The consensus analyst forecast for AI capex in 2025 was off by $200 billion. Not a rounding error. Not a reasonable miss. Off by an order of magnitude.
Azeem Azhar argues the pattern isn't new - and it isn't random. Analyst forecasts on AI have been systematically wrong because the models they're using were built for a different kind of market.
The Error Isn't Pessimism
This isn't about analysts being too conservative or missing the hype. It's structural. The tools analysts use to model markets - linear regression, mean reversion, historical comparables - assume the future looks like the past, just slightly more so.
That works for industries in steady state. It breaks completely in exponential regimes.
When a technology is improving exponentially, the curve doesn't fit a straight line. When adoption is doubling year-on-year, historical averages become irrelevant. When the companies driving the market are private and don't publish quarterly earnings, your data is incomplete from the start.
Analysts are fitting exponential curves to linear models. The result is forecasts that look reasonable on paper and collapse on contact with reality.
The Private Company Problem
Azhar points to another gap: the most important companies in AI don't report public financials.
OpenAI is private. Anthropic is private. Mistral is private. xAI is private. Perplexity is private. Most of the labs doing frontier model work are private.
That means analysts are building forecasts from incomplete data. They can see Microsoft's Azure revenue, Google Cloud's AI contribution, NVIDIA's GPU sales. But they can't see what OpenAI is spending, what Anthropic is raising, or what the next wave of model training will cost.
They're estimating a market where the core players don't publish numbers. It's like trying to forecast the smartphone market in 2010 if Apple didn't report iPhone sales.
The result is forecasts anchored to what's visible (cloud provider revenue, chip sales) rather than what's actually happening (the capex explosion happening inside private labs).
Mean Reversion When the Mean Is Irrelevant
The other failure mode Azhar identifies: mean reversion models assume things return to average. Markets overshoot, then correct. Outliers regress to the mean.
But in an exponential market, the mean is always wrong. Last year's "outlier" growth becomes this year's baseline. What looked like overshoot was actually the new trajectory.
Analysts see explosive capex growth, assume it's unsustainable, and model a correction. Then the correction doesn't come. Because the market isn't overshooting - it's accelerating.
This is what happened with cloud infrastructure in the 2010s. Analysts repeatedly forecast AWS growth would slow. It didn't, because the entire category was expanding faster than the models could accommodate. The same dynamic is playing out now with AI infrastructure.
Why This Matters for Decision-Makers
If you're a business leader relying on analyst forecasts to plan AI strategy, you're working from data that's systematically lagging reality.
That doesn't mean analyst reports are worthless. But it does mean you can't treat them as ground truth. The forecasts are rear-view mirrors, not windshields.
The companies that understand this - that recognise we're in an exponential regime and plan accordingly - are the ones making bets that look too aggressive to consensus analysts and turn out to be correct.
The ones anchoring to consensus forecasts are the ones that will be surprised in twelve months when the numbers are off by another order of magnitude.
What Actually Works
Azhar doesn't just critique the models - he suggests what to watch instead.
Track capex commitments from the companies actually building infrastructure. When Microsoft, Google, and Amazon all raise their data centre budgets by 50%, that's not noise - that's signal.
Watch model improvement curves, not revenue forecasts. If models are improving faster than expected, demand will follow. If they plateau, the market corrects.
Pay attention to adoption velocity in real use cases. How fast are developers integrating AI into production systems? How many companies are moving from pilot to deployment? These are leading indicators, not lagging ones.
And critically: assume the forecasts are wrong. Not wrong in detail - wrong in structure. Plan for scenarios where the market moves faster than consensus expects, because that's what keeps happening.
The AI market isn't behaving like a normal market because it isn't one. The sooner decision-makers stop treating it like one, the better their strategies will hold up against what's actually coming.