A rapidly growing bill with almost zero visibility. That’s the paradox Gartner lays out: by 2028, companies may spend more on AI coding assistants than they pay the developers who use them. It’s a number that reshuffles the return on investment of tools like GitHub Copilot, Codeium or Cursor, too often perceived as low-cost commodities.

The hidden spending sink

At the core of the issue isn’t just per-user license pricing. The Gartner analysis underscores that most organizations “cannot even see what they are spending.” Pricing models have fractured into per-user subscriptions, token-based consumption, and premium tiers for advanced features. Combined with wide and often ungoverned adoption, these costs slip past traditional IT control points.

Many companies realize too late they’ve rolled out services at scale without clear cost attribution. AI coding turns into an operational expense that inflates monthly cloud bills, mirroring the dynamics already seen in the shift from on-premise datacenters to pay-as-you-go.

What on-premise changes in the equation

For those evaluating self-hosted LLM stacks, the calculation shifts sharply. Inference costs on dedicated hardware (GPUs with adequate VRAM, owned servers) don’t follow per-token metering: TCO moves toward upfront CapEx and energy, but delivers predictability and full control over data. The trade-off is familiar: higher initial investment, fewer end-of-month surprises.

It’s not just about the bill. In regulated environments or where sensitive data is involved, self-hosting becomes a sovereignty lever that’s hard to ignore. Cost analysis, however, must include maintenance, model updates, and the need for in-house expertise. AI-RADAR provides analytical frameworks at /llm-onpremise for anyone wanting to compare deployment scenarios beyond the price of a single subscription.

Transparency first

Gartner’s wake-up call isn’t only for CFOs. It’s for CTOs and engineering teams pushing generative AI adoption without building a shared cost metric. Ungoverned growth risks eroding the competitive advantage: if every code suggestion carries an invisible cost, productivity measured in lines of code could be deceptive.

Looking toward 2028, the message is broader. AI coding has entered a maturity phase where price is no longer a minor detail. For companies, tracking spend, assessing real impact, and, where it makes sense, considering alternative architectures that bring cost levers back under control has become urgent. Without that discipline, the risk is the bill overtaking the salary — and nobody noticing.