Week of May 25, 2026 — a synthesis of 4 daily investment briefs and the underlying blended pricing data.
The week's most kinetic move was TRAINIUM, which closed 11.6% lower at $0.74/hr blended. Across all instruments, the median GPU rental rate was unchanged 0.3% week-on-week.
Week in Briefs
Across 4 daily intelligence briefs this week, the analysis surfaced 21 investable hypotheses. Here's what built conviction.
Week at a Glance
| Date | Headline Hypothesis |
|---|---|
| Tue May 26 | The B200 "Demand Confirmation" Trade: Ohio Validates, Virginia Diverges |
| Wed May 27 | The H200→B200 Pricing Stall Is a B200 Demand Disappointment Signal, Not a Soft Landing |
| Fri May 29 | GCP IS FORCE-UPGRADING ITS IOWA FLEET, AND LEGACY GPU HOLDERS WILL GET SQUEEZED ON BOTH SIDES |
| Sun May 31 | AWS ASIC Spot Tightening Is the Leading Indicator for a Reserved-Instance Re-Pricing Event |
Convictions That Built This Week
Inferentia — appeared in 3 briefs (Tue May 26, Wed May 27, Fri May 29)
This geographic bifurcation is highly informative: Virginia Inferentia capacity is being vacated as AWS routes production inference workloads to newer Inferentia2/Trn1 instances, while the EU and APAC regions are seeing gen-1 Inferentia spot prices tighten as local demand exceeds available capacity reallocation. The key insight is that Inferentia gen-1 is a 6.5-year-old ASIC (launched December 2019) that should be in secular decline — yet it's posting the strongest spot price momentum of any asset class in the dataset outside of B200.
Framing across the week: The Inference ASIC Scarcity Trade: AWS Inferentia Spot Is a Demand Proxy for Production AI; EU ASIC Inference Is Repricing Higher Structurally; Asia Is Flushing; This Is a Policy Sto; AWS INFERENTIA'S HONG KONG COLLAPSE IS THE FIRST PRICING ADMISSION THAT TPU V6E IS WINNING
H100 — appeared in 3 briefs (Tue May 26, Fri May 29, Sun May 31)
EU/Canada markets (Stockholm +42.4%, Madrid +53.1%, Montreal +57.8%, Amsterdam +18.5% over 30 days) are tightening sharply, while core US (Oregon -9.1%, Dublin -49.8%, Virginia +6.3%) is loosening or flat. This bifurcation implies: (a) regulatory/permitting delays are hitting European H100 supply more than US, and (b) neocloud players who own EU colocation capacity are sitting on pricing power they haven't yet exercised on their 1-year RIs.
Framing across the week: The European GPU Supply Fracture: H100 Market Structure Collapse as an Early Warning of Po; EUROPE'S H100 SPOT MARKET IS FRAGMENTING INTO TWO TIERS, AND THE AMSTERDAM–DUBLIN DIVERGEN; _THE NEOCLOUD EQUITY RALLY IS OUTRUNNING THE UNDERLYING PRICING DATA, CREATING A VALUATION _
A100 — appeared in 2 briefs (Wed May 27, Fri May 29)
The endgame is a bifurcated fleet where frontier GPUs command premium rates and anything older becomes economically unviable to hold. Any neocloud or enterprise that locked in A100-class GPU clusters at 2023–2024 contract prices faces margin compression from two directions: customers migrating to H100/H200, and spot market comps for legacy hardware collapsing.
Framing across the week: The H200→B200 Pricing Stall Is a B200 Demand Disappointment Signal, Not a Soft Landing; GCP IS FORCE-UPGRADING ITS IOWA FLEET, AND LEGACY GPU HOLDERS WILL GET SQUEEZED ON BOTH SI
Open Questions Into Next Week
- GCP IS FORCE-UPGRADING ITS IOWA FLEET, AND LEGACY GPU HOLDERS WILL GET SQUEEZED ON BOTH SIDES (raised May 29)
- EUROPE'S H100 SPOT MARKET IS FRAGMENTING INTO TWO TIERS, AND THE AMSTERDAM–DUBLIN DIVERGENCE IS AN ARBITRAGE SIGNAL (raised May 29)
- AWS INFERENTIA'S HONG KONG COLLAPSE IS THE FIRST PRICING ADMISSION THAT TPU V6E IS WINNING THE ASIC INFERENCE RACE IN ASIA (raised May 29)
- H100 SPOT Is Bifurcating: EU/Canada Signal New Capacity Starvation While Core US Loosens (raised May 31)
Get the Weekly Pulse
This analysis is part of Signwl's weekly research. Subscribe free — Tuesday delivery, no spam.
Market Snapshot
Blended GPU compute costs remained stable 0.3% this week across major cloud providers, led by TRAINIUM's 11.6% decline.
Top Movers
| GPU | Blended Price | WoW Change | Class |
|---|---|---|---|
| TRAINIUM | $0.74/hr | ▼ 11.6% | Training |
| ALVEO_U30 | $0.19/hr | ▲ 5.3% | General |
| TPU_V5LITEPOD | $0.17/hr | ▼ 2.4% | General |
| TPU_V5P | $1.06/hr | ▼ 1.9% | Training |
| B200 | $7.96/hr | ▼ 1.8% | Training |
| L4 | $0.45/hr | ▼ 1.6% | Inference |
| V100 32GB | $1.56/hr | ▲ 1.6% | Training |
Blended pricing = average of spot, on-demand, and 1-year reserved rates across major cloud providers.
Training vs Inference
Training-class GPU pricing fell this week (avg $4.27/hr, -0.8% WoW), while inference-class pricing held steady (avg $0.52/hr, -0.3% WoW).
The training-to-inference price ratio stands at 8.2x — stable compared to last week. The elevated spread suggests strong demand for training compute relative to inference, consistent with ongoing large model training activity.
Regional Spotlight: North America
North America trades at a 10% discount to global averages this week, with 38 GPU types available across the region. The most expensive GPUs in the region are GB200 ($12.47/hr), GB300 ($12.32/hr), B300 ($10.31/hr). The 10% discount makes North America one of the more cost-effective regions for GPU deployment this week.
For detailed pricing data across all North America sub-regions, see the full regional profile.
Implications
For cloud buyers: Asia Pacific continues to offer the lowest average GPU pricing ($2.04/hr blended average). For workloads with regional flexibility, the gap between Asia Pacific and Middle East is $2.06/hr — a 101% premium. Compare regional pricing →
For semiconductor analysts: GPU pricing trends remain broadly stable this week. H100 (+0.0% WoW) and MI300X (+0.0% WoW) are tracking within normal ranges. Blackwell (B200) blended pricing at $7.96/hr (-1.8% WoW) provides an early read on next-generation adoption curves. View all GPU profiles →
For GPU investors: Stable pricing supports predictable returns for existing deployments. Model scenarios with the GPU ROI Calculator →