When Zillow’s artificial intelligence started overvaluing homes, the damage wasn’t just technical. The inflated “Zestimate” fed straight into its buying algorithm, leading the company to overpay for thousands of properties. By the end of 2021, Zillow had shut down its home-flipping business, taken an $881 million loss, and laid off 2,000 employees. A painful lesson in what happens when you scale AI without properly accounting for AI costs and modeling its true cost structure.

That lesson now plays out across every boardroom wrestling with their AI strategy. IBM’s Institute for Business Value found that enterprise computing costs are projected to surge 89 percent between 2023 and 2025, with seven in ten executives pointing to generative AI as the primary driver . For startups, the math is even worse. Kruze Consulting tracked their AI clients through a 300 percent year-over-year spike in compute and hosting costs, while traditional SaaS companies saw only 53 percent growth . Two years ago, these AI companies allocated a quarter of revenue to compute. Today, it consumes nearly half.

The problem is that compute costs are buried in plain sight. Manufacturers count steel by the ton. Retailers check inventory every morning. But most companies bury compute expenses deep in IT budgets, invisible to CFOs and boards. This blind spot distorts pricing models, margin analysis, and capital allocation decisions. It creates the myth that automation always saves money, while each model run and fine-tuning cycle quietly erodes profit.

Here’s the core issue: a boost in output is only half the equation. When compute costs grow faster than the value created, margins shrink even as revenue expands. This single dynamic will define the next phase of corporate AI adoption.

The impact varies by industry. In telecom and banking, the cost to serve is shifting from payroll to GPU cycles. In healthcare and insurance, compliance-driven computation is rising so fast that Deloitte now questions whether automation delivers net savings at all . In media and software, companies once celebrated for near-zero marginal costs now face rising cloud bills with every new user.

Efficiency gains are real. New model architectures, hardware improvements, and better orchestration can slash inference costs dramatically. But these gains are being overwhelmed by scale: the sheer number of models, queries, and retraining cycles multiplying across organizations. The efficiency improvements can’t keep pace with adoption rates.

For most organizations, there’s roughly a year before these costs become permanent fixtures in the cost structure. Once embedded, they’re extremely difficult to reverse. Early adopters learned this the hard way: when you build inefficiency into an algorithm, it compounds with every execution.

A small but growing group of finance chiefs is responding with operational discipline. Select Fortune 500 CFOs now track compute the same way they track physical inventory, using live dashboards that monitor usage patterns and unit economics in real time. Others treat model retraining like factory maintenance, with scheduled downtime and allocated budgets. Governance advisors like Spencer Stuart recommend that boards should either add directors with a broad understanding of AI or build an AI advisory board, in order to meet their fiduciary duty around the impact of AI strategies.

Most executives still celebrate small productivity gains without recognizing that the economics cut both ways. The companies that will win this era will think like operators, not just innovators. They will forecast GPU demand with the same precision a steelmaker applies to tonnage or a retailer brings to inventory turns. They will track AI adoption operationally first and financially second, comparing the two every quarter. They will know their actual cost per inference, per model, per customer.

For investors, this goes beyond budgeting details. Compute inflation will show up in quarterly earnings as margin compression, especially in data-heavy sectors like insurance, healthcare, and logistics. The cost of compute has become a new form of energy exposure, moving with GPU supply chains, cloud pricing, and energy markets.

The irony is stark. Many organizations adopted AI specifically to cut costs and boost efficiency. Yet as models grow more powerful and complex, those savings are being consumed by the very infrastructure designed to deliver them. Imagine if the industrial revolution had somehow made factories less efficient with each new steam engine.

But we now have the data to get ahead of the problem, which makes this moment different from Zillow’s failure. The companies building sustainable AI operations today are proving that compute costs can be managed, not just endured. They’re running smaller, task-specific models instead of deploying massive general-purpose systems for every use case. X They’re caching frequent queries, batching inference requests, and setting hard budget limits that force teams to optimize before they scale. They’re asking a simple question before every new AI project: does the value created justify the compute consumed?

Speed up financial literacy around AI instead of slowing down adoption. Organizations that treat compute as a strategic resource rather than an IT expense will build AI systems that actually improve margins instead of eroding them. The tools exist. The frameworks are proven. What’s missing is the mindset shift from viewing AI as magic to understanding it as manufacturing.

Zillow’s expensive education taught a clear lesson, and the opportunity has only expanded since then. Compute costs have moved from invisible line item to manageable input that, when properly measured and optimized, can deliver the productivity gains AI promised from the start.

What separates tomorrow’s AI leaders from tomorrow’s cautionary tales won’t be who moves fastest. It will be who moves with their eyes open to the true cost of intelligence at scale and builds accordingly. A competitive advantage waiting to be claimed in a moment when most competitors are still blind to the economics, creating wide open space for those who see clearly.

By Chiara Marcati, Chief AI Advisory & Business Officer, ai71

Related post

View all