Daily Investment Brief — September 27, 2026

All charts rendered. Here is the full brief.

SIGNWL DAILY COMPUTE BRIEF — 27 September 2026

Model run: 07:31 UTC. Prices are model-decomposed $/GPU-hr unless otherwise noted. Neocloud prices are independent market reads, not unified model outputs. Confidence scores reflect post-stress-test assessments.

Market Pulse

The dominant structural story today is sub-market decoupling at the provider and regional level — the aggregate GPU market is not moving in one direction; it is fracturing. Azure's H100 in Zurich has reached $15.24/hr on-demand (+54% over 30 days), while GCP's identical chip on preemptible in Frankfurt is clearing at $0.167/hr, a 91× spread on the same nominal product. Neither number is an artifact: the Zurich price is confirmed across multiple Azure SKU families (ND96isr_H100_v5 at $17.57/GPU, NC40ads at $11.82/GPU), and the Frankfurt floor is a stable, nine-provider-corroborated GCP preemptible price that has held for 30 days. Simultaneously, a significant share of today's apparent market volatility — a wave of 100–700% 24-hour ticker moves across Azure's A10, T4, M60, and A100 — is a Stage 1 decomposition artifact, not real repricing. Azure's vCPU/RAM baseline shifted during the September 27 model run, inflating residuals across every Azure GPU component, at a time when zero Azure SKU-level prices actually changed in the billing catalog. The practical risk: every Azure GPU ticker moved strongly today, almost none of it is real, and any system that auto-trades on ticker deltas without a raw-SKU cross-check fired a false signal this morning.

Key Movers

Note: All Azure GPU ticker moves >50% on September 27 should be treated as potentially contaminated by Stage 1 baseline instability. Cross-reference query_underlying_instances before acting. The anomaly diagnostics table shows Azure Stage 1 GPU_type dimension at 64.4% anomaly rate for this run — the highest in the dataset.

Investable Insights

H1 — H100 Is Running Two Entirely Separate Markets: A 91× Spread That Is Real, Widening, and Structurally Grounded Confidence: 4 / 5

Thesis: The H100 global market has bifurcated into two products that share a name but serve fundamentally different commercial functions. Azure Zurich's H100 on-demand at $15.24/hr is a premium colocation product — confirmed across three distinct SKU families with genuine multi-tier list pricing ranging from $1.996/GPU (NC40ads shared-tenancy) to $17.57/GPU (ND96isr with InfiniBand fabric), all in the same region, the same billing catalog, simultaneously. The 30-day move of +54% in Zurich is not model drift: it is Azure building a structured price ladder in the only Swiss H100 zone, where no other provider competes. At the other end, GCP Frankfurt preemptible H100 has stabilized at $0.167/hr after a –71% 30-day collapse — a floor that has held for 30 consecutive days with n_instance_types=7 throughout, meaning there is no catalog churn; GCP simply set this floor and held it. The cross-sectional depreciation model shows 18.6% annual decay as the macro average (p=0.017, R²=0.24), but the empirical 30-day range across H100 OD and SPOT regions in this dataset runs from –71% to +54% — a distribution so wide that any DCF model applying a single depreciation rate to H100 hardware capex is carrying unquantified regional basis risk that likely exceeds the depreciation rate itself.

The mechanism behind the bifurcation is straightforward: where only Azure operates (Zurich, certain Middle East zones), Azure has pricing power and is exercising it structurally. Where GCP operates preemptible capacity alongside Azure OD (Frankfurt, US-West, US-Central), the preemptible floor sets a reference price that anchors buyer expectations, compresses spreads, and drives OD prices toward multi-provider equilibrium. The two markets are not arbitrageable in practice — compute SLAs prevent workload migration from OD to preemptible for production inference — so the spread can persist for as long as Azure remains the sole provider in high-demand European inference markets. The catalyst for compression is Azure deploying H200 or GB200 in Zurich, which would force H100 into the secondary inference tier and compress OD pricing; that event has not yet appeared in any price alert or SKU catalog.

Key Evidence:

1. Azure Zurich H100 OD confirmed at $15.24/hr median via ticker_prices_v2, with underlying instances Standard_NC40ads_H100_v5 at $9.981/hr and $11.821/hr and Standard_ND96isr_H100_v5 at $140.60/hr total ($17.57/GPU) — multi-tier catalog confirmed, not a decomposition artifact (underlying instance data, Sep 27, 2026)
2. GCP Frankfurt H100 SPOT at $0.167/hr decomposed, stable floor for 30 days, a3-highgpu-8g / a3-megagpu-8g / a3-edgegpu-8g preemptible confirmed as source instances; n_instance_types=7, no churn (instance price history, 30d window, Sep 27)
3. Cross-sectional depreciation model: 18.6% annual decay, R²=0.24 — macro average masks regional variance from –71% to +78% on 30d horizons (depreciation analysis, Sep 27)
4. Osaka H100 OD –55.5% over 30d, consistent with APAC multi-provider entry; Middle East H100 OD +7.6% 30d, consistent with Azure-dominant single-provider market — directional pattern validates provider concentration hypothesis
5. No Azure H200 SKU present in Zurich in any pricing table or alert feed — compression catalyst absent as of today's data

Implied Action:

• Infrastructure capex buyers: Do not apply a uniform 18–19% H100 depreciation curve to hardware investment decisions. The true range is –55% to +54% over 30 days depending on region and provider concentration. Zurich-based H100 reserved instances are holding value; Frankfurt spot capacity is at a floor that implies zero scarcity value.
• Long reserved: Buyers with inference SLAs in EU-Central or Middle East should lock Azure H100 1-year reserved rates before Azure's H200 deployment in those zones triggers the re-rate — the RI price typically moves after OD but before the list-price announcement.
• Short Frankfurt spot: Neocloud operators who built P&L models assuming Frankfurt H100 SPOT would recover to $1.50–$2.50/hr face structurally stranded capacity. The GCP preemptible floor at $0.167/hr, held for 30 days, is not a transient clearing price — it is GCP's stated floor for a3 preemptible in Europe.
• Watch: Any UPCOMING_PRICE_CHANGE alert for Azure europe-central-zurich with accelerator=H200 is the primary compression catalyst. Set a standing alert.

H2 — Nscale's IPO Creates the First Observable Vera Rubin Pricing Event — and a 6–18 Month Window to Position Ahead of It Confidence: 4 / 5

Thesis: There is no Vera Rubin GPU-hr ticker in the market. Not on any of the 15 providers tracked, not on CoreWeave (which submitted Vera Rubin MLPerf results), not on Nebius (which benchmarked 591K DeepSeek-R1 tokens/second on Rubin silicon). The hardware exists — the component tracker confirms Vera Rubin NVL72 at status=AVAILABLE, with Monarch Compute Campus listed as an active deployment site, 20.7 TB HBM4 per rack, TSMC 3nm fabrication. Ward County Texas is already receiving 104,000 GB300/VR200 units for Microsoft. The benchmarks are public — MLPerf Inference v6.1 (September 16) shows 3.7× the throughput of GB300 on interactive workloads. The contracts are disclosed — Nscale's S-1 reveals $103.4B in total contracted value, $44.6B from Anthropic and $43.8B from Microsoft. The price does not exist yet. This is a structurally quantifiable pre-announcement gap: the first Vera Rubin $/GPU-hr market print will instantly obsolete the current GB200 value-per-TFLOP benchmarks and trigger a re-rating cascade across GB200 and H100 in training markets.

The math is direct. GB200 ON_DEMAND in US Central clears at $0.176/TFLOP-hr today (Azure only, the sole GB200 price-maker). Vera Rubin delivers approximately 3.5× the training performance and 5× the inference performance per GPU relative to Blackwell. A competitively priced Vera Rubin should clear at $0.035–$0.05/TFLOP-hr equivalent for inference workloads, implying a per-GPU-hr rate that, despite being higher in absolute dollars, would make GB200 at $0.176/TFLOP immediately appear overpriced. Any neocloud or hyperscaler currently charging more than $0.17/TFLOP on GB200 — which includes Azure's entire GB200 portfolio — faces a repricing cliff when the first Vera Rubin rate card is published. The asymmetry is structural: Vera Rubin's price discovery is delayed, not absent, and the entities that control the disclosure timeline (Nscale, CoreWeave, NVIDIA's cloud partners) have economic incentives to price-discover as publicly as possible given pending IPO and S-1 requirements.

The Nscale IPO introduces a specific risk that did not appear in the scan step: H1 2026 net losses approaching $1 billion against $140.6M revenue. At a targeted $35B IPO valuation, Nscale is pricing at roughly 124× annualized revenue — an aggressive multiple that requires the road show to successfully anchor investors to the $103.4B TCV rather than the current P&L. If the IPO clears below $20B, Nscale's debt capacity for the Monarch 8 GW build is materially compromised, which would delay the first commercial Vera Rubin price signal by 12–24 months. This is the primary execution risk, not hardware availability.

Key Evidence:

1. Vera Rubin NVL72 component tracker: status=AVAILABLE, Monarch Compute Campus as deployment site, 20.7 TB HBM4 per rack, confidence=0.82 (component tracker, Sep 27)
2. Ward County TX datacenter tracker: 104,000 GB300/VR200 units, Microsoft as anchor tenant, Q3 2026 delivery beginning — Rubin-class hardware already deploying, no ticker exists (datacenter tracker, confidence=0.93)
3. GB200 price_performance: US Central OD $0.176/TFLOP-hr — the current generational benchmark that Vera Rubin will obsolete at 5× inference multiplier (unified analytics price_performance table, Sep 27)
4. Nscale S-1 (NYSE: NSCL, filed Sep 18, 2026, source: TradingView, scraped_ok=true): $103.4B TCV, 1,252% H1 YoY revenue growth, $140.6M H1 revenue, ~$1B H1 net loss, $35B IPO target valuation
5. MLPerf Inference v6.1 (Sep 16, 2026): Vera Rubin NVL72 3.7× throughput vs. GB300, 99% scaling efficiency at 288 GPUs — benchmarks are public, pricing is not
6. No Vera Rubin ticker in any of 15 tracked providers: confirmed absence across all available_components data, Sep 27

Implied Action:

• Long NSCL at IPO: The $103.4B TCV is the demand anchor, not the P&L. Nscale is structurally analogous to a net-lease REIT with 6-year contracts from Anthropic and Microsoft — the cash flow waterfall is contracted; the risk is execution, not demand. The Monarch microgrid certification (West Virginia's first state-certified AI microgrid) provides regulatory arbitrage that no Virginia-based competitor can replicate within 12–18 months.
• Short GB200 OD exposure: Any neocloud or operator currently building GB200-based revenue models using today's $0.176/TFLOP pricing as a floor should model downside scenarios from Vera Rubin price discovery by Q2–Q3 2027. The compression will be faster in training markets than inference, because inference buyers are stickier (latency, fine-tune dependencies).
• Watch: First Vera Rubin ticker appearance on any of the 15 tracked providers (most likely CoreWeave or Nebius first, given their MLPerf submissions). Also watch Nscale's post-IPO 10-Q — the first quarterly filing must disclose material contract pricing terms that will be the first public Vera Rubin per-GPU-hr data point.

H3 — GCP's TPU v5p Amsterdam Spot Is a Supply Valve, Not a Demand Signal — And It Points to an Unannounced Product Transition Confidence: 3 / 5

Thesis: The TPU v5p spot price in Amsterdam executed one of the most technically legible ASIC pricing patterns in this dataset: nine months of step-function decay from $0.575/hr in January 2026 to a $0.0145/hr trough by mid-September (a 97.5% decline), followed by a hard reversal on September 22 to $0.135/hr (+831% over 7 days). Each step in the descent was a hard shelf — the pattern of a price-setter adjusting allocation deliberately, not an auction market clearing at progressively lower prices. GCP's fixed spot pricing model for TPUs means this is an administrative decision made by GCP's internal capacity team, not a market outcome. The critical diagnostic is what OD did in parallel: Amsterdam TPU v5p OD sat at $1.4503/hr with a –6.85% 30-day decline, and Dublin TPU v5p OD at $1.2045/hr was flat. OD softening while SPOT recovers is the fingerprint of supply withdrawal, not demand recovery — GCP pulled back spot allocation because clearing prices approaching zero threatened to erode OD revenues from customers who might switch from OD to spot.

The timing of the reversal — six days after MLPerf Inference v6.1 (September 16) confirmed GPU architectures outperforming v5p on key benchmarks — is consistent with a scenario where GCP observed reduced spot take-up immediately post-benchmark and responded by tightening allocation. This is a counterintuitive but mechanically sound sequence: fewer buyers at the spot floor → GCP withdraws spot supply → cleared price recovers, not because demand returned but because the supply curve shifted left. The Dublin SPOT divergence (–11.1% 7d, declining while Amsterdam recovers) confirms this is zone-specific rebalancing rather than a system-wide ASIC supply withdrawal — weakening the broadest version of the thesis but not the core mechanism. Virginia OD at +4.93% 30d adds a hint of genuine demand recovery in the US that Amsterdam doesn't show, leaving the picture mixed.

The structural implication is more important than the price level: GCP setting a $0.0145/hr spot floor for a ASIC with an OD rate of $1.4503/hr — a 99% discount that GCP maintained for weeks — represents an aggressive utilization-maintenance posture consistent with a product approaching end-of-commercial-life. GCP does not run 99% spot discounts on capacity it believes will be in demand. The Amsterdam pattern is most consistent with a product transition toward TPU v6, where GCP needs to keep v5p utilization up to justify its datacenter footprint while simultaneously managing down customer price expectations before the next-generation transition.

Key Evidence:

1. TPU v5p SPOT Amsterdam: trough $0.0145/hr, reversal to $0.135/hr on Sep 22 (+831% 7d), step-function descent from $0.575/hr in January — all confirmed at n_providers=1, n_observations=94 (ticker history, Sep 27)
2. TPU v5p OD Amsterdam: $1.4503/hr, –6.85% 30d — OD declining concurrent with SPOT reversal confirms supply withdrawal, not demand recovery (ticker data, Sep 27)
3. Dublin TPU v5p SPOT divergence: $0.4653/hr, –11.1% 7d — zone-specific pattern, breaks system-wide withdrawal narrative (ticker data, Sep 27)
4. Virginia TPU v5p OD: $1.7120/hr, +4.93% 30d — small OD recovery in US but not EU, mixed signal (ticker data, Sep 27)
5. MLPerf Inference v6.1 (Sep 16, 2026): GPU architectures outperform v5p on key benchmarks — independent corroboration for reduced ASIC spot demand post-publication
6. Zero UPCOMING_PRICE_CHANGE alerts for TPU v5p across all regions — GCP has not pre-announced any rate change; transition, if occurring, is unannounced (price alerts, 7d window, Sep 27)

Implied Action:

• Buyers of GCP TPU v5p capacity: The September trough represented the optimal window to lock multi-year OD contracts; that window has now likely closed as GCP withdraws the spot subsidy. Buyers who didn't lock are exposed to OD re-rating if GCP raises Amsterdam OD rates within 30 days — the primary confirmation signal for this hypothesis.
• Watch for TPU v6: Any GCP announcement of TPU v6 pricing or availability would be the smoking gun for the product-transition thesis. It would also be the largest ASIC price discovery event of 2026–27, resetting the ASIC vs. GPU TCO comparison that today's data puts firmly in GPU's favor.
• Pairs posture: A long TPU v5p OD forward position (if available) hedged against short H100 GCP preemptible (which faces the same demand softness in GCP's own infrastructure) is the tightest expression of the ASIC supply-withdrawal thesis. Both legs depend on GCP's internal capacity reallocation decisions rather than external demand — which makes the trade endogenous to GCP's product roadmap.

H4 — The Regulatory Ratchet Is Real, Five-State Synchronized, and Will Set a Supply Floor — But Only in 2027–2028 Confidence: 3 / 5

Thesis: Within a five-day window ending September 26, five independent regulatory actions arrived simultaneously: Maryland's Utility RELIEF Act of 2026 defining "Large Load Customer" procurement constraints (Sep 26), Milwaukee's CPC moratorium hearing (Sep 25), Lysander NY advisory committee meeting on data center power draw (Sep 18), CAISO's FERC large-load interconnection re-rule (Sep 21), and Oregon's cooling limit consideration. This is not a policy gradient — it is a synchronized state-level response to grid stress that has reached the legislative stage. The CAISO FERC re-rule specifically adds procedural steps to the interconnection queue for large loads, which historically translates to 12–18 months of additional timeline for projects that have not already cleared. The Maryland Act is the most operationally specific: it defines a statutory framework for what counts as a "Large Load Customer," which is the legislative prerequisite for load interconnection fees that have historically added 15–25% to the effective cost of power for data centers in constrained markets.

Construction wage inflation of 125–164% year-over-year in electrician and pipefitter categories — confirmed in news feed across multiple sources — creates a parallel cost pressure that applies even to projects that clear permitting. The cost spiral has two components that are mutually reinforcing: interconnection delays push more projects into the same constrained labor market, which drives wages higher, which increases the all-in cost of capacity that eventually does connect. The projects that escape this dynamic are the ones already off-grid: Nscale's Monarch campus in West Virginia, with state certification as the US's first AI microgrid under House Bill 2014, bypasses the FERC/PUC interconnection queue entirely. Starwood Digital Ventures' $12B, 2.16 GW project in Mason County WV is competing for the same regulatory arbitrage structure — confirming it has real economic value.

The transmission to GPU pricing, however, is lagged and geographically bounded. The current pricing data shows no H100 or H200 price increases in US East Virginia on any pricing type — the regulatory constraint hasn't yet materialized as a capacity shortage. The mechanism requires: (1) the interconnection rule changes to take legal effect and accumulate queue backlogs (6–12 months), then (2) existing pipeline projects to exhaust available capacity in constrained zones (12–24 months), then (3) GPU prices to reflect the reduced new supply (18–30 months from today). The near-term actionable element is not GPU pricing — it is the valuation premium embedded in off-grid, regulatory-arbitrage infrastructure like Monarch.

Key Evidence:

1. Maryland Utility RELIEF Act (Sep 26, 2026): "Large Load Customer" statutory definition, direct constraint on data center electricity procurement — confirmed from intel feed, scraped_ok=true
2. CAISO FERC large-load interconnection re-rule filing (Sep 21, 2026): new procedural requirements for large-load interconnection; historical precedent adds 12–18 months to queue (intel feed, Sep 21)
3. Milwaukee CPC moratorium hearing (Sep 25, 2026), Lysander NY power advisory meeting (Sep 18), Oregon cooling limits under consideration — five distinct state-level actions within five days (intel feed, Sep 18–26)
4. Construction wage inflation 125–164% YoY for data center electricians/pipefitters: confirmed across multiple sources in news feed, Sep 2026
5. Nscale Monarch (WV): off-grid microgrid, state-certified, natural gas + battery storage, HB2014 microgrid program — only US project confirmed fully clear of grid interconnection dependency (datacenter tracker, confidence=0.95)
6. Starwood Digital Ventures $12B, 2.16 GW Mason County WV project: citing same HB2014 program, rapid imitation confirms the regulatory arbitrage has real economic value (datacenter tracker)
7. Zero H100/H200 OD price alerts for US_EAST_VIRGINIA: no near-term pricing reflection of the regulatory constraint (price alerts, Sep 27)

Implied Action:

• Long: Off-grid, regulatory-arbitrage infrastructure assets — Monarch (Nscale), and any future projects that secure WV/similar microgrid certification before the program becomes oversubscribed. This is the cleanest long-duration, infrastructure-rate play on AI compute infrastructure that currently trades at a discount to its embedded option value.
• Structural caution: Hyperscalers with large Virginia/Maryland/California pipeline (all three have major NOVA and Bay Area footprints) face cost headwinds in 2027–28 that don't appear in their current capex models. For Azure specifically — already the price-setter in constrained EU markets — a simultaneous US East supply constraint could push OD pricing power into a single region, creating exactly the geographic pricing power Azure is already demonstrating in Zurich.
• Watch: Any FERC order approving or rejecting a specific Virginia data center interconnection application will set precedent for the queue timeline and trigger the first observable pricing response. Also watch for AWS, GCP, or Azure announcing new capacity commitments in Texas, Ohio, or mountain-west states — geographic flight from constrained markets is the most direct evidence of the regulatory floor operating as predicted.

Inventory & Channel Health

The composite channel regime reads as MIXED, driven by a sharp bifurcation between the flash/NAND channel and the broader IC/logic channel. Flash imports (HS 852351) surged +499% year-over-year and +4.65% month-over-month in July 2026, with a 75% three-month momentum — a RISING signal at extreme magnitude. Simultaneously, broad IC imports (HS 8542) fell –13% month-over-month in the same period, meaning the semiconductor channel is bifurcating: NAND/flash is absorbing at an extreme rate while logic and compute ICs soften. This divergence historically precedes a NAND inventory build by one to two quarters — when flash import volumes detach from IC volumes at this magnitude, distributor DIO for NAND-adjacent suppliers typically follows upward within 90 days. The channel stuffing flag is not yet triggered, but the conditions are assembling.

The read is structurally important for the Vera Rubin thesis: training runs at Monarch-class scale require massive flash/NVMe staging for dataset loading and checkpointing. A concurrent surge in flash imports alongside the deployment of 104,000 GB300/VR200 units at Ward County TX is consistent with pre-deployment data staging, not demand-side speculation. If flash imports maintain this trajectory through Q4 2026 while IC imports remain soft, the most likely outcome is a NAND glut signal in Q1 2027 — which would pressure storage costs for new data center builds but not directly affect GPU pricing.

No RED DIO flags were returned from the distributor layer in this cycle. The absence of distributor-level RED signals provides a partial check against the glut thesis, but the trade flow divergence is large enough to warrant re-checking distributor DIO (ARW, SNX, AVT) in 30–45 days.

Supply & Demand Pipeline

Supply — Nscale's Ward County TX Is Already in Motion: The most material supply-side development is not in the pipeline — it has started. The Ward County TX datacenter (Microsoft tenant, 275 MW, 104,000 GB300/VR200 units, $14B contracted) began delivery in Q3 2026. This is Vera Rubin-class hardware entering real deployment with no corresponding pricing ticker anywhere in the market. The Monarch Compute Campus (WV, 2 GW current build toward 8 GW, Anthropic $45B contract) is the follow-on wave, targeting H1 2028. Both are off-grid or microgrid, which insulates them from the CAISO/Virginia interconnection constraints discussed under H4.

Supply — European Regulatory Arbitrage: Nscale's Sines Portugal facility (200 MW, 66,000 Vera Rubin chips, end-2027 target) and Narvik Norway (1 GW planned, Microsoft primary tenant) are the European legs of the same off-grid arbitrage. The Norway site targets hydroelectric power, which places it outside the grid stress dynamics affecting CAISO and mid-Atlantic utilities. Neither project is in the pricing tickers — both represent dedicated sub-markets where pricing will be governed by long-term contracts rather than spot markets.

Demand — Anthropic's Committed TCV as the GPU Demand Anchor: Anthropic's aggregate compute commitments have reached $517B across all announced contracts (news feed, Sep 13, 2026 via Crypto Briefing) — a figure that implies near-doubling of GPU demand annually. The Nscale-specific piece ($44.6B over 6 years) is the largest single dedicated compute contract in the dataset and will be the primary demand anchor for Vera Rubin pricing when it materializes.

Demand — Rubin Ultra on the Horizon: The component tracker shows Rubin Ultra (512 GB HBM4e, TSMC 3nm, 576 GPUs per Kyber NVL576 rack, 600 kW power draw) at status=ANNOUNCED for H2 2027 — creating a second generational price discovery event beyond the base Rubin. This matters for the GB200 short thesis: if Rubin base pricing appears in Q2–Q3 2027 and Rubin Ultra follows in H2 2027, the generational compression cycle for GB200 is compressed into an 18-month window, which is faster than the H100→H200 transition took.

Pre-Announced Price Changes

No UPCOMING_PRICE_CHANGE alerts were detected in the 7-day scan window across any provider. This is notable: given the level of actual market activity (H100 Zurich +54% 30d, Frankfurt SPOT –71% 30d, TPU v5p +831% 7d), the absence of any pre-announced changes means providers are implementing adjustments without publishing future effective dates — making the pricing landscape harder to front-run systematically. The implication for institutional buyers is to increase monitoring frequency rather than rely on alert pipelines: changes are happening between model runs, not ahead of them.

Risk Flags

Immediate (0–30 Days)

R1 — Azure Stage 1 Baseline Instability: All Azure GPU Tickers Contaminated on September 27 The Azure Stage 1 GPU_type dimension ran at a 64.4% anomaly rate in today's model run, versus a 4% overall Stage 1 rate. The result is that every Azure GPU component ticker — A10, T4, M60, A100_40GB, V710, H100, A100_80GB — moved 100–700% in 24 hours, while zero Azure SKU-level billing prices actually changed (confirmed by query_price_changes returning no records with ≥50% Azure list price changes). Any automated system that ingested today's Azure GPU tickers as real signals and acted on them fired a false signal. The risk persists until the next model run re-establishes a stable vCPU/RAM baseline. Watch anomaly_diagnostics.stage1_anomaly_rate for Azure falling below 15% before treating any Azure GPU ticker delta as actionable.

R2 — Israel Central A10 Spot Inversion: Real Signal, Unknown Duration The Azure Israel Central A10 SPOT price at $2.51/hr is 16% above the OD price of $2.16/hr — the only confirmed spot-above-OD pricing in the entire dataset for a mature GPU. This is a genuine market anomaly (it survives the Stage 1 stress-test) and is consistent with a capacity-saturated market where someone — government, defense contractor, or large enterprise — is absorbing preemptible capacity at above-list rates. It has been in place for at least 24 hours. If it persists for 7 days, it becomes a structural signal; if it resolves in 48 hours, it was a brief allocation imbalance. Do not build a trade around it until the 7-day check clears.

Near-Term (