Daily Investment Brief — October 5, 2026

Market Pulse

The GPU rental market is experiencing a structural pricing fragmentation unlike anything in the post-H100 era: three non-communicating price layers — spot, on-demand, and reserved — are now pricing the same physical chip at radically different rates with no arbitrage mechanism connecting them. H100 SPOT clears at $0.19/hr in Amsterdam and $0.58/hr in Frankfurt while ON_DEMAND list remains frozen at $7.98/hr globally and no provider has published a single pre-announced price change. The reserved tier is simultaneously inverting: Frankfurt RESERVED_1YR at $4.72/hr prices 8× above spot but well below ON_DEMAND, while Jakarta RESERVED_1YR fell –30% over 30 days to $5.42/hr and Sydney RESERVED_1YR sits above ON_DEMAND at $8.35/hr — a rational pricing equilibrium between layers has collapsed entirely. The dominant narrative framing for today is sub-market decoupling: the elite frontier-training layer (100,000-GPU committed clusters, ASIC-custom silicon, multi-year capex) and the commodity spot layer (H100s clearing at 50–92% discount to list) have fully separated into distinct markets that no longer share a pricing signal, and the press is covering only the former while the data screams from the latter.

Key Movers

† The L4 Oregon –50.3% 24h move is a decomposition artifact (baseline churn); the –17.9% 30d trend is real and directional. The V100 and K80 moves are thin-market noise with no investable signal.

Investable Insights

H1 — The H100 Is Already Priced as a Depreciating Asset; List Price Is the Last Thing Holding Confidence: 5 / 5

Thesis: Every pricing layer below ON_DEMAND list has capitulated. H100 SPOT has collapsed to clearing levels between $0.19/hr (Amsterdam, single neocloud) and $4.45/hr (London, multi-provider) while the $7.98/hr global list price stands frozen — not because there is sufficient demand at that price, but because no provider has yet pulled the trigger on a public cut. The mechanism is visible in the reserved tier: Jakarta RESERVED_1YR fell –30% in 30 days to $5.42/hr (from $7.70/hr); Frankfurt RESERVED_1YR at $4.72/hr has detached from ON_DEMAND entirely; Sydney RESERVED_1YR at $8.35/hr now exceeds ON_DEMAND at $7.53/hr — a term-structure inversion that signals providers locked in commitments at above-market rates months ago and are now unable to clear new reservations at any rational premium. This is not a temporary liquidity event. The 60-day AWS H100 SPOT minimum traced a documented arc: $2.63/hr on August 6 → $0.74/hr on September 17 (72% collapse in 42 days, with n_observations stable at 1,600-1,700/day — deep-liquidity price discovery, not thin-market noise) → partial recovery to $1.72/hr today as B200 spot tightened slightly. The B200 mechanics are the proximate cause: p6-b200.48xlarge is live on AWS in Virginia, Ohio, and Oregon at $113.93/hr per 8-GPU node ($14.24/GPU); as operators migrate committed workloads to Blackwell, H100 slots are released into spot. B200 is live on AWS but not yet on Azure or GCP — which is precisely why AWS H100 markets show more aggressive SPOT compression than Azure markets in comparable regions.

Key Evidence:

1. H100 SPOT prices across 20 active global regions range $0.19–$9.50/hr; 19 of 20 markets price at a discount of 50–98% to the $7.98 ON_DEMAND list. The outlier (São Paulo at $9.50/hr, +0.5% 30d) reflects single-provider geographic monopoly, not genuine demand (live ticker, Oct 5, 2026).
2. AWS H100 SPOT minimum fell from $2.63/hr (Aug 6) to $0.74/hr (Sep 17) — 72% collapse — with n_instance_types stable at 3 and n_observations at 1,600-1,700+ per day, confirming genuine market clearing rather than catalog churn (60-day price history, query_instance_price_history, Oct 5, 2026).
3. Frankfurt RESERVED_1YR ($4.72/hr) is 8.2× the current SPOT price ($0.576/hr) but 2.7× below ON_DEMAND — a three-layer inversion without historical precedent in the dataset (live RESERVED_1YR ticker, Oct 5, 2026).
4. Jakarta RESERVED_1YR fell –30.0% over 30 days (from $7.70 to $5.42/hr) while ON_DEMAND remained flat — providers cutting reserved rates faster than list, consistent with pricing in long-duration excess capacity (live ticker, Oct 5, 2026).
5. AWS p6-b200.48xlarge confirmed live in three US regions at $113.93/hr/node; B200 is catalogued as active on AWS but inactive on Azure and GCP — the asymmetry explains the geographic discount dispersion across H100 SPOT markets (depreciation analysis + underlying instance data, Oct 5, 2026).
6. Qualifying: The UPCOMING_PRICE_CHANGE alert queue is empty — no provider has published a forward-dated ON_DEMAND list cut. The cut will arrive unannounced when it comes, and the reserved-rate decline is the typical 60–90 day leading indicator.

Implied Action: For buyers: negotiate H100 RESERVED_1YR contracts now — the negotiating leverage has fully shifted to buyers and will not improve meaningfully before a list-price cut (which could reset the floor). Jakarta's –30% reserved decline in 30 days is the template for how quickly the conversation shifts. Watch H100|RESERVED_1YR|US_EAST_VIRGINIA for the first US reserved cut — that is the institutional trigger for enterprise repricing conversations. For neocloud equity holders (CoreWeave, Lambda, Hyperstack): the H100 resale margin is compressing simultaneously from the wholesale side (B200 availability at hyperscalers releases H100 supply) and the resale side (spot clearing below $2/hr in key markets). The peak H100 margin window is closing. For institutional GPU cluster owners: any A100 or H100 cluster contracted at 2024 reserved rates ($3–5/hr) represents a mark-to-market impairment that has not been formally recognized — the spot market implies those assets are worth 40–70 cents on the contract dollar.

H2 — FPGA Collapse Is the Leading Indicator for A100 Catalog Exit — Timeline Has Compressed Confidence: 4 / 5

Thesis: FPGA cloud rental pricing is in structural freefall — not cyclical softening, but architectural displacement. VIRTEX_VU47P SPOT is down –47.7% in Frankfurt, –39.3% in Virginia, and –36.7% in Oregon over 30 days, with all three regions continuing to decline in the 24-hour delta (–5.9%, –9.5%, –4.2% respectively). The only exception — London, +4.6% over 30 days — is a single-provider catalog hold at zero spread, not a demand signal. The mechanism is permanent workload migration: inference pipelines that ran on FPGAs for attention acceleration and custom routing have been rewritten for dedicated ASICs (OpenAI Jalapeño, confirmed June 2026 launch; Google TPU v5; AWS Inferentia2). Once a workload is ported to ASIC, the FPGA demand for that workload does not return. The Jalapeño timing — June 2026 launch — is precisely concurrent with the FPGA Q2 2026 acceleration of decline. The critical inference: A100 plays a structurally analogous role to FPGA in the current market — a "good enough" workhorse being displaced by more efficient purpose-built alternatives (L40S, L4, inference-optimized Blackwell). The lifecycle data confirms A100 is 5.88 years post-launch, the same window at which A10 (5.48 years) is already inactive on AWS and GCP, and A10G (4.85 years) is already gone from Azure and GCP. The A100 is entering the catalog-exit pressure window now. Crucially, Azure A100_80GB SPOT in Virginia is already printing ~$0.74/GPU/hr — the sub-$1/GPU threshold that, in the FPGA lifecycle, preceded catalog removal by 2–3 quarters.

Key Evidence:

1. VIRTEX_VU47P SPOT decline is synchronized across Frankfurt (–47.7%), Virginia (–39.3%), and Oregon (–36.7%) over 30 days — three independent regions, same direction, same magnitude. This eliminates regional catalog explanations and confirms market-wide architectural repricing (live tickers, Oct 5, 2026).
2. FPGA spot discounts now stand at 62–89% across all active regions, with Frankfurt at 72.4% discount and Virginia at 68.4%. These are providers attempting to liquidate FPGA capacity at any clearing price (live tickers, Oct 5, 2026).
3. OpenAI Jalapeño ASIC confirmed launched June 2026 — 700W, 2,048 inference pods, 27.5TB HBM4 per 128-chip rack. This is precisely the purpose-built inference hardware that replaces FPGA workloads (GPU news feed, Oct 4, 2026). The 4-month temporal coincidence between Jalapeño launch and FPGA price acceleration is the best-timed causal narrative in the dataset.
4. A100_80GB is active on all three hyperscalers at 5.88 years post-launch. A10 (5.48 years) is inactive on AWS and GCP; A10G (4.85 years) is inactive on Azure and GCP; A30 (5.48 years) is inactive everywhere. The lifecycle pattern is deterministic — A100 is next (depreciation analysis, Oct 5, 2026).
5. Azure A100_80GB SPOT in Virginia has already crossed the sub-$1/GPU/hr threshold at ~$0.74/hr (underlying instance data, Oct 5, 2026). For context: this is the floor level that preceded FPGA catalog removal by 2–3 quarters in prior cycles.
6. Frankfurt A100_80GB SPOT fell –18.9% in 7 days and –18.7% in 30 days — the most rapid A100 SPOT decline in the dataset, consistent with multi-provider competition accelerating the clearing process (live ticker, Oct 5, 2026).
7. Qualifying: A100's training utility (a workload that inference ASICs cannot replicate at scale) provides a genuine durability argument that FPGA and A10/A10G lacked. The A100 could survive in the training niche longer than pure lifecycle analysis suggests. The FPGA analogy is directionally correct but may be off by 2–4 quarters on timing.

Implied Action: For operators with A100 reserved contracts entered in 2024-2025 at $3–4/hr: the economic NPV of those contracts is already negative against current spot clearing rates. Use the next contract negotiation window to seek early termination provisions or rate amendments — the hyperscaler counterparty has incentive to clear A100 capacity ahead of formal catalog retirement. For investors: any managed-GPU-service business billing clients at A100 rates while paying hyperscaler reserved prices faces a compressing margin as customers observe the spot market and demand competitive rates. Short the A100 managed-service margin thesis. Watch A100_80GB|SPOT|US_EAST_VIRGINIA — if it drifts below $0.74/hr at any major hyperscaler other than Azure, the catalog-exit clock has started.

H3 — The Capex Narrative Is Confirmed, But "Demand Outrunning Supply" Is Empirically False — The Spot Market Says So Confidence: 4 / 5

Thesis: The press consensus — that AI infrastructure spending is "unstoppable," that demand is outrunning supply, that GPU lead times of 9–12 months reflect genuine scarcity — is contradicted by the most objective pricing signal available: real-money spot clearing rates. If H100 supply were genuinely constrained by demand, spot would not print $0.57/hr in Frankfurt, $1.38/hr in Dublin, or $2.50/hr in Seoul against a $7.98 list. These are not thin-market anomalies — Frankfurt has 9 distinct providers in the SPOT tier, Dublin has 15, Seoul has 8, and AWS alone shows 1,700 daily H100 SPOT observations. The correct framing is two non-overlapping markets: (1) a concentrated elite demand layer — OpenAI GPT-6 Astra training at 100,000+ GPUs at Stargate Texas, Amazon's confirmed $8 billion Nvidia sale-leaseback, Broadcom projecting $58B AI revenue for FY2026 driven by custom ASICs for Google, Meta, and Anthropic — which operates entirely in multi-year committed capacity and never touches spot; and (2) a commoditizing spot layer where H100s clear at 50–92% below list and the B200 arrival is releasing more capacity hourly. The capex is real and provable — 179 datacenter projects UNDER_CONSTRUCTION including the 1.4 GW Stargate Michigan campus, Amazon's 2.25 GW New Carlisle Indiana build, and AWS West Texas permitted for 7.65 GW of natural gas capacity. But capex is being built for the elite layer, and it does not prevent the commodity spot layer from softening simultaneously. Conflating the two is the single most dangerous investment mistake in this space today.

Key Evidence:

1. H100 SPOT across 19 of 20 active global markets trades at 50–98% discount to list. At 9-provider Frankfurt with 30-day volume, this is liquid-market price discovery, not a data artifact (live tickers, Oct 5, 2026).
2. AWS H100 SPOT n_observations has grown from ~1,550/day in mid-August to ~1,700/day today. If elite demand were absorbing all excess capacity, n_observations would be falling as H100s move from the spot pool to committed clusters. The opposite is happening — the commodity pool is expanding (60-day instance price history, Oct 5, 2026).
3. 179 datacenter projects confirmed UNDER_CONSTRUCTION in the tracker, including 1.4 GW (Stargate Michigan, $56B, Oracle/OpenAI), 2.25 GW (Amazon New Carlisle Indiana, $11B), and 1.56 GW (xAI Shelby County). The capex signal is real and large (datacenter tracker, Oct 5, 2026).
4. GPT-6 Astra confirmed trained on 100,000+ GPUs at Stargate Texas — a single training run consuming roughly $70–90M/month in GPU capacity at committed rates. The elite layer is real and growing (GPU news feed, Oct 4, 2026).
5. Amazon confirmed in negotiations for $8B Nvidia sale-leaseback — a committed-capacity transaction orthogonal to the spot market (GPU news feed, Oct 4, 2026).
6. Qualifying: The bifurcation hypothesis requires that the elite tier is growing while commodity softens — the spot data confirms commodity softening clearly, but we have no direct pricing signal from the elite tier to confirm it is growing proportionally. The risk is that even elite demand slows (training efficiency improving via MoE, quantization, Jalapeño ASIC substitution), in which case the commodity supply glut becomes the whole story.

Implied Action: Distinguish sharply between NVDA equity (exposed to both layers — custom ASIC revenue from elite lab procurement plus GPU chip revenue from hyperscaler purchases, both growing) and neocloud equity (exposed only to the commodity spot layer where pricing is softening). The bull thesis on Nvidia is the elite demand story, not spot-market tightness. The bear thesis on neoclouds is spot-layer commoditization accelerating as B200 supply normalizes. For corporate buyers: the "9–12 month GPU lead times" cited in trade press are for large-cluster committed orders — spot-market H100s are available today at $0.57–4/hr. If your workload is price-elastic and can tolerate spot interruption, the rental market is not scarce.

H4 — Open-Weight + Neocloud Is Building a New Price-Elastic GPU Demand Floor — But Azure Spot Is the Threat Confidence: 3 / 5

Thesis: The combination of Qwen3 30B A3B (Apache 2.0, self-hostable, available on Nebius at $0.20/$2.40/MTok input/output) against GPT-6 Sol ($2/$10/MTok) is a 10× input-token price gap for models that perform comparably on many structured enterprise tasks. This creates a genuine second demand curve for GPU-hours: price-elastic enterprise users who self-host open-weight models on neocloud clusters rather than paying per-token API rates. Claude Opus 5's 100% score on accounting benchmarks at $0.21/criterion (vs. $10.35/criterion for licensed accountants) is the demand catalyst — enterprises crossing professional-quality thresholds will self-host to avoid API cost lock-in at scale. The economics work: an 8× H100 neocloud cluster at $24/hr (Hyperstack, Lambda pricing) can serve Qwen3 30B at ~150-200 tokens/second, undercutting API pricing by 70–90% at enterprise volumes. Falling H100 spot rates at neoclouds ($3–4/hr today) are stimulating, not reflecting the absence of, this demand cohort. The key threat: Azure H100 SPOT implied at $1.48/hr (UK West, corroborated SPOT discount data) directly undercuts neocloud H100 pricing in those regions. If Azure makes spot broadly accessible to enterprise users without hyperscaler contract requirements, the neocloud value proposition in the self-hosting segment could be structurally undercut.

Key Evidence:

1. Qwen3 30B (Apache 2.0) available via Nebius at $0.20/$2.40/MTok vs. GPT-6 Sol at $2/$10/MTok — 10× input token gap with comparable benchmark performance on structured tasks (AI model tracker + news feed, Oct 5, 2026).
2. Azure H100 SPOT implied prices in UK West region at ~$1.48/hr — the primary structural threat to neocloud H100 pricing floors, confirmed by 20+ Azure SPOT observations in the spot discount table (spot_discounts analytics, Oct 5, 2026).
3. Frankfurt H100 SPOT spread of 557% between min ($0.15/hr) and max ($1.00/hr) across 9 providers confirms hyperscaler-neocloud pricing has not yet converged — neocloud margins are currently protected by market fragmentation (live ticker, Oct 5, 2026).
4. B200 neocloud pricing (CoreWeave hgx-b200-8x, independent read): $43.92/hr ($5.49/GPU/hr) in Virginia — at this rate, enterprise self-hosting of Qwen3 on B200 hardware implies inference costs orders of magnitude below API pricing (underlying instance data, Oct 5, 2026).
5. Qualifying: The self-hosting demand thesis is structurally sound but not yet visible as a stabilizing floor in the spot data — neocloud H100 SPOT prices continue declining in most markets. The demand cohort may be growing but is not yet large enough to absorb the supply release from B200 migration.

Implied Action: Monitor Nebius, Lambda, and Hyperstack monthly revenue metrics for signs that GPU utilization is rising even as per-GPU prices fall — that would confirm the elastic demand cohort is absorbing supply. The stabilization signal to watch: neocloud H100 SPOT prices holding or recovering even as hyperscaler H100 SPOT continues softening. Weekly check of the H100 Frankfurt and Dublin tickers (both multi-provider, high liquidity) will be the earliest indicator. Avoid binary neocloud positions until the Azure spot threat is resolved — the hyperscaler undercutting risk is present and not theoretical.

Inventory & Channel Health

The formal inventory health pipeline (DIO signals, composite regime) returned no active records this cycle — the DIO system has not refreshed with the latest earnings cycle, so no layer-level signal (buyers, distributors, chip suppliers) is current. Treat the absence as data absence, not a GREEN.

What the trade flows make unambiguous: US flash/SSD imports (HS 852351) are running at +499% year-over-year with a $7.1 billion monthly import value as of July 2026, with 3-month momentum accelerating at +75%. Exports are flat (+6% MoM). The import surge without a corresponding export rise is the NAND pre-loading signature — hyperscalers building storage buffers ahead of large cluster deployments, not a distribution channel glut forming. Broad IC imports (HS 8542, which captures GPUs and other chips) are up +40% YoY but decelerating at –13% MoM — consistent with the initial H100/H200/Blackwell import wave having passed through the trade pipeline in Q1-Q2 2026 and now deployed or in datacenter staging. Forward demand is a function of the Vera Rubin and B300 ramp cycles, not a continuation of the current wave.

Read: The NAND import surge corroborates the capex buildout and confirms datacenter deployments are active. The IC import deceleration suggests the GPU hardware wave is now deployed, not in transit — the next pricing inflection waits for the next hardware generation to clear customs.

Supply & Demand Pipeline

Supply — Capacity Landing:

The most material supply-side development is not a new project announcement but a project failure: Project Jupiter (Stargate New Mexico, 4 GW planned, $165B, Oracle/OpenAI) has force majeure invoked with construction at 26% complete and on hold due to a natural gas pipeline rejected repeatedly by state regulators. Debt is trading below 90 cents. This is not an isolated event — it is the first confirmed instance of a mega-campus failing to resolve power procurement in the current pipeline, and its failure mode (regulatory, not financial) is replicable. Simultaneously, three corroborating friction points have emerged: Amazon's Ohio expansion was paused by the state specifically around energy service agreements; Google's proposed 859-acre South Carolina campus triggered litigation and county-level development approval pauses; and the TVA created a new datacenter rate structure specifically for xAI's Shelby County 1.56 GW facility, including a 10% average increase for facilities over 5 MW — a first-of-kind regulatory precedent. The intel feed shows active large-load proceedings at Utah (300 MW single-facility caps), Pennsylvania (25 MW triggers), and Missouri (protective orders for large-load contracts). These are not theoretical risks — they are live regulatory proceedings. The power constraint is a 2027-2030 problem for GPU pricing, but it is a now problem for datacenter construction timelines and project financing.

Demand — Model Velocity:

Five major model releases in 10 days establish the fastest frontier release cadence on record: GPT-6 Sol ($2/$10/MTok, 1M context, Sept 29); Claude Sonnet 5.5 (30% faster, Sept 28); Gemini 4 Argon (security-specialized, Sept 30); Meta Muse (multimodal, restricted EEA/UK); and Apple AFM 3 Cloud Pro (running on Blackwell GPUs via GCP). Token economics are deflationary across all: GPT-6 Sol is one-fifth the price of its predecessor GPT-6 Astra ($10/$50/MTok), and Qwen3 30B (Apache 2.0, MoE) at $0.20/MTok represents a 100× token efficiency gap versus GPT-6 Astra at the flagship end. The inference efficiency gains embedded in these models — MoE architectures reducing active parameter count per query, speculative decoding, improved quantization — are a direct mechanical driver of H100 spot softness: each efficiency gain serves more queries per GPU-hour, reducing the demand signal for additional GPU capacity at any given query volume.

Component — B200 Asymmetric Availability:

The B200 is catalogued as active (is_active: true) on AWS only among the three major hyperscalers — Azure and GCP show inactive. This asymmetry is not a data error; it reflects AWS moving faster on Blackwell deployment than its peers. The practical consequence is a geographic two-speed market: AWS regions have B200 supply pressure driving H100 release into spot; Azure and GCP regions do not yet. This explains why the H100 spot discount range is so wide (50–92%) — it is partly a function of which hyperscaler dominates the regional market, not just local demand.

Pre-Announced Price Changes

No UPCOMING_PRICE_CHANGE alerts are present in the system. No provider (AWS, Azure, GCP, or Oracle) has published a forward-dated price move via API as of this morning's pipeline run. The pre-announcement channel is silent — all future moves will arrive without warning.

Risk Flags

Immediate (0–30 days)

R1 — H100 ON_DEMAND List Cut Arrives Without Warning The UPCOMING_PRICE_CHANGE queue being empty is not reassurance — it is the expected state before an unannounced cut. The model for how these moves happen: reserved rates fall first (Jakarta –30% in 30 days is live), spot clearing rates collapse (documented), then ON_DEMAND list follows 60–90 days after the reserved move. Jakarta's reserved decline started roughly 30 days ago; the Virginia reserved rate has not yet moved. If it moves in the next 2–4 weeks, the ON_DEMAND list cut is within the immediate window. Any enterprise with GPU-cost-sensitive pricing models (SaaS AI features, GPU resellers, neocloud operators) should be running sensitivity scenarios against a 20–40% ON_DEMAND list cut now, not after announcement.

R2 — São Paulo H100 SPOT at $9.50/hr Is a Hidden Concentration Risk The only H100 SPOT market globally that prices above ON_DEMAND list ($9.50/hr vs. $7.98 list) is São Paulo — a single-provider market with zero spread and only 2 observations. This is a geographic monopoly print, not demand scarcity. Any workload in Brazil paying SPOT rates is paying above list for the same chip available globally at 50–92% discount. This is a procurement inefficiency, not a market signal, but it is material for any operator with latency requirements routing Brazilian workloads exclusively to in-region capacity.

R3 — A100 Virginia SPOT Already at Sub-$1/GPU/hr — Catalog Exit Is Not a Future Risk Azure A100_80GB SPOT in Virginia has crossed ~$0.74/GPU/hr. This is the threshold I identified as the structural inflection signal in the hypothesis step. The risk is immediate: operators billing A100 managed-service workloads at 2024 contracted rates face customer churn to spot alternatives within the current contract cycle, not at renewal. Any managed-GPU business with A100-based products should be tracking whether client churn is accelerating in the next 30 days.

Near-Term (30–90 days)

R4 — Blue Owl Capital / Stargate New Mexico Earnings Disclosure Project Jupiter (4 GW, $165B, 26% complete, on hold, debt below 90 cents) has not yet triggered a public disclosure from Blue Owl Capital, the project landlord. The next earnings call is a material event risk. If force majeure provisions trigger rent deferral or covenant relief language, it will be the first confirmed datacenter REIT write-down tied to regulatory power failure — and it will reprice the entire sector's regulatory risk premium. The datacenter REIT sector (EQIX, DLR) is currently pricing near-zero regulatory failure probability into its pipeline NAV. A Stargate NM disclosure changes that calculus. Watch for New Mexico regulatory filings on the gas pipeline — a third rejection triggers a formal default clock.

R5 — B200 Spot Normalization Compresses H100 Market Further The AWS B200 SPOT floor rose modestly from $14.3/hr (Sep 15) to $14.8/hr (Oct 5), and the partial H100 spot floor recovery from $0.74/hr to $1.72/hr appears to be a B200/H100 substitution dynamic: when B200 SPOT tightened, some operators reverted to H100 spot. If B200 SPOT normalizes downward as production scales — the expected trajectory as TSMC CoWoS-L capacity ramps — the substitution support for H100 spot will disappear, driving H100 spot to new floors. The risk window is 60–90 days as Blackwell production volumes grow.

R6 — Neocloud Margin Destruction if Azure Spot Becomes Broadly Accessible Frankfurt's 557% H100 SPOT spread between min and max across 9 providers is the current protection for neocloud margins: market fragmentation prevents hyperscalers from systematically clearing the market at $0.15–0.19/hr. But Azure's implied H100 spot at $1.48/hr in UK West and the documented US East Virginia Azure-only SPOT pool represent latent price-leadership capability. If Azure makes H100 spot accessible without hyperscaler contract prerequisites — a feasible policy choice — it would undercut neocloud H100 floor pricing in 30 days across most major markets. Monitor Azure's SPOT accessibility policies in the next earnings cycle for any signal.

Structural (6–24 months)