GenAI Model Spending Is Growing 80.8% This Year. I Want to Know Who Is Paying and Why.

Gartner's April 2026 IT spending forecast includes a figure that I find hard to put in context: GenAI model spending is growing at 80.8 percent in 2026. Not overall AI spending, not software spending in general, but specifically the spending on GenAI models. That growth rate is running faster than almost any other technology category in Gartner's forecast. Total IT spending is projected at $6.31 trillion for 2026, which is already an extraordinary number. Within that, the GenAI model segment is the fastest-moving line item. When I see 80.8 percent growth in a single category in a single year, my first question is not "is the technology exciting?" My question is "what is actually being purchased, by whom, and what determines whether it produces value?"

The spending on GenAI models is not one thing. It breaks into at least three distinct buckets that have different cost structures, different beneficiaries, and very different relationships to business outcomes.

The first bucket is training frontier models. This is where the largest absolute dollars are concentrated, among a very small number of organizations: the handful of labs and hyperscalers that are actually training models at frontier scale. Training a frontier model requires enormous compute, measured in thousands of specialized GPUs running for weeks or months. The cost is real and it is enormous. This part of the spending is mostly invisible to enterprise buyers. It happens upstream, at the infrastructure level, and it is what makes the APIs those enterprises call possible. When a company pays OpenAI or Anthropic per token, part of that fee is amortizing the training cost. When they buy compute from a cloud provider, they are also contributing to the infrastructure that runs the inference. The 80.8 percent growth number aggregates all of this together.

The second bucket is inference at scale. Every time an enterprise application calls a language model, that call consumes compute. Inference costs are real and they scale with usage in ways that surprise many organizations who priced their AI projects based on development costs. A model that performs well in a pilot with a hundred users behaves very differently as a cost item when it is handling a million requests a day in production. Organizations that are scaling GenAI from pilot to production are discovering that the inference bill is an ongoing operating expense, not a one-time capital cost. This is a meaningful shift for procurement and finance teams that are used to thinking about software as a license fee rather than a variable consumption model.

The third bucket is enterprise customization. Organizations that need a model to know about their internal policies, their product catalog, their customer data, or their regulatory environment cannot simply point a general-purpose model at their use case and expect it to work. They either fine-tune a model on their data, use retrieval-augmented generation to pull relevant context at inference time, or build some combination. Each of these approaches requires its own investment: compute for fine-tuning runs, vector databases and embedding infrastructure for retrieval, engineering time for building and maintaining the pipelines. This is real spending that does not show up in the "AI model" budget line but is directly driven by the decision to adopt GenAI.

McKinsey's State of AI 2025 report found that 88 percent of organizations say they use AI, but only 7 percent report having fully scaled AI across their enterprise. The gap between those two numbers is one of the most important data points in the AI adoption landscape right now. Most of the 80.8 percent spending growth is concentrated in the organizations that have crossed into serious scaling, plus the labs doing frontier training. For the 81 percent who use AI but have not scaled it, the spending is real but the value realization is still mostly a promise. McKinsey puts the economic potential of GenAI at $2.6 trillion to $4.4 trillion annually across all sectors. If only 7 percent of organizations have achieved full scale, the gap between realized value and potential value is enormous.

This is where the DeLone and McLean (2003) IS Success Model becomes a useful frame for thinking about what all this spending is actually buying. DeLone and McLean argued that IS success requires system quality (the technical capability works reliably), information quality (the outputs are accurate and relevant), and service quality (the support and integration are good), and that all three must be present before use translates into net benefit. Applied to GenAI spending, the 80.8 percent growth is mostly purchasing system quality. Organizations are buying access to capable models. What many of them have not yet invested in is the information quality side (the data pipelines, the retrieval infrastructure, the evaluation frameworks that tell you whether the model's outputs are accurate for your use case) or the service quality side (the integration work, the change management, the training, the support structures). A capable model with poor information quality produces confidently wrong outputs. A capable model with poor service quality sits unused after the pilot.

The energy question sits underneath all of this spending and rarely gets enough attention. Training large models at frontier scale and running inference at enterprise scale both consume substantial electricity. The major cloud providers are building data centers at a pace not seen in decades, and in many geographies the power infrastructure to support those centers is genuinely constrained. That constraint is not fully priced into most AI market analyses, and I think it will become a meaningful limiting factor for growth in the model segment faster than most 2026 forecasts assume.

As an IS researcher what I find most interesting about the 80.8 percent figure is not the size of the growth. It is the question of what IS research frameworks can tell us about which organizations will convert that spending into business value. The DeLone and McLean model was developed for evaluating information systems in organizations, and the questions it asks, does the system work reliably, are its outputs accurate, is the supporting service adequate, are exactly the right questions to ask about GenAI deployments. I am not sure the organizations driving the spending are asking those questions systematically. And if they are not, the productivity paradox Brynjolfsson identified in the 1990s for IT spending generally could repeat itself here: enormous investment, delayed or uneven value realization, and a long period of confusion about why the technology that seemed so capable is not showing up in the performance numbers.

---
claims_checked:
- "GenAI model spending growth 2026: 80.8%": "https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-forecasts-worldwide-it-spending-to-grow-9-8-percent-in-2026"
- "Total IT spending 2026: $6.31T": "https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-forecasts-worldwide-it-spending-to-grow-9-8-percent-in-2026"
- "McKinsey State of AI 2025: 88% use AI, 7% fully scaled": "https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai"
- "McKinsey GenAI economic potential: $2.6T-$4.4T annually": "https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai"
claims_unverified:
- "DeLone and McLean (2003) IS Success Model: drawn from IS theory background, not re-fetched this session"
- "Brynjolfsson productivity paradox: drawn from IS theory background, not re-fetched this session"
- "Frontier training cost details: described directionally from widely reported estimates, not a single primary source"
sources_used:
- "https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-forecasts-worldwide-it-spending-to-grow-9-8-percent-in-2026"
- "https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai"
word_count: 1070