Corporate Investment in Generative AI Remains Low, Accounting for Less Than 1% of Expenditures
-
In 2023, generative AI has garnered significant attention as a breakthrough technology capable of driving transformation across multiple fields, even raising hopes of changing human life. However, Menlo Ventures' report indicates that despite the breakthroughs in generative AI in 2023, it has largely been a hype.
The report points out that generative AI's share in enterprise cloud spending is "relatively meager," accounting for less than 1%. In contrast, traditional AI occupies an 18% share in the cloud market, reaching a scale of $40 billion. Derek Xiao, an investor at Menlo, told VentureBeat in an interview: "Many people believe generative AI will quickly take over the world, viewing AI as a fundamental leap. But the reality is that this will take time, especially in the enterprise sector."
Image credit: AI-generated image, licensed by Midjourney
While some projections estimate the generative AI market will reach $76.8 billion by 2030 with a compound annual growth rate (CAGR) of 31.5%, or create at least $450 billion in value across 12 vertical industries within the next seven years, Menlo's survey found that as of 2023, half of enterprises had already implemented some form of AI according to Menlo's "State of Enterprise AI" report.
However, Menlo's research reveals strong corporate hesitation towards generative AI. Menlo partner Naomi Ionita stated in an interview: "We initially thought generative AI would be an overnight success story, but 2023 turned out to be a year of 'experimentation and exploration'." Derek Xiao added: "2024 will be the year of hard work in implementing generative AI."
The survey shows that leaders of large-scale enterprises should find comfort in these findings and recognize that moving forward slowly is acceptable. Tim Tully, a partner at Menlo, stated: "Smart people are taking their time," pointing out that the rapid development of generative AI has led to hesitation in adoption and, in many cases, "underfunding."
Corporate hesitation towards generative AI primarily centers around unverified return on investment and the 'last mile problem'. Other concerns include data privacy, AI talent shortages, insufficient organizational bandwidth, compatibility with existing infrastructure, and limitations in interpretability and customization.
Menlo's report indicates that generative AI solutions 'have yet to achieve meaningful transformation', failing to create new workflows and behaviors, with perceived limited production benefits. Until buyers can see real value, they will remain skeptical.
However, early adopters of generative AI have achieved significant benefits in data utilization and reducing 'tedious, painful workflows.' Ionita stated: 'It satisfies users in ways that were previously impossible.' Tully noted that users were able to create 'exceptionally outstanding tools' within 20 minutes, 'transforming workflows,' 'replacing teams, making people's work easier and more successful. It truly creates value and revenue.'
With the continuous development of the generative AI market, Menlo sees significant opportunities in both vertical (industry-specific) and horizontal (broader) applications. Ionita points out that the AI world will be hybrid: many enterprises are already using multiple foundational platforms, with smaller models being employed for various specialized purposes.
Regarding the standardization of the modern AI stack, Menlo found that enterprises have invested $11 billion in this field this year, making it the largest new market in generative AI. Buyers indicated that 35% of infrastructure spending goes to foundational models like OpenAI and Anthropic. These closed-source models continue to dominate, accounting for over 85% of production models.
Additionally, most models are used off-the-shelf, with only 10% of enterprises pretraining them. Most businesses adopt multiple models to achieve higher controllability and lower costs, with 96% of expenditures allocated to inference. Prompt engineering is the most popular customization method, while human review is the preferred evaluation approach.
Additionally, Retrieval-Augmented Generation (RAG) is becoming the standard. This framework enhances large language models (LLMs) by retrieving information from external knowledge bases, overcoming the limitations of fixed datasets to generate the latest, contextually relevant responses. According to Menlo's survey, 31% of enterprises are using this method, while 19% are employing fine-tuning approaches, 18% are implementing adapters, and 13% are applying reinforcement learning through human feedback.
Derek Xiao from Menlo stated that startups shouldn't just be 'ChatGPT wrappers,' but should focus on providing new workflows, next-generation reasoning, chain-of-thought, and proprietary data analysis tools. "It's not just about being a 'ChatGPT wrapper,' but more about the ability to create new markets within existing ones. This is a warning to startups - differentiation truly matters."