The economics underpinning the AI boom are starting to wobble, and the numbers are getting harder to defend with a straight face. According to The Information, the margin math behind tech’s AI bets is getting messier as hyperscalers and model labs pour record sums into chips, data centers, and power deals while revenue growth struggles to keep pace with the capital burn.
This is the story that’s been hiding behind every glowing earnings call. Microsoft, Google, Amazon, and Meta have collectively committed hundreds of billions in AI capex over the past 18 months. The pitch to Wall Street was simple: spend big now, win the platform shift, print money later. The Information’s reporting suggests that ‘later’ keeps moving, and the spreadsheet is starting to creak.
What’s actually changing
Three forces are squeezing the math at once:
- Compute costs aren’t falling fast enough. Nvidia’s pricing power is still intact, and the next-gen Blackwell and Rubin systems carry premium price tags. Even with efficiency gains per token, total spend per customer keeps climbing because usage climbs faster.
- Model prices keep dropping. API pricing for frontier models has fallen roughly 10x over two years. Great for buyers, brutal for the gross margin line of anyone selling inference.
- Depreciation is showing up. Those GPU clusters bought in 2023 and 2024 are now hitting the income statement. Useful lives on AI hardware are being debated, and any shortening of that window slams reported earnings.
What stands out here is the gap between top-line AI revenue (which everyone loves to quote) and the contribution margin once you back out hosting, power, and amortized silicon. The Information’s framing suggests investors are finally asking the second question.
Two camps forming
On one side: the believers. Satya Nadella, Sundar Pichai, and Mark Zuckerberg keep arguing that underinvesting is the bigger risk than overinvesting. Their bet is that AI becomes the operating layer for every enterprise workflow, and whoever owns the substrate captures decades of rent.
On the other side: a growing chorus of analysts and short sellers pointing out that circular revenue (Nvidia invests in CoreWeave, CoreWeave buys Nvidia chips, OpenAI buys CoreWeave compute, Microsoft funds OpenAI) makes the whole ecosystem look healthier than the underlying cash flows suggest. When the same dollar shows up in three different companies’ revenue lines, margin analysis gets fuzzy fast.
Why this matters now
The timing is what makes this significant. We’re heading into a stretch where:
- 2024-2025 capex commitments start landing as 2026 depreciation
- Enterprise AI deployments are being scrutinized for actual ROI, not just pilot enthusiasm
- Power and grid constraints are forcing longer build cycles and higher all-in costs per megawatt
- Sovereign AI deals and Gulf-state funding are propping up demand, but at terms that may not flatter unit economics
If reported AI revenue growth slows even slightly while depreciation accelerates, the operating margin compression will be visible to anyone with a calculator.
What practitioners should do
For founders and operators building on top of this stack, the takeaways are practical:
- Don’t assume API prices will keep falling at the same rate. The race to the bottom on inference pricing only works while hyperscalers are willing to subsidize it. Model your worst case at current prices.
- Diversify model providers. Single-vendor lock-in is more dangerous when the vendor’s margin structure is under pressure and pricing power may return.
- Watch open-source carefully. If hyperscaler margins compress, open-weight models running on commodity infrastructure become more attractive for cost-sensitive workloads.
- Build cash-generative products, not demos. The era of burning compute to chase user growth without revenue is closing fast.
The AI build-out isn’t stopping. But the next 12 months will reveal which companies were building a business and which were building a story. The full breakdown of the numbers is in The Information’s reporting.