03 May GEN AI – Do returns justify costs?
Banner image generated using ChatGPT, OpenAI, May 2nd, 2025
AI, GenAI is very costly. Despite the sensation caused by Deepseek, there still needs to be a lot of investment to make AI happen. This is clear from USA based The Stargate project, a new company by Sam Altman of OpenAI along with others which seeks to invest $500 Billion to build AI infrastructure in USA [1]. In other places, the total estimated cost is up to $1 Trillion on AI capex [2].
However, it is not clear as to whether there will be sufficient returns to justify this investment. To quote a Goldman Sachs (GS) report, which will form the basis of this article, “this spending has little to show for it so far beyond reports of efficiency gains among developers” [3]. This is clear from a lack of good AI consumer products in market. In fact, well know tech reviewers, Marques Brownlee last year had labelled some of the products like Humane AI pin and Rabbit R1 as “The Worst Products I’ve Ever Reviewed..” and ‘Barely Reviewable’ [4].
Goldman Sachs report titled “Gen Ai: Too Much Spend, Too Little Benefit?” does a deep dive, with experts from both sides of the debate and highlights the potential risk and return.
Skeptical View
Among the skeptical voices are MITs Daron Acemoglu and GS’ Jim Covello. Daron Acemoglu, Institute Professor at MIT, is especially skeptical. He has published an article titled “The Simple Macroeconomics of AI*” (we will discuss the article in a future blog) which, “..evaluates claims about large macroeconomic implications of new advances in AI” [5]. According to Daron Acemoglu the amount of task AI can automate in cost effective manners are limited (1/4 of total task it can automate). This in turn represents a mere 5% of all tasks where AI will have an impact. With this low base, the total productivity gain will be 0.5% in US and will contribute just 0.9% of US GDP in 10 years. One can assume it will be lower in a low-cost country like India. To quote Daron, “Many tasks that humans currently perform, for example in the areas of transportation, manufacturing, mining, etc., are multifaceted and require real-world interaction, which AI won’t be able to materially improve anytime soon. So, the largest impacts of the technology in the coming years will most likely revolve around pure mental tasks, which are non-trivial in number and size, but not huge, either” [6].
In a similar vein, GS Head of Global Equity Research Jim Covello points out that to generate adequate returns on this massive investment, AI needs to be able to solve complex problems. According to Jim Covello, presently GenAI is not designed to solve complex problems. Moreover, according to him, truly life changing invention like interest was technology which enabled a low-cost solutions to disrupt solution which were expensive even when internet was in early stage. To quote him, “Amazon could sell books at a lower cost than Barnes & Noble because it didn’t have to maintain costly brick-and-mortar locations. Fast forward three decades, and Web 2.0 is still providing cheaper solutions that are disrupting more expensive solutions, such as Uber displacing limousine services” [7]. He remains skeptical of both the cost of AI technology and its ultimate transformative potential.
He also doubts the speed at which AI costs will drop enough to make automation costs effective. This is because of the complexity of building critical inputs—like GPU chips and consequent lack of competition in the GPU market. Also, there is fierce competition as very tech company must adopt AI (as demanded by other stakeholders), but there are no clear revenue models for AI. Hence AI’s impact on tech companies’ valuation may be limited or even negative [8].
Optimist view:
Not everyone is this pessimistic. GS’ Joseph Briggs, Kash Rangan, and Eric Sheridan remain more optimistic about AI’s economic potential. While recognizing that, “automating many AI-exposed tasks isn’t cost-effective today” he maintains that costs will fall and AI will ultimately add up to 9% to US productivity and add 6.1% to US GDP by automating 25% of all tasks. He also estimates that improved productivity will lead to labour reallocation and new task creation on the horizon contributing to growth. He gives examples of how IT created new jobs like webpage designers, software developers, etc. who in turn created more demand for service jobs[9].
In a similar vein, lack of Killer AI apps like ERP was in the 1990s and E-Commerce in the 2000s, doesn’t dampen the enthusiasm of Kash Rangan and Eric Sheridan who are Senior Equity Research Analysts at GS. According to Kash, “…this shouldn’t come as a surprise given that every computing cycle follows a progression known as IPA—infrastructure first, platforms next, and applications last. The AI cycle is still very much in the infrastructure buildout phase, so finding the killer application will take more time.” He also argues as incumbent tech giants are leading the Capex, unlike in tech boom of late 90s, chances that AI does not become mainstream are low despite risks. Eric Sheridan also is unwilling to call AI boom a bubble, pointing out that, “ ..longer-than-expected payoff process won’t kill this tech cycle. I’m loathe to use the word “bubble” because I don’t believe that AI is a bubble, but most bubbles in history ended either because the cost of capital changed dramatically or end demand deteriorated and affected companies’ ability to deploy capital, not because companies retreated from investing in a technology where the payoff was taking longer than expected.” Eric also points out that Return on invested capital (ROIC) is low at present but the same is not true for the future for multiple reasons like improved productivity and efficiency. Plus, as smart phones show that resisting new tech is difficult once the usage of a new product becomes mainstream [like Smartphones apps like uber become popular] [10].
Where does this leave us?
While this report was published in August 2024, the core point from both sides is still developing. DeepSeek has shown that the cost of AI maybe an overestimate. At the same time, GenAI’s ability to solve complex problems cost effectively is still an open question.
As Marques Brownlee points, another question is whether GenAI will really become an AI Product or AI Feature, in which AI case will be just another feature that consumer expects in near future. In fact, AI as a feature is already happening in Smart phones, Email, LinkedIn etc. This does raise a question, what potential revenue source are likely to be with increase its general adoptions by consumers [11].
Conclusion
The debate around whether the returns from Generative AI justify its massive costs remains unresolved. While optimistic analysts foresee long-term productivity gains, new job creation, and GDP growth once AI matures beyond the infrastructure phase, sceptics caution that current technological limitations, high infrastructure costs, and unclear revenue models cast doubts on its transformative potential. The Goldman Sachs report captures this divide—emphasizing both the promise and the pitfalls. With no breakthrough consumer applications yet, and questions about whether AI will evolve as a standalone product or merely a supporting feature, the future impact of GenAI continues to hinge on how these uncertainties unfold.
Footnotes:
1https://www.bbc.com/news/articles/cy4m84d2xz2o
4 https://www.youtube.com/watch?v=TitZV6k8zfA; https://www.youtube.com/watch?v=ddTV12hErTc
5 https://economics.mit.edu/sites/default/files/2024-05/The%20Simple%20Macroeconomics%20of%20AI.pdf