90% of Enterprises are Exploring Generative AI, But 70% of Projects will Fail to Deliver

Fully 88 percent of organizations surveyed by S&P Global Market Intelligence on behalf of Weka for its 2024 Global AI Trends Report are now actively investigating generative AI, ahead of other AI applications such as prediction models (61 percent), classification (51 percent), expert systems (39 percent) and robotics (30 percent).  

Access to graphics processor unit capabilities is an issue. About 40 percent of respondents surveyed suggest access to AI accelerators is a leading consideration in their infrastructure decision-making, and 30 percent cite GPU availability among their top three most serious challenges in moving AI models into production.

That seems to be a bigger issue in the Asia-Pacific region than elsewhere, where lack of access to GPUs is restricting organizations from deploying AI. For example, 38 percent of organizations in India see accelerator access among their top three challenges to moving AI projects into production.

In other regions, access to “GPU as a service” might obviate such concerns. 

For these end user enterprises, though, the greatest proportion of respondents (35 percent) indicate storage and data management are the primary infrastructure issues hindering AI deployments, significantly greater than compute (26 percent), security (23 percent) and networking (15 percent).

That noted, it remains the case that many AI projects fail to reach “deployment at scale” status. According to one Gartner study, more than half of AI projects never are deployed at scale. The S&P Global Market Intelligence survey of about 1,500 respondents tends to confirm that finding, as few of the respondents have more than a handful of AI projects in production, at scale.  

source: Weka

And Gartner analysts suggest that about half of all IT projects fail to reach their financial goals. That might be equally true for GenAI projects as well, given the difficulty of quantifying success for the wide range of AI use cases being implemented. 

source: Weka 

And AI projects are at least as complicated as other IT initiatives and projects, all of which often fail to meet objectives, for any number of reasons. 

source: Weka

Of course, since generative AI and other machine learning applications remain relatively new, perhaps that should not come as a surprise. But Gartner research suggests as many as 85 percent of AI projects fail, and made that prediction about 2018, as nearly as I can ascertain. 

That would not be wildly out of line with the general industry rule of thumb that about 70 percent of IT projects also fail to deliver their intended results. 90% of Enterpris

Study Name	Date	Publisher	Key Conclusions
Standish Group Chaos Report	Annual	Standish Group	Consistently reports high failure rates for IT projects, often exceeding 70%, with common causes including unrealistic expectations, poor communication, and inadequate planning.
Project Management Institute (PMI)	Ongoing	PMI	While not always quoting the exact 70% figure, PMI's research frequently highlights the challenges and risks associated with IT projects, contributing to high failure rates.
Harvard Business Review	Various articles	Harvard Business Review	Numerous articles have discussed the high failure rates of IT projects, attributing them to factors such as organizational culture, lack of executive sponsorship, and poor change management.
IT Governance Institute	Various reports	IT Governance Institute	Provides insights into the factors contributing to IT project failures, such as inadequate governance, insufficient resources, and unclear objectives.

Can AI Improve Mobile Account Churn Detection and Enable Better Retention Efforts?

The high-performance server market illustrates an important principle in the computing markets, which is that a sufficiently-important function or appliance with high price tags creates incentives for a few big end users to shift some procurement to “do it yourself” alternatives. 

In high-performance server markets, for example, a few big end users might already have about 20 percent share. In 2023, for example, the U.S. commercial value of such products was about $53 billion annually, if we included the value of the DIY production as well as commercial sales. 

Supplier	Share (%)
Dell	25-30
HP Enterprise	20-25
Lenovo	10-15
Supermicro	5-10
Other Vendors ( IBM, Fujitsu)	15-20
DIY Solutions
Meta	5-7
Google	5-7
Other Major End Users (Amazon, Microsoft)	10-15

In the AI chip business, possibly 15 percent to 20 percent of total market share is held by DIY suppliers including Meta, Google, Amazon and Microsoft, for example. 

Supplier	Market Share (%)
NVIDIA	50-60
Intel	10-15
AMD	10-15
Google (TPU)	5-10
Other Traditional Vendors (e.g., Qualcomm, MediaTek)	5-10
DIY Solutions
Meta	5-7
Google	5-7
Other Major End Users (e.g., Amazon, Microsoft)	5-10

One might note the same trend in the area of generative AI models, where a few “end users” also are suppliers of models. 

Supplier	Share (%)
OpenAI (GPT-4, GPT-3.5)	30-35
Google (PaLM 2, LaMDA)	20-25
Meta (LLaMA)	10-15
Other Major Vendors (e.g., Anthropic, Cohere)	10-15
DIY Solutions
Meta	5-7
Google	5-7
Other Major End Users (e.g., Amazon, Microsoft)	5-10

Some Problems Have No Obvious Solutions

The EU is losing ground in research and development and in the creation of innovative technology companies with global reach, says a report commissioned by the European Commission. That is unlikely to surprise anybody familiar with the matter, as such concerns have been in place for many decades.

The report says the EU lags in artificial intelligence, cybersecurity, the internet of things (IoT), blockchain and quantum computers. 

The EU has generated fewer new lead innovators in the past decade than the United States, and that the share of EU firms in the top 2,500 global R&D companies has fallen compared to other blocs,” the study argues. 

“For instance, among leading companies in software and internet, EU firms represent only seven percent of R&D expenditure, compared with 71 percent for the U.S. and 15 percent for China; similarly, the EU only accounts for 12 percent of R&D expenditure among leading companies producing technology hardware and electronic equipment, compared with 40 percent for the U.S., and 19 percent  for China.

The study notes that the EU is home to only four of the fifty largest digital marketplaces worldwide, while the ten largest platforms serving EU citizens are owned by U.S. or Chinese companies (Alphabet, Amazon, Meta, Apple, Microsoft, X, Tencent, Alibaba, Byte Dance and Baidu). 

On the other hand, the report notes that “the EU has important capabilities, in particular, in green technologies, advanced manufacturing and advanced materials, the automotive industry and biotechnology.” 

Among the perhaps-obvious recommendations are to increase research and development spending. But many of the other recommendations are less directly related, such as increasing the quality of world-class research institutions.

“According to the Nature Index in 2022, which ranks institutions based solely on the volume of

publications in a selected list of top academic science journals, the EU has only three research institutions among the top fifty globally,” the report says. 

While praiseworthy, the effort to grow more world-class research institutions is a tough and long-term goal that many would argue is unlikely to happen in a world where leadership begets leadership. 

The computer science, semiconductor and biology industries  are typically concentrated in a small number of science and technology clusters, with leading clusters accounting for a large share of overall innovation in a country, the report notes. The EU simply has such clusters, but few in the top 10. 

According to the WIPO classification of world clusters (2023 Global Innovation Index), the EU has a similar number of clusters in the top 100 as the US and China, but only one in the top 20. None of the EU clusters appear among the top ten, while the United States has four and China has three, the report says.

But there are lots of other issues, ranging from a weaker venture capital system to the degree to which academics create private companies; overall commercialization of research and regulatory and bureaucratic obstacles. 

A slower pace of technology adoption; smaller firm size; quality of digital infra and skills also are cited as obstacles. But the list of issues is numerous, including the cost of energy; access to raw materials and digital infrastructure in general. 

For some observers, the report simply restates what critics have said for decades about EU competitiveness, 

The common litany of concerns includes the fragmentation of the Single Market impedes scale. It always is argued that the EU makes insufficient investment. Regulatory barriers are said to be too high. There are talent shortages as well.

Digital infra is not well-developed enough and competitors do better in such areas. Most of the proposed remedies, and they are numerous, would require both a huge shift of resources and significant time, something perhaps unavailable to the EU as the artificial intelligence and computing innovation cycles either appear or continue. 

Some problems might not have solutions, at least not solutions that are politically or economically feasible. And even if successful, EU “catching up” to China and the United States in many technology fields would take a long time, many would likely argue.

GenAI is Not Machine Learning

Generative artificial intelligence is different from machine learning, so the value will be different as well. GenAI value will be distinguishable from machine learning, since GenAI is optimized for content creation, where machine learning is optimized for seeing patterns in data. 

Industry	Generative AI Value	Machine Learning Value
Healthcare	Personalized medicine, drug discovery, automated medical documentation	Predictive diagnostics, patient management, anomaly detection
Financial Services	Algorithmic trading, automated report generation, fraud detection	Risk assessment, customer segmentation, credit scoring
Manufacturing	Product design, prototyping, generative design for optimization	Predictive maintenance, process optimization, quality control
Retail	Personalized marketing, product recommendations, automated content creation	Inventory management, demand forecasting, recommendation systems
Education	Custom learning content, adaptive learning platforms, automated grading	Student performance analysis, personalized learning paths, dropout prediction
Media & Entertainment	Content creation (e.g., scripts, music, art), virtual actors, automated editing	Audience analysis, content recommendation, trend prediction
Transportation & Logistics	Route optimization, autonomous vehicle development, dynamic scheduling	Fleet management, logistics planning, demand forecasting
Real Estate	Property design, virtual staging, automated property descriptions	Market analysis, property valuation, predictive maintenance
Legal	Contract generation, legal research, automated case summarization	Document review, case management, legal analytics
Marketing & Advertising	Campaign creation, content generation, personalized advertisements	Customer segmentation, ad targeting, sentiment analysis
Energy	Renewable energy solutions, smart grid development, automated reporting	Predictive maintenance, energy management, demand forecasting
Technology	Software development (e.g., code generation, bug fixing), AI model training	Predictive analytics, system optimization, anomaly detection

And It is not easy these days keeping track of artificial intelligence use cases that are examples of machine learning and which are examples of generative AI. 

Task Type	Machine Learning	Generative AI
Data Analysis	Excels at analyzing large datasets to find patterns and make predictions	Can summarize and interpret data, but not its primary strength
Prediction	Strong at making predictions based on historical data (e.g. sales forecasting, risk assessment)	Can make predictions, but often less accurate than specialized ML models
Classification	Very effective for categorizing data into predefined classes (e.g. spam detection, image classification)	Can perform classification tasks, but typically not as accurately as specialized ML models
Anomaly Detection	Excellent at identifying unusual patterns or outliers in data	Can describe anomalies, but less effective at detecting them compared to ML
Content Creation	Limited capabilities in generating new content	Excels at creating various types of content (text, images, code, etc.)
Natural Language Processing	Good at tasks like sentiment analysis and language translation	Superior at understanding context and generating human-like text responses
Decision Support	Provides data-driven insights to assist human decision-making	Can offer more nuanced, context-aware recommendations and explanations
Automation	Automates specific, well-defined tasks based on patterns in data	Can automate more complex, creative tasks that require understanding and generation
Personalization	Effective at providing personalized recommendations based on user data	Can create highly personalized content and interactions
Problem Solving	Solves specific problems it's trained for within defined parameters	Can approach novel problems creatively and propose innovative solutions