Daily Investment Brief — May 26, 2026

Compute Market Intelligence Daily Brief

May 26, 2026 | Cross-Provider Pricing, SEC, News & Regulatory Intelligence

Market Pulse

The compute market is bifurcating along two fault lines simultaneously: generational displacement and geographic fragmentation. On the generational axis, Blackwell spot capacity in US core markets is being absorbed at an accelerating rate — Ohio B200 spot has recovered from a February trough of $1.51/hr to $5.70/hr today, a 4x move in 100 days, while its spot discount has compressed 31 percentage points to 56.8%, the cleanest real-money demand signal in the current dataset. On the geographic axis, European H100 infrastructure is undergoing a structural price collapse that has no parallel anywhere else: Frankfurt H100 on-demand at $1.78/hr and Amsterdam at $0.43/hr sit 85–97% below the $8–11/hr global median, and the Frankfurt market has developed a structural inversion where spot ($2.45/hr) now clears 38% above on-demand — a textbook sign of a provider pricing at marginal cost to prevent utilization collapse. The dominant narrative is supply-demand decoupling by geography and generation: Blackwell scarcity in the US is happening simultaneously with Hopper abandonment in Europe, driven by the compounding pressure of B200 generational displacement, EU power-and-permitting constraints, and early evidence of inference workload migration pulling demand away from legacy training tiers.

Key Movers

Investable Insights

H1 — Blackwell Demand Absorption Is Real, Regional, and Accelerating

Confidence: 4 / 5

Thesis: AWS's B200 spot market is the only genuine real-money Blackwell demand proxy in existence — Azure and GCP both show B200 as inactive in catalog survival data, meaning every spot bid is a live training workload placing a direct monetary bet on Blackwell capacity. After a -58% drawdown from launch pricing to the February trough ($1.51/hr), the Ohio fleet has absorbed demand at an accelerating rate, reaching $6.35/hr on May 20 before easing to $5.70/hr as of today — still representing a 31 percentage-point compression in spot discount (from a max ~87.9% to 56.8%). The $12.83/hr on-demand anchor has not moved a basis point across all three US B200 regions, which is critical: price movement is entirely demand-driven, not a provider repricing event. The regional divergence — Ohio/Oregon surging while Virginia declined -24.7% over 30 days — is not contradiction; it is sequencing. Ohio absorbed first (newer fleet, initial workload placement), Oregon followed, and Virginia's 7-day +7.4% reversal after hitting its trough suggests the demand wave is now reaching the larger Virginia node pool. If Virginia converges toward Ohio over the next 14 days, it confirms fleet-wide Blackwell tightening and eliminates the single most credible bear case against this thesis.

The AMD MI350P PCIe announcement (+40% FP16 vs. H200 NVL) does not threaten B200 directly — there is no PCIe B200, and MI350P targets inference/RAG workloads, not the large NVLink training clusters where B200 dominates. B200's training moat is reinforced by AMD's own product positioning decision. The RI-implied decay rate on GB200/B200 at Azure (24.5%/yr) already bakes in hard competition from GB300/Vera Rubin by 2027, so the clock is ticking — but the current 12–18 month window represents peak Blackwell pricing power.

Key Evidence:

1. B200 SPOT Ohio: $5.70/hr today, +87.4% over 30 days, peak $6.35/hr on May 20; spot discount compressed from ~87.9% max to 56.8% — 31pp tightening (live ticker data, May 26, 2026)
2. B200 SPOT Oregon: $4.98/hr, +61.4% over 30 days — independent corroboration of Ohio signal across two AWS regions (live ticker data, May 26, 2026)
3. B200 SPOT Virginia: $3.01/hr, -24.7% over 30 days BUT +7.4% over 7 days — first reversal signal after trough; 68.7% spot discount, 12pp above Ohio (live ticker data, May 26, 2026)
4. B200 ON_DEMAND: frozen at $12.83/hr across all three US regions — OD anchor unmoved, confirming all price action is demand-driven (live ticker data, May 26, 2026)
5. No B200 listings active at Azure or GCP (catalog survival: is_active: false) — single-provider market, every bid is a real training job (depreciation data, May 26, 2026)
6. Azure RI-implied decay on GB200/B200 at 24.5%/yr — market already pricing Vera Rubin competition into long-dated reservations (depreciation data, May 26, 2026)
7. AMD MI350P PCIe +40% FP16 vs. H200 NVL targets inference/RAG, not NVLink training — does not compete with B200 NVL (news feed, May 7–8, 2026)
8. Qualifying: Ohio 7-day delta turned negative (-8.9%) — the May 20 peak may represent a completed training run, not a permanent new floor; sustaining $5.50+ is the confirmation test

Implied Action: Long APLD (Applied Digital, $45.87, analyst target $61.10, +33% upside, next earnings July 29). APLD's CoreWeave lease at Polaris Forge 1 provides direct B200 capacity exposure with contracted revenue underpinning. Beat by 143% last quarter, +139% YoY revenue to ~$127M/quarter, consensus strong buy. Beta 5.7 means this is not a low-volatility hold — size accordingly. Trigger: Ohio B200 spot holding at or above $5.50/hr through June; Virginia recovering above $3.50/hr within 14 days is the bull escalation signal. Avoid direct CRWV exposure despite being the most obvious beneficiary: $35.1B in debt, -$8.6B FCF TTM, trading 44% below its 52-week high of $187 — balance sheet stress is dominating price action even as the underlying demand signal strengthens.

H2 — H100 Accelerated Depreciation: The EU Market Is Setting the Global H100 Price Floor

Confidence: 4.5 / 5

Thesis: The H100 on-demand price map has split into two irreconcilable regimes. In the US, APAC, and select Northern European markets (Warsaw $12.18/hr, Zurich $9.93/hr, Stockholm $8.54/hr), H100 trades at $7.50–$14/hr — consistent with the historical norm and reflecting healthy demand in markets where AI infrastructure buildout continues to absorb capacity. In Frankfurt ($1.78/hr) and Amsterdam ($0.43/hr), H100 has entered a distressed regime where pricing has decoupled from global rates by a factor of 5–28x, and the Frankfurt market has developed a structural inversion: spot clears at $2.45/hr — 38% above the $1.78/hr on-demand floor. This inversion is the single cleanest structural signal in the dataset. It can only persist when a provider has priced on-demand below historical spot clearing rates to maintain utilization, effectively advertising desperation.

The mechanism is not temporary: Frankfurt's 696% OD spread (min $0.40/hr, max $3.16/hr across providers) and 30-day -0.98% drift have been stable for a full month, and Amsterdam is frozen at $0.43/hr with no movement. These aren't flash events — they are the equilibrium state of a market where B200 generational displacement is compounded by EU power-and-permitting constraints that prevent capacity from being redeployed or absorbed into new build-outs. The H200/H100 Frankfurt ratio tells the rest of the story: H200 Frankfurt at $13.04/hr represents a 7.3x premium over H100 Frankfurt, versus the ~1.5–2x premium observed everywhere else globally. This ratio confirms H100 is being abandoned while H200 holds value — generational displacement is not theoretical in Frankfurt, it is already the market clearing price.

The RI-implied decay models reinforce the trajectory: AWS prices H100 with a 22.3%/yr implied decay rate, Azure at 16.1%/yr — a 6pp divergence indicating AWS is writing off H100 faster. The cross-sectional training decay model shows 17.4%/yr RAW secular decline (R²=0.265, p=0.041, statistically significant). At these rates, a $10/hr H100 fleet today implies a $2.50–3.00/hr equilibrium in 5 years — and the EU is already there today. The EU is the leading indicator; US markets will follow with an 18–36 month lag as B200 capacity scales domestically.

Key Evidence:

1. Frankfurt H100 OD: $1.78/hr, 696% spread (min $0.40, max $3.16); -0.98% over 30 days — slow, persistent floor-setting (live ticker data, May 26, 2026)
2. Amsterdam H100 OD: $0.43/hr, single provider, frozen; 0.0% change over 30 days — abandoned fleet, below commercial rates (live ticker data, May 26, 2026)
3. Frankfurt H100 SPOT: $2.45/hr — 38% ABOVE Frankfurt OD of $1.78/hr; structural inversion confirming forced OD repricing (live ticker data, May 26, 2026)
4. Frankfurt H200 OD: $13.04/hr — 7.3x premium over Frankfurt H100 ($1.78/hr), vs. ~1.5–2x globally; H100 abandonment while H200 holds value (live ticker data, May 26, 2026)
5. Dublin H100 SPOT: 18,294% spread, -50% over 7 and 30 days — market structure fracture; spread has persisted for a full 7-day window, confirming a new low-price regime has entered (live ticker data, May 26, 2026)
6. H100 RI-implied decay: AWS 22.3%/yr, Azure 16.1%/yr; cross-sectional training decay 17.4%/yr (R²=0.265, p=0.041 statistically significant) (depreciation data, May 26, 2026)
7. H100→H200 GENERATION_GAP: 23.3%/yr decay over 1.65 years, price ratio 0.647 (H100 at 65% of H200 pricing) — displacement is already priced generationally (depreciation data, May 26, 2026)
8. MI300X (Azure RI-implied decay): only 14.6%/yr — lowest of any GPU in the dataset; Azure has more confidence in AMD's long-term pricing than NVIDIA's H100 (depreciation data, May 26, 2026)
9. Qualifying: H100 spot in US/select EU is recovering — Ohio +37.8%, London +31%, Madrid +60.8% over 30 days; the depreciation thesis is geographic (EU distressed), not a global short (live ticker data, May 26, 2026)

Implied Action: Long AMD ($467/share, +138% over 3 months, forward P/E 36x) as the structural beneficiary of H100 displacement. MI300X's 14.6%/yr RI-implied decay (vs. H100's 16–22%/yr at Azure/AWS) reflects platform confidence that MI300X has a longer useful pricing life — a key advantage for enterprises making 3-year reserved commitments. The 3yr reserved MI300X at Azure ($26.34/hr vs. $42.30/hr OD) represents a compelling entry for workloads seeking cost stability without Blackwell premium exposure. Catalyst: MI400 (432GB HBM4, H2 2026 launch) benchmark announcements. Structural short thesis on EU H100 neocloud revenue — any operator with unhedged European H100 fleets at $7+/hr contract rates is facing forced renegotiations toward $1.78/hr Frankfurt market reality. Monitor: H100|ON_DEMAND|ie-dublin at $5.77/hr (6,886% spread, min $0.16/hr) is the next EU market likely to fracture toward Frankfurt levels.

H3 — AWS Inferentia Spot Surge: A Margin Signal for AWS's Own Inference Business

Confidence: 3.5 / 5

Thesis: AWS Inferentia gen-1 spot prices are surging in Frankfurt (+175% over 30 days to $0.249/hr), Sydney (+97% to $0.474/hr), and Tokyo (+52.4% to $0.268/hr) while collapsing in Virginia (-57.9% to $0.027/hr) and China-adjacent markets (Hong Kong -71.3%, Singapore -47.3%). A 6.5-year-old ASIC posting the strongest spot momentum outside of B200 is analytically counterintuitive — unless the mechanism is AWS itself, not external customers.

The Virginia collapse is the interpretive key. Virginia is AWS's primary US compute region and home to its heaviest production workloads (Alexa, advertising ML, recommendation systems). A -57.9% collapse in Virginia Inferentia spot means AWS has evacuated its own gen-1 Inferentia capacity there, freeing it as spot — almost certainly because Inferentia2 and Trn1 have absorbed those production workloads in Virginia. Meanwhile, in Frankfurt, Sydney, and Tokyo, where Inferentia2/Trn1 deployments are likely less dense, gen-1 Inferentia remains in production use, and rising spot prices reflect real-time internal workload bid pressure competing with external customers for the same capacity. The RI-implied decay on Inferentia is 20.8%/yr — theoretically in secular decline — yet spot prices in these regions are moving in the opposite direction. Demand is outrunning the depreciation curve locally.

The geographic split also validates Hypothesis 5's mechanism: HK and Singapore Inferentia collapsing (-71.3% and -47.3%) is the demand destruction footprint of DeepSeek API substitution. Workloads in China-adjacent markets are migrating to sub-$1/M token API endpoints rather than running owned Inferentia capacity. The Inferentia geographic map is simultaneously a AWS internal utilization gauge (Frankfurt/Sydney/Tokyo rising = production inference demand) and a DeepSeek substitution map (HK/Singapore falling = API migration underway). The L4 cross-check weakens the broad-market inference demand narrative: EU L4 spots are broadly declining (-14% to -33% across Frankfurt, Paris, Zurich, Stockholm, Amsterdam). The Inferentia surge is ASIC-specific and AWS-internal, not a pan-EU inference demand wave.

Key Evidence:

1. Frankfurt INFERENTIA SPOT: $0.249/hr, +175% over 30d, +122.8% over 7d — accelerating, not decelerating; multi-week confirmation (live ticker data, May 26, 2026)
2. Sydney INFERENTIA SPOT: $0.474/hr, +97% over 30d; Tokyo: $0.268/hr, +52.4% over 30d; Mumbai: $0.126/hr, +27.8% over 30d — geographic breadth across APAC (live ticker data, May 26, 2026)
3. Virginia INFERENTIA SPOT: $0.027/hr, -57.9% over 30d — capacity release consistent with Inferentia2/Trn1 migration in primary US region (live ticker data, May 26, 2026)
4. Hong Kong INFERENTIA SPOT: $0.018/hr, -71.3% over 30d; Singapore: $0.031/hr, -47.3% over 30d — DeepSeek API substitution pattern in China-adjacent markets (live ticker data, May 26, 2026)
5. Inferentia RI-implied decay: 20.8%/yr — theoretical secular decline CONTRADICTED by EU/APAC spot strength; demand outrunning the depreciation curve (depreciation data, May 26, 2026)
6. Inference-class cross-sectional decay: 12.6%/yr vs. training 17.4%/yr — inference ASICs depreciate 38% more slowly; stickier demand cycle (depreciation data, May 26, 2026)
7. Alibaba MaaS: +1,500% revenue growth in 5 months (KrASIA, May 25, 2026) — token volume driving growth, not price; inference throughput demand is structurally accelerating
8. Qualifying: EU L4 spot broadly declining (Frankfurt -14.4%, Zurich -32.9%, Amsterdam -18.6% over 30d) — no cross-GPU inference demand corroboration; Inferentia surge is ASIC-specific, likely AWS-internal (live ticker data, May 26, 2026)

Implied Action: Long AMZN via AWS inference margin thesis. Each Inferentia chip carries an OD rate of $1.475/hr; AWS capturing production inference workloads through internal cost-plus pricing while rising spot prices signal falling excess capacity represents operating leverage on AWS's highest-margin workloads. This is not a primary trade but a supporting pillar in the AWS segment re-rating story. More direct: Long MRVL (Marvell, $678M datacenter revenue, +42% YoY). The Inferentia spot surge in production-relevant markets validates customer demand for hyperscaler-proprietary inference ASICs — exactly MRVL's product pipeline, with 12–24 month ramp cycles and multi-year contracts already signed. Trigger: Any AWS announcement of Inferentia2/3 capacity expansion in EU or APAC confirms demand exists to justify next-gen builds and will initially push spot prices higher before stabilizing.

H4 — European GPU Market Fracture: A Leading Indicator for US Data Center Moat Pricing

Confidence: 4.5 / 5

Thesis: The European H100 market has reached a state of structural dysfunction that is analytically more important for its forward-looking implications than its current absolute values. Three layers of evidence converge: (1) a 5–28x intra-continent price spread with Frankfurt and Amsterdam at distressed levels that have not mean-reverted in 30+ days; (2) a Frankfurt spot-above-OD inversion that identifies a provider pricing at or below marginal cost; (3) a Dublin spot market fracture where an 18,294% spread has persisted for a full 7-day window, confirming a new entrant has entered at near-zero pricing and is holding.

This is not a temporary demand trough. It is the end-state of a market where: (a) B200 generational displacement has made H100 the inferior training option; (b) European power and permitting constraints prevent new H100 capacity from being redeployed into next-gen infrastructure; and (c) the cost of carrying idle H100 inventory exceeds the revenue from distressed OD pricing, pushing providers toward floor-pricing strategies. The European energy constraint is the structural amplifier — what in the US takes 18–36 months to manifest through permitting queues and NERC reliability guidelines is already the EU's current reality. NERC's May 6 draft reliability guideline on large-load entities and NV Energy's decision to cut residential power to prioritize AI data centers are the US analogs of what European operators encountered 12–18 months ago.

The implication for US markets: operators with existing grid interconnection positions — particularly in PJM and MISO territories where B200 capacity is actively being absorbed — hold a structural moat against new entrants. Power queue positions can take 3–5 years to secure in constrained markets. NERC's new compliance requirements add cost and timeline to new builds. This creates an asymmetric advantage for established US data center REITs and neoclouds with legacy power contracts: their capacity is physically scarce in a way that cannot be resolved by manufacturing more silicon.

Key Evidence:

1. Frankfurt H100 OD: $1.78/hr (min $0.40/hr) vs. Warsaw H100 OD: $12.18/hr — 6.8x intra-EU spread; stable over 30+ days with no mean reversion (live ticker data, May 26, 2026)
2. Amsterdam H100 OD: $0.43/hr, single provider, frozen; 28x below Warsaw, 20x below global median (live ticker data, May 26, 2026)
3. Frankfurt H100 SPOT ($2.45/hr) trading 38% ABOVE Frankfurt H100 OD ($1.78/hr) — structural inversion; incumbent pricing below spot clearing rates (live ticker data, May 26, 2026)
4. Frankfurt H200 OD at $13.04/hr = 7.3x H100 Frankfurt, vs. ~1.5–2x ratio globally — H100 abandoned at distressed prices while H200 holds global rack rates (live ticker data, May 26, 2026)
5. Dublin H100 SPOT: 18,294% spread, -50% over 7d and 30d — new low-price entrant held position for full 7-day window; not a blip (live ticker data, May 26, 2026)
6. NERC draft reliability guideline for large-load entities: frequency/voltage trip settings required; first formal compliance framework for AI data center grid impact (intel feed, May 6, 2026)
7. NV Energy prioritizing AI data center power over Lake Tahoe residential customers; Quanta Services +127% YoY on grid infrastructure demand (news feed, May 15, 2026)
8. 14 states considering data center legislation; Virginia resident approval of new data centers down 34 percentage points to 35% (news feed, May 2026)
9. Qualifying: H100 spot in London (+31%), Madrid (+60.8%), Stockholm (+17.6%), and Ohio (+37.8%) over 30 days shows demand-healthy markets absorbing capacity; the fracture is Frankfurt/Amsterdam/Dublin-specific, not pan-European (live ticker data, May 26, 2026)

Implied Action: Long Quanta Services as the grid infrastructure pure-play beneficiary (+127% YoY already, but the regulatory formalization of NERC large-load requirements creates a durable compliance-driven capex pipeline). Long US data center REITs with PJM/MISO power queue positions — the moat from existing grid interconnections becomes structurally unassailable under new NERC requirements. Short European H100 neocloud revenue models: any operator with European H100 fleet contracts written above $3/hr faces forced renegotiation toward the Frankfurt market floor; Dublin's fracture suggests $0.01–0.10/hr clearing prices are achievable if the new entrant holds. Monitor: H100|ON_DEMAND|ie-dublin at $5.77/hr with $0.16/hr minimum — if the spread normalizes downward (vs. upward) over the next 30 days, it confirms the EU oversupply regime is propagating north and west, and the London $8.21/hr OD canary comes into play.

H5 — DeepSeek/Inference Compression: Premature Signal, Real Mechanism

Confidence: 3 / 5

Thesis: The "missing middle" hypothesis — that DeepSeek's permanent 75% token price cut (V4-Pro from ¥24 to ¥6/M tokens, May 25) will destroy demand for mid-tier H100 training rentals in the $7–10/hr OD bracket — is real as a mechanism but has not yet materialized in pricing data. The H100 OD global prices at London ($8.21/hr), Seoul ($8.63/hr), Tokyo ($8.82/hr), and Dubai ($8.93/hr) are catatonic: 0.0% change over 24h, 7d, and 30d in most cases. This pricing stasis is ambiguous — it could reflect supply-demand equilibrium, managed list-price stickiness, or the calm before a demand attrition inflection that is building slowly.

The geographic evidence from the Inferentia data provides the most tangible support for the mechanism: Hong Kong (-71.3%) and Singapore (-47.3%) Inferentia spot collapses are the only current examples of inference demand destruction visibly manifesting in spot prices — and they are happening specifically in China-adjacent markets where DeepSeek API access is cheapest and most reliable. These are the early-adopter markets for API substitution. If the pattern propagates to Western Europe and the US, H100 OD pricing in the $8–11/hr range becomes the next shoe to drop. The 60–150 day time horizon is analytically correct — this is a lagging-indicator thesis by design. The AMD MI350P (+40% FP16 vs. H200 NVL, PCIe form factor) introduces a competing mechanism: hardware-led H100 displacement from the supply side in H2 2026, distinct from but complementary to API-led demand destruction.

Key Evidence:

1. DeepSeek V4-Pro permanent price cut: ¥24→¥6/M tokens (75% reduction), effective May 25, 2026 (news feed, Daily Tech News Show, May 25, 2026)
2. Alibaba MaaS: +1,500% volume in 5 months — inference throughput demand is growing but economics are deflationary (KrASIA, May 25, 2026)
3. London H100 OD: $8.21/hr — 0.0% change over 24h, 7d, AND 30d; pricing catatonia suggesting managed list prices (live ticker data, May 26, 2026)
4. Seoul H100 OD: $8.63/hr (0.0% all horizons); Tokyo: $8.82/hr (-0.09% 30d); Dubai: $8.93/hr (-0.025% 30d) — identical stasis globally (live ticker data, May 26, 2026)
5. Hong Kong INFERENTIA SPOT: -71.3% over 30d; Singapore: -47.3% — China-adjacent market inference demand destruction already visible in spot data (live ticker data, May 26, 2026)
6. H100 spot recovering in US/select EU: Ohio +37.8%, London +31%, Madrid +60.8% over 30d — directly contradicts near-term demand destruction at the $7–10/hr OD level (live ticker data, May 26, 2026)
7. AMD MI350P PCIe: +40% FP16 vs. H200 NVL, drop-in upgrade — competing H100 displacement mechanism from supply side, arriving H2 2026 (news feed, May 7–8, 2026)
8. Qualifying: No H100 OD price moves yet in primary hypothesis markets; the thesis is ahead of the data by 60–90 days minimum

Implied Action: Long Nebius Group (NBIS) ($214.77, +113% over 3 months, +36.7% over 1 month, PEG ratio 0.63 — cheapest in the neocloud cohort). Nebius operates a European-focused inference-optimized GPU cloud that is positioned on the right side of the demand shift. At 0.63 PEG it is materially cheaper than CRWV or APLD on a growth-adjusted basis, and its explicit inference-optimization strategy aligns with the structural inference demand durability (12.6%/yr cross-sectional decay vs. 17.4% for training). Watch signal: H100|ON_DEMAND|gb-london — the first sustained break below $7.50/hr is the thesis activation event. Until then, position sizing should be modest. Avoid: CRWV at $105.49, -10.2% over 30 days — 738x debt-to-equity and -$8.6B FCF make it the most levered expression of a thesis that hasn't triggered yet; risk/reward is unfavorable relative to NBIS.

Risk Flags

Immediate (0–30 Days)

R1 — B200 Ohio Peak-and-Fade Risk The May 20 Ohio peak at $6.35/hr followed by a 7-day pullback (-8.9%) is the most urgent monitoring item in the entire dataset. If this represents a completed large-scale training run (e.g., a frontier model pre-training job) rather than structural demand absorption, the spot price could revert toward $3–4/hr within weeks. At that level, the APLD re-rating thesis loses its primary catalyst. The distinction between a single-job spike and structural absorption should be visible within 14 days: sustained bid pressure above $5.50/hr is absorption; a rapid decline toward $4/hr is a completed run. This is binary and time-sensitive.

R2 — Dublin H100 Market Structure Ambiguity The 18,294% spread in Dublin H100 spot (incumbent at $2.54/hr, entrant implied near $0.01/hr) has held for 7 days but remains structurally unstable. The risk is bidirectional: either the incumbent exits and the spread collapses to near-zero (confirming EU oversupply deepening and propagating to London's $8.21/hr OD canary), or the new entrant retreats and the 2-provider structure is restored (denying the EU fracture thesis). Both resolutions are plausible within 30 days. Positions predicated on EU H100 distress (H4) should account for the possibility of resolution in either direction.

R3 — Frankfurt H100 Spot-Above-OD Inversion The structurally unstable Frankfurt H100 spot ($2.45/hr) above OD ($1.78/hr) must resolve. Either spot corrects downward — confirming the oversupply is deepening and spot bidders are recognizing the OD floor — or OD corrects upward as the distressed provider raises prices, suggesting the floor-pricing strategy is being abandoned. The former confirms H