The AI infrastructure story is quietly shifting from “how much are they spending” to “how fast does it pay back” — and the first hard numbers are starting to answer that question. Nebius says its Q2 AI cloud deals averaged more than $20 million in annual contract value per megawatt, with four landmark agreements landing between $20 million and $25 million per MW, and an estimated payback period of roughly 1 year and 10 months. That’s the most concrete data point yet in a debate that has mostly run on vibes and capex headlines.
Why it matters
For two years, the AI infra story has been told almost entirely through spending totals — bigger data centers, bigger chip orders, bigger capex lines. That framing is starting to crack. Investors and builders alike are now asking the follow-up question: is this infrastructure actually generating revenue, or just sitting there depreciating? Nebius’ disclosure — annual contract value per MW, prepayment coverage, payback timeline — is one of the first attempts to answer that in concrete unit economics rather than aggregate dollars. Whether this holds up as a durable metric or turns out to be a favorable snapshot from one vendor is exactly the kind of thing worth tracking over multiple quarters, not just reporting once.
The numbers so far
| Metric | Figure | Source |
|---|---|---|
| Nebius Q2 avg. annual contract value per MW | >$20 million | Nebius disclosure |
| Top 4 Nebius deals, ACV per MW | $20M–$25M | Nebius disclosure |
| Nebius deals with prepayment | ~70%, covering 50%–60% of capex | Nebius disclosure |
| Estimated payback period | ~1 year, 10 months | Nebius disclosure |
| Oracle cloud infra revenue growth (YoY) | +121%, to $7.4B | Oracle reporting summary |
| Oracle GPU utilization | 98% | Oracle reporting summary |
| Q2 cloud capex (major players) | $54.21B, +68.44% YoY | Cloud capex summary |
Key developments
- Revenue per MW is becoming the metric to watch. Nebius’ $20M+ per MW figure gives the market its first vendor-disclosed benchmark for what an AI cloud deal is actually worth on a capacity basis, rather than just total contract size.
- Prepayments are doing real work. With ~70% of deals including prepayment covering half to 60% of associated capex, Nebius is effectively getting customers to co-fund the buildout — a structural detail that matters more than the headline contract value.
- Utilization claims are entering the narrative. Oracle’s reported 98% GPU utilization alongside 121% cloud infrastructure revenue growth to $7.4 billion suggests some hyperscalers are starting to pair capex spend with usage evidence, not just growth guidance.
- Aggregate capex is still enormous. One recent summary put Q2 2026 capex across major cloud players at $54.21 billion, up 68.44% year over year — the spending race hasn’t slowed, it’s just being asked to justify itself.
Living data: trackers are becoming the format, not the footnote
The same shift toward evidence over narrative is showing up in AI accountability journalism generally. One litigation tracker says it follows 113 major AI lawsuits daily; a separate enforcement tracker lists 263 tracked actions tied to official sources. Neither is a one-off story — both are designed to be revisited weekly as filings, rulings, and enforcement actions accumulate. That’s the same instinct behind treating AI capex and cloud GPU economics as a living dataset rather than a single earnings-day headline: the interesting signal is the trend line, not the snapshot.
What it means for builders
If you’re shipping products on top of rented AI infrastructure, these numbers matter more directly than they might seem. A vendor claiming sub-two-year payback and 98% utilization is telling you something about pricing stability: infrastructure that pays for itself quickly and runs near capacity is less likely to see abrupt price hikes or capacity throttling than infrastructure still burning cash on low utilization. Watch for repricing risk from vendors whose capex-to-revenue ratio looks unsustainable — that’s often where surprise cost increases or service degradation shows up six to twelve months later. Track the same figures your provider is disclosing: contract value per MW, utilization rate, and payback period, and ask for them directly if they’re not public.
The honest counter-argument
These figures are self-reported by the companies with the most incentive to make AI capex look justified. Nebius’ payback estimate assumes current contract terms hold and doesn’t account for GPU depreciation cycles or the possibility that early landmark deals are priced above what the broader market will bear. Oracle’s 98% utilization figure, similarly, is a single-quarter snapshot without independent verification of methodology. None of this means the numbers are wrong — it means they’re the opening data points in a story that needs several more quarters of disclosure before the capex-versus-revenue debate is actually settled.
LLM Watch will keep tracking AI capex vs. revenue, cloud GPU economics, and litigation/enforcement trackers as new figures land.