Enterprise AI Cost Curve: The 29% ROI Gap Explained

The Gap Everyone Misreads

Fifty-nine percent of organizations pour at least $1 million a year into AI. Twenty-nine percent report significant returns. The consensus reads that spread as failure.

The consensus has the frame wrong. This is a pricing artifact of the AI compute cost curve, and it dissolves on a schedule.

Enterprises bought foundation model access at 2024 prices while the cost of inference descends along a solar-like path. According to the WRITER 2026 enterprise AI adoption survey, 29% of companies see significant ROI, alongside individual productivity gains of 5X. That paradox tells you where value hides.

What The Consensus Watches Versus What Predicts

Analysts watch quarterly spend and quarterly return. They chart the ratio and call it a verdict. That ratio measures the present with precision.

The present is the wrong variable. The variable that predicts enterprise AI outcomes is the cost per unit of intelligence, and its second derivative.

Ninety percent of analysts read the current data correctly. They read the pace of change poorly. When input cost collapses across an 18-month window, every ROI calculation built on today's token price becomes obsolete before the contract renews.

Wall Street desks publish the pessimistic read and call it rigor. Those desks lag the curve. They price the present, and the present carries the least information about a market defined by falling marginal cost.

The AI Compute Cost Curve

Call something a trajectory when three data points anchor it. Inference cost for GPT-4 class performance has fallen sharply across successive model releases and hardware generations since 2023.

The pattern mirrors solar photovoltaics, where module cost dropped roughly 90% across the decade from 2010 to 2020, per IEA figures. Compute follows the same shape: each doubling of deployed capacity drives cost down a predictable percentage.

The curve dictates a clear conclusion. Generic and undifferentiated foundation model access becomes a commodity. Margin migrates. The layer that captures value shifts from the model to the vertical application with a proprietary data moat.

The Cliff Event

Adoption tends to jump rather than climb. Canadian business AI use tripled across two years, moving from 6.1% in Q2 2024 to 19.2% in Q2 2026, per Statistics Canada figures cited by Tech Insider.

Cliff event: GPT-4 equivalent inference reaches roughly $0.001 per token by the close of 2026. At that price, the ROI arithmetic inverts.

The 79% of firms reporting adoption challenges face a cost obstacle today. Once the input price crosses that threshold, the obstacle vanishes and the reported ROI share climbs. The wall was economic all along.

The mechanism is mechanical. Cheaper tokens let a workflow call the model ten times where budget once allowed a single call. Quality per task rises, error rates fall, and the productivity that workers already report converts into measured return.

Why The Company Is Tearing Itself Apart

Fifty-four percent of C-suite respondents said AI adoption is tearing their company apart internally, driven by fights over budget ownership and tool sprawl.

Those fights are a governance vacuum wearing a budget costume. When the cost of a model ran high, procurement centralized it. Central control created the turf war.

As the compute cost curve pushes prices toward zero, the calculus flips. AI governance stops being about rationing scarce capacity and becomes about routing abundant capacity toward proprietary data. Firms that reframe governance around data moats resolve the internal war. Firms that keep fighting over generic access keep bleeding.

The internal war reveals a deeper truth. Value was migrating while the org chart argued over an expiring asset.

The Canadian Signal

A separate RSM survey released in July 2026 found Canadian firms trailing American peers on both AI integration and measured returns.

The consensus reads that lag as a competitiveness problem. Read the curve instead. The lag is a timing gap, and timing gaps close fast when input costs fall for everyone at once.

Canadian adoption tripling across two years already signals the jump ahead. The measured-return gap reflects earlier contracts at higher prices, and it compresses as the compute cost curve delivers cheaper intelligence to laggards and leaders alike.

The lesson for any executive outside the United States: the window to build a data moat opens wider precisely because the model layer commoditizes uniformly across borders.

My Position

My position is direct. The enterprise AI ROI gap is a temporary artifact of the compute cost curve, and it closes as foundation models commoditize through 2026, shifting margin to vertical applications with proprietary data.

The evidence sits in the descent itself: three generations of falling token prices, the solar analogue, and adoption curves that jump rather than crawl.

What would change my mind? A structural floor under inference cost. Should token prices flatten for three consecutive quarters while hardware supply tightens, the commoditization thesis breaks and generic model access retains pricing power. I watch that floor closely.

Three Categories That Change Shape by 2028

Three enterprise categories that vanish in their current form:

• Generic AI capacity resellers, whose product becomes water
• Horizontal copilots with no data moat, absorbed into the platform layer
• Procurement teams locking multi-year contracts on generic access, stranded on obsolete pricing

For the CTO, the move is to rebuild the stack around a data moat now, before the price signal makes it obvious. For venture capital, the impossible-looking bet sits in vertical applications with proprietary data, priced as though the model layer stays expensive.

For the Chief Strategy Officer, any three-year plan that assumes today's token economics assumes a world about to disappear. For technology procurement, the danger runs concrete: a vendor lock on generic capacity becomes a liability the moment the market reprices the same capacity toward zero.

More market trajectories from this desk sit on the AGORA blog index.

The Prediction

Prediction: by the close of 2026, GPT-4 equivalent inference reaches roughly $0.001 per token, and the share of enterprises reporting significant AI ROI rises above 45%, up from the 29% the WRITER survey recorded.

Confidence: high on the technology, medium on the exact ROI figure. The compute cost curve carries the first half. Adoption lag introduces variance in the second.

Kill signal: token prices flatten across three consecutive quarters while the reported ROI share holds below 35%. That combination would falsify the pricing-artifact thesis and hand the argument back to the skeptics.