Daily Investment Brief — June 11, 2026

Market Pulse

The defining story in today's GPU compute market is geographic bifurcation driven by power availability, now visible in live pricing data with mathematical clarity. Power-secure regions — Montreal, Oregon, Stockholm, Amsterdam, Frankfurt — are experiencing 70–300%+ spot price surges over 30 days, while power-congested core US markets (Virginia, Texas, Illinois) sit essentially flat. This is not noise: it is the market pricing the 14-year Northern Virginia grid interconnection queue, ComEd's 12% rate hikes across 80+ Illinois data centers, and NERC's formal escalation of AI loads as a systemic grid risk — all resolving into a single factor that now dominates regional GPU price differentiation more than GPU generation itself.

Simultaneously, the accelerator obsolescence cycle is running hot and faster than models imply. AMD's MI25 is being actively cleared out of 13 of 14 global regions in a coordinated rationalization, while Intel's GAUDI2/3 are already fully inactive across all three major hyperscalers. In Oregon, a provider exit has converted a competitive multi-provider L4 market — where spread was 2,304% just 30 days ago — into a single-provider monopoly now charging $0.567/hr vs. $0.295/hr prior. The dominant market regime is: capacity is tightening where power is available; legacy silicon is clearing everywhere else; and the operators sitting on power-secure GPU infrastructure have structural pricing power that equities markets have not yet credited.

Key Movers

Thin-market noise to disregard: M60, V710, and VIRTEX_VU47P moves reflect single-provider catalog activity with no volume significance. The MI25 Singapore -82% is analytically meaningful for the rationalization thesis but does not represent a tradeable spot opportunity.

Investable Insights

H1 — Oregon's L4/H100 Markets Have Converted to Single-Provider Monopoly Pricing

Confidence: 5 / 5

Thesis: Oregon's GPU compute market has undergone a clean provider-exit event across multiple GPU classes simultaneously. Thirty days ago, both the L4 ON_DEMAND and H100 ON_DEMAND markets in Oregon had competitive spreads exceeding 2,300% — the signature of one provider dramatically undercutting another. Today, L4 ON_DEMAND shows n_providers=1, spread=0%, price=$0.567/hr. H100 ON_DEMAND has repriced +40% in the last 7 days alone, with spread collapsing from 2,306% to 81% as the surviving provider raises to monopoly rates. The cheap provider — almost certainly a mid-tier cloud operator rather than a hyperscaler — has exited the Oregon market for both GPU classes. L4 SPOT at $0.35/hr tracking +99% in 24h confirms this is a genuine capacity event, not just list-price catalog management: you cannot fake spot prices. The surviving provider now charges $0.567/hr in Oregon for L4 ON_DEMAND vs. $0.29/hr in Virginia for the same GPU — a 96% premium sustained entirely by provider concentration, not GPU scarcity.

Key Evidence:

1. L4 ON_DEMAND Oregon: n_providers=1, spread=0%, price=$0.567/hr vs. $0.295/hr 30 days ago; prior spread was 2,304% (live ticker data, June 11, 2026)
2. H100 ON_DEMAND Oregon: +40% in 7 days ($5.96→$8.37/hr), spread collapsed from 2,306% to 81% — second provider repricing toward market leader, not yet converged (live ticker data, June 11, 2026)
3. L4 SPOT Oregon +127% over 30 days, +99% in 24h at $0.35/hr — spot markets confirm real capacity withdrawal, not just list price catalog change (live ticker data, June 11, 2026)
4. L4 ON_DEMAND Virginia: $0.29/hr, 2 providers, spread 4,435% — Virginia's competitive structure intact, confirming Oregon repricing is local, not market-wide (live ticker data, June 11, 2026)
5. Iowa L4 ON_DEMAND: $0.024/hr — astonishingly cheap single-provider capacity-clearing price; likely a neocloud dumping inventory, not a sustainable floor, but a near-term arbitrage for cost-sensitive workloads
6. Qualification: A10 standard ON_DEMAND Oregon (not dev tier) shows n_providers=1 but only -0.15% 30d change at $0.729/hr — A10 market was already single-provider, so no exit signal there; the scan's A10 headline was for the dev tier, which has less pricing weight

Implied Action: Compute buyers should immediately reroute L4 inference workloads from Oregon to Virginia (96% cheaper at $0.29/hr) or benchmark Iowa's $0.024/hr capacity-clearing price for non-latency-sensitive batch inference. For investors tracking GPU-as-a-service operators with Pacific Northwest exposure: the cost basis for Oregon L4 capacity has doubled; any neocloud underwriting infrastructure contracts on Oregon L4 at pre-exit pricing is now significantly margin-compressed. Monitor whether the exiting provider reappears in Washington or California — a West Coast withdrawal would confirm a strategic market exit rather than a temporary capacity pause.

H2 — Montreal's H100 Surge Is Early-Stage Hydro-Power Premium; Stockholm and Amsterdam Co-Signal Confirm It's Structural

Confidence: 4 / 5

Thesis: Montreal's H100 SPOT price has risen +141% over 30 days to $4.98/hr, and the mechanism is geographic rather than GPU-specific: it is the market beginning to price Quebec's clean hydro grid as a premium compute location in a world where the alternative — Northern Virginia — carries a 14-year grid interconnection queue and Illinois data center operators face 12% rate hikes from ComEd. The hypothesis is validated by geographic specificity: Stockholm H100 SPOT is up +71% over 30 days, Amsterdam H100 SPOT is up +68%. These three regions — Quebec hydro, Nordic hydro/wind, Netherlands gas-to-renewable grid — are the three power-secure North Atlantic compute zones. Meanwhile, Virginia, Texas, and Illinois H100 SPOT are within ±1% over 30 days. The 197% spread between Montreal's two SPOT providers ($2.51/hr min vs. $7.45/hr max) reveals one operator has recognized the structural premium and raised aggressively; the other remains at prior pricing. This spread will compress within 60–90 days as the cheaper provider reprices up or exits — either outcome is price-positive. Crucially, H100 ON_DEMAND Montreal has stayed essentially flat (-0.24%) during this SPOT surge, confirming this is a real-time utilization squeeze in spot markets, not a hyperscaler list price revision — spot precedes list price changes by 60–90 days on average.

Key Evidence:

1. H100 SPOT Montreal: $4.98/hr, +141% over 30 days, +19% in the last 7 days; min/max spread of 197% ($2.51–$7.45/hr) expanded dramatically from 55% 30 days ago (live ticker data, June 11, 2026)
2. H100 SPOT Stockholm: +71% over 30 days; H100 SPOT Amsterdam: +68% over 30 days — three power-secure North Atlantic zones moving in unison, confirming a structural geographic premium (live ticker data, June 11, 2026)
3. H100 SPOT Virginia, Texas, Illinois: -0.9%, -0.2%, 0.0% over 30 days — power-congested core US markets flat or declining (live ticker data, June 11, 2026)
4. H100 ON_DEMAND Montreal: -0.24% over 30 days at $7.67/hr — hyperscaler list prices stable while spot is surging; this is the early-stage signal before a structural reprice (live ticker data, June 11, 2026)
5. A100_40GB SPOT Montreal: +107% over 30 days — multi-GPU-class co-movement confirms regional phenomenon, not GPU-specific supply shock (live ticker data, June 11, 2026)
6. Forbes (May 31, 2026): Data center grid interconnection in Northern Virginia can take 14 years; ComEd warning 12% electric bill increases in Illinois with 80+ data centers already operating there
7. NERC draft reliability guideline on "Risk Mitigation for Emerging Large Loads" — grid operators formally treating AI data centers as systemic grid risk, escalating permitting timelines (intel feed, June 2026)
8. Qualification: The denial condition was partially triggered — SPOT/ON_DEMAND divergence suggests inventory squeeze rather than a full structural reprice. EQIX/DLR confirmation awaits July 29/23 earnings, respectively

Implied Action: Long EQIX ($1,038, analyst target $1,197, -4.4% over 1 month) and DLR (-7.9% over 1 month) as the primary public equity expressions of the power-secure infrastructure premium. Both are being sold in the sector-wide neocloud risk-off move, but both have material Montreal and Nordic colocation exposure — precisely the regions where realized pricing power is accelerating. The trade is a 3–6 month position with July earnings as the first catalyst confirmation. Short-term SPOT arbitrage: the $2.51/hr Montreal floor (cheapest provider) vs. $7.45/hr ceiling represents a 197% spread that has historically compressed — buy the floor provider before it reprices. Structural warning for compute buyers: any inference fleet deployed in Montreal on 30-day-old pricing contracts is now carrying costs that have nearly tripled.

H3 — SRAM Architecture Disruption Is Real But 12–36 Months Out; Inference GPU Catalog Survival Data Reveals the Timeline Is Compressing

Confidence: 3 / 5

Thesis: The cross-sectional depreciation model assigns inference GPUs (L4, A10G) an 8-year implied useful life at 12.55%/year decay — meaningfully slower than training GPUs' 5.8 years at 17.26%/year. But real-world catalog survival data contradicts the 8-year assumption with mounting evidence: Azure has already dropped the A10G (4.53 years old) and L4 (3.23 years old) from active catalogs; GCP has dropped the A10G; most strikingly, AWS's own Inferentia2 inference ASIC is already inactive on all three major hyperscalers after just 3.53 years — barely 44% of its model-implied life. Architecture news amplifies the signal: d-Matrix's Corsair (now in full production) delivers 150 TBps memory bandwidth vs. 1.2–2 TBps for GPU HBM — a 75–125× advantage at the inference memory wall. Nvidia acquired Groq for $20B and launched Groq 3 LPX SRAM inference racks. Google has booked Intel to package 3 million TPUs in 2028 using EMIB technology at below-TSMC-CoWoS cost. The near-term pricing data does not yet show SRAM displacement — inference GPU spot prices are surging (Inferentia Frankfurt +288%, L4 SPOT Oregon +127%), confirming GPU inference demand remains firm through at least 2026. But the catalog survival data and architecture announcements collectively argue the 8-year inference GPU useful life assumption used in asset valuations will need to be revised to 5–6 years — a 25–33% reduction — within 24 months. Investors with neocloud positions underwriting GPU book value at 8-year depreciation curves face a meaningful write-down risk in 2027–2028 models.

Key Evidence:

1. Inference GPU annual decay model: 12.55%/yr (implied 8.0-year life), regression p=0.059, R²=0.475 — marginally significant on 8 data points; directionally clear but statistically fragile (depreciation model, June 2026)
2. Training GPU annual decay model: 17.26%/yr (implied 5.8-year life), p=0.028, R²=0.600 — more robust statistically; confirms training/inference depreciation gap is real (depreciation model, June 2026)
3. A10G catalog survival: AWS Active, Azure Inactive, GCP Inactive — 2/3 hyperscalers dropped at age 4.53 years, vs. 8.0-year model-implied life (catalog survival data, June 2026)
4. Inferentia2 catalog survival: Inactive on ALL providers at age 3.53 years — AWS's own inference ASIC failed to survive 4 years; strongest evidence for compressed inference hardware lifecycle (catalog survival data, June 2026)
5. L4 catalog survival: AWS Active, Azure Inactive, GCP Active — Azure already dropping at 3.23 years (catalog survival data, June 2026)
6. d-Matrix Corsair in full production: 256 MB SRAM, 150 TBps bandwidth, 2,400 TFLOPS at 8-bit — 75–125× memory bandwidth advantage over GPU HBM at the inference memory wall (HPCwire, June 10, 2026)
7. Nvidia acquired Groq for $20B; launched Groq 3 LPX SRAM inference racks — Nvidia hedging its own inference disruption (AOL/news feed, June 2026)
8. AWS Inferentia SPOT Frankfurt +288% over 30 days to $0.32/hr — AWS tightening proprietary ASIC supply in EU, not expanding; consistent with concentrating on gen-1 silicon while transitioning platform (live ticker, June 11, 2026)
9. Qualification: Near-term GPU inference spot prices are surging, not falling — SRAM displacement is not yet in the price signal; 12–36 month call, not immediate

Implied Action: No immediate trade, but critical model risk action: stress-test neocloud asset valuations (particularly inference-GPU-heavy fleets) by running a 15–18% annual decay scenario vs. the 12.55% model base case — the delta represents 2–3 years of compressed useful life and materially alters IRR on GPU-leasing models underwritten in 2024–2025. For NVDA specifically (forward P/E 15.7×, $200.42): the Groq acquisition is underappreciated as a hedge — Nvidia now controls both GPU inference and SRAM inference architecture, which makes it structurally less disruptable than consensus expects. The $20B Groq acquisition cost at Nvidia's scale is equivalent to roughly 1.3 quarters of net income — cheap portfolio insurance.

H4 — AMD MI25 Global Rationalization Confirms a Coordinated Non-NVIDIA Accelerator Exit That's Faster Than the Market Prices

Confidence: 5 / 5

Thesis: The AMD MI25's price moves across 14 global regions are not organic market fluctuations — they are a centralized end-of-life inventory sweep with one anomalous exception that confirms rather than denies the thesis. Thirteen of fourteen regions are declining simultaneously: Singapore -82% (effectively zero at $0.02/hr), Montreal -57%, Milan -47%, Stockholm dev -37%, London -30%, Illinois -27%, Tokyo -26%, Korea -25%, Australia -18%, Mumbai -11%. The one outlier — Texas at +83% over 30 days to $1.14/hr — had its entire move compressed into the last 7 days (7d delta equals 30d delta), from a single provider with zero spread. This is a textbook distressed-inventory absorption event: one specific Texas customer is being invoiced at above-market rates for end-of-life GPU inventory, likely because they have existing MI25 infrastructure and need additional capacity to complete a project cycle. The temporal coincidence with Hut 8's $4.25B senior secured bond (8-K, June 10) financing a 352MW Texas data center for an AA-rated tenant is circumstantially notable. Intel's GAUDI2 and GAUDI3 are already fully inactive across all three hyperscalers — the non-NVIDIA accelerator rationalization is industry-wide. AMD has correctly identified that its real opportunity in the agentic AI cycle is not MI25 GPU competition with Nvidia, but Venice (256-core, 2nm TSMC) CPU deployment as the GPU:CPU ratio in data centers shifts from 4:1 toward 1:1 per Intel's own CFO commentary.

Key Evidence:

1. MI25 ON_DEMAND Singapore: -82% in 24h to $0.02/hr, n_providers=1 — effective catalog exit (live ticker, June 11, 2026)
2. MI25 ON_DEMAND Montreal: -56.8% over 30 days — major NA market wind-down (live ticker, June 11, 2026)
3. MI25 ON_DEMAND Texas: +82.8% over 30 days to $1.14/hr; 7d delta = 30d delta (entire move in last 7 days), n_providers=1, spread=0% — single-operator absorption event, not organic demand (live ticker, June 11, 2026)
4. MI25 pricing above comparable markets: Texas $1.14/hr vs. Ohio $0.529/hr vs. Virginia $0.219/hr — premium pricing for end-of-life GPU is distressed-inventory pattern (live ticker comparison, June 11, 2026)
5. GAUDI2 and GAUDI3: Inactive on AWS, Azure, and GCP — Intel's accelerator ecosystem fully wound down across all three major hyperscalers (catalog survival data, June 2026)
6. Intel CFO commentary: GPU:CPU data center ratio shifting from 4:1 toward 1:1 as agentic AI scales — structural tailwind for CPU plays (AMD Venice, Graviton), not GPU challengers (AOL, June 10, 2026)
7. AMD identified $120B agentic AI CPU opportunity vs. Nvidia's $200B estimate — AMD's public positioning confirms the strategic pivot away from GPU competition (news feed, June 2026)
8. Hut 8 $4.25B Senior Secured Notes at 6.129% due 2042 for 352MW Texas data center with AA- or better tenant (8-K, June 10, 2026) — temporal and geographic coincidence with TX MI25 absorption; cannot confirm connection but warrants monitoring
9. Qualification: The MI25 is a single GPU generation and may not be representative of MI300X trajectory; AMD MI300X catalog status (still active on AWS and Azure) confirms rationalization is generation-specific

Implied Action: Avoid any neocloud or GPU-leasing asset with material MI25 exposure — these are heading to zero catalog value within 12 months, confirmed by 13/14 regions declining. For AMD equity: the MI25 story is noise; the real catalyst is Venice CPU adoption in agentic AI workloads and the GPU:CPU ratio shift. AMD is not well-covered in our equity feed but the structural repositioning is confirmed by their own public statements. For distressed GPU infrastructure buyers: the Texas MI25 absorption pattern ($1.14/hr, single operator) suggests there may be sub-market MI25 inventory available in exiting geographies (Singapore, Montreal) — a potential opportunistic HPC or government acquisition at clearance prices before catalog removal. Track Hut 8's next disclosure for tenant identification — if AA-rated and Texas-based, the MI25 connection becomes investable context for the broader HUT bond thesis.

H5 — Neocloud Equity Selloff Is a Forced-Liquidation Overshoot for Power-Secure Large-Caps; Bifurcation Within the Sector Is the Actual Trade

Confidence: 3 / 5

Thesis: The simultaneous selloff in neocloud equities — Applied Digital (APLD) -12.7% over 1 month, CoreWeave (CRWV) -16.6%, IREN -6.6%, EQIX -4.4%, DLR -7.9% — is being driven by macro risk-off on the back of the Crusoe Project Jade halt and SMCI's -17% dilution reaction, not by deteriorating GPU pricing fundamentals for the sector's better-positioned operators. The pricing data makes this distinction sharply: power-secure regions where large-cap neoclouds concentrate capacity are repricing up significantly (H100 SPOT Montreal +141%, L4 SPOT Oregon +127%, Inferentia Frankfurt +288%), while the operators are being sold as if capacity pricing is universally declining. The analyst consensus has not revised down — every major name in the sector carries buy or strong-buy ratings with 15–82% upside to targets. APLD at $38.92 vs. a $71 analyst target (strong_buy) represents an 82% consensus discount for an operator concentrated in North Dakota and Texas — both repricing-up markets. CRWV at $95.61 vs. $140 analyst target (+46%) is pure-play Nvidia H100/H200 capacity in US markets where Virginia H100 SPOT has held steady at -0.9% over 30 days. The bifurcation within the sector is what separates this from a uniform shorting opportunity: micro-cap, pre-revenue neocloud vehicles (AGPU, AIB, DTCX) are pricing in genuine existential risk and should be avoided entirely. HIVE Digital at $3.50 vs. $6.93 analyst target (+98%) with no clear power-secure footprint is a value trap, not an opportunity.

Key Evidence:

1. APLD (Applied Digital): $38.92, analyst target $71, recommendation strong_buy; -12.7% over 1 month, -7.1% in 24h — 82% discount to consensus on a power-secure footprint (equities feed, June 11, 2026)
2. CRWV (CoreWeave): $95.61, analyst target $140, recommendation buy; -13.8% over 5 days — pure-play H100/H200 operator in US markets with flat underlying spot pricing (equities feed, June 11, 2026)
3. EQIX (Equinix): $1,038, analyst target $1,197, -4.4% over 1 month — significant Montreal and Nordic colocation exposure in the exact regions where GPU SPOT is +71–141% over 30 days (equities feed, June 11, 2026)
4. DLR (Digital Realty): analyst target above current price, -7.9% over 1 month — same power-secure geographic exposure argument as EQIX (equities feed, June 11, 2026)
5. Oracle Q4 FY2026: OCI revenue +93% to $5.8B, RPO +363% YoY to $638B, GPU utilization 97.5% — hyperscaler-level demand confirming GPU capacity is scarce, not oversupplied (ERP Today, June 11, 2026)
6. Hut 8 $4.25B Senior Secured Notes at 6.129% for 352MW Texas facility with AA- tenant — institutional project finance at investment grade for data center infrastructure despite broader sector stress (8-K, June 10, 2026)
7. SMCI -17% on $7B financing deal; Bloom Energy -9% on Crusoe Project Jade halt — market negative sentiment drivers are financing dilution and a single demand pause, not GPU pricing deterioration (news feed, June 11, 2026)
8. Critical risk — denial condition partially triggered: Goldman Sachs explicitly warned the Crusoe halt "triggers AI capex demand reassessment." If this is systemic rather than isolated, the entire sector re-rates lower regardless of spot prices. This is the binary risk event for the hypothesis (KuCoin/Goldman Sachs, June 11, 2026)
9. Nebius £1.7B UK investment in three Blackwell Ultra facilities by 2027 — continued aggressive European neocloud buildout confirms demand commitment at the institutional level despite headline noise (W.Media, June 11, 2026)

Implied Action: Selective long positions in power-secure large-caps: APLD (82% analyst upside, strong_buy, North Dakota/Texas footprint) and CRWV (46% analyst upside, US H100/H200 capacity) are the highest-conviction asymmetric setups in the sector if Crusoe proves isolated. EQIX and DLR are lower-beta expressions of the same thesis with July earnings catalysts. Strictly avoid micro-cap pre-revenue neocloud vehicles — the pricing environment favors operators with scale and power-security, not speculative capital formation. Key binary event: Monitor the next 2–3 weeks for any additional major compute project pauses or hyperscaler capex guidance reductions — a second Crusoe-style halt would confirm systemic demand softening and invalidate the overshooting thesis. July 29 APLD earnings is the sector confirmation date.

Risk Flags

Immediate (0–30 Days)

R1 — Oregon L4/H100 Monopoly Pricing Creates Buyer Disruption Risk Any compute buyer or neocloud with Oregon L4 ON_DEMAND contracts anchored to pre-exit pricing (~$0.295/hr) is now facing a 92% cost increase with no competitive backstop — n_providers=1 means there is no alternative Oregon supplier to renegotiate against. The H100 ON_DEMAND market in Oregon is mid-repricing (+40% in 7 days, spread still at 81%) — the second provider has not yet fully matched the leader, creating a narrow 7–14 day window to lock contracts with the cheaper remaining provider before convergence. After convergence, expect H100 ON_DEMAND Oregon to settle meaningfully above $8.37/hr. Any SLA or rate-card agreement referencing Oregon pricing should be flagged for immediate renegotiation.

R2 — Montreal H100 SPOT Spread (197%) Is Unstable and Will Resolve Within 60 Days A 197% spread between two providers in the same region ($2.51/hr vs. $7.45/hr) is not a stable market structure. Resolution paths: (a) the cheap provider raises to match, which is price-positive; (b) the cheap provider exits, also price-positive for the remaining operator; or (c) new capacity enters, compressing both. Path (c) is the risk for neocloud operators who have underwritten Montreal capacity at the current high watermark ($7.45/hr). If spread compression happens via new entrant rather than floor repricing, operators holding Montreal H100 inventory at near-peak rates face margin compression before the next contract cycle.

R3 — Crusoe Project Jade Pause Is an Unresolved Binary Goldman Sachs' explicit warning that the 1.8GW Crusoe halt "triggers AI capex demand reassessment" has not been resolved. If a single additional major compute project pauses in the next 30 days, the narrative shifts from isolated client decision to systemic softening — and the entire sector re-rates. This is the highest-impact near-term risk for Hypothesis 5 and directly relevant to neocloud equity positions. Watch Crusoe's public communications and any AWS/Azure/GCP guidance adjustments to hyperscaler build schedules.

Near-Term (30–90 Days)

R4 — AMD MI25 Texas Absorption Ends Without Warning The single Texas operator absorbing MI25 inventory at $1.14/hr (above-market for end-of-life silicon) will eventually complete their inventory draw. When they stop, Texas becomes the 14th declining region and the MI25 global price collapses to near-zero uniformly. Any infrastructure investor with MI25 Texas exposure has a limited window — weeks to months — before the absorption event concludes. The risk is not directional (decline is certain) but timing: don't mistake the Texas premium for a signal of MI25 demand recovery.

R5 — SRAM Inference Architecture Crossing the Volume Threshold d-Matrix Corsair is "in full production" per HPCwire (June 10, 2026) but from an unknown base. If any major hyperscaler or neocloud announces a volume SRAM inference deployment contract in the next 90 days — whether with d-Matrix, Groq (now Nvidia-owned), or Cerebras — it will trigger an immediate reassessment of inference GPU spot pricing and depreciation assumptions. The L4/A10G SPOT prices currently look strong, but they are supported by demand that has no SRAM alternative at scale. That changes when the first volume SRAM contract is announced publicly. Watch for AWS/Azure inference ASIC procurement RFPs and neocloud job postings for SRAM-inference-specific engineering roles as leading indicators.

R6 — NERC Grid Guidelines Converting to Hard Interconnection Reform The NERC draft reliability guideline on large-load risk mitigation is currently a draft. If it advances to a formal interconnection standard — likely within the 90-day comment and adoption window — it becomes a mandatory compliance item that adds 12–36 months to data center permitting timelines in PJM and MISO territories. This is directly bullish for existing capacity (scarcity premium rises) but bearish for greenfield projects and operators relying on capacity expansion for revenue growth. Monitor NERC's comment deadline and FERC's 2025 State of Markets implementation guidance.

Structural (6–24 Months)

R7 — Power Availability Becomes a Binding Constraint on GPU Capacity Growth, Creating a Two-Tier Market The pricing divergence between power-secure regions (+71–326% SPOT in 30 days) and power-congested regions (flat to -1%) is not a temporary dislocation — it is the early-stage emergence of a structurally bifurcated GPU market where power availability, not GPU generation, is the dominant pricing determinant. JLL's projection of 100 GW of new data center capacity by 2030 is physically impossible to execute at current grid interconnection timelines (14 years in Virginia; CAISO queue heavily subscribed). The implication: GPU capacity in power-secure regions (Montreal, Stockholm, Amsterdam, Oregon, portions of Texas with wind access) will command persistent premiums over power-congested regions. Operators and investors who treat GPU pricing as a uniform commodity market will systematically underestimate returns in the former and overestimate them in the latter. This structural thesis has a 12–36 month full expression horizon but the price signals are already live.