What Happened
In early May 2026, a tweet about the Neocloud GPU rental market sparked discussion in the AI infrastructure community:
“No Neocloud ever imagined they’d be renting out H100s today at higher prices than 3 years ago.”
This tweet, which received 432 likes and 50K+ views, revealed a counterintuitive market phenomenon: against the backdrop of expanding GPU capacity and intensifying AI chip competition, H100 rental prices are rising rather than falling.
Why GPU Rentals Are Getting More Expensive
The logic behind this can be broken down into three layers:
1. Demand Side: Frontier Firms Lock Orders Early
Even if you have funding, frontier labs and Neolabs have already locked up most of the 2026 GPU supply. Hyperscalers (Microsoft, Google, Meta, Amazon) are projected to spend $725 billion on AI capital expenditure in 2026, up 77% year-over-year. These massive orders are taking priority in NVIDIA’s capacity allocation.
2. Supply Side: HBM Capacity Bottleneck
GPUs are not standalone chips—they require HBM (High Bandwidth Memory). Micron’s CEO previously confirmed that 2026 HBM supply is completely sold out, meeting only 50-65% of customer demand. The HBM capacity bottleneck directly limits total GPU output.
| Bottleneck | Current Status | Impact |
|---|---|---|
| GPU wafers | NVIDIA capacity continuously expanding | Not the main bottleneck |
| HBM memory | Three manufacturers capacity locked | Core bottleneck |
| CoWoS packaging | TSMC capacity tight | Secondary bottleneck |
| Power and cooling | Data center siting constraints | Long-term constraint |
3. Market Structure: Neocloud Pricing Power Lost
The Neocloud (emerging cloud computing providers) business model was originally “bulk purchase GPUs, retail at prices below AWS/Azure.” But in a tight supply market:
- Neoclouds cannot get enough GPUs to build scale effects
- Even when they do get GPUs, rising HBM costs push up overall TCO
- Demand far exceeds supply, Neoclouds have no motivation to lower prices
Where Every $1M of AI Spending Flows
Understanding why GPU rents are rising requires looking at the overall AI infrastructure spending structure:
| Spending Category | Amount (per $1M) | Percentage | Key Suppliers |
|---|---|---|---|
| GPUs and accelerators | $520K | 52% | NVIDIA, AMD, Broadcom |
| Networking and optical | $150K | 15% | Arista, Coherent |
| Data center infrastructure | $200K | 20% | Power, cooling, racks |
| Memory and other | $130K | 13% | Micron, SK Hynix, Samsung |
Over half of AI investment flows to GPUs and accelerators—whoever controls the GPU supply chain controls AI infrastructure pricing power.
Landscape Assessment
Rising GPU rents reveal a deeper trend: AI compute is returning from “commoditization” to “scarcity.”
In 2023-2024, the market generally expected that with GPU capacity expansion and increased competition, compute costs would drop rapidly. But the reality is:
- Demand growth far outpaces capacity growth
- Supply chain bottleneck shifted from GPU to HBM
- Early order locking by frontier players squeezes smaller players’ access
This means the “compute democratization” narrative faces challenges in 2026—having GPU access is becoming a competitive advantage, not infrastructure.
Actionable Recommendations
- AI entrepreneurs: Do not assume GPU rents will naturally fall. Calculate inference costs in your business model at current or higher prices, reserving supply risk buffer.
- Model selection strategy: When GPU costs are high, choosing models with higher inference efficiency (such as MoE architecture, quantized versions) makes more economic sense than pursuing the largest parameter scale.
- Watch alternatives: AMD MI300 series, Google TPU, AWS Trainium and other alternative chips’ availability and cost-effectiveness are improving. Multi-chip strategy can reduce dependency on a single supplier.
- Long-term perspective: HBM capacity expansion cycle is 18-24 months. 2026 supply tightness may begin easing in the second half of 2027, but only if all major manufacturers’ expansion plans proceed as scheduled.