Daily Investment Brief — June 17, 2026

Market Pulse

The GPU compute market is not consolidating — it is fragmenting along three independent fault lines simultaneously: generation (H100 vs. Blackwell), geography (cheap-core US/Frankfurt vs. premium-edge APAC/Canada/Nordics), and product lifecycle (active ASIC deprecation of AWS Inferentia vs. deliberate margin extraction in EU). The clearest single-market data point capturing all three dynamics: H100 SPOT in US-Iowa is locked at $1.25/hr — a committed structural floor set by one aggressive low-cost entrant since April 16 — while the identical chip in Montreal sits at $4.98/hr (+93% over 30 days) as a new premium entrant entered and widened the ceiling. That 4× gap, confirmed stable over 60 days in Iowa and widening in premium markets, is the dominant narrative. Overlaid on this: AWS Inferentia SPOT in US-Virginia has collapsed to $0.005/hr — effectively zero — confirming a deliberate product deprecation underway, while Frankfurt Inferentia has simultaneously risen 5.5× to $0.44/hr since March. The macro backdrop — $725B in projected 2026 AI infrastructure spend, 14-year Northern Virginia interconnection queues, and 14 states actively legislating data center constraints — ensures that geographic access to power and fiber is now the rate-limiting variable in this market, not silicon supply.

Key Movers

Investable Insights

H1 — APAC Legacy GPU Scarcity: Cyclical Premium, Actively Tradeable Confidence: 4 / 5

Thesis: End-of-life GPUs command a genuine and recurring scarcity premium in APAC markets that is structurally disconnected from US pricing. This premium is not permanently stable — the 90-day V100 Tokyo history reveals two distinct demand waves, with a parabolic first peak of $4.69/hr on March 3 followed by a trough near $0.28/hr in mid-April before re-ascending to today's $1.67/hr (+477% over 30 days). The mechanism is real: Japan's AI buildout (SoftBank, NTT, Sakura Internet) is absorbing legacy V100/A100 inventory in a market with no domestic foundry alternative and data residency constraints preventing migration to cheaper US-based H100 spot. The A100_40GB market is structurally more stable than V100: Tokyo has been anchored near $1.92/hr for most of the 90-day window — a ceiling set by two providers holding price — while a third discounts aggressively, producing a 287% intra-market spread. This A100_40GB price is 4.6× the US-Texas equivalent ($0.42/hr) and is holding because the discount entrant has not yet eroded the two-provider ceiling. The key distinction from noise: three providers participate in A100_40GB Tokyo, confirming real market demand rather than a single-SKU repricing event.

Key Evidence:

1. V100_32GB Tokyo SPOT: $1.67/hr today, peak $4.69/hr March 3, trough $0.28/hr mid-April — confirmed two-wave cyclical pattern over 90 days (ticker history, June 17, 2026)
2. A100_40GB Tokyo SPOT: $1.92/hr, n=3 providers, spread_pct 287% — multi-provider participation confirms demand validity, not artifact (live ticker, June 17, 2026)
3. A100_40GB cross-regional comparison: Tokyo $1.92/hr vs. US-Texas $0.42/hr (4.6×), Sydney $1.37/hr (+69% over 30d), Montreal $1.10/hr (+120% over 30d) — APAC premium is geographically consistent, not Tokyo-specific (live ticker scan, June 17, 2026)
4. QUALIFYING: The V100 Tokyo cycle shows price can collapse 94% in 6 weeks (from $4.69/hr to $0.28/hr) — cyclical, not structural; the premium requires active supply management to monetize
5. QUALIFYING: A discount entrant in A100_40GB Tokyo ($0.58/hr min vs. $2.01–$2.24/hr max) is actively competing — if this provider scales, it can erode the $1.92/hr ceiling rapidly

Implied Action:

• Operators with APAC V100/A100 inventory: The current V100 re-ascent is the second wave. If the prior cycle's timing holds, the next peak may approach $3–4/hr within 3–4 weeks — a high-return short-dated window for operators who can selectively list capacity onto spot markets. A100_40GB is the more durable play given multi-provider ceiling discipline.
• Workload buyers in APAC: Running legacy A100_40GB spot in Tokyo at $1.92/hr when US-Oregon H100 SPOT sits near $1.70/hr is economically irrational unless hard data residency or latency constraints apply. The $/TFLOP arbitrage is 20–40×.
• Monitor trigger: A100_40GB|SPOT|jp-tokyo spread_pct — if it begins compressing below 100%, the discount entrant is winning and the premium ceiling is being eroded. That is the exit signal.

H2 — H100 Two-Tier Bifurcation: The Clearest Actionable Arbitrage in the Dataset Confidence: 5 / 5

Thesis: The H100 spot market has bifurcated into structurally distinct pricing tiers that are not converging — they are diverging. A "cheap core" tier consisting of US-Iowa ($1.25/hr), DE-Frankfurt ($1.79/hr), US-Virginia ($2.62/hr), and US-Ohio ($2.27/hr) stands in contrast to a "premium edge" tier of SE-Stockholm ($3.59/hr), US-California ($4.84/hr), CA-Montreal ($4.98/hr), and AU-Sydney ($5.12/hr). The gap between cheapest and most expensive spot markets — Iowa to Montreal — is 4×, and critically, it is widening: Montreal is up 93% in 30 days while Iowa has been stable at $1.24–$1.25/hr for 60 consecutive days. The mechanism behind Iowa's cheapness is now confirmed: on April 16, a single aggressive entrant — almost certainly OCI Supercluster — entered the market and drove the median from $3.04/hr to $1.25/hr in one day, then committed to that pricing floor continuously. Meanwhile, premium-edge regions are rising because no equivalent competitive entrant has appeared: Stockholm +61%, Montreal +93%, California +25% — all in 30 days. Frankfurt is the most compelling "next Iowa" candidate in EU: currently at $1.79/hr with 0% spread (committed single provider, down -27% over 30d) — becoming the structural cheap-core anchor for European workloads. The H100 on-demand ceiling of $8.69/hr in US-Virginia is completely flat (0.00% delta over 30 days) — providers are holding list prices despite Blackwell availability, confirming they see no demand erosion that would justify discounting.

Key Evidence:

1. H100 SPOT us-iowa: $1.25/hr, stable for 60+ consecutive days post-April 16 structural discontinuity — median dropped $3.04→$1.25 in a single day; spread_pct 4,620–8,797% confirms two-provider bipolar structure with one aggressive floor-setter (ticker history, June 17, 2026)
2. H100 SPOT ca-montreal: $4.98/hr (+93% over 30d); spread_pct_30d_delta_pct +3,612% — a new premium entrant entered the market and pushed the ceiling up, the exact inverse of Iowa's dynamic (live ticker, June 17, 2026)
3. H100 SPOT se-stockholm +61% over 30d ($3.59/hr), us-california +25% ($4.84/hr), au-sydney $5.12/hr — premium-edge regions uniformly appreciating, zero competitive entrant pressure (live ticker scan, June 17, 2026)
4. H100 ON_DEMAND us-virginia: $8.69/hr, 0.00% delta over 30 days — deliberate list-price rigidity despite Blackwell availability confirms providers see no substitution pressure (live ticker, June 17, 2026)
5. H100 SPOT ie-dublin: spread_pct 6,580% (min $0.04, max $2.53) — one provider near-dumping while another holds $2.50; confirms intra-market fragmentation is the structural reality, not a temporary anomaly (live ticker, June 17, 2026)
6. NVIDIA Blackwell GB300 NVL72 MLPerf Training 6.0 clean sweep at 8,192-GPU scale — Blackwell supply going to hyperscalers on 1–3yr reserved commitments, not into spot market (news feed, June 16, 2026)

Implied Action:

• Immediate workload arbitrage: Any ML team running H100 in Stockholm, Montreal, or California without hard data residency requirements can cut compute costs 60–75% today by migrating to Iowa or Frankfurt. This is operationally actionable within a sprint cycle.
• Provider positioning: OCI's Iowa presence and GCP's Oregon/Frankfurt concentration place them as the structural beneficiaries of the cheap-core tier. Neoclouds anchored in California, Canada, or Nordics face margin pressure as buyer geographic flexibility increases.
• APLD (Applied Digital, next earnings July 29): Strong-buy consensus, +68% over 3 months, US-centric footprint — watch earnings call for realized GPU-hour pricing commentary as a bellwether for cheap-core tier profitability.
• Trigger to monitor: H100 SPOT us-ohio at $2.27/hr (+23% over 30d) — if Ohio continues rising toward $3+/hr, it signals even secondary US markets are tightening as pre-secured power runs out, expanding the cheap-core tier's scarcity value.

H3 — AWS Inferentia: A 90-Day Product Abandonment Signal in US-East, Margin Extraction in EU Confidence: 4 / 5

Thesis: AWS is not managing Inferentia spot capacity dynamically — it is executing a deliberate regional succession strategy. In US-Virginia, Inferentia SPOT has collapsed from $0.125/hr on January 8 to $0.005/hr today — a 96% decline over 160 days with n=1 provider — a trajectory that is too smooth and sustained to be a spot market fluctuation. AWS is pricing near-zero to accelerate migration of inference workloads off Inf1 and onto Trainium2 in US-East, where Trainium2 capacity is concentrated. In Frankfurt, the identical product has done the exact opposite: after a mid-cycle trough of $0.064/hr in February, it made a clean, consistent ramp to $0.44/hr today — a 5.5× appreciation since March, with a smooth curve that is the signature of deliberate pricing action, not capacity noise. The mechanism: Frankfurt H100 SPOT at $1.79/hr is cheap enough that AWS European customers have less urgency to migrate away from Inferentia than US-East customers (where H100 SPOT sits at $2.62/hr), allowing AWS to extract margin before Trainium2 EU supply arrives — likely 1–2 quarters out. The INFERENTIA2 (Inf2) catalog shows inactive status at AWS, Azure, and GCP — confirming this is a generation-end lifecycle event, not a temporary capacity management decision.

Key Evidence:

1. INFERENTIA SPOT us-virginia: $0.005/hr (June 17, 2026), down from $0.125/hr (Jan 8, 2026) — 96% decline over 160 days, n=1 provider, smooth monotonic descent confirming deliberate pricing action (ticker history, June 17, 2026)
2. INFERENTIA SPOT de-frankfurt: $0.44/hr (June 17), up from $0.064/hr trough (Feb 10) — 5.5× appreciation over ~4 months; smooth upward curve rules out spot artifact (ticker history, June 17, 2026)
3. INFERENTIA2 catalog status: inactive at AWS, Azure, and GCP — generation-end lifecycle signal confirmed (depreciation catalog data, June 17, 2026)
4. Frankfurt H100 SPOT: $1.79/hr (-27% over 30d) — relatively cheap European H100 alternative reduces urgency for EU customers to migrate, giving AWS the window to extract Inferentia margin (live ticker, June 17, 2026)
5. AMD MI300X gaining inference design wins at top-4 hyperscalers per $725B 2026 infrastructure spend projections — eroding the moat that made purpose-built ASIC inference economically defensible (news feed, June 16, 2026)
6. QUALIFYING: If the Frankfurt spike reverses sharply in 7–14 days, it could indicate a spot capacity artifact; 90-day history showing a clean ramp rather than a spike substantially reduces this risk

Implied Action:

• Enterprise buyers with EU Inferentia reserved capacity: Do not renew or extend Inferentia RI contracts in Frankfurt — the current elevated pricing reflects near-term margin extraction before the product cycle ends. The ramp will reverse when Trainium2 EU supply arrives, expected within 1–2 quarters. Negotiate short-term on-demand commitments only.
• US-East buyers: AWS is effectively subsidizing migration off Inf1 in Virginia at $0.005/hr — any enterprise still running production workloads on reserved Inf1 in US-East should immediately calculate conversion economics to Trainium2. The migration window is being incentivized by AWS pricing itself.
• Investor angle (AVGO): The Inferentia story is a cautionary signal for pricing in cloud-operator ASIC moats at a premium. Broadcom (+17% over 3 months) is the beneficiary of the counter-thesis: purpose-built custom silicon designed by specialists (Google TPU, Meta MTIA) rather than cloud operators captures the durable ASIC inference opportunity.

H4 — Power Moat: Grid Access Is Now the Rate-Limiting Variable in GPU Cloud Deployment Confidence: 4 / 5 (structural) | 3 / 5 (direct pricing causation)

Thesis: The energy infrastructure chokepoint has crossed from near-term risk to permanent structural reality. Northern Virginia's grid interconnection queue is running 14 years (confirmed by Forbes, May 2026) — not a forecast, an active queue measurement. Florida SB 484 has been enacted, creating the first state-level statutory local government veto over new data center approvals and explicitly prohibiting utilities from passing data center load costs to residential customers. NERC is drafting "Emerging Large Loads" reliability guidelines specifically targeting data centers, adding compliance hardware requirements to new interconnection applications. The 14-state legislative wave means the geographic diversification playbook that operators relied on through 2023 is increasingly closed: every secondary market that attracted data center migration (Missouri, Utah, Indiana) is now developing its own opposition ecosystem within 18–24 months of buildout announcements. The critical investment nuance revealed by the Iowa H100 history: power access alone does not guarantee cheap GPU pricing. Iowa's post-April cheapness required both pre-secured power and a competitively aggressive provider (OCI) committing to low-cost spot pricing. Regions with secured power but limited competition (Ohio at $2.27/hr, +23% over 30d) are still appreciating. The cleanest formulation of the moat: geographic entry is now structurally impossible for new operators in most US coastal and EU markets, creating durable oligopoly dynamics wherever capacity is installed — and the pricing evidence is visible in constrained markets like Stockholm (+61% over 30d) and California (+25%), where incumbents face zero competitive threat from new builds.

Key Evidence:

1. Northern Virginia grid interconnection queue: 14 years (Forbes, May 2026) — not a forecast, an active measurement; no new operator can secure power in the region's primary GPU cloud market at any price within a reasonable investment horizon
2. Florida SB 484 enacted: local government veto power over new data center approvals, utility residential cost protection — first state-level enacted constraint; template risk for 13 other states (news feed, June 16, 2026)
3. NERC "Emerging Large Loads" reliability guidelines drafted: frequency/voltage compliance hardware requirements for new large load interconnections — adds $10–50M+ compliance cost to new builds (intel feed, May 2026)
4. H100 SPOT pricing differential: Iowa $1.25/hr vs. Stockholm $3.59/hr (+61% over 30d), California $4.84/hr (+25%), Montreal $4.98/hr (+93%) — constrained regions uniformly appreciating with no competitive entrant relief (live ticker scan, June 17, 2026)
5. Amazon announcing multi-billion Missouri data center campus as part of $200B 2026 capex plan — major hyperscaler forced to secondary market due to primary market saturation, confirming the constraint is operational, not theoretical (news feed, June 16, 2026)
6. QUALIFYING: Iowa cheapness reflects provider competition, not power cost alone — H100 pricing in a pre-secured power region can still be $2.45/hr if only one provider operates there; the moat requires both secured power AND competition

Implied Action:

• Long infrastructure (established operators): Equinix, Iron Mountain, Digital Realty — operators with pre-2024 power secured in US Midwest corridors hold a compounding structural moat. The value of their interconnect agreements increases each year the queue lengthens.
• On-site generation as moat bypass: Centrica-style DC microgrids achieving 15–20% efficiency gains while bypassing grid interconnection queues are the highest-leverage pick-and-shovel play. Hadron Energy's S-1 (filed June 15, 2026) is speculative but consistent with the thesis.
• Short risk — new-build neoclouds: Any GPU cloud operator with heavy concentration in post-2024 builds in Northern Virginia, California, or regulated EU markets faces compound exposure: rising power costs, permitting delays, and active local opposition. Scrutinize APLD's July 29 earnings call specifically for disclosure of which builds have pre-secured versus post-2024 power agreements.
• Monitor: PJM interconnection queue filings for Ohio, Indiana, and Pennsylvania — accelerating timelines in secondary US markets would signal hyperscaler power scramble is widening geographically.

H5 — H100 On-Demand Rigidity vs. Spot Volatility: The Real Pricing Structure Confidence: 3 / 5

Thesis: The original hypothesis — that H100 spot floors are "rising" — is not confirmed by the 90-day data, but it points to a more precise and actionable observation: on-demand prices are perfectly sticky downward while spot prices are highly supply-elastic in both directions. The Virginia on-demand price of $8.69/hr has not moved by a single cent over 30+ days — a deliberate margin hold by providers who see no competitive reason to discount, validated by the same NVIDIA Blackwell MLPerf sweep that was supposed to create substitution pressure. Meanwhile, spot prices in Virginia have been stable but not rising: the min has oscillated between $2.21–$2.25/hr for the entire 90-day window, and the median has drifted only modestly from ~$2.27/hr to $2.62/hr (+15% over 5 months). The Montreal history reveals the critical mechanism: spot prices can collapse from $5/hr to $1.93/hr in weeks when a cheap entrant arrives, and re-spike to $4.98/hr when that entrant reduces capacity. This is supply-elastic spot pricing, not a demand floor. The investment read-through: the on-demand/spot gap in Virginia (spot $2.62/hr vs. on-demand $8.69/hr = 70% discount) persists because providers maintain list prices as an anchor, not because they genuinely believe they can sell at $8.69/hr at scale. Any Reserved Instance (RI-1YR) negotiation anchored off on-demand list pricing will contain a large extractable discount — the on-demand price is the ceiling, not the clearing price.

Key Evidence:

1. H100 ON_DEMAND us-virginia: $8.69/hr, 0.00% delta over 30 days — complete price rigidity over a full month during a period of active Blackwell supply scaling (live ticker, June 17, 2026)
2. H100 SPOT us-virginia min_price range: $2.21–$2.25/hr for 90-day window; median drift +15% Jan→Jun ($2.27→$2.62/hr) — stable floor, not a rising floor (ticker history, June 17, 2026)
3. H100 SPOT ca-montreal: collapsed from $5/hr to $1.93/hr (late April/May) then re-spiked to $4.98/hr on entrant withdrawal — confirms spot is supply-driven, not demand-floored (ticker history, June 17, 2026)
4. $725B projected 2026 AI infrastructure spend (nearly double 2025) — demand side strong enough to prevent on-demand discounts; Blackwell supply going into reserved hyperscaler contracts, not spot (news feed, June 16, 2026)
5. QUALIFYING: Virginia spot max price showed volatility between $2.24–$4.01/hr (Feb–Mar spike likely from large training job bursts) — on-demand rigidity holds even during these demand surges, confirming the ceiling is a strategic hold, not a clearing price

Implied Action:

• RI negotiation strategy: The 70% gap between Virginia on-demand ($8.69/hr) and spot ($2.62/hr) represents the negotiation floor for any 1-year reserved capacity contract. Buyers who accept on-demand anchoring in RI negotiations are leaving substantial value on the table; spot pricing should be the reference for RI floor pricing.
• Don't short on-demand H100 prices: The flat on-demand price is a strategic hold, not a pre-collapse plateau. Providers see no reason to lower list prices when spot capacity absorbs price-sensitive buyers — the two-tier market structure is intentional.
• Early indicator for H100 displacement: Monitor L40S|SPOT|us-virginia as the leading signal — if AMD MI300X inference traction materializes, L40S (the purest inference GPU) will soften before H100 does. No H100 on-demand cut is credible until L40S shows weakness first.

Risk Flags

Immediate (0–30 Days)

R1 — Thin-Market Pricing Spikes Creating False Signals The ALVEO_U30 Oregon +2,256% and Dublin +107% over 30 days are near-certain artifacts of n=1 provider markets where a single SKU repricing event dominates the median. These are noise. The more dangerous version of this risk is in actively watched tickers: the H100 Dublin spread of 6,580% (min $0.04, max $2.53) means the median is being computed from a two-provider market where one participant is behaving anomalously. Any workload cost model anchored to Dublin H100 median pricing will be systematically wrong. Validate all thin-market (n≤2 provider) pricing against natex data before making capacity decisions — and check whether the natex_median confirms or diverges from the reported median.

R2 — V100 Tokyo Second-Wave Peak Timing Risk The V100 Tokyo second wave is currently at $1.67/hr and ascending. The first wave peaked at $4.69/hr on March 3 then collapsed 94% in 6 weeks. Operators with APAC V100 capacity who are not actively managing spot availability windows risk being caught holding inventory when supply is restored and the wave collapses. The prior collapse was supply restoration (GCP/OCI reinstating capacity), not demand destruction — meaning it can happen without any demand-side warning signal. Set a hard exit threshold and monitor weekly, not monthly.

R3 — Inferentia Frankfurt Reversal Risk Inferentia Frankfurt at $0.44/hr is the result of 4 months of deliberate AWS pricing action. When AWS announces Trainium2 EU availability — which could come at any major event (re:Invent pre-announcements, SDK updates, or AWS EU Summit) — the Frankfurt ramp will reverse rapidly, potentially retracing to the Virginia near-zero level within weeks. Any enterprise customer holding Frankfurt Inferentia spot exposure above $0.20/hr is carrying event-driven reversal risk.

Near-Term (30–90 Days)

R4 — GCP A10 DEV Tier Repricing Spillover The +125% (Oregon) and +105% (Mexico) simultaneous surge in A10 ON_DEMAND_DEV over 30 days, confirmed at +133% and +115% over 7 days, is a platform-wide GCP tier repricing event affecting two geographies simultaneously. The DEV tier repricing typically precedes standard tier adjustments by 30–60 days in historical GCP pricing behavior. If this follows prior patterns, A10 standard ON_DEMAND pricing in Oregon and Mexico could see 30–60% adjustments by August. Monitor A10|GPU|ON_DEMAND|us-oregon weekly. The A10 is also a primary inference GPU — any repricing here will accelerate AMD MI300X competitive benchmarking discussions at enterprises.

R5 — CRDO Customer Concentration Post-Acquisition Integration Risk Credo Technology's 10-K (filed June 2026) disclosed two acquisitions — Hyperlume (Sep 2025) and CoMira Solutions (Feb 2026) — alongside material customer concentration risk. In AI infrastructure connectivity, customer concentration at a component supplier creates binary revenue risk: the loss of a single hyperscaler design win can eliminate 20–40% of revenue in a quarter. The integration of two acquisitions simultaneously while managing hyperscaler-level technical requirements is a compounding operational risk. CRDO is a bellwether for the AI connectivity layer — any earnings miss driven by concentration or integration execution will ripple into the broader AI infrastructure supply chain sentiment.

R6 — H100 Ohio Scarcity Inflection H100 SPOT Ohio is at $2.27/hr (+23% over 30 days). Ohio has been positioned as a secondary cheap-core market with pre-secured power from major hyperscaler builds. A continued rise toward $3/hr would signal that even secondary pre-secured US markets are absorbing capacity without competitive relief — narrowing the cheap-core tier from five regions to effectively two (Iowa and Frankfurt). This would be a significant tightening signal and would strengthen the power moat thesis beyond its current 4/5 confidence.

Structural (6–24 Months)

R7 — The 14-State Legislative Wave Reaches Critical Mass Florida SB 484 is the first enacted state-level constraint. Maine passed a ban (vetoed by the governor — this time). The pattern from prior infrastructure regulatory cascades (offshore wind permitting, broadband regulation) is that a first enacted state law in a major market creates a template that accelerates adoption in 3–5 additional states within 18 months. The investable implication is not "data center construction stops" — it is that the regulatory premium for pre-approved, pre-built capacity in compliant jurisdictions will compound. Operators with existing permitted capacity in Missouri, Iowa, Ohio, and Indiana hold a structural option value that does not yet appear on their balance sheets. The risk for new entrants attempting builds in any of the 14 actively legislating states is a potential permitting halt mid-construction — the most destructive capital allocation outcome.

R8 — Blackwell Supply Release Creating Sudden H100 On-Demand Pressure The current on-demand rigidity ($8.69/hr Virginia, flat for 30+ days) holds because Blackwell (B200/GB300) supply is going into hyperscaler reserved contracts rather than the open spot market. If NVIDIA or a major hyperscaler releases excess B200 capacity into the spot market — even a modest allocation — the psychological anchor on H100 on-demand pricing would break. The leading indicator to monitor is whether B200/H200 SPOT pricing appears in Iowa or Ohio at any level below $6/hr: that would signal Blackwell is competing for the same budget as H100, not just the premium segment. Given NVIDIA's MLPerf clean sweep at 8,192-GPU scale, the performance differential is large enough that early B200 spot availability could immediately reset buyer willingness to pay for H100 on-demand contracts.

R9 — NERC Compliance Costs Becoming a Material Barrier to Entry NERC's draft "Emerging Large Loads" reliability guidelines introduce frequency and voltage compliance hardware requirements for new large load interconnections (intel feed, May 2026). These are not discretionary — NERC reliability standards are mandatory for grid-connected facilities. For existing operators with grandfathered interconnects, this creates zero new cost. For new entrants, this adds a material layer of technical and financial compliance burden to every new interconnection application, effectively raising the minimum viable capex for a new GPU cloud build by $20–100M depending on interconnection scale. Combined with the 14-year Virginia queue and 14-state legislative pressure, the structural barrier to new GPU cloud market entry is compounding in three simultaneous dimensions: power access, regulatory permitting, and compliance cost. This is the most durable long-term moat for incumbents — and the most durable headwind for the "10 new GPU clouds will emerge" bull case.

Brief compiled June 17, 2026. All prices from live cross-provider ticker data and 90-day history pulls. News and regulatory signals sourced from feeds dated June 15–17, 2026. Charts use inline data from query steps. This brief is analytical output, not financial advice.