Daily Investment Brief — May 27, 2026

Market Pulse

The GPU spot market is bifurcating along two simultaneous fault lines: generational displacement and regulatory geography. L4 inference GPUs are in freefall across three continents — down 66% in Zurich, 56% in Oregon, 45% in Brazil — as the 1.3-year implied useful life of the L4→L40S generational gap expires in real time. Meanwhile, AWS Inferentia SPOT pricing has split into two irreconcilable markets: Frankfurt at $0.260/hr (+194% over 30 days) and Tokyo at $0.266/hr — while Hong Kong has collapsed 74% to $0.045/hr and Singapore 47.5% — a 6× spread between the most and least expensive regions for identical chips. That divergence is not a demand story; it maps cleanly onto EU and Japan AI governance frameworks creating jurisdiction-specific compliance-pull. Overlaying both trends is an escalating legislative siege on datacenter siting: NC SB 730 has advanced out of committee to the full House, NERC is formalizing large-load grid requirements, and 71% of Ohio residents now oppose local datacenter construction — more than nuclear plant opposition. The binding constraint on AI infrastructure is no longer silicon; it is permitted land, connected power, and political tolerance.

Key Movers

Thin-market noise alert: Alveo U30 Oregon (+7,537%), T4G Tokyo (+410%), and Paris RAM (+53% 1-day) are all single-provider markets with ≤2 observations. They are directional indicators only — no liquidity for position-sizing.

Investable Insights

H1 — Power Constraint Legislation Is a Phase Transition, Not an Incremental Headwind

Confidence: 5 / 5

Thesis: The datacenter siting risk landscape crossed a qualitative threshold this week. NC SB 730 — which would impose tiered tariffs on 100MW+ loads, restrict evaporative cooling, and codify siting rules — has advanced out of committee to the full North Carolina House, covering Duke Energy's disclosed pipeline of 15.4 GW under active discussion. That 15.4 GW figure alone represents roughly 15% of current total US datacenter power consumption. The mechanism matters: SB 730 doesn't ban datacenters; it makes every new hyperscale campus above 100 MW economically marginal through punitive rate structures. Since virtually every tier-1 hyperscale facility exceeds 100 MW, the law functions as a de facto moratorium on economic new builds in North Carolina — the state with the largest existing concentration of hyperscaler infrastructure in the Eastern Seaboard.

This is not isolated. Ohio ballot initiative organizers have secured 71% public opposition polling (exceeding nuclear plant opposition levels) against new datacenter construction, driven by a documented 50% electricity cost increase already attributed to existing loads. NERC's Large Load Reliability Guideline is formalizing voltage and frequency trip requirements for large consumers across the entire US interconnect. NV Energy has literally cut residential power supply to redirect capacity to datacenters. The pattern is: policy is now catching up to demand, and it is catching up fast. For existing, permitted, grid-connected campus operators, this is a moat-deepening event. For new-build development pipelines, it is a terminal risk at current underwriting assumptions.

The power infrastructure side of this trade has its own confirmation signal: Vesari Inc. launched this week with venture backing on the explicit thesis that "power availability, not semiconductor supply, is the primary constraint" (W.Media, May 27). The Forgent Power Solutions S-1 filed May 26 at $47.90/share for a company premised entirely on dedicated AI datacenter power delivery. VC and IPO markets are pricing the bottleneck — public equity markets in datacenter REITs have not fully adjusted.

Key Evidence:

1. NC SB 730 advances out of House panel (Data Center Knowledge, May 26) — covers Duke Energy's 7.6 GW active contracts and 15.4 GW pipeline under discussion; 100MW+ tariff structure targets every tier-1 hyperscale facility.
2. Ohio: 71% community opposition polling + active ballot initiative effort; 50% electricity cost increase already realized; further 50% increase projected (TiffinOhio.net, May 26).
3. NERC Large Load Reliability Guideline (intel_feed, power_grid): Formal interconnection standards being drafted for loads >100 MW — raises capex and timeline for new grid-connected builds across all US regions.
4. NV Energy cutting residential power supply to accommodate datacenter load (Quartz, May 15) — concrete evidence of grid reallocation with direct political consequence.
5. Blue Owl $2.96B acquisition of three Gainesville, VA hyperscale facilities (8-K, May 22): Two 72 MW facilities, 100% pre-leased — confirms existing-campus scarcity premium is being bid up by institutional capital.
6. IREN equity: $59.78, +31.5% over 3 months, buy-rated, analyst target $75 — the market's cleanest proxy for "build dedicated power + compute where others can't permit" is already re-rating.
7. 14 states considering datacenters bans or moratoria; Maine passed one (governor vetoed) — legislative momentum is broadening beyond NC and OH.
8. Vesari Inc. VC launch (W.Media, May 27) + Forgent Power Solutions S-1 (SEC, May 26) — private and IPO capital explicitly pricing the power bottleneck as an investable thesis.

Implied Action:

• Long IREN ($59.78, target $75): The clearest listed expression of dedicated power-first AI compute infrastructure. The +31.5%/3m move has not overshot the $75 analyst target; the SB 730 advance is a fresh catalyst not yet priced.
• Long EQIX ($1,077, target $1,197) and DLR ($193.67, target $218.72): Existing-campus datacenter REITs whose moats deepen as new supply is politically constrained. These aren't cheap (56x and 67.6x forward P/E respectively), but the valuation is for monopoly-quality cash flows on infrastructure that is now genuinely difficult to replicate. Their development pipeline deserves scrutiny — their operating campus does not.
• Avoid new-build datacenter development vehicles without (a) pre-permitted sites, (b) executed grid interconnection agreements, and (c) demonstrated community acceptance processes. Blue Owl's Virginia deal was smart because it was already built and leased — that is the correct template.
• Monitor trigger: SB 730 full House vote (likely June-July 2026). Passage is the catalyst; defeat means the soft-friction scenario where existing operators still benefit, but new-build risk is reduced.

H2 — EU + Japan ASIC Inference Is Repricing Structurally; This Is a Compliance Story

Confidence: 4.5 / 5

Thesis: The AWS Inferentia SPOT market has bifurcated in a way that cannot be explained by raw compute demand. Frankfurt has surged from $0.088/hr to $0.260/hr in 30 days (+193.6%), Tokyo sits at $0.266/hr (+40%), Dublin at $0.131/hr (+7.4%). Meanwhile, Hong Kong has collapsed 73.7% to $0.045/hr, Singapore -47.5% to $0.057/hr, and Virginia -56.6% to $0.027/hr — a 10× price gap between Frankfurt and Virginia for the same chip on the same provider. The splitting mechanism is regulatory: EU AI Act enforcement timelines and Japan's data residency framework are creating jurisdiction-specific demand pull for cloud-native, certified inference infrastructure. Operators running production inference APIs in EU and Japanese jurisdictions are buying Inferentia specifically because it is AWS-managed, auditable infrastructure that satisfies model provenance and data residency requirements — and they are being priced accordingly.

The Frankfurt signal passes the on-demand cross-check: Inferentia Frankfurt on-demand sits at $0.434/hr, flat over 30 days, confirming this is a genuinely high-cost Inferentia market, not a spot artifact. The SPOT/OD ratio of 60% is a normal discount — the SPOT price has risen because the entire demand level in the market has risen. Tokyo's parallel move (+40%, reaching $0.266/hr — nearly tied with Frankfurt) complicates a pure "EU premium" framing but actually strengthens the thesis: Japan's AI governance framework has created a second compliance-pull jurisdiction, and Tokyo is the clearest APAC market confirmation. The flip side — Virginia SPOT at $0.027/hr, down 57% — suggests Trainium is displacing Inferentia in AWS's domestic US market where compliance requirements are lighter, tightening the supply available for EU/Japan while flushing US capacity.

Key Evidence:

1. Frankfurt Inferentia SPOT: $0.260/hr, +193.6% over 30d, +52.8% over 7d (live ticker, May 27) — confirmed by on-demand at $0.434/hr (zero 30d movement), validating the demand-pull interpretation.
2. Tokyo Inferentia SPOT: $0.266/hr, +40.0% over 30d (live ticker, May 27) — Japan AI governance framework aligns with EU compliance-pull thesis; second jurisdiction confirming the pattern.
3. Hong Kong Inferentia SPOT: $0.045/hr, -73.7% over 30d — same chip, same provider, 5.7× cheaper than Frankfurt. No demand signal from China-adjacent capacity.
4. Virginia Inferentia SPOT: $0.027/hr, -56.6% over 30d — 10× cheaper than Frankfurt; Trainium displacement in the domestic US market is the structural explanation.
5. Cross-sectional model: Inference GPU annual decay = 12.63% (implied useful life 7.9 years) — inference hardware holds value precisely because it becomes embedded in long-lived production API contracts with compliance requirements.
6. EQIX $1,077, +13.2% over 3 months, buy-rated, target $1,197 — Frankfurt's DE5/DE6 campus is the world's largest neutral interconnection hub in continental Europe; Inferentia demand concentration near-shores to exactly this infrastructure.
7. US AI Export Program includes Japan as allied expedited-licensing nation (Washington Examiner, May 17) — Japanese exemption from export controls is durable US policy, supporting Japan AI buildout thesis.

Implied Action:

• Long EQIX directly plays the Frankfurt campus demand uplift — EU inference compliance workloads route through Equinix Frankfurt as the neutral interconnection layer. The current $1,077 price vs. $1,197 target is a clean 11% gap with a structural catalyst behind it.
• Watch trigger: Frankfurt Inferentia SPOT breaking above $0.300/hr (currently $0.260) confirms the repricing has not peaked. Conversely, a 2-week reversion below $0.150/hr would suggest the +194% was a listing event, not sustained demand.
• Avoid: APAC-general inference infrastructure bets that include Hong Kong or Singapore capacity — the Asia flush is real, and the 73.7% HK collapse is not mean-reverting; it is directional.
• Secondary long: DLR EU exposure — Digital Realty has significant Frankfurt and EU presence. The same compliance-demand pull that benefits EQIX's Frankfurt campus benefits DLR's colocation tenants choosing EU-jurisdiction inference deployments.

H3 — The L4→L40S Clock Has Expired: Global Mid-Tier Inference Is Stranded

Confidence: 4.5 / 5

Thesis: The L4 SPOT collapse is not a regional anomaly; it is a generational clock striking midnight across three continents simultaneously. The generation-gap depreciation model assigns L4→L40S an annual decay rate of 77.5% — the steepest in the entire dataset by a wide margin, double the already-steep A100→H100 transition at 53.4%. That 77.5% annualized rate implies an L4 useful life of 1.3 years from L40S launch (~mid-2024). We are now in month 12–18 of that clock. What a 77% annualized rate looks like on a month-by-month basis is not a smooth glide — it is a step-change repricing when capacity holders mark inventory as stranded. That step-change is now visible: Zurich -66%, Oregon -56%, Brazil -45%, Mumbai -24%, Stockholm -18%, Amsterdam -19%, Frankfurt -14%, Paris -12% in 30 days, alongside a 3,126% spread in Zurich ($0.004/hr vs. $0.144/hr) and a 20,505% spread in Mumbai ($0.002/hr vs. $0.484/hr). The spread structure is definitive — it does not represent competitive pricing; it represents one provider marking L4 as essentially worthless while another hasn't yet caught up. The market will clear toward the low price.

Critical refinement: This is not a "short L4, long L40S everywhere" trade. The L40S market is not uniformly strong. L40S Stockholm is down 37.6% over 30 days; L40S Seoul is down 35.7%; L40S Oregon is down 18.2%. The Nordic and Asia L40S markets are also contracting — workloads are not simply rotating from L4 to L40S in those regions; they are leaving those regions altogether (likely migrating to cheaper US or India capacity). The Frankfurt exception — L40S up 11.2% in the last 7 days while L4 Frankfurt falls 14.1% over 30 days — is the clean pair trade: Frankfurt is the one EU market where L4→L40S rotation is actually happening, driven by the same compliance-pull that is inflating Inferentia pricing there.

Key Evidence:

1. L4 SPOT Zurich: $0.074/hr, -65.7% over 30d, -49% over 7d, 3,126% spread ($0.004–$0.144) — textbook provider-level stranding; spread resolves downward (live ticker, May 27).
2. L4 SPOT Mumbai: 20,505% spread ($0.002–$0.484/hr) — most extreme case; one provider has already marked L4 inventory at near-zero (live ticker, May 27).
3. L4 SPOT Oregon: $0.174/hr, -56% over 30d, 7,898% spread — US West confirms the pattern is not EU-isolated.
4. Generation-gap model: L4→L40S annual decay = 77.51%, implied useful life 1.3 years (vs. A100→H100 at 53.4% and 1.9 years; H200→B200 at 3.05% and 32.8 years) — fastest depreciation pair in dataset.
5. L40S Frankfurt SPOT: $2.04/hr, +11.2% in 7 days — the displacement accelerator in the one EU market where rotation is clean (live ticker, May 27).
6. L40S Stockholm SPOT: $0.417/hr, -37.6% over 30d; L40S Seoul -35.7% — Nordic/APAC collapse in both generations confirms these are workload-migration events, not generational rotations.
7. L4 SPOT Ohio: +31.4% over 30d — counter-trend confirms the supply-constraint floor: Ohio's contested siting environment has created genuine scarcity even for aging mid-tier inference hardware.
8. Qumulo cloud AI report: Enterprise GPU utilization averages 5% (HPCwire, May 26) — structural oversupply of enterprise-owned mid-tier GPUs feeds the spot dump.

Implied Action:

• Short L4-heavy operators in Zurich, Oregon, and Mumbai — the $0.004/hr Zurich floor price is not a bottom; it is liquidation. The high-price provider's $0.144/hr hasn't repriced yet. When it does, the median falls further. Any entity with unhedged L4 capacity in these three markets should expect an additional 30–50% markdown.
• Long Frankfurt L40S — the cleanest paired trade. L4 Frankfurt falling while L40S Frankfurt rises is confirmed displacement, not general inference demand collapse. The +11.2% 7-day move is accelerating.
• Explicit avoid: Do NOT pair the Frankfurt L40S long with Stockholm or Seoul L40S positions — those markets are experiencing outright workload flight, not a generational rotation.
• REIT implication: Any colocation REIT with L4-heavy tenant rosters should be stress-tested for lease renewal risk in 2026–2027 — tenants will demand concessions or migrate to L40S racks. Flag in any DLR/EQIX position review.
• Ohio exception: L4 Ohio SPOT +31.4% is a real signal — Ohio's datacenter opposition is creating supply scarcity that benefits existing capacity holders, including aging L4 fleets. The paradox: the most politically hostile environment for new builds is the most economically hospitable environment for existing operators.

H4 — The H200/B200 Pricing Stall Is Geographically Bounded; CRWV Premium Regions Are the Exposure

Confidence: 4 / 5

Thesis: The H200→B200 generation gap shows a near-zero annual decay rate of 3.05% — implying a 32.8-year useful life for H200, which is absurd relative to historical norms. The correct interpretation is not that H200 will hold value forever, but that B200's rollout is so geographically contained (three US regions only: Ohio, Oregon, Virginia) and so expensive (2× the price of H200 in cheap regions) that it has not yet forced an H200 markdown. The bifurcation is stark: in Ohio and Oregon, H200 sits at $6.22–$7.23/hr while B200 demands $12.83–$14.60/hr — B200 is not a competitive substitute at those prices; it's a premium tier. In Virginia, the spread nearly vanishes ($12.61 H200 vs. $12.82 B200), and in Oslo/Frankfurt/London, H200 sits at $12.40–$14.36/hr with no B200 listed at all. The stall is a premium-region phenomenon where B200 offers marginal performance improvement at near-parity pricing — meaning any B200 supply expansion in premium-region markets will land directly on H200 unit economics in those markets.

CoreWeave's business is disproportionately concentrated in US-East premium markets (Virginia-class regions where the B200/H200 spread is tightest). At $105.89/share — already -43% from its 52-week high of $187 — CRWV is priced at roughly 6.6× forward P/S with a market cap of $57.8B. That multiple is pricing extended H100/H200 scarcity rents in exactly the markets most vulnerable to B200 normalization. Applied Digital ($45.14, PEG 1.83, strong_buy) presents a better risk/reward: it operates in cheaper US regions where B200's 2× premium provides a durable H200 buffer, and its earnings growth thesis doesn't depend on perpetual scarcity pricing.

Key Evidence:

1. B200 on-demand: Ohio $12.83/hr, Oregon $14.60/hr, Virginia $12.82/hr — zero non-US listings; highly controlled rollout from a single provider (live ticker, May 27).
2. H200 on-demand: Ohio $6.22/hr, Oregon $7.23/hr — B200 premium of 2.06× and 2.02× respectively; not a near-substitute at those prices (live ticker, May 27).
3. H200 Virginia: $12.61/hr vs. B200 Virginia $12.82/hr — 1.7% spread; B200 is essentially parity in premium US-East markets. Supply expansion here is directly deflationary for H200 in CRWV's core market.
4. H200 Oslo: $14.36/hr, Frankfurt ~$12.50/hr — no B200 listed in EU; premium-region H200 pricing is entirely uncontested by B200.
5. Generation-gap model: H200→B200 annual decay = 3.05%, implied life 32.8 years (vs. A100→H100 at 53.4%) — confirms stall, but the 32.8-year number signals something structural, not sustainable (live depreciation data, May 27).
6. H200 export approvals to China with 25% government fee (Let's Data Science, May 17) — H200 is being monetized through multiple channels; catalog price discipline is deliberate.
7. US-China summit leaves semiconductor controls unresolved (Chosun, May 17) — B200 will not be supply-released via China export sales; the export ceiling on B200 is intact.
8. CRWV: $105.89, -43% from 52-week high of $187, 6.6× forward P/S (equities feed, May 27) — already significantly derating; but the specific B200/H200 parity in its core Virginia market suggests further earnings risk if B200 supply normalizes.
9. APLD: $45.14, strong_buy, analyst target $61.10, PEG 1.83 — near 52-week high of $48.81; cheaper-region exposure provides B200 buffer (equities feed, May 27).

Implied Action:

• Long APLD ($45.14, target $61.10) — H200 scarcity rents in cheap US regions are protected by B200's 2× premium; PEG of 1.83 and earnings growth are credible at current H200 utilization rates.
• Risk-monitor CRWV for a B200 SPOT tier appearing in US-East. That event — a B200 spot listing in Virginia or Ohio — is the single trigger to reduce neocloud long exposure. B200 SPOT anywhere in the US would confirm supply is loosening, and the Virginia near-parity pricing means H200 in that market would face immediate discount pressure.
• Watch CIFR ($23.02, strong_buy, target $30.53, +26.5% over 1m) as the quality neocloud proxy — 34× forward P/E implies earnings visibility that CRWV's multiple doesn't have; CIFR near its 52-week high of $25.52 with fundamental support.

H5 — Japan Legacy Capacity Tightness Is Real but Time-Limited

Confidence: 3.5 / 5

Thesis: The V100 32GB SPOT price in Tokyo has doubled from ~$0.30/hr to $0.617/hr over 30 days and is still accelerating (+12.5% in the last 24 hours). This is a single-provider, zero-spread market with one observation — thin by any measure — but the signal is corroborated by parallel moves: Tokyo Inferentia +40%, Tokyo L4 showing a 4,996% spread with the high-price provider holding at $0.346/hr as the dumping provider appears to be beginning to withdraw cheap capacity (spread compressing from 5,704% to 4,996% over the past 7 days). Japan is experiencing genuine legacy GPU tightness. The mechanism is clear: export controls have redirected H100/H200 supply to specific allied nations, creating a two-tier market in Japan where operators who cannot access or afford H200 are bidding up V100-class and L4-class capacity aggressively. The US AI Export Program (Washington Examiner, May 17) explicitly confirms Japan's expedited-licensing status — Japan's exemption from hardware restrictions is durable US policy, not administrative oversight, which means Japanese AI operators have long-term strategic investment incentives that justify paying up for available cloud compute.

The thesis carries a 3.5/5 confidence rather than higher because single-provider V100 data is inherently noisy — each new observation can move the ticker materially — and the mean-reversion risk within 72 hours is real. The structural argument (Japan AI buildout + legacy GPU tightness) is sound; the entry timing signal is imprecise.

Key Evidence:

1. V100_32GB SPOT Tokyo: $0.617/hr, +112.5% over 30d, +12.5% in 24h — single provider, zero spread, 1 observation (live ticker, May 27).
2. V100_32GB SPOT Hong Kong: $0.581/hr, -0.05% over 30d; Singapore $0.536/hr, -0.0001% — flatline confirms Tokyo is Japan-specific tightness, not a global V100 reprice.
3. Inferentia SPOT Tokyo: $0.266/hr, +40.0% over 30d — corroborating cross-chip demand signal in same region.
4. L4 SPOT Tokyo: 4,996% spread (one provider $0.007, another $0.346/hr), spread compressing from 5,704% over 7 days — precursor to violent convergence upward if the dumping provider withdraws.
5. US AI Export Program: Japan as expedited-licensing allied nation (Washington Examiner, May 17) — Japan's H100/H200 access is durable; long-term APAC AI investment incentive supports legacy compute demand as near-term bridge.
6. Qualcomm × ByteDance chip deal (Bloomberg, May 26) — APAC chip spend acceleration from non-Nvidia vendors confirms broad Asia AI buildout, not just Japan.
7. HIVE Digital: $4.10, +62.7% over 1m, strong_buy, target $5.80 — miner-to-AI pivot with legacy GPU fleet; BTBT $1.99, +22.8% over 1m, target $4.60, beta 3.96 (equities feed, May 27).

Implied Action:

• Speculative long HIVE ($4.10, target $5.80) and BTBT ($1.99, target $4.60) — both miner-to-AI pivots trading well below analyst targets with legacy GPU fleets positioned for exactly this APAC capacity tightening environment. Size small given data quality limitations.
• Conviction upgrade trigger: V100_32GB Tokyo price holding above $0.55/hr for 5 consecutive days would confirm structural tightness rather than a spike artifact. That is the signal to add to HIVE/BTBT positions.
• Stop signal: A reversion below $0.35/hr within 72 hours indicates the +112% move was a data update artifact, not sustained demand. Exit speculative positions immediately on that trigger.
• Policy tail risk: Any US executive action extending export control restrictions to Japan (currently exempt) would collapse this thesis. This is a non-trivial risk given the velocity of US-China tech policy changes in 2025–2026.

Risk Flags

Immediate (0–30 Days)

R1 — Zurich L4 Spread Resolution: Forced Markdown Incoming The 3,126% spread between $0.004/hr and $0.144/hr in Zurich L4 SPOT cannot persist. The -49% 7-day move confirms the low-price provider is setting direction. When the high-price provider ($0.144/hr) reprices down to match, the market median will collapse further — currently at $0.074/hr, the math implies a floor somewhere near $0.010–$0.020/hr. Any entity holding Zurich L4 capacity at near-spot prices has a mark-to-market event arriving within days. Watch for the spread to compress below 500% as the signal that repricing has occurred.

R2 — Mumbai L4/L40S Simultaneous Collapse: Workload Flight, Not Rotation Mumbai is the market where the L4→L40S rotation thesis fails. L4 Mumbai's 20,505% spread ($0.002–$0.484/hr) and L40S Mumbai's -36% over 30 days are happening simultaneously. Workloads are not rotating from L4 to L40S in Mumbai — they are leaving the region entirely, likely migrating to cheaper US or European capacity. Anyone building an India-centric inference deployment thesis on Mumbai cloud pricing data from Q1 2026 needs to re-underwrite: the market has repriced radically. This is a near-term operational risk for enterprises with India-jurisdiction AI API deployments.

R3 — Single-Provider SPOT Data Noise (V100 Tokyo, Alveo U30, Paris RAM) Three major movers this week — V100_32GB Tokyo (+112%), Alveo U30 Oregon (+7,537%), and RAM Paris (+53% in 24h) — are all single-provider markets with ≤2 observations. These are not tradeable signals at any reasonable position size. The V100 Tokyo move has a structural thesis behind it, but the 24h increment (+12.5%) can flip violently in either direction. Treat these as qualitative directional indicators, not precision price signals, and do not size positions based on the magnitude of the delta alone.

Near-Term (30–90 Days)

R4 — B200 SPOT Tier Launch: The H200 Derating Trigger B200 exists today only in on-demand format across three US regions. The moment a B200 SPOT tier appears — particularly in Virginia or Ohio — it signals that supply is normalizing and the controlled-rollout phase is ending. That event would directly compress H200 on-demand pricing in premium regions where B200 is currently near-parity ($12.82 B200 vs. $12.61 H200 in Virginia). CoreWeave's unit economics in US-East markets are most exposed. The watch signal is simple: set an alert for B200|GPU|SPOT in the ticker database. Its appearance is the trigger to reduce neocloud exposure.

R5 — EU AI Act Timeline Slippage Reverses Frankfurt Inferentia Premium The Frankfurt Inferentia +194% repricing thesis depends on EU AI Act compliance timelines creating procurement urgency for certified inference infrastructure. If enforcement timelines slip (court challenges, member-state implementation delays, or Commission guidance revisions), the compliance-pull demand signal weakens. The Frankfurt on-demand anchor at $0.434/hr provides a floor, but the 60% spot/OD discount could widen back toward historical norms if spot demand softens. Monitor EU AI Act implementation news for any 6-month timeline extension — that is the primary denial signal for H2.