Daily Investment Brief — August 2, 2026

GPU & Cloud Compute Daily Brief — August 2, 2026

Cross-provider pricing intelligence | News, regulatory, and SEC synthesis

Market Pulse

The AI compute market has entered a structurally new phase: flagship GPU on-demand list prices are stable or declining (H100 Frankfurt ON-DEMAND has compressed from ~$4.40/hr in January to a pinned $1.76/hr median, while B200 ON-DEMAND is anchored at $12.82–14.60/hr with zero 30-day delta), but spot markets — the real-time utilization signal — are repricing hard and simultaneously across geographies, GPU tiers, and provider ASICs. The dominant narrative is a demand-outrunning-deployed-capacity squeeze, not a supply-chain event: the chip scarcity era of 2023–2025 has given way to a power-and-real-estate scarcity era where the bottleneck is watts in the ground, not wafers in the fab. B200 SPOT in US-Virginia has made 20 consecutive days of gains to reach $5.05/hr — up 111% from its May trough — while its spot discount has compressed from 82% to 63% in six weeks, the tightest reading since Blackwell launched. Simultaneously, AWS ASIC spot pools are bifurcating along a transatlantic axis: Inferentia SPOT has collapsed -98% in Montreal and surged +135% in Stockholm in the same week, a geographic fingerprint consistent with deliberate capacity pool reallocation rather than demand variation. The infrastructure layer is also moving: SSD Provisioned IOPS surged +1,146% in Warsaw and +215% in Doha in 24 hours, a consistent leading indicator for GPU on-demand repricing 2–4 weeks out.

Key Movers

Thin-market noise to disregard: A100 80GB SPOT Tel Aviv (+621% / 30d) — n=1 provider, very low absolute price ($1.13/hr), small market snapping to new clearing price. Statistically valid but not investable signal. RTX PRO 6000 SPOT dual-region launches (Stockholm/California, +95% / 7d) — new product launch artifact, not a demand signal.

Investable Insights

H1 — B200 Structural Repricing: The Spot Market Is Making New Highs and the Discount Is Still Compressing Confidence: 5 / 5

Thesis: The B200 Blackwell GPU spot market in US-Virginia has executed a textbook demand-absorption cycle and is now in an unambiguous second-phase squeeze. After initial Blackwell supply hit the market in late April and drove the spot price down to a $2.39/hr trough by July 3 (with the spot discount widening to 82% as providers competed for placement), demand absorbed that supply and reversed the trend with remarkable consistency: every single trading day from July 11 to July 31 — 20 consecutive sessions — saw the Virginia spot price rise, from $3.43/hr to $5.05/hr. The spot discount compressed monotonically from 71% to 63% over the same window. At 63%, this is the tightest the B200 spot market has been since Blackwell launched — and the current pricing trajectory has no reversal signal. The geographic breadth confirms this is structural, not regional: B200 SPOT is simultaneously +41.6% in Oregon (30d), +20.7% in Ohio (30d), and has emerged in Mumbai at $6.14/hr — the highest absolute B200 SPOT print in the dataset. The mechanism is straightforward: hyperscalers are absorbing Blackwell capacity into committed workloads (training runs, inference serving at scale) faster than NVIDIA can supply new instances to spot pools. AMD's Helios/MI455X entering full production (AOL/Reuters, Jul 29) paradoxically tightens near-term B200 demand — buyers who need Blackwell capacity today cannot wait for MI450 deliveries starting H1 2027. This pricing data is the most direct leading indicator available for NVIDIA's data center GPU average selling prices, and it is trending at multi-month highs heading into August 26 earnings.

Key Evidence:

1. B200 SPOT Virginia: $5.05/hr on July 31 vs. $2.39/hr trough on July 3 — +111% in 28 days (ticker history, Aug 2, 2026)
2. Spot discount compressed from 82.1% (July 3) to 63.0% (July 31) — tightest reading across full 90-day history (ticker history, Aug 2, 2026)
3. 20 consecutive daily price increases with zero reversal days since July 11 (ticker history, Aug 2, 2026)
4. B200 SPOT Mumbai at $6.14/hr, +72.2% over 30 days — confirms global not regional squeeze (live ticker, Aug 2, 2026)
5. B200 ON-DEMAND anchored at $12.82–$14.60/hr across US regions with zero 30-day delta — spot is the only price discovery venue (live tickers, Aug 2, 2026)
6. Qualifying: Phase 1 (Jan–Apr 16) showed two-provider competition with 65.6% discount — the current single-provider dominance makes the market more susceptible to mean-reversion if a second provider re-enters (ticker history)
7. NVDA at $200.75, 15.6x forward PE, strong_buy consensus, earnings August 26 (equities feed, Aug 2, 2026)

Implied Action: Long NVDA into August 26 earnings. The B200 spot pricing trajectory is a real-time leading indicator for data center GPU ASP. At 15.6x forward PE — among the lowest in NVDA's modern history — the market is not pricing in the Blackwell spot premium compression signal. Pair with a short CRWV (CoreWeave, currently -35.7% over 3 months, deeply negative FCF, trades at an extreme negative forward PE): B200 spot cost increases are NVDA's revenue signal but CRWV's input cost signal — if B200 SPOT reaches 55% discount or below before CRWV's next earnings, margin compression becomes acute. Monitor: Watch the B200 SPOT Virginia discount daily; a break below 60% is the trigger for conviction add. A reversal >15% over 7 days (or a second provider re-entering the market, which would expand spot supply) is the stop-out signal.

H2 — AWS ASIC Spot Pool Withdrawal: Monopoly Repricing, Not Demand Surge Confidence: 4 / 5

Thesis: The simultaneous cross-regional explosion in Trainium SPOT prices (+459% São Paulo, +353% Melbourne, +338% Oregon over 30 days) and the geographic near-collapse of Inferentia SPOT (Montreal -98%, Virginia -43%, Oregon -39% over 30 days) are not organic demand signals — they are the fingerprint of a deliberate AWS capacity pool management decision. The São Paulo Trainium history makes this case forensically: for six consecutive months (January through July 23), the spot price was administered in a tight band of $0.61–$1.06/hr, pinned near $0.658/hr for the final two months with sub-penny daily variation. That is not market clearing; it is a managed price. On July 24, a single-day +33% snap to $0.871/hr was followed by a near-vertical move to $3.74/hr over 9 days — a 470% increase from the administered floor. Every single Trainium SPOT ticker has exactly n=1 provider (AWS) and zero spread. There is no competitive tension; AWS is the sole price setter. The Inferentia data provides the reallocation tell: in the same week, five EU/UK markets surged (Stockholm +135%, Frankfurt +80%, London +63%, Dublin +43%, Paris +15% over 7 days) while four North American markets fell (-98% Montreal, -55% Ohio, -43% Virginia, -39% Oregon). The same product, five up in Europe, four down in North America, simultaneously — this is a pool rebalancing decision made at the AWS infrastructure layer, not a geographic demand differential. The economic logic is clear: AWS earns higher margins on reserved and committed-use ASIC capacity (Bedrock, SageMaker) than on spot, and Amazon's Q2 10-Q confirms it has simultaneously invested in both OpenAI and Anthropic — meaning AWS needs its own silicon cost-competitive to avoid dependence on third-party GPU supply chains.

Key Evidence:

1. Trainium SPOT São Paulo: 6-month administered price at ~$0.658/hr snapped to $3.74/hr in 9 days starting July 24 — +470% from managed floor (ticker history, Aug 2, 2026)
2. Trainium SPOT Melbourne at $6.28/hr, +353% over 30d; Oregon at $0.18/hr, +338% over 30d — cross-regional simultaneity (live tickers, Aug 2, 2026)
3. Every Trainium/Inferentia SPOT ticker: n_providers = 1, spread = 0% — zero competitive tension, pure monopoly pricing (live tickers, Aug 2, 2026)
4. Inferentia SPOT: 5/5 EU markets positive (7d), 4/4 NA markets negative (7d) — perfect geographic split with zero exceptions (live tickers, Aug 2, 2026)
5. Inferentia Montreal: -98% over both 7d and 30d, reaching $0.001/hr — capacity pulled, not consumed (live ticker, Aug 2, 2026)
6. Amazon Q2 10-Q confirms investments in both OpenAI and Anthropic plus heavy PP&E/lease capex for data center infrastructure (SEC feed, filed Jul 31, 2026)
7. Qualifying: Ohio Inferentia SPOT shows +372% over 30d but -55% over 7d — the surge-then-partial-reversal pattern indicates AWS is actively managing these pools, which means the directional repricing can be partially reversed at will

Implied Action: Long AMZN on the AWS AI Services margin thesis. If Trainium/Inferentia capacity is successfully migrated from low-margin spot to higher-margin reserved/on-demand Bedrock and SageMaker workloads, this is a Q3 2026 revenue quality upgrade. AWS carries >60% of Amazon's operating income — even a 200–300bps AWS segment margin expansion from ASIC monetization optimization materially moves consolidated earnings. Short APLD (Applied Digital, $27.39, -22.9% over 1 month) and similar small GPU hosting neoclouds as an asymmetric hedge: if AWS prices its Trainium capacity 3-4x higher while maintaining software-layer stickiness through Bedrock, third-party GPU hosting operators lose their cost arbitrage and face direct competition from AWS's own higher-margin ASIC offering. Monitor: Watch for Trainium/Inferentia ON-DEMAND price increases mirroring the spot repricing — that would signal demand successfully migrated to higher-margin products. Watch for n_providers > 1 appearing on any Trainium/Inferentia spot ticker, which would break the monopoly pricing thesis.

H3 — Mid-Tier Inference Mispricing: L4/L40S Spot Is Screaming What On-Demand Lists Won't Say Confidence: 4 / 5

Thesis: The L4 SPOT market is delivering an inference demand signal that GCP's on-demand list prices — frozen at 0% 30-day delta across all 15 regions — are not yet reflecting. In Ohio, the L4 SPOT price has risen for 22 consecutive days, from a post-provider-exit floor of $0.136/hr on April 16 to $0.457/hr on July 31, a 236% increase over 3.5 months. Critically, this is happening while the L4 ON-DEMAND list price in Ohio sits unchanged — meaning the gap between real-time demand pricing and published list price is widening by the day. The spot discount has compressed from 88% in mid-April to 74% today, and the trajectory has not flattened. The same pattern appears in Singapore (+277% over 30d) and Iowa (+183% over 30d). The mechanism is Jevons' paradox operating at inference scale: more efficient models are driving volume expansion that outpaces efficiency gains, consuming more L4 and L40S capacity per dollar of revenue for model API providers. The L4 is GCP's primary cost-optimized inference workhorse — the chip behind the vast majority of cost-sensitive LLM serving, image generation, and embedding workloads. GCP holds a monopoly on L4 spot supply (n=1 provider in all surging markets), which means this tightening cannot be competed away by AWS or Azure. The on-demand list price freeze is a temporal lag, not a denial — GCP list prices update infrequently, and spot markets are running 2–6 weeks ahead of list price discovery, as they always do in constrained markets. L40S in Mumbai at $1.33/hr SPOT (+110% over 30d) extends the signal to NVIDIA's next-tier inference chip, confirming it is not L4-specific.

Key Evidence:

1. L4 SPOT Ohio: $0.457/hr on July 31 vs. $0.136/hr on April 16 — +236% over 3.5 months; 22 consecutive days of increases since July 11 (ticker history, Aug 2, 2026)
2. L4 ON-DEMAND: 0.0% 30-day delta across all 15 regions (live tickers, Aug 2, 2026) — spot/on-demand divergence is widening, not narrowing
3. Spot discount compression: 88% (Apr 16) → 74% (Jul 31) — 14 percentage points in 3.5 months (ticker history, Aug 2, 2026)
4. L4 SPOT Singapore +277% over 30d; Iowa +183% over 30d — multi-region corroboration (live tickers, Aug 2, 2026)
5. L40S SPOT Mumbai: $1.33/hr, +110% over 30d — mid-tier inference scarcity confirmed across both GPU tiers (live ticker, Aug 2, 2026)
6. L4 n_providers = 1 in all surging markets — GCP monopoly on spot supply; cannot be competed away (live tickers, Aug 2, 2026)
7. April 16 as a shared inflection date across L4 Ohio, H100 Iowa, and H100 Frankfurt suggests a coordinated single-provider exit from multiple markets — the same secondary provider (likely Oracle or an AWS reseller) may have withdrawn simultaneously from all three (cross-ticker analysis, Aug 2, 2026)
8. Qualifying: L4 SPOT prices in absolute terms remain very low ($0.457/hr vs. ~$1.75/hr on-demand Ohio list) — the compression is meaningful percentagewise but the absolute dollar move is modest; a list price reset by GCP would normalize the spread quickly

Implied Action: Long GOOGL/Alphabet on the GCP inference pricing power thesis. As L4 utilization reaches physical limits with a frozen list price, GCP faces a binary choice: raise list prices (capturing the spread as revenue), or accept lower-margin spot-clearing at elevated prices. Either outcome benefits GOOGL. The AI efficiency gains narrative (SK Hynix, Seoul Economic Daily, Jul 29) — Jevons' paradox in action — is directly confirmed by mid-tier spot surges: this is a bullish volume signal for both GCP and NVDA (as L4 replacement cycles accelerate toward L40S and L4 successors). Avoid long positions in mid-market model API providers with fixed-price contracts to end users — their inference cost basis is rising faster than they can reprice. Monitor: Watch for the first L4 ON-DEMAND price increase in any region; that is the confirmation event that GCP is catching up to spot pricing reality and would validate the mispricing thesis fully. Also watch L4|SPOT|asia-southeast1 (Singapore) — if it holds elevated for 14+ more days, the demand signal is structural.

H4 — AMD Infrastructure Optionality vs. H100 Depreciation: Long-Term Confirmed, Near-Term Binary Confidence: 3 / 5

Thesis: AMD's announcement of a $14B/15-year Core Scientific deal (530 MW across 5 campuses, exclusive rights to reserve up to 2 GW additional) combined with an Anthropic multiyear commitment for up to 2 GW of MI450 chips starting H1 2027 (Indiatimes, Jul 29; AOL/Reuters, Jul 29) marks a structural inflection: AMD is no longer a chip company competing with NVIDIA on silicon specs alone; it is vertically integrating chips → servers → power → real estate in a single strategic move. This optionality is not priced in — AMD at $476.15 (-11.97% over 1 month, -8.77% over 5 days) is being sold ahead of August 4 earnings despite arguably the most strategically significant announcements in the company's history. But the H100 data in Iowa tells the concurrent story of what AMD is threatening: a secondary provider entered the H100 Iowa spot market on April 16 at ~$0.05/hr against an incumbent at $2.45/hr, then progressively lowered its price to $0.014/hr by July — offering H100 capacity at 99% below the institutional clearing price. This is a provider that accumulated H100 inventory it cannot place at market rates. By July 31, the spread between the two H100 Iowa spot providers has reached 8,600%+. That is the H100 depreciation curve playing out in real-time market data, not a theoretical model. Enterprises writing 3-year H100 reserved contracts today are implicitly wagering that this secondary market pricing ($0.014/hr) remains an outlier through 2029 — a bet that AMD MI450 mass deployment and GB300/Vera Rubin availability will not normalize H100 to near-zero spot value by maturity.

Key Evidence:

1. AMD Core Scientific deal: $14B, 530 MW, 15-year, 5 campuses, exclusive 2GW reservation rights (Indiatimes, Jul 29) — $19B+ committed revenue pipeline including Anthropic deal
2. AMD Helios/MI455X confirmed in full production; OpenAI "massive scale" deployment ongoing; Anthropic MI450 commitment for H1 2027 (AOL/Reuters, Jul 29)
3. H100 SPOT Iowa: 2-provider spread of 8,600%+ ($2.45/hr vs. $0.014/hr) persisting 3.5 months since April 16 (ticker history, Aug 2, 2026) — secondary provider progressively lowering from $0.05 to $0.014
4. H100 ON-DEMAND Frankfurt: $1.76/hr median (pinned since April) vs. $4.40/hr in January — 60% compression already executed in EU's most liquid H100 market (ticker history, Aug 2, 2026)
5. H100 ON-DEMAND Seoul: $15.14/hr vs. Frankfurt: $2.98/hr — 5x geographic spread signals regional supply/demand asymmetries not explained by chip scarcity alone (live tickers, Aug 2, 2026)
6. AMD at $476.15, -11.97% over 1 month, earnings August 4 — near-term binary event (equities feed, Aug 2, 2026)
7. Qualifying: H100 is still active on AWS, Azure, and GCP per depreciation data; MI450 delivery starts H1 2027 — AMD infrastructure thesis is 12–24 months from inflecting spot pricing; H100 on-demand prices may hold firm through that period

Implied Action: Avoid new H100 3-year reserved capacity commitments. The Iowa bifurcation data is the most direct evidence that H100 will have a near-zero marginal clearing price by 2028–2029 as AMD MI450/Helios and GB300/Vera Rubin mature — a 3-year reserved commitment written today matures into that environment. For equity, this hypothesis resolves on AMD earnings August 4: if Helios delivery timelines and Anthropic commitments are confirmed without delay, AMD re-rates to Tier 2 alongside GOOGL on the infrastructure optionality thesis. A delivery slip of any kind drops the hypothesis to 2/5. Long AMD on weakness post-earnings if the Core Scientific/Anthropic deal execution is confirmed. The current sell-off (-11.97% in 1 month) on the back of structurally positive strategic news is a classic "sell the news before earnings" dynamic that tends to reverse on confirmation.

H5 — US-East Power Constraint Exporting to Europe: A 3-Year Structural Price Floor Confidence: 4 / 5

Thesis: The US-East-Virginia compute market is physically constrained for 3–4 years — not by chip supply, but by power infrastructure. Dominion Energy's interconnection queue is processing multiple simultaneous large-load applications (Culpeper Tech Zone expansion, Carmel Church/Ruther Glen substation allocations, the 5-volume GoldenMars transmission project filed March 2025 and still in process). Amazon is now lobbying Virginia regulators to fund its own transmission infrastructure via CIAC rather than wait in the standard queue (Data Center Knowledge, Jul 29) — an admission that the normal regulatory pathway is too slow even for the world's largest cloud provider. At the federal level, FERC's AD24-11 docket on large-load integration is actively contested, with the Data Center Coalition filing as recently as February 2026, and FERC issuing new show-cause orders in June 2026 on co-located loads — regulatory friction that further extends timelines. The pricing implication is a structural floor on US-East compute costs for at least 36 months. But the more actionable near-term signal is the second-order effect that is already visible in the data: compute capacity is migrating to Europe. In the same week, AWS Inferentia SPOT collapsed in Montreal (-98%), Virginia (-43%), and Oregon (-39%) while surging in Stockholm (+135%), Frankfurt (+80%), and London (+63%). GCP's TPU v5LitePod mirrored this: -89% in Montreal, +76% in Stockholm, simultaneously. This is not geographic coincidence — it is the physical manifestation of hyperscalers reorienting their expandable capacity toward EU markets where power is available. The Arizton report (Yahoo Finance Singapore, Jul 29) forecasting 18% CAGR to $140B for the European hyperscale market through 2031 is the demand-side confirmation of the supply-side reorientation visible in the ASIC spot data.

Key Evidence:

1. Dominion Energy queue: Multiple simultaneous large-load filings (GoldenMars 5-volume, Culpeper Tech Zone, Carmel Church/Ruther Glen) all processing simultaneously (intel feed, multiple dates 2025-2026)
2. Amazon lobbying for CIAC model in Virginia — willing to pay to jump the interconnection queue (Data Center Knowledge, Jul 29)
3. FERC AD24-11 actively contested as of February 2026; new show-cause orders June 2026 on co-located loads (intel feed, Jun 2026)
4. EU Inferentia SPOT: 5/5 markets positive (Stockholm +135%, Frankfurt +80%, London +63%, Dublin +43%, Paris +15% / 7d); NA Inferentia: 4/4 markets negative (Montreal -98%, Virginia -43%, Oregon -39% / 7d) — perfect geographic split (live tickers, Aug 2, 2026)
5. TPU v5LitePod: Montreal ON-DEMAND -89% / 7d, Stockholm +76% / 7d — GCP corroborating the AWS reallocation signal independently (live tickers, Aug 2, 2026)
6. H100 EU premium: Stockholm ON-DEMAND DEV $12.25/hr, Dublin $11.05/hr, Madrid $13.07/hr vs. Singapore $3.16/hr, KL $3.02/hr — 3-4x EU/SEA premium stable for 90+ days (live tickers, Aug 2, 2026)
7. Digital Realty Trust (DLR): $188.52, analyst target $221.68 (+17.6% upside), 10-Q filed July 31, next earnings October 22 (equities feed, Aug 2, 2026)
8. NextEra/Brookfield Paducah 2GW campus (AJOT, Jul 29): 2032 completion — confirms meaningful US capacity relief is 6 years away, not 2-3
9. Qualifying: EU energy regulation risk is real — potential data center power consumption caps in Germany, Netherlands, Ireland could constrain EU expansion; FERC CIAC approval for Amazon would partially reduce US-East queue friction sooner than expected

Implied Action: Long Digital Realty Trust (DLR) as the pure-play beneficiary of structural EU data center scarcity. DLR's pan-Atlantic footprint directly monetizes the supply/demand imbalance now confirmed by the ASIC spot reallocation data; the company files its Q2 10-Q July 31 (this week) and reports Q3 earnings October 22, providing a long runway to accumulate ahead of what should be a bullish European colocation demand quarter. The 17.6% gap to analyst target ($188.52 current vs. $221.68 target) is achievable if the EU capacity migration thesis plays out over 6–18 months. Supplementary long: Equinix (EQIX) as the second-largest pan-Atlantic data center operator with dense EU/Nordic presence — if DLR benefits from EU scarcity, EQIX is a closely correlated beneficiary. Watch SSD Provisioned IOPS in EU regions (Warsaw +1,146% in 24h) as a leading infrastructure-layer indicator — sustained elevation in EU storage IOPS typically precedes GPU on-demand repricing by 2–4 weeks. The next confirmation event is H100 SPOT or ON-DEMAND price increases in Frankfurt or Stockholm over the next 30 days, which would validate that the EU capacity migration is driving pricing power for operators, not just spot pool management.

Risk Flags

Immediate (0–30 Days)

R1 — B200 Single-Provider Concentration Risk The entire B200 SPOT repricing thesis (H1) rests on a one-provider market in Virginia since April 16. If a second provider re-enters the B200 SPOT pool — which requires only a commercial decision, not new infrastructure — it would expand spot supply and potentially unwind the 20-day consecutive price rise within days. Monitor n_providers on B200|SPOT|US_EAST_VIRGINIA daily. The Phase 1 two-provider dynamic (Jan–Apr 16) produced a 65.6% discount and oscillating prices; re-entry would likely push the discount back toward 75%+, which would undercut the NVDA ASP thesis.

R2 — Oregon Multi-Directional 24h Volatility Signals Provider Repricing Instability The US-Oregon market showed H100 ON-DEMAND -29%, L4 ON-DEMAND -49%, T4 ON-DEMAND +43%, and L4 SPOT -47% all within a single 24-hour window, while n_providers = 2–3 across these tickers. This is not organic demand volatility — it is a provider-level list price reset (likely GCP or Oracle updating tariffs), which can temporarily break pricing relationships across tickers. Any model or trade calibrated to Oregon prices in the last 24 hours should be treated with caution until the next 24–48 hours of data confirm the new stable level.

R3 — AWS ASIC Pool Management Is Reversible The Trainium/Inferentia spot pool strategy (H2) is a deliberate operational decision, not structural scarcity. AWS reversed Ohio Inferentia from +372% (30d) back down -55% (7d) in a single management cycle. The same capacity that was pulled from Montreal can be re-listed at any time. The thesis is valid as a directional signal for AWS margin strategy, but the spot price levels themselves should not be relied upon as stable inputs to compute cost models.

Near-Term (30–90 Days)

R4 — AMD August 4 Earnings Binary The AMD infrastructure optionality thesis (H4) has a near-term binary resolution: if AMD reports any delivery delays on Helios/MI455X, any pushback from Core Scientific on the $14B deal structure, or any softness in MI300 data center revenue, the stock — already down -11.97% in a month — faces further weakness that would delay the institutional re-rating. The Anthropic 2GW MI450 commitment (AOL/Reuters, Jul 29) is contingent on H1 2027 delivery starting on schedule; any supply-chain friction with TSMC on the advanced packaging for MI450 would cascade to AMD's cloud-native pricing ambitions. Watch AMD earnings August 4 as a 3/5 → 4/5 or 3/5 → 2/5 confidence swing event.

R5 — GCP L4 ON-DEMAND List Price Reset Would Collapse the Mispricing Spread The L4 inference mispricing thesis (H3) requires a sustained divergence between spot (rising) and on-demand (frozen). GCP updates list prices infrequently but does update them — a retroactive list price adjustment that catches up to the spot level would instantly eliminate the spread. This is not a denial of the underlying demand signal, but it would remove the actionable mispricing. Watch the L4 ON-DEMAND Ohio or Singapore prices for any non-zero daily delta; the first list price move resets the spread arithmetic.