ACIS · AI INFRASTRUCTURE · WEEKLY · ISSUE 004
Orders Are Still Surging, but AI Infrastructure Is Entering a Delivery-and-Funding Shakeout
AI Infrastructure Weekly | Issue 004 | 2 September 2026
Fresh data across chips, servers, cloud, power and cooling confirm continued expansion. The new dispersion is not in orders but in delivery timing, funding structure and free-cash-flow conversion.
AI BRIEFING · ISSUE 004
The core view in 90 seconds
Demand · Delivery · Funding · Action
Healthy Expansion
Score held at 81
EXECUTIVE SUMMARY
Healthy Expansion holds; delivery and funding move to the first line of analysis
We keep the ACIS AI Infrastructure score at 81/100. NVIDIA, Dell, Microsoft and Vertiv provide further evidence of strong demand and spending, insufficient for a downgrade to Balancing or Overcapacity. CoreWeave's large backlog alongside losses shows that orders do not automatically become shareholder free cash flow.
Orders create revenue visibility; delivery and funding determine cash-flow quality.
01 · EVIDENCE CHECK
Five fresh data points, one consistent conclusion
| Node | Confirmed fact | Research implication |
|---|---|---|
| NVIDIA | Data Center revenue $89.0B, +117% YoY | Compute demand is still accelerating |
| Dell | $60.9B AI-server orders; $95B ending backlog | Order visibility strengthened materially |
| Microsoft | Next-quarter CapEx expected above $50B | Capacity and finance leases remain key variables |
| CoreWeave | Revenue backlog ~$104B; adjusted net loss $567M | Demand is exceptional; capital structure needs a separate test |
| Vertiv | Sales +24% YoY; adjusted FCF $925M | Power and cooling are converting demand into cash |
Figures are company disclosures; research implications are ACIS Research inferences.
02 · TRANSMISSION
Six layers: from silicon to monetizable compute
NVDA, AMD, custom ASICs and AVGO: test accelerator demand, silicon economics and usable compute.
ANET, AVGO, COHR and LITE: test scale-up, scale-out and optical connectivity bottlenecks.
MU, HBM, DRAM and NAND: track bandwidth, supply discipline, pricing and the memory cycle; MU is P1 Watch / No Chase.
CRWV, NBIS, ORCL and hyperscalers: test delivered capacity, utilization and contract quality.
VRT, ETN and GEV: test grid access, power delivery, thermal limits and time to service.
ARES, KKR, APO and private credit: test cost of capital, leverage, refinancing and cash conversion.
03 · RISK FRAMEWORK
Not a blanket bearish call—stop treating all AI infrastructure as one risk
- Prioritize proof: Backlog conversion, on-time delivery, utilization and free-cash-flow improvement.
- Funding separation: Separate net-cash suppliers, investment-grade platforms and refinancing-dependent neoclouds.
- Valuation discipline: A strong industry trend does not make every entry price safe; stage additions and wait for pullbacks.
DOWNGRADE RULEDowngrade the cycle only if at least three of orders/backlog, hyperscaler CapEx, delivery utilization and pricing weaken together. If funding worsens without demand deterioration, downgrade leveraged nodes rather than the whole industry.
04 · GLOSSARY
Terms used in this report
Contracted business not yet recognized as revenue; still subject to delivery and service availability.
A new cloud provider centered on GPU capacity for AI workloads.
A long-term lease that finances data-center assets and creates interest and balance-sheet obligations.
The ability to turn accounting profit into operating and free cash flow.
05 · SOURCES
Primary sources
The score is held rather than mechanically raised: stronger demand evidence is offset by delivery, funding and valuation constraints. Not investment advice.