For three years, every CIO budget meeting I have heard about has included some version of the same conversation: AI is in the innovation line, sitting next to blockchain pilots and metaverse experiments, with a CFO sharpie hovering over it whenever quarterly margins compress. AI was a project. AI was an experiment. AI was, structurally, optional.
JPMorgan Chase has now ended that argument inside the world's largest bank — and, by example, inside the rest of the Fortune 500. In its 2026 budget cycle, JPMorgan reclassified AI spending from discretionary innovation into core infrastructure, alongside data centers, payment systems, and core risk controls. The bank's total technology budget is now $19.8 billion, of which roughly $2 billion — about 10% — is tagged as AI. The incremental year-over-year increase is approximately $1.2 billion, almost all of it directed at AI and modernization.
The accounting change sounds dry. The implication is not. When AI moves from "innovation" to "core infrastructure" in a bank with $4 trillion in assets and the most-cited enterprise AI maturity score in the industry (79.0 on the Evident AI Index), the rest of the financial sector — and any enterprise that benchmarks against JPMorgan — now has a budget defense problem. Why is your bank's AI still in the R&D line? That is the next board question, and most CIOs do not have a good answer for it.
This piece is about three things: what the JPMorgan reclassification actually means architecturally, what the bank has already shipped that justifies the reclassification, and the layer JPMorgan has not built yet — the layer that will determine whether $2 billion a year produces a durable advantage or just temporary parity.
What "Core Infrastructure" Actually Means
The reclassification is not a marketing exercise. Core infrastructure inside a bank carries specific accounting, governance, and operational treatment.
Accounting. Core infrastructure is base operating cost, capitalized over multi-year horizons, defended in every budget cycle by reference to baseline service levels rather than ROI calculations. Innovation budgets get cut when margins compress; infrastructure does not. The reclassification means that even if JPMorgan misses a quarter, the AI line is structurally protected.
Governance. Core infrastructure reports through the CIO and the Operating Committee, with formal SLAs, change-management processes, audit trails, and regulatory reporting. JPMorgan's AI leadership now reports to the Operating Committee under Teresa Heitsenrether — a business veteran, not a technologist — making AI a peer to payments and lending rather than a subordinate of IT.
Operations. Core infrastructure is treated as undifferentiated heavy-lifting, run with the same reliability discipline as the core banking ledger. That means 24/7 monitoring, capacity planning, vendor lock-in scrutiny, and disaster-recovery requirements. AI is no longer allowed to "fail interestingly" — it is required to fail predictably, with rollback plans and incident reviews.
The architectural consequence: AI is now a substrate. Every business unit can assume it is there, the way they assume the data center is there. Product roadmaps stop asking "should we use AI?" and start asking "which workflow do we move onto the AI substrate this quarter?" That is a different operating tempo than the experimental phase, and it is the operating tempo the rest of enterprise AI has been waiting for.
The Numbers Behind the Reclassification
Treating AI as infrastructure only makes sense if the deployment numbers justify it. JPMorgan's do.
LLM Suite penetration: the bank's proprietary internal LLM platform — a secure interface to external frontier models — is deployed to 250,000 employees. Over 100,000 use it daily. Roughly one in three JPMorgan employees opens an AI tool every workday. For comparison, that is more daily active AI users inside one bank than most enterprise SaaS vendors have customers.
Production use cases (doubled in 2025):
- COiN reviews 12,000 commercial agreements in seconds — work that previously consumed an estimated 360,000 lawyer-hours annually.
- Investment Bank Copilot generates pitch decks in 30 seconds. The first-draft research cost on an IB pitch dropped roughly 90%.
- Proxy IQ manages shareholder voting decisions for $3 trillion in assets under custody.
- Fraud detection systems running on machine-learning analytics drove material revenue and loss-avoidance gains, which CFO Jeremy Barnum credited in investor briefings.
- Developer productivity tools across the engineering org, in the same family as GitHub Copilot deployments at peer banks.
- Call-center summarization, marketing personalization, and middle/back-office processing automation rolled into routine operations.
Realized value: the bank reports approximately $2 billion in realized annual AI value, roughly equal to its annual AI spend. That break-even number is what unlocks the infrastructure reclassification: if AI is paying for itself in observed dollars, the CFO can defend the reallocation without leaning on speculative future ROI.
CEO Jamie Dimon framed it bluntly in his investor commentary: institutions that fall behind on AI risk losing ground to competitors. Cutting tech spend may improve short-term margins but erodes long-term competitiveness. CFO Barnum acknowledged that AI returns are difficult to quantify — and then defended the spend anyway, on strategic-necessity grounds. That is the language of infrastructure, not innovation.
What JPMorgan Has Built That Other Banks Have Not
Three architectural decisions separate JPMorgan from the rest of the banking sector — and they are the decisions other CIOs need to study before their next budget cycle.
One: a sovereign LLM Suite, not a public-AI passthrough. JPMorgan does not let employees use ChatGPT, Claude.ai, or Gemini directly. The LLM Suite is a proprietary internal platform that brokers calls to external frontier models with hard data-loss-prevention rails: customer data, deal data, and material non-public information stay inside the bank's perimeter. The LLM Suite also gives JPMorgan the option to swap underlying model providers — Anthropic, OpenAI, Google, internal models — without changing the user experience. That optionality is worth more than any single model partnership.
Two: $billions of pre-AI data engineering. JPMorgan migrated 65% of workloads to cloud, decommissioned 2,500 legacy applications, and standardized data contracts across business lines before the LLM Suite went enterprise-wide. The unsexy, multi-year cleanup is what makes AI deployment possible; banks that skipped it are now discovering that their AI initiatives die at the data-quality boundary.
Three: business-line ownership of AI outcomes. In February 2026, JPMorgan appointed Chief Data & Analytics Officers inside major business units — investment banking, asset management, consumer, commercial. These roles own AI P&L, not just AI capability. The structural lesson: AI accountability lives where the revenue lives, not in a central CoE that issues recommendations no one is required to follow.
These three decisions explain why JPMorgan's deployment numbers are an order of magnitude ahead of peers. Bank of America, Citi, Wells Fargo, and Goldman Sachs are all spending material money on AI — but each is, at this writing, still navigating one or more of the three decisions above. The gap is not budget. The gap is operating model.
The Layer JPMorgan Has Not Built Yet
Here is the part that does not show up in the press releases.
For all of JPMorgan's leadership, the bank has not yet built the AI layer that turns parity into durable advantage. Call it Intelligence Capital — institutional knowledge that is captured, encoded, and accumulated by AI systems with every decision the organization makes. Today's deployments automate existing processes; they do not systematically capture the reasoning behind decisions.
When a JPMorgan partner overrides COiN's contract recommendation, that override is a learning signal — the kind of human-judgment correction that, captured at scale, encodes institutional wisdom that no competitor can replicate. Today, COiN does not store the override and the reasoning. The next decision starts from the same baseline. The bank automates the work but does not compound the knowledge.
This is the Parity Problem. Every efficiency gain JPMorgan unlocks is, in principle, replicable by any peer that buys the same vendor stack and migrates the same data. The COiN savings are reproducible at Wells Fargo. The Copilot productivity gains are reproducible at Goldman. As of today, JPMorgan's AI advantage is broadly defensible only as long as the bank stays a year or two ahead in deployment maturity. Once peers catch up — and they will — the strategic moat closes.
Intelligence Capital is the layer that does not close. A deliberation system that captures the reasoning, dissent, and override patterns of a bank's senior partners over a decade compounds into a model that no competitor can clone, because no competitor has access to those decisions. That is durable advantage. JPMorgan has not built it yet, and neither has any other large enterprise. Whoever builds it first is on the right side of the next ten years.
The reason it has not been built is that the architectural primitives are unsettled. There is no widely accepted pattern for capturing deliberation records. Vendor contracts often allow customer interaction data to be used for general model improvement, which means deliberation data leaks into the competitive substrate. Most enterprise AI platforms do not provide portable, auditable storage of decision rationale. The category is still pre-product.
For CIOs, the Intelligence Capital question is the next architectural decision after the substrate is in place. Where do we store deliberation? Who owns it? How do we keep it out of the vendor's training data? What is our migration path if we change platforms? These are not questions any major enterprise software vendor answers cleanly today.
What This Means for Non-Bank Enterprises
Three concrete implications for the rest of the Fortune 500.
One: AI moves to infrastructure inside the next two budget cycles. If JPMorgan can defend the reclassification, every CFO peer-benchmarking against banks now has cover to do the same. Expect the AI line to migrate out of the innovation budget at most large enterprises by the FY27 planning cycle. The CIOs who arrive at that meeting with realized-value evidence — not pilot summaries — win the budget.
Two: the LLM Suite pattern is the right enterprise architecture. Direct employee use of public AI is governance theater; sooner or later it leaks. A sovereign internal platform that brokers external models, enforces DLP, and preserves provider optionality is the pattern enterprises should be building. The vendors furthest along on this — IBM watsonx, Microsoft Azure AI Foundry, Anthropic enterprise — are competing for the substrate role. The pattern matters more than the vendor choice.
Three: data engineering is the binding constraint. The reason JPMorgan can deploy LLM Suite to 250,000 employees is that it spent the prior decade making its data fit for cloud-native AI. Enterprises that skipped that work are discovering AI initiatives stall at the data-quality boundary. The first-quarter AI rollout failures of 2026 will, on inspection, almost all trace back to legacy-data hygiene. Budget reform without data reform produces expensive disappointment.
Where the Bet Is Fragile
Three risks worth tracking.
Concentration risk in the LLM Suite. A single internal platform brokering all frontier-model access is a tempting attack surface and a single point of operational failure. The architecture assumes the platform team can keep pace with frontier-model release velocity, which is currently a six-week clock. Falling a quarter behind on a critical model upgrade would degrade competitiveness across 250,000 employees at once.
Vendor data leakage in deliberation records. As JPMorgan moves toward systems that capture override reasoning, the standard vendor contract terms become a strategic liability. If interaction data is being used to improve the vendor's general model, JPMorgan is funding a competitive substrate that benefits its rivals. The contract negotiations on this are happening now, behind closed doors, and will be one of the most consequential procurement battles of the next 18 months.
Regulatory whiplash. Treating AI as core infrastructure means it falls under existing operational-risk and third-party-risk frameworks. The Federal Reserve, OCC, and CFPB will, by Q4 2026, almost certainly issue guidance treating AI substrates as critical operations subject to BCP and resolution-planning requirements. JPMorgan is well-positioned for this; banks still in the innovation-budget phase will be retrofitting governance under deadline pressure.
What CIOs Should Do Before Q3
Three things, in order.
First: stop arguing about whether AI is strategic and start arguing about which budget line it belongs in. The JPMorgan reclassification is the cover you need to move it to infrastructure. Build the realized-value evidence base that supports the reclassification, even if your numbers are smaller than $2 billion. Break-even-vs-spend is the threshold the CFO will defend; aim for it.
Second: design or buy a sovereign LLM Suite. Direct public-AI use across the workforce is a transition-state pattern, not an end-state. By end of 2026, every Fortune 500 enterprise should be brokering frontier-model access through an internal platform with hard DLP and provider optionality. Whether you build this on watsonx, Azure AI Foundry, Bedrock, or a custom stack is a vendor question; the architecture is non-negotiable.
Third: start the Intelligence Capital conversation now, even if you cannot finish it. Deliberation capture is not yet a productized category, but the procurement decisions you make on AI substrate today determine whether you have data sovereignty over deliberation records in three years. Insist on contractual carve-outs for interaction data. Insist on portable, auditable deliberation storage. Insist on architectural — not contractual — protection. These are the terms of the next decade's competitive moat.
JPMorgan made AI core infrastructure not because the technology has matured into a stable category — it has not — but because the bank decided that arguing about whether AI was strategic was costing more than the technology itself. That is the right reason. The rest of enterprise AI is now playing on JPMorgan's clock.
Sources: Banking Exchange, AI News, Prism News, Fortune, AI-Risk.co analysis, JPMorgan investor briefings.
