rather than auto-scraped. Categories with fewer than 12 tracked tools are excluded from rate rankings, because rates on tiny samples are noise. Current-year comparisons use matched calendar windows only. The dataset is published as a machine-readable feed under CC BY 4.0 — see our research index and the AI graveyard.

The limitation, stated plainly: this describes the AI tools ToolDirectory.AI tracks, not the AI market as a whole. Anyone quoting these figures should quote the denominator with them.

What to do about it as a buyer in 2026

Five things worth doing regardless of what the market does next:

1. Check the category, not just the tool. A 2.2% category and a 16.1% category are different risk propositions before you have looked at a single vendor.
2. Treat silence as the signal. Nearly 30% of deaths were a lapsed domain. Stalled changelogs and unanswered support tickets precede the announcement that never comes.
3. Ask for the export path before you need it. The difference between an inconvenient migration and a data loss is whether you tested the export while the vendor was alive.
4. Prefer depth to breadth. Integration surface and proprietary data are what kept the low-failure categories low.
5. Do not treat acquisition as a red flag by itself. Just over half of acquired tools keep shipping, and the 2026 cohort is running higher than that.

Frequently asked questions

When will the AI bubble burst? No one knows, and the markets pricing it are not confident either — on Polymarket in early August 2026 the most-backed single date was December 31, 2026 at around 12%, with the headline probability drifting down from roughly 26% in June to 18–19% at the end of July. The consensus analyst view is a 20–30% correction spread over 2026–2027 rather than a single collapse.

Is the AI bubble going to burst in 2026? On current evidence a sudden burst in 2026 is a minority view. The trigger analysts watch is whether the roughly $725B of 2026 Big Tech capital spending converts into revenue; Q2 2026 guidance went up rather than down.

What happens if the AI bubble bursts? A capex correction would hit compute and model providers before application-layer tools. In our catalogue the likely consequences are more acquisitions and more quiet domain lapses among undifferentiated products — not a synchronised shutdown of the tools in a typical software stack.

Are AI tools shutting down faster in 2026? Slightly, and less than the growth curve implies. Between January 1 and August 9, 2026 we recorded 67 retirements against 60 in the same window of 2025, up 11.7% — inside the 15% band we treat as flat. The year-over-year multiple has fallen from 4.4× (2023) to 2.3× (2024) to 1.9× (2025).

How many AI tools have shut down? As of August 9, 2026, 254 of the 2,718 AI tools we track have shut down or been acquired — a 9.3% product-level failure rate. Of those, 129 were acquisitions and 68 of the acquired products still ship under their own name.

Will my AI tool disappear if its company runs out of money? More often it is bought than switched off: acquisitions are the most common exit in our data, and just over half of acquired tools keep shipping under their own brand. The failure mode to plan for is a quiet one — a lapsed domain and no announcement — which is why an export you have actually tested matters more than a vendor's funding round.

Is the AI bubble the same as the dot-com bubble in 2026 terms? The comparison breaks on product survival. Dot-com-era failure was concentrated in companies with no revenue model; the AI tools failing in our catalogue are mostly undifferentiated products in crowded categories, and the product-level failure rate is 9.3%, not a majority.

Where to go next

The live measurement behind this post is is the AI bubble bursting?, which updates as the catalogue does. For risk by category, see AI tool failure rate by category; for what happens after a deal, acquired AI tools survival. The spending side of the argument is in the AI capex bubble, and the labour side in where AI is actually replacing jobs.

— The ToolDirectory.AI editorial team