$200 billion a year invested, 95% of AI projects failing: what the gap says about the AI bubble

What is the investment gap behind $200 billion a year and 95% failures?

The gap is the distance between money going in and results coming out. About $200 billion a year is invested in AI research and deployment, and 95% of generative AI projects deployed in companies fail to reach production. Put simply, for each 20 projects started, only 1 reaches production, and the other 19 consume budget, time, and attention without changing how the business operates.

A gap of that size matters for the AI bubble question because bubbles form when spending ignores results. If investment grew each year while results stayed flat, expectations would be carrying the market. The gap gives you a way to test this claim. Watch whether investment reacts to the failure rate, or keeps rising regardless. Reaction is healthy and shows buyers learning, while indifference is the pattern that earlier bubbles displayed.

Where do the $200 billion and 95% AI project figures come from?

The 95% figure comes from a 2025 MIT study of generative AI projects deployed in companies, which the EQOS Asia AI study cites alongside MIT Technology Review, Gartner, McKinsey, Vectara, and Bain data (April 2026). The $200 billion is the study’s estimate of yearly global spending on AI research and deployment. Both are estimates, and the study presents them as orders of magnitude.

Read the definition carefully. Failure here means a deployed project did not reach production, which is a different thing from a model that does not work. A pilot that never scales counts as a failure even if the technology performed. This definition explains why failure rates can be high while the underlying tools improve, and it tells you to ask what a project must achieve before it counts as a success.

Why do 95% of AI projects fail in production?

The study ranks the causes. Insufficient or biased data appears in 42% of failures, no alignment with strategy in 31%, team resistance in 24%, ROI not demonstrated in 22%, and ethical issues or bias in 18%. Projects often show several causes at once, so the shares add up to more than 100%.

None of these causes is technical in the narrow sense. They describe organizations that buy tools before preparing data, goals, and people. This is the study’s central point: the problem is the structural alignment between AI and the organization. A larger budget for chips or models does not fix an organization that has not defined what success looks like, and the next budget cycle repeats the same pattern.

Is the 95% failure figure enough evidence of an AI bubble?

A high failure rate alone proves nothing. Early stages of many technologies show high failure and still lead to lasting industries. The bubble question is whether money keeps flowing at the same pace regardless of the failures. A market that learns from failure slows spending on weak projects and moves it to strong ones. A bubble keeps spending in both.

So read the gap in 2 steps. First, check whether failing projects get new budgets. Second, check whether the money comes from customer revenue or from fresh investors. Failing projects funded by customers’ own earnings correct themselves over time. Failing projects funded by outside capital can continue until the capital stops, which is when a bubble ends. For that reason the funding source matters more than the failure rate itself.

What failure rate does the EQOS Asia model predict for AI projects?

At a φ(S) score of 0.337 the EQOS Asia model predicts a failure rate of about 94% for complex deployments, which leaves a 6% chance of success. The observed MIT rate is 95%, so prediction and observation sit within 1 point of each other. The study reads this as a sign that the score captures the structural weakness behind the failures.

The score sits in the fragile zone, where the system works and a shock can tip it. The study lists Communication at 0.305 as the weakest dimension, only 0.005 above the stop threshold of 0.3, and Energy at 0.588 as the strongest. The weakest dimension drags the whole system down, because the aggregation is a weighted geometric product, not an average. Raising the weakest area therefore pays more than polishing the strongest one.

How much of the AI investment gap can be recovered?

The study recommends 4 corrections: measure the data signal before deploying, require explainability for critical decisions, audit the alignment between AI strategy and ground reality, and fund a retraining plan for impacted jobs at 2% to 5% of AI investment. Together they lift φ(S) from 0.337 to 0.401.

At 0.401 the predicted failure rate falls from 94% to 82%. The study turns that into about 40% of the yearly spend recovered, or about $80 billion a year for the world market. Bain estimates $2.3 trillion of digital transformation projects fail each year, and generative AI is now a main component of that total, which shows the scale of what is at stake.

How can a business measure its own AI investment gap?

Start with your own numbers. Divide AI spending by the number of projects in production, and note how many pilots stalled. Then score the 4 areas above for your organization. Your score may sit above or below the market’s 0.337. The study uses 0.5 as the viability threshold for a successful AI deployment, so aim there first.

The EQOS Asia method explains how the score is built, and the Λ simulator estimates your φ(S) in 10 minutes. For a decision with real money behind it, a 30-minute session on the contact page applies the scoring to your case in Singapore, Hong Kong, or ASEAN and returns a ranked list of actions.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile