GPU Market Pulse — Week of June 08, 2026

Week of June 08, 2026 — a synthesis of 7 daily investment briefs and the underlying blended pricing data.

The week's most kinetic move was K80, which closed 47.4% lower at $0.12/hr blended. Across all instruments, the median GPU rental rate declined 1.5% week-on-week.

Week in Briefs

Across 7 daily intelligence briefs this week, the analysis surfaced 36 investable hypotheses. Here's what built conviction.

Week at a Glance

Convictions That Built This Week

H100 — appeared in 6 briefs (Tue Jun 9, Wed Jun 10, Thu Jun 11, Fri Jun 12, Sat Jun 13, Sun Jun 14)

| # | Hypothesis | Conviction | Urgency | Primary Instrument | |---|---|---|---|---| | 1 | APAC H100 Spot Structural Premium | High | Medium | Long APAC reserved H100; Long APAC-positioned neoclouds | | 2 | Montreal Spot → On-Demand Convergence | High | Immediate | Lock reserved now; CRWV pricing-power validation | | 3 | AWS Inferentia Deprecation / Trainium2 Ramp | Medium-High | Medium | Long MRVL on Trainium2 confirmation | | 4 | Power-Sovereign vs Grid-Dependent Two-Tier Market | High | Slow-Build | Long NBIS, BE; Short APLD/IREN on margin compression | | 5 | H100 On-Demand

Framing across the week: The Montreal H100 Spot Surge is a Leading Indicator of a Permanent Secondary-Market Premiu; The H100 Depreciation Clock Is Running Faster Than Reserved Contracts Are Pricing In — A H; The Montreal H100 Spot Squeeze Is a Structural Demand Signal, Not a Recovery

Inferentia — appeared in 4 briefs (Mon Jun 8, Fri Jun 12, Sat Jun 13, Sun Jun 14)

This is the pricing fingerprint of a deliberate architectural transition: AWS is flooding Inferentia spot capacity as it migrates Inf1 workloads toward Inf2/Trainium2, while simultaneously maintaining Trainium spot scarcity to signal demand strength and protect on-demand pricing. Inferentia has been in catalog since December 2019 (6.5 years), and the CATALOG_SURVIVAL data confirms it remains "active" on AWS despite being deprecated at Azure and GCP — the definition of an end-of-life platform being drained of commercial viability.

Framing across the week: AWS Is Monetizing a Custom Silicon Monopoly — The Inferentia/Trainium SPOT Tightening Is D; H2 AWS Inferentia SPOT in EU: Utilization Surge Ahead of Policy-Driven Inference Buildou; _ CONFIRMED_

NVIDIA — appeared in 3 briefs (Mon Jun 8, Tue Jun 9, Thu Jun 11)

Thesis: The Korea buildout narrative (NVIDIA + SKT gigawatt DC, LG AI factory, NAVER expansion) has created a reflexive pricing anomaly: H100 RESERVED_1YR in kr-seoul is $9.55/hr — the highest reserved rate in the entire dataset, above Japan Osaka at $11.02/hr and significantly above US Virginia at $6.13/hr. Yet H100 SPOT in kr-seoul is only $2.69/hr with a 70% spot discount — implying the reserved premium cannot be explained by current demand. This divergence between reserved rates ($9.55/hr) and spot ($2.69/hr) is a 255% gap, the widest anywhere globally.

Framing across the week: Korea Is the Most Asymmetric APAC AI Infrastructure Play — First-Mover Compute Capacity Wi; Korea Is Becoming the World's Most Expensive Spot GPU Market — A Mispricing Trap for Naive; The AMD MI25 Collapse Signals a Broader Non-NVIDIA Accelerator Rationalization That Invest

AMD — appeared in 2 briefs (Tue Jun 9, Fri Jun 12)

Conviction: Medium | Horizon: 6–12 months ### Thesis The flood of AMD MI25 (Fiji, 2017) and M60 listings appearing simultaneously across 15+ global regions — all single-provider, all zero-spread, all in the new ON_DEMAND_DEV pricing tier — represents a specific provider (most consistent with Oracle Cloud's rapid global SKU expansion pattern) building a global budget GPU catalog without any price competition discovery. The pricing is incoherent: MI25 in Dublin is $0.019/hr, in Madrid it's $1.34/hr — a 70× spread for identical hardware across two EU regions.

Framing across the week: AMD's £2B UK Commitment Is a Sovereign AI Arbitrage Play — Cloud Pricing Divergence Signal; H3 AMD Legacy GPU Monetization: The Sub-$1/hr Market Has No Competitive Pricing Yet — An

Open Questions Into Next Week

• ** CONFIRMED — PARTIALLY** (raised Jun 13)
• ** CONFIRMED** (raised Jun 13)
• The APAC H100 Spot Arbitrage Is a Structural, Not Cyclical, Premium (raised Jun 14)
• Montreal's H100 Spot Spike Is the Canary for a North American Spot Supply Crisis — and 30 Days to On-Demand Convergence (raised Jun 14)

Market Snapshot

Blended GPU compute costs fell 1.5% this week across major cloud providers, led by K80's 47.4% decline.

Top Movers

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 $4.23/hr, -3.1% WoW), while inference-class pricing fell (avg $0.52/hr, -0.6% WoW).

The training-to-inference price ratio stands at 8.2x — narrowing 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: Asia Pacific

Asia Pacific trades at a 8% premium to global averages this week, with 32 GPU types available across the region. The most expensive GPUs in the region are GB200 ($14.86/hr), H200 ($10.20/hr), TRAINIUM2 ($5.93/hr). The 8% regional premium reflects regional demand dynamics.

For detailed pricing data across all Asia Pacific sub-regions, see the full regional profile.

Implications

For cloud buyers: Asia Pacific continues to offer the lowest average GPU pricing ($1.99/hr blended average). For workloads with regional flexibility, the gap between Asia Pacific and Middle East is $2.09/hr — a 105% premium. Compare regional pricing →

For semiconductor analysts: GPU pricing trends remain broadly stable this week. H100 (-0.6% WoW) and MI300X (-0.3% WoW) are tracking within normal ranges. Blackwell (B200) blended pricing at $7.53/hr (-1.1% WoW) provides an early read on next-generation adoption curves. View all GPU profiles →

For GPU investors: Declining rental rates are compressing returns — monitor payback period shifts closely. Model scenarios with the GPU ROI Calculator →