JP Morgan's recent infrastructure report sent ripples through the investment community: the AI industry must generate 650 billion dollars in annual revenue to deliver even a 10 percent return on the unprecedented capital expenditure expected through 2030. The bank estimates more than 5 trillion dollars will pour into data centres and AI infrastructure over the next five years—the largest capital market event in history.
The figure sounds staggering. JP Morgan contextualised it as equivalent to every iPhone user paying an extra 35 dollars monthly, or every Netflix subscriber contributing 180 dollars per month, in perpetuity. Critics have seized on this as evidence of an AI bubble, with some analysts warning the economics simply don't work.
But this analysis may be missing the forest for the trees.
Why Agentic AI Changes Everything
Agentic AI—systems that autonomously execute multi-step tasks, use tools, and reason through complex problems—represents a fundamentally different compute paradigm. Unlike simple chatbot interactions, agentic workflows consume dramatically more resources. A single agent researching a topic might execute dozens of web searches, analyse results, synthesise findings, and iterate on its output. Each step requires a full model pass.
The compute intensity is staggering: agentic tasks consume 10 to 100 times more resources than traditional AI queries.
[Chart_1]
This multiplier effect explains why agentic AI will drive compute demand far beyond what current usage patterns suggest. When a chatbot becomes an agent, compute requirements don't increase linearly—they explode.
The Research Houses Are Converging
The major research houses are arriving at a striking conclusion: agentic AI alone could dwarf JP Morgan's 650 billion dollar threshold.
Gartner projects that 33 percent of enterprise software will incorporate agentic AI by 2028, up from less than 1 percent in 2024. By that same year, 15 percent of day-to-day work decisions will be made autonomously, and 70 percent of AI applications will use multi-agent systems. Perhaps most remarkably, Gartner forecasts that AI agents will intermediate over 15 trillion dollars in B2B spending by 2028.
McKinsey's analysis is equally bullish. The firm estimates agentic commerce—AI agents shopping, negotiating, and transacting on behalf of consumers—could generate 3 to 5 trillion dollars in global retail revenue by 2030. In advanced industries alone, McKinsey projects agentic AI will deliver 450 to 650 billion dollars in additional annual revenue by decade's end.
PwC's estimates align closely, suggesting AI agents could contribute 2.6 to 4.4 trillion dollars annually to global GDP by 2030.
The Adoption Curve Is Steeper Than You Think
Today, roughly 10 percent of AI usage qualifies as agentic. By 2028, as enterprise adoption accelerates and Gartner's forecasts materialise, we expect agentic workflows to represent half of all AI compute consumption. By 2030, that share could reach 70 percent.
[Chart_2]
The revenue implications follow directly. In 2025, agentic AI generates approximately 19 billion dollars in annual compute revenue. By 2026, that figure reaches 81 billion dollars. The inflection arrives around 2028: with agentic workflows handling half of all AI tasks, annual compute revenue climbs to 252 billion dollars. By 2030, we project agentic AI will consume 219 quadrillion tokens annually, generating 608 billion dollars in compute revenue.
That's nearly JP Morgan's entire 650 billion dollar threshold from this single category alone.
Putting the 650 Billion in Context
Layer in McKinsey's 3 to 5 trillion dollar agentic commerce forecast, Gartner's 15 trillion dollars in B2B intermediation, and adjacent applications in healthcare, autonomous vehicles, and enterprise automation, and the total addressable market comfortably exceeds 1 trillion dollars annually—potentially multiples of that.
The Risks Are Real
JP Morgan is right that the path won't be smooth. The bank warns of spectacular losers alongside spectacular winners. Gartner echoes this caution: over 40 percent of agentic AI projects will be cancelled by 2027 due to escalating costs, unclear business value, or inadequate risk controls.
Power constraints present another bottleneck. Current lead times for natural gas turbines have ballooned to three to four years, and nuclear plants historically take over a decade to build. JP Morgan notes that adding 150GW of power capacity is a remarkable challenge.
The Real Question
But JP Morgan's framing may inadvertently undersell the opportunity. The 650 billion dollar figure assumes AI must extract value through subscriptions and fees. Agentic AI inverts this model: these systems don't just answer questions—they complete tasks, automate workflows, and generate direct economic value.
The question isn't whether AI can generate 650 billion dollars to justify today's infrastructure investments. It's whether the infrastructure being built today will prove sufficient for the compute demands that lie ahead.
Assumptions & Methodology
Methodology: Derivation of Compute Revenue and Adoption Forecasts
The compute revenue and adoption projections presented in this analysis are derived from a bottom-up examination of AI usage patterns, token consumption volumes, and infrastructure cost structures. This section provides a detailed account of the analytical approach and underlying assumptions.
Estimation of Current Usage Baseline
Global AI message volume was estimated through aggregation of publicly reported usage data from major platforms. OpenAI's ChatGPT processes approximately 2.5 billion messages per day, serving 700 million weekly active users. Anthropic's Claude processes in excess of 25 billion API calls per month, with 45 percent attributable to enterprise customers. Incorporating estimated volumes from Google Gemini, Microsoft Copilot, and other platforms, aggregate global AI usage is estimated at approximately 5.3 billion messages per day as of 2025.
Classification of Agentic and Non-Agentic Interactions
AI interactions were classified into two categories based on computational intensity.
Non-agentic interactions comprise simple question-and-answer exchanges, averaging 3,000 tokens per message (2,000 input tokens plus 1,000 output tokens). This category represents approximately 90 percent of current AI usage by volume.
Agentic interactions involve multi-step reasoning, tool utilisation, and iterative workflows. These interactions average 35,000 tokens per message for routine tasks and may reach 300,000 tokens for complex workflows involving multiple web searches, code execution, and self-correction loops. This category represents approximately 10 percent of current usage by volume but accounts for a disproportionate share of compute consumption.
Token-to-Revenue Conversion Methodology
Token consumption was converted to compute revenue using the following assumptions:
GPU throughput was modelled at 200 tokens per second per NVIDIA H100 GPU under typical inference workloads.
Annualised GPU cost was set at 17,500 dollars per H100 per year, reflecting a blended rate across hyperscale cloud providers and owned infrastructure deployments.
On this basis, processing one quadrillion tokens annually requires approximately 158,000 H100-equivalent GPUs, generating approximately 2.8 billion dollars in compute revenue.
Construction of the Adoption Curve
Adoption projections were triangulated against forecasts published by leading technology research firms.
Gartner projects that 33 percent of enterprise software applications will incorporate agentic AI capabilities by 2028, compared with less than 1 percent in 2024. The firm further forecasts that 15 percent of day-to-day work decisions will be executed autonomously by 2028, and that 70 percent of AI applications will utilise multi-agent architectures by the same year.
McKinsey reports that 62 percent of organisations are currently experimenting with AI agents, with 23 percent actively scaling deployments. The firm further notes that 45 percent of Fortune 500 companies are operating pilot programmes or early-stage production systems.
Utilising these benchmarks, adoption was modelled as an S-curve, with the period of steepest growth occurring between 2026 and 2028 as enterprise adoption accelerates. The projections assume agentic workflows will constitute 50 percent of total AI compute consumption by 2028 and 70 percent by 2030.
Summary of Annual Projections
Combining projected usage growth, the adoption curve, and token intensity assumptions, the following estimates were derived:
2024: 1 percent agentic adoption, 5 billion dollars compute revenue
2025: 10 percent agentic adoption, 19 billion dollars compute revenue
2026: 25 percent agentic adoption, 81 billion dollars compute revenue
2027: 38 percent agentic adoption, 155 billion dollars compute revenue
2028: 50 percent agentic adoption, 252 billion dollars compute revenue
2029: 60 percent agentic adoption, 410 billion dollars compute revenue
2030: 70 percent agentic adoption, 608 billion dollars compute revenue
Key Assumptions and Limitations
These projections assume continued advancement in model capabilities without a corresponding reduction in compute costs per unit of output. Efficiency gains arising from techniques such as quantisation, model distillation, and improved inference architectures could materially reduce compute requirements per task. Conversely, the emergence of more capable models may drive increased token consumption as users delegate increasingly complex workflows to AI systems.
The projections do not incorporate adjustments for potential market consolidation, regulatory intervention, or macroeconomic disruption. The figures presented represent a base case scenario predicated on the continuation of current trends.
Infrastructure constraints—particularly power availability and capital financing—may limit actual deployment below theoretical demand levels, as highlighted in JP Morgan's analysis. These supply-side factors represent material downside risks to the projections.
Sources: JP Morgan AI CAPEX Report (November 2025), Gartner Press Releases (2025), McKinsey Global Institute, PwC Global AI Study, Signwl Analysis