When AI Growth Looks Like a Bubble: Revenue, Costs, and Adoption Compared to the Dot‑Com Era

How AI revenue growth behaves in the wild

Headline AI revenue numbers tell a seductive story—300% year-over-year growth isn't uncommon. But peel back the layers and you'll find most of that growth comes from three sources: pilot programs (typically 6-12 month engagements), usage spikes from a handful of large customers, and professional services disguised as product revenue.

In practice, this means AI firms report SaaS-like recurring revenue while actually running what amounts to a consulting business with worse margins. The Stanford AI Index found that among companies reporting AI revenue gains in 2024, the most common impact was just 5% uplift—far below the triple-digit growth touted in press releases.

What breaks first: customer concentration. We've seen AI providers where 60% of Q2 revenue came from two enterprise clients scaling up temporary model testing. When those projects conclude or budgets tighten, the revenue cliff arrives faster than sales can backfill.

Takeaway: Track paying customer count and revenue dispersion alongside growth rates—concentration risk appears long before financials show it.

'300% growth' often means '3 customers experimenting'

Where profitability disappears: the hidden cost centers

Gross margins tell the real story. While software companies routinely achieve 70-80% margins, AI providers average just 30% after accounting for: - Continuous fine-tuning (typically 15-20% of COGS) - Human-in-the-loop validation (8-12%) - Customer-specific infrastructure (varies wildly, but often 25%+ for regulated industries)

The killer is that these costs scale with usage. Unlike software where marginal costs approach zero, every additional AI inference burns more compute. IBM's 2025 data shows enterprise computing costs jumped 89% in two years—70% of surveyed executives cited generative AI as the primary driver.

What breaks first: unit economics. We've modeled cases where a 10% increase in customer usage drops margins from 35% to -2%. This isn't theoretical—it's why 42% of companies abandoned most AI initiatives in 2025 when they ran the real numbers.

Takeaway: Model margin erosion at 110% and 150% of forecast usage—the breakpoints often surprise teams.

AI's dirty secret: more usage can mean less profit

Modeling AI as a capital‑intensive service, not pure software

AI infrastructure behaves like industrial plants, not software. Fixed training capex grows 60% YoY, while variable serving costs consume 23% of revenue—a ratio unseen since telecom buildouts in the 1990s. Hyperscalers now spend $660B-$690B annually on AI/cloud infrastructure, with projections hitting $1.6T by 2031.

The trap: assuming economies of scale will materialize. In reality, serving costs remain stubbornly variable. Top providers now allocate 34% of revenue to capex versus 15% for 1990s tech firms. This explains why Big Tech's cash reserves ($490B in Q3 2025) matter more than their AI revenue growth.

Takeaway: Compare capex/revenue ratios to dot-com era benchmarks—when they exceed 20%, growth becomes capital-constrained.

AI runs on steel, not code

How adoption frictions turn pilots into dead weight

Enterprise AI adoption looks like a leaky pipe: 88% of organizations now use AI in some function, but only 5.5% report meaningful EBIT impact. The attrition happens in three phases: 1. Pilot enthusiasm (0-6 months) 2. Integration quagmire (months 6-18) 3. Measurement crisis (months 18-24)

IBM's 2025 research found 95% of generative AI pilots fail to deliver promised impact. The killers are mundane: data quality issues (42% of cases), workflow incompatibility (33%), and inability to measure ROI (25%).

What breaks first: procurement cycles. We've tracked deals where enthusiastic POCs got approved, only to stall for 9+ months on data governance reviews. By then, the champion often leaves, budgets shift, and what looked like an imminent rollout becomes shelfware.

Takeaway: Map your customer's internal approval chain before celebrating a pilot win—the real sale happens 12-18 months later.

95% of AI pilots fail—not on technology, but on paperwork

Repricing the market: what a disciplined comparison to dot‑coms implies

Valuation multiples now cluster in two groups: firms trading at 8-12x revenue with >50% gross margins and 20%+ pilot conversion rates, and those at 25x+ with sub-30% margins and <10% conversion. The latter group mirrors dot-coms that traded at 10x sales while burning 50% of revenue on capex.

The signal: watch capex/revenue ratios. When Big Tech's AI investments hit 0.8% of GDP (projected for 2027), historical parallels suggest a repricing. The 1999-2000 correction began when telecom capex reached 1.2% of GDP—a threshold where capital intensity overwhelms growth narratives.

Takeaway: Monitor the capex/GDP ratio—crossing 0.8% triggers reassessment of growth assumptions in capital-intensive markets.

Bubbles pop when capex becomes a percentage point of GDP