Daily Investment Brief — June 18, 2026

Market Pulse

The compute market is bifurcating along two axes simultaneously: geography and workload type. In training silicon, H100 spot prices are recovering sharply in APAC — Tokyo's 57-day unbroken uptrend and Stockholm's +60% over 30 days contrast starkly with Iowa spot still anchored at $1.24/hr, 14% of its own 1-year reserved rate of $8.60/hr. US core remains flooded with inventory while APAC is genuinely cycle-turning. In inference silicon, the signal is cleaner: mid-tier GPUs (A10, L4) are repricing upward in frontier and EM markets while softening in EU/US core, as production AI deployments consume capacity in geographies that built last. Overlaying both: AWS Inferentia spot in Frankfurt has completed a textbook four-month trough-to-scarcity recovery — from $0.064/hr at the February 10 floor to $0.449/hr today — with a mirror-image collapse in Virginia to $0.003/hr, confirming this is deliberate geographic inventory management, not random volatility. The dominant structural narrative is the supply glut is regional, not global — and the regions running out of capacity are exactly the ones growing fastest.

Key Movers

Entries marked are confirmed thin-market artifacts. T4 São Paulo's 1,332% spread_pct disqualifies it as a tradeable signal. Disregard ALVEO_U30 and V100 Tokyo as noise.

Investable Insights

H1 — The H100 Geographic Bifurcation: APAC Is Cycling Up, US Core Is Still Flushing Confidence: 4 / 5

Thesis: The H100 training GPU market is not uniformly bearish or uniformly recovering — it is splitting cleanly along geographic lines in a way that has direct implications for neocloud margin trajectories and enterprise contracting strategy. APAC is in a confirmed multi-month demand recovery: H100 spot in Tokyo hit a trough of $2.32/hr on April 22, reversed without interruption, and has now posted 57 consecutive days of appreciation to $2.85/hr, with the two-provider spread compressing from 66% to 5.2% — the hallmark of genuine demand equilibration, not a single-provider reprice. Stockholm is even more aggressive: from a February trough of $1.72/hr to $3.61/hr today, a +60% move over 30 days accompanied by spread widening to 82%, which means a second, more bullish provider has repriced into the market — a different and arguably stronger signal than spread compression. Meanwhile, US core is structurally oversupplied: Iowa spot at $1.24/hr represents just 14.4% of its own 1-year reserved rate of $8.60/hr, and California spot at $4.93/hr is 73.7% of a $6.69/hr reserved floor that has not moved a single dollar in 91 consecutive days. Reserved is an administered list price, not a market-clearing mechanism — the inversion will only close when spot rises, not when reserved falls, and in US core that recovery has not started.

The TSMC CoWoS packaging constraint (new line qualification takes years, not quarters) is the structural ceiling on supply response. Even with Oracle's stated $90–95B in combined FY2027 capex and customer prepayments, new H100-class supply cannot ramp fast enough to further flood an already-flooded US market. The direction of travel is spot recovery — the question is timing and geography.

Key Evidence:

1. H100 SPOT jp-tokyo: $2.32/hr trough April 22 → $2.85/hr today, +22% over 30 days; 57-day continuous appreciation with spread compression from 66% to 5.2% (live ticker data, June 18, 2026)
2. H100 SPOT eu-stockholm: +60% over 30 days to $3.61/hr; spread widened from 32% to 82%, signaling second-provider repricing entry (live ticker data, June 18, 2026)
3. H100 RESERVED_1YR us-california: locked at $6.69/hr for 91 consecutive days; no repricing in full 90-day history window (ticker history, June 18, 2026)
4. H100 SPOT us-iowa: $1.24/hr vs. $8.60/hr reserved = 14.4% ratio — the widest inversion in the dataset (live ticker data, June 18, 2026)
5. H100 SPOT de-frankfurt: -26.9% over 30 days to $1.79/hr — EU core still deflating, confirming the APAC/EU split is geographic, not cyclical (live ticker data, June 18, 2026)
6. NatEx data unavailable for H100 (no matched-pair comparison possible); decomposition prices are unvalidated but internally consistent across regions
7. QUALIFYING: Montreal spot at $4.98/hr (+91% 30d) is a powerful APAC-analog signal in North America — but this is a single large recovery move that may reflect a specific customer, not broad market tightening

Implied Action:

• Watch signal: H100 SPOT us-california is the most actionable: at $4.93/hr and +26% over 30 days, it is approaching the $6.69/hr administered reserved floor. If California spot crosses $5.50/hr over the next 30 days, the reserved/spot inversion in the most liquid US market is closing — that is the signal that the US recovery has begun. Set a price alert.
• CRWV Q2 earnings (August 12) is the key catalyst: any geographic utilization breakdown showing APAC >85% with strengthening contract renewal rates will confirm the cycle is turning in the markets that matter most for near-term margin recovery. Forward P/E of -142x means this is a gross margin trajectory story, not a net income story. Watch gross margin guidance closely.
• Enterprise contracting: if you are buying H100 compute in US/EU for H2 2026, the 30-day spot premium over reserved is still so wide in Iowa and Oregon that short-duration spot remains economically superior. In Tokyo and Stockholm, the widening spreads suggest on-demand or short reserved contracts are worth paying up for before the market equilibrates fully.
• Short-side risk for US neocloud operators (CRWV, APLD) heavily weighted toward US inventory: their cost basis of $5–9/hr in reserved contracts against $1.24–$4.93/hr spot clearing continues to destroy per-GPU margins on underutilized inventory. The recovery thesis requires believing Oracle's 97.5% utilization is industry-wide — the US spot data says it is not.

H2 — AWS Inferentia Frankfurt: EU ASIC Scarcity Creates an Immediate Migration Arbitrage Window Confidence: 4 / 5

Thesis: The Frankfurt Inferentia SPOT price trajectory is one of the cleanest structural stories in current pricing data, and it carries a sharp, time-sensitive operational implication. From a January 16 peak of $0.31/hr, the price collapsed in near-perfect linear descent to a $0.064/hr floor on February 10 — AWS flooding EU with Inferentia v1 capacity. Then, without interruption, it reversed and has climbed for four straight months. The acceleration in June is the critical development: from $0.124/hr on May 20 to $0.449/hr today — a $0.32/hr move in 28 days, with daily increments of $0.05–0.08/hr in recent sessions. The observation count doubled from 1 to 2 on June 4, confirming a second Inferentia workload entered the Frankfurt market — this is utilization evidence, not a pricing artifact.

The geographic structure of the surge validates it as a genuine capacity constraint rather than a local anomaly: London +45.5% over 30 days (currently $0.065/hr, still cheap but accelerating), Hong Kong +137% over 7 days, and Ohio -53.4% plus Virginia -91.6% all trending simultaneously. AWS is not randomly varying Inferentia prices — it is executing a deliberate geographic rebalancing, pricing US east-coast inventory to liquidate (near-free at $0.003/hr in Virginia) while European and APAC markets exhaust their allocations. The parallel AWS Trainium2 ramp in US-east is the likely explanation: as Trainium2 displaces Inferentia v1 capacity domestically, excess US inventory is being cleared at near-zero cost, concentrating residual Inferentia v1 supply in markets where Trainium2 hasn't yet landed.

At today's prices, the arbitrage is narrowing but still actionable: Frankfurt Inferentia at $0.449/hr vs. L4 SPOT Frankfurt at approximately $0.26/hr. The GPU is cheaper today. At the current trajectory — roughly $0.05–0.08/hr per day of appreciation — Frankfurt Inferentia will cross A10 ON_DEMAND Frankfurt ($0.975/hr) within 2–3 weeks, and the AWS ASIC cost advantage in EU inference will have fully evaporated.

Key Evidence:

1. INFERENTIA SPOT de-frankfurt: $0.064/hr floor on Feb 10 → $0.449/hr today, +164% over 30 days; four months of unbroken daily appreciation (ticker history, June 18, 2026)
2. n_observations jumped from 1 to 2 on June 4 in Frankfurt market — independent utilization corroboration, not just repricing (ticker history, June 18, 2026)
3. INFERENTIA SPOT us-virginia: $0.038/hr in January → $0.003/hr today, -91.6% over 30 days (live ticker data, June 18, 2026)
4. INFERENTIA SPOT us-ohio: -53.4% over 30 days; us-oregon flat — structured US capacity liquidation pattern (live ticker data, June 18, 2026)
5. INFERENTIA SPOT gb-london: +45.5% over 30 days to $0.065/hr — Frankfurt analog following with a 3–4 week lag (live ticker data, June 18, 2026)
6. INFERENTIA SPOT hk-hongkong: +137% over 7 days — APAC Inferentia scarcity confirming the pattern is cross-regional, not EU-specific (live ticker data, June 18, 2026)
7. Inferentia2 marked inactive across all providers in catalog survival data — concentrates residual demand on Inferentia v1, reducing the safety valve that would otherwise cap v1 repricing
8. QUALIFYING: L4 SPOT Frankfurt currently ~$0.26/hr, already 58% cheaper than Inferentia in the same region — the GPU migration window is open today but narrowing (live ticker data, June 18, 2026)

Implied Action:

• Immediate operational call for EU AI workloads: route EU inference from Inferentia to L4 SPOT Frankfurt now. The $0.449/hr vs. $0.26/hr gap is actionable today; at the current Inferentia appreciation rate, this window closes within 2–3 weeks.
• Watch signal: INFERENTIA SPOT gb-london at $0.065/hr is the leading indicator for the next leg. London typically follows Frankfurt with a 3–4 week lag. If London crosses $0.20/hr, EU Inferentia scarcity is broadening and the migration window is fully closed.
• Strategic risk for AWS: if Trainium2/Inferentia2 EU capacity doesn't land before Frankfurt Inferentia crosses GPU price parity, enterprise EU customers building AI applications will default to NVIDIA GPUs — potentially cementing NVIDIA's position in a market AWS specifically designed ASICs to contest. Every week of delay is a competitive positioning loss in EU, the world's second-largest enterprise AI market.
• Monitor: AWS EU infrastructure announcements for Trainium2 availability — that is the event that would break the Frankfurt repricing trend and invalidate the thesis.

H3 — EM Inference Premium: Durable but Static, Not Self-Reinforcing Confidence: 3 / 5

Thesis: The A10 inference GPU in Hyderabad trades at $2.61/hr — a 257% premium over Oregon's $0.73/hr for identical silicon. This premium has persisted without meaningful variation for five months in the 90-day history: fewer than five price changes from January through mid-June, with the most recent a modest step-down from $2.69 to $2.56/hr on June 11. The mechanism matters: this is a single-provider (GCP) administered tariff, not a market-clearing price under competitive pressure. There is no price discovery happening in Hyderabad's A10 market — there is one provider, one price, and it moves when GCP decides to move it. The critical downgrade from the initial hypothesis is that this premium will not accelerate due to demand; it will persist at elevated levels for as long as competitive entry remains 12–18 months away. CPP Investments' $741M into India's CtrlS and Blackstone/AirTrunk's ~$30B India commitments are building that competitive capacity now — but data center HVAC and power lead times alone are 12–18 months, and Comfort Systems USA's backlog growth from $6.89B to $12.45B (+81% YoY) confirms the construction pipeline is just now entering execution. The Hyderabad premium is a durable rent, not a dynamic signal.

Where the EM thesis does hold is in the few markets with genuine price discovery: A10 in São Paulo +10.8% and Doha +10.3% over 30 days are multi-provider markets where spot dynamics are real. L4 SPOT in Seoul +19.9% and Paris +12.5% over 30 days represent genuine demand-driven repricing, not administered tariffs. The geographic split within inference GPUs — EU/NA core softening vs. EM/frontier appreciating — is real; it is the mechanism that is more nuanced than originally characterized.

Key Evidence:

1. A10 ON_DEMAND in-hyderabad: $2.61/hr today vs. $0.73/hr Oregon — 257% premium; <5 price changes in 90-day history (ticker history, June 18, 2026)
2. n_providers = 1 for virtually all A10 on-demand tickers globally — no competitive price discovery in any A10 market (live ticker data, June 18, 2026)
3. A10 ON_DEMAND in-hyderabad: step-down from $2.69 → $2.56/hr on June 11 — GCP-initiated tariff adjustment, not demand response (ticker history, June 18, 2026)
4. A10 ON_DEMAND br-saopaulo: +10.8% over 30 days; qa-doha: +10.3% — genuine EM appreciation in multi-provider competitive markets (live ticker data, June 18, 2026)
5. L4 SPOT kr-seoul: +19.9% / 30d; fr-paris: +12.5% / 30d; se-stockholm: +8.6% / 30d — inference demand expanding geographically (live ticker data, June 18, 2026)
6. Comfort Systems USA backlog: $6.89B → $12.45B (+81% YoY), Q1 revenue +57% YoY — confirms EM/frontier data center construction pipeline in multi-year execution (news feed, June 17, 2026)
7. CPP Investments $741M into India's CtrlS; Blackstone/AirTrunk India commitments — 12–18 month competitive entry horizon (news feed, June 17, 2026)
8. QUALIFYING: Administered tariffs create price stability, not price acceleration — the Hyderabad premium is durable but won't compound

Implied Action:

• Avoid: long positions premised on Hyderabad A10 prices rising further — the administered tariff mechanism means appreciation requires a GCP decision, not demand pressure. The premium is real but static.
• Long infrastructure contractors with EM exposure: FIX (Comfort Systems, backlog 4.3x quarterly revenue, UBS target $2,125) is the cleanest expression. The premium's persistence attracts capex and construction — that capex is already committed and executing. FIX has no GPU obsolescence risk.
• Long EM-positioned neoclouds when they IPO: the 257% India premium is a strong demand pull signal attracting new entrants. The first 2–3 providers to enter India A10/L4 at below-Hyderabad pricing will capture outsized market share. Monitor CtrlS and Airtel/Jio pricing announcements — competitive entry, not price appreciation, is the signal here.
• Watch: n_providers for A10 in-hyderabad — if it rises from 1 to 2, competitive entry has begun and the premium will start compressing. That is the signal to exit infrastructure-long, not to enter.

H4 — The On-Demand Premium Is Anchored Fair Value, Not an Inflating Premium Confidence: 5 / 5

Thesis: The T4 London data delivers a counterintuitive but high-conviction conclusion: the on-demand market for production inference compute is not inflating — it is institutionally locked. T4 ON_DEMAND in London has held at $0.3634/hr for 89 of the past 90 days across three providers (AWS, Azure, GCP), with the single exception being a June 17 data-collection artifact (3 vs. normal 223 observations, 2 vs. normal 3 providers). The natural experiment cross-check provides independent validation: natex_median for T4 London is $0.35/hr from 330 matched pairs across providers — within 4% of the $0.3634/hr administered rate. This is the most important finding: the on-demand price is not an artificial premium over market value; it is exactly what independent price decomposition says the compute is worth. Spot is the discount. What's changing is not the on-demand price — it's the mix of workloads between spot and on-demand. As Vultr's CEO stated explicitly in their June 17 GB300 NVL72 announcement, AI demand has shifted from "experimentation" to "production customer-facing services." Production services do not run on preemptible spot. They never did. The T4 on-demand floor is durable precisely because it serves a workload class — always-on inference serving — that has no spot-compatible substitute.

Oracle's $20–25B in FY2027 customer prepayments (up from $8B in FY2026) is the enterprise demand side of this equation: customers are locking in dedicated capacity at on-demand rates in advance, accepting Oracle's premium over spot because availability certainty is worth more than cost optimization when serving production AI applications. CoreWeave's entire business model is built on this dynamic — long-duration on-demand contracts with enterprise AI customers — and their +39% stock move over three months reflects the market pricing this shift as structural and durable, not cyclical.

Key Evidence:

1. T4 ON_DEMAND gb-london: $0.3634/hr for 89/90 days; 3 providers; zero price movement in 90-day history (ticker history, June 18, 2026)
2. natex_median for T4 London: $0.35/hr from 330 matched pairs across providers — within 4% of administered rate, confirming on-demand is fair value, not artificial premium (ticker data NatEx fields, June 18, 2026)
3. T4 SPOT gb-london: drifted from $0.113/hr to $0.097/hr trough in January, currently $0.117/hr — spot declining as on-demand flat, confirming workload migration direction (ticker history, June 18, 2026)
4. Oracle FY2027 customer prepayments guidance: $20–25B vs. $8B in FY2026; GPU utilization 97.5% — enterprise production workload commitment at scale (news feed, June 17, 2026)
5. Vultr/HPE partnership: GB300 NVL72 deployment specifically for "production customer-facing AI services," with HPE Spectrum-X and liquid cooling — hardware-level confirmation of on-demand workload class (Data Center Knowledge, June 17, 2026)
6. T4 (launched September 2018, 7.76 years old) still active across all three major clouds — longevity in on-demand catalog confirms it serves locked-in production workloads with no migration pressure (catalog survival data, June 18, 2026)
7. CRWV: +39% over 3 months, analyst target $140 vs. current $115; Q2 earnings August 12 — equity market pricing in the production-shift narrative (equities feed, June 18, 2026)
8. APLD: strong_buy, analyst target $73 vs. current $45.57; Q2 earnings July 29 — nearer-term catalyst for the same thesis (equities feed, June 18, 2026)

Implied Action:

• Long CRWV into the August 12 earnings catalyst. The thesis requires utilization guidance >85% with stable or rising average contract price per GPU-hr. Any indication of APAC geographic concentration (where H100 spot is recovering) would double-confirm the hypothesis. The -142x forward P/E is a growth-stage distraction — watch gross margin trajectory and contract book duration. Analyst target of $140 implies 22% upside from current $115.
• Long APLD into the nearer July 29 catalyst at lower conviction. Strong-buy consensus, $73 analyst target vs. $45.57 current = 60% upside. Earlier catalyst and more asymmetric setup.
• Long CIFR as the momentum expression: +105% over 3 months, strong-buy, recently +38% over one month. Higher volatility, but the production-shift narrative is already moving the stock. Size accordingly.
• Avoid: spot-centric GPU resellers without long-term contract books. Spot for T4/A10/L4 in core US/EU is softening as production workloads migrate to on-demand; these operators face margin erosion from both sides — lower spot clearing prices and shrinking spot workload pool.
• Key confirm signal: any hyperscaler announcing minimum contract durations or deprecation of spot GPU tiers would be a decisive confirmation that the production-shift is at hyperscaler scale, not just neocloud scale.

H5 — The Power Moat: Grid Constraints Are a Multi-Year Structural Advantage for Power-Secured Operators Confidence: 4 / 5

Thesis: The Northern Virginia grid interconnection timeline — up to 14 years for new large-load connections in some CAISO/PJM queues — is not a headline risk. It is a structural moat already being priced into construction backlogs and power-secured site valuations. The mechanism is straightforward: every new data center that cannot connect to the grid creates incremental demand for existing data centers that already have power. NERC's draft reliability guideline specifically naming data centers as "emerging large loads" requiring new frequency and voltage trip settings is working through a 12–24 month formal standards process — once binding, it will impose compliance costs on new builds while existing infrastructure is grandfathered. Fourteen US states are actively considering legislation to restrict new data center development, Florida has already signed SB 484 prohibiting cost pass-through to residential utility customers, and Virginia resident comfort with new data centers has dropped 34 percentage points since 2023 — in the state hosting the world's largest concentration of them. The regulatory and community pressure is real, multi-directional, and accelerating. The investable implication is not that data center growth stops — it is that the operators and contractors who have already secured power and permits hold a durable competitive advantage that is compounding with every month the regulatory environment tightens. Iowa's power advantage ($1.24/hr H100 spot, among the cheapest globally) reflects available capacity and cheap wind power — the moat is latent, not yet priced as scarcity. Texas (ERCOT) and Nevada similarly benefit from lower interconnection friction than Virginia or California.

The second-order play is even more compelling: Oklo's amended 10-Q (reflecting the Atomic Alchemy acquisition, February 2025) is an SEC-level commitment to advanced nuclear reactor development targeted at data centers. This is not a concept paper — it is a disclosed corporate acquisition with data center co-location as the stated application. Nuclear co-location is the 5-year solution to the 14-year grid queue problem: grid-independent power supply that sidesteps interconnection entirely. Marvell's optical-networking play ($2B NVIDIA strategic investment, 90% optical-networking sales growth projected) identifies the secondary constraint: once power is secured, network bandwidth becomes the next bottleneck in AI cluster architecture. NVIDIA doesn't make $2B strategic investments in interchangeable commodity suppliers.

Key Evidence:

1. Northern Virginia grid interconnection: up to 14-year queue for new large-load data center connections in CAISO/PJM queues (intel feed, power_grid, June 2026)
2. NERC draft reliability guideline: specifically names data centers as "emerging large loads" requiring new frequency/voltage trip settings; 12–24 months to binding (intel feed, federal_regulatory, June 2026)
3. 14 states considering data center legislation; Florida SB 484 signed prohibiting residential cost pass-through; Virginia resident comfort -34pp since 2023 (intel feed, federal_regulatory, June 2026)
4. Comfort Systems USA: backlog $6.89B → $12.45B (+81% YoY), Q1 revenue $2.87B (+57% YoY), prefab capacity +30%; UBS target raised to $2,125 (equities feed / news feed, June 2026)
5. H100 SPOT us-iowa at $1.24/hr — among the cheapest H100 prices globally; latent power moat reflected in available capacity (live ticker data, June 18, 2026)
6. Oklo 10-Q/A: Atomic Alchemy acquisition disclosed, advanced nuclear reactors for data center/AI infrastructure co-location (SEC feed, June 2026)
7. NVIDIA $2B strategic investment in Marvell; Marvell 90% optical-networking sales growth projection; analyst targets raised $105 → $150 (news feed, June 17, 2026)
8. Box Elder County, Utah: 40,000-acre data center approval consuming 2× the state's total energy — crystallizes community opposition dynamics; a regulatory risk accelerant (intel feed, June 2026)
9. QUALIFYING: Iowa's power moat is latent, not yet manifested as pricing power — $1.24/hr spot reflects oversupply, not scarcity premium. The moat requires demand to catch supply before it converts to pricing power

Implied Action:

• Long FIX (Comfort Systems USA): 4.3x quarterly revenue in backlog, multi-year execution queue, no GPU obsolescence risk, no AI model cycle risk. UBS target $2,125. Cleanest risk/reward in the infrastructure stack. Q2 earnings report is the confirm signal — watch backlog expansion continuation.
• Long MRVL (Marvell): NVIDIA's $2B strategic investment establishes a floor on customer concentration risk. The 90% optical-networking growth projection is now backstopped by NVDA's own balance sheet. Analyst target raised to $150. The network constraint thesis is structural — every additional GPU cluster is a Marvell switch sale.
• Long OKLO as a 5-year power option: the grid queue problem creates the market failure that makes nuclear co-location economically rational. The SEC-level corporate commitment (Atomic Alchemy acquisition) has been made. This is not a speculative thesis — it is a disclosed strategy from a public company backed by grid-queue math. Size as an option, not a core position.
• Monitor: building_permits and power_grid intel feeds for Mortenson Construction, Turner, and Hensel Phelps hyperscaler data center filings — these are the primary GCs and serve as 12-month leading indicators for capacity additions. Any slowdown in new filings validates the moat thesis.
• FERC risk: if FERC expedites large-load interconnection rules or creates a fast-track process for data center connections, the queue-constraint moat compresses rapidly. Watch for FERC notice-of-proposed-rulemaking on large-load interconnection — that is the primary thesis-denial event.

Risk Flags

Immediate (0–30 days)

R1 — Frankfurt Inferentia Price Trajectory: Window Closes in Days, Not Weeks At $0.449/hr today and appreciating $0.05–0.08/hr per day, Frankfurt Inferentia spot will cross A10 ON_DEMAND Frankfurt (~$0.975/hr) within approximately two to three weeks. EU workloads currently running on Inferentia for cost reasons will hit economic parity with GPU alternatives at that point, removing the AWS ASIC cost advantage in EU inference permanently until new Inferentia/Trainium2 capacity arrives. Any AWS EU customer still on Inferentia-optimized workloads faces a forced migration event on a short timeline. The risk is operational, not speculative — failure to migrate before parity results in overpaying for ASIC compute that was cheaper than GPU for the past four months.

R2 — T4 São Paulo Spread Event: Thin Markets Can Front-Run Real Supply Changes The $0.083/hr intraday T4 SPOT São Paulo spike with a 1,332% spread_pct (three wildly divergent providers) is technically a liquidity artifact, but thin-market events in EM regions sometimes precede structural supply changes by 2–4 weeks. If São Paulo spot doesn't normalize within 5–7 days, recheck whether a provider has withdrawn capacity — that would be a genuine signal of EM inference supply tightening consistent with the broader EM thesis.

R3 — UNMAPPED Pricing Type Classification Artifact Masking Real 7-Day Moves The near-total (-99.9997%) 7-day collapses across Dublin, Singapore, Ohio, Sydney, Seoul, Texas, and Virginia are confirmed data classification artifacts (previously-mapped pricing types re-bucketed to UNMAPPED). However, the reclassification itself may indicate a data pipeline change that will affect how certain spot tiers are reported going forward — any analysis using 7-day deltas for these regions should be treated with caution until the reclassification is confirmed stable.

Near-Term (30–90 days)