AI: The Information Era’s Apex Technology

In a new report from Citi Research, Head of Tech and Communications Research Heath Terry explores the prospects for artificial intelligence (AI), which he calls the apex technology of the information era. We think the market is underestimating the pace of acceleration in the AI technology cycle, which is being fueled by advancements in state-space models, world model-based reinforcement learning, neuro-symbolic reasoning and other research paths beyond transformer technology. Given such advancements, we use our Citi AI Industry Model to estimate that AI revenues will grow nearly 80% annually over the next five years to reach $780 billion in 2030 from $43 billion this year.

Three primary drivers inform this belief.

The first is infrastructure. Capex spending on infrastructure in 2025 alone is above $400 billion, according to our estimates. Such spending levels should ease capacity constraints that currently limit AI adoption, increasing the return on investment (ROI) for enterprise adoption of AI just as use cases pass the proof-of-concept phase and enter production.

The second driver is models. A combination of agentic AI with verifiable reasoning chains, domain-specific fine-tuning, and native multi-model reasoning is creating additional possibilities for use cases that require consistency and reliability for higher-stakes workloads, particularly when combined with more advanced applications of machine learning and other forms of AI beyond generative AI (GenAI).

The third driver is AI Squared, the point at which AI can effectively develop and improve AI itself. This is AI’s most powerful accelerant; while it’s still early, we’re seeing signs of this development becoming reality.

The ROI question

The single most important debate among AI investors remains AI’s ability to generate a return on the hundreds of billions of dollars being invested in infrastructure, research and development, and human capital. AI startups are generating tens of billions in losses, a scale beyond anything venture capital has ever seen, but the overarching philosophy of seemingly everyone in AI was captured by Alphabet CEO Sundar Pichai’s comment last summer that “the risk of under-investing is dramatically greater than the risk of over-investing for us here … not investing to be at the front here I think definitely has much more significant downsides.”

We see this mindset, echoed by similar comments from the CEOs of Meta, Amazon and Microsoft, as the inevitable outcome of a generation of management trained on The Innovator’s Dilemma, the 1997 book by Clayton Christensen that suggested a blueprint to help large companies avoid being overtaken by competing startups. Few of this generation’s founders or CEOs will risk a future in which their company fades away by focusing on margins and capital returns only to be passed by new technologies and business models. Both public- and private-market investors have embraced this strategy, but continue to ask from where and in what form these returns will come.

So far, AI’s answer has largely been about cost efficiencies. Surveys of initial AI deployments offer a mixed message: On the one hand, CIOs are recognizing 3.7x returns on investment; on the other, 95% of enterprise applications are producing no return. We estimate that today’s big use cases — customer service, code development and knowledge retrieval — will combine for close to $275 billion in annual cost savings, based on just the ~30% efficiency gains currently cited by companies that have deployed AI tools for call centers, development organizations and knowledge workers. 

Beyond that, we have emerging use cases in healthcare (such as drug discovery, record keeping and patient monitoring), industrials (preventative maintenance, product design, operations monitoring, etc.) and media (marketing personalization, campaign management and content creation) that could combine for multiples of that $275 billion figure given these industries’ relative scale. We believe as AI technologies improve, cost of training and inference continue to decline, and new use cases are found, questions regarding AI’s ROI will disappear, driving incremental acceleration.

The future of the tech stack

Cloud Computing and Software as a Service (SaaS) drove the last major redefining of the enterprise technology stack; so will AI, in our view. While we’re still very early in this process, with most of the current impact being felt at the infrastructure and development layer, we recently surveyed 90 tech startups to understand the choices of companies that are starting with a blank slate. At a high level, 18% of startups (but no established companies surveyed) consider their tech stack to be AI-native. Eighty percent of companies (78% of startups and 87% of established firms) consider that stack to be AI-enabled. And 13% of established companies and 4% of startups surveyed say they aren’t using AI.

Based on that survey and our conversations with senior technologists, we think AI will profoundly rewrite the technology stack, accelerating modernizations and migrations, introducing new products and services, and displacing existing core components. 

Three risks to the AI transition

Massive tech transitions don’t happen in a straight line. While we acknowledge that there are countless unknown unknowns, we focus on three primary risks to how AI and companies around it will develop.

The first is GenAI’s limitations. Fundamental concerns exist around the technology, including the “stochastic parrot” criticism that large language models statistically mimic text without a real understanding of it, with their apparent reasoning often just pattern recognition. An incremental .001% error rate is more than acceptable in customer service, code development or advertising, but not in air traffic control, cryptography, financial trading or medical diagnoses. A scalable, general-purpose reasoning architecture remains elusive. Until these limitations are addressed, AI will remain strong in language and pattern recognition but limited in areas requiring grounded reasoning, abstraction and flexible decision-making amid uncertainty.

Second, we note numerous bottlenecks to AI’s development, ranging from the availability of electrical equipment and grid reliability to privacy and ethical concerns. Such bottlenecks could slow development, raise costs and restrict the adoption of AI.

Third, the cost to deliver AI may in many cases outstrip the value companies get from it. Today’s tens of billions in annual losses for AI startups are being borne by corporate cashflows and venture capital, subsidizing the adoption of AI and masking its true economics. 

But these risks appear to be having little impact on venture capital’s enthusiasm for AI: So far this year we’ve seen $104 billion in investments, growth of nearly 180% annualized. This has driven private-market valuations among the largest AI companies. When we look at the ecosystem of venture-backed AI startups, it’s clear we’re still in the early stages of development. Most venture funding is going to the hardware, networking and infrastructure layers, with applications just beginning to see seed and series A funding. There’s no real evidence yet of the kind of dedicated consumer companies that marked prior cycles of tech innovation.

AI is definitely a “big company technology,” with most of its startups unlikely to survive the consolidation that comes for all venture-capital-driven booms. That said, we expect that we’ll see durable new giants emerge from the advancements driven by such investment, and we believe AI’s impact on innovation and efficiency will yield returns proportionate to the scale of investment.