NVIDIA AI Survey 2026: Breakthrough or Marketing?

When the world's largest seller of AI infrastructure publishes a survey on AI adoption, the fine print deserves close scrutiny. The sixth annual "NVIDIA State of AI in Financial Services" report, based on a survey of more than 800 industry professionals, reports record levels of usage, return on investment (ROI) and budgets. 89 per cent say AI is increasing revenue and reducing costs. 73 per cent of executives describe AI as critical to their future success. Nearly 100 per cent intend to increase or maintain their AI budgets. A touch too convenient to accept uncritically.

The Headline Figures – and What Lies Beneath

The report delivers impressive headlines. 65 per cent of respondents say they are actively using AI – up from 45 per cent in the previous year. 61 per cent are deploying or evaluating generative AI, a 52 per cent year-on-year increase. Use cases range from document processing to fraud detection, algorithmic trading and risk management. 52 per cent cite creating operational efficiencies as the greatest improvement, 48 per cent point to increased employee productivity.

Then the ROI figures: 64 per cent report that AI has increased annual revenue by more than five per cent – including 29 per cent who cite an increase of more than ten per cent. On the cost side, 61 per cent say AI has reduced annual costs by more than five per cent, with 25 per cent reporting savings exceeding ten per cent.

These figures are valuable as a trend indicator. As a basis for investment decisions, they are insufficient – for four reasons.

Thesis 1: ROI Without Full Cost Accounting

Self-reporting is not an audited result

The 89 per cent figure is based on respondents' self-assessment. These are not audited financial results but estimates from industry participants. What matters is what the calculation omits: a total cost of ownership (TCO) that includes the actual expenditure on GPU infrastructure, specialised ML personnel, data preparation, governance and ongoing model maintenance.

The costs of GPU clusters are substantial. A single NVIDIA H100 GPU costs between 25,000 and 40,000 US dollars depending on configuration. Productive large language model (LLM) inference at financial institutions typically requires multiple such units. Add energy costs, cooling, maintenance and the specialist staff who operate these systems. When an institution reports that AI has increased revenue by five per cent but does not offset the TCO for the underlying infrastructure, the ROI statement is incomplete.

There is also the question of causality. Does the revenue increase actually correlate with AI deployment – or with concurrent market movements, product innovations or regulatory changes? The report draws no distinction.

Thesis 2: Open Source as a Trojan Horse for GPU Lock-in

Model freedom, infrastructure dependency

84 per cent of respondents consider open-source models and software important to their AI strategy, with 43 per cent rating them as very to extremely important. The report presents this as a sign of flexibility and freedom of choice. The reality is more nuanced.

Organisations that fine-tune and operate open-source models themselves require significantly more compute capacity than those consuming APIs. Fine-tuning a 70-billion-parameter model on proprietary institutional data demands dozens of high-end GPUs over weeks. And this is precisely the business model: NVIDIA sells not AI models but the hardware on which they run. The more institutions embrace open source and train in-house, the more GPU capacity they demand.

The real lock-in is not the licence – which is free by definition in open source – but the ecosystem. NVIDIA's CUDA platform (Compute Unified Device Architecture) is the de facto standard for GPU-accelerated computing. Once code, workflows and toolchains are built on CUDA, switching to AMD or Intel is far from straightforward. The open-source trend in models paradoxically cements dependency at infrastructure level.

Thesis 3: Agentic AI – Hype vs. Reality

42 per cent are using or assessing, 21 per cent have deployed

The report states that 42 per cent of respondents are "using or evaluating" agentic AI. This sounds like a technology approaching mainstream adoption. But the breakdown tells a different story: only 21 per cent have actually deployed AI agents in production. A further 22 per cent plan deployment "within the next year and beyond" – a remarkably vague timeframe.

The gap between "using or assessing" and "deployed" is not a semantic quibble. A proof of concept (PoC) in a sandbox is something fundamentally different from an autonomous agent processing payments, preparing credit decisions or conducting compliance checks in a regulated production environment. The jump from 21 to 42 per cent consists largely of institutions that are exploring the topic – not those that have mastered it.

For European institutions, a specific challenge arises: the regulatory framework for autonomous AI agents remains largely undefined. Who is liable when an agent makes a flawed credit decision? How do autonomous agents fit within the Minimum Requirements for Risk Management (MaRisk)? How should operational risk under the Digital Operational Resilience Act (DORA) be assessed when a third-party agent fails? The NVIDIA report does not pose these questions.

Thesis 4: The Regulatory Vacuum in the Report

EU AI Act, DORA, MaRisk – all absent

The report's most conspicuous blind spot: regulation is virtually absent. For a US-centric hardware manufacturer, this may be understandable. For European financial institutions seeking to use the report as a compass, it is a serious shortcoming.

The EU AI Act classifies numerous AI applications in the financial sector as high-risk systems – including credit scoring, insurance pricing and fraud detection. From August 2026, stringent documentation, transparency and monitoring obligations apply. DORA has established binding requirements for digital operational resilience since January 2025, including the management of ICT (information and communications technology) third-party risks – and GPU cloud providers such as NVIDIA potentially fall within this category.

The ninth MaRisk amendment, currently in consultation, is expected to formulate explicit requirements for the use of AI in risk management. Every AI investment decision that European institutions make today must be assessed against these regulatory requirements – regardless of whether the NVIDIA report mentions them.

Recommendations for European Financial Institutions

The NVIDIA report is useful as a trend indicator but insufficient as a basis for AI investment decisions in regulated financial institutions. European institutions should rigorously assess the figures against regulatory reality and an honest full-cost calculation. Four concrete action areas: