Daily Investment Brief — September 18, 2026

SIGNWL DAILY BRIEF — September 18, 2026

Investment thesis edition. All claims grounded in pricing data, instance-level attribution, and model diagnostics collected at 07:31 UTC. Model pipeline health confirmed clean: Stage 2 GPU anomaly rate 0.17%; ticker-delta anomaly rate 0.26%. TPU signals excluded — GCP Stage 1 GPU_type anomaly rate 34.7% renders TPU decomposition unreliable.

Market Pulse

The dominant narrative today is supply-tier bifurcation: frontier silicon is simultaneously glut-forming in neocloud emerging markets and finding a new performance-per-dollar floor in hyperscaler-anchored EU/US spot auctions, while the inference compute layer — long assumed to follow training hardware economics — is decoupling into a permanent structural discount. The sharpest single signal is the H100 SPOT Stockholm ticker completing a textbook nine-month market cycle, with spread compressing from 95% at its August peak to just 13.6% today as the two incumbents reconverge toward a $2.57/hr clearing price — simultaneously capped by AWS's newly live B200 SPOT listing in the same region at $1.67/hr, the cheapest decomposed frontier-GPU spot price on earth. Layered beneath the hyperscaler view is a more consequential hidden market: Nebius is offering H200 ON_DEMAND in Stockholm at $4.30/GPU/hr, invisible to every automated procurement system that reads only the cross-provider ticker anchored to Azure's $13.18/hr monopoly price. The convergence of that hidden arbitrage, the B200 ceiling forming in the Nordics, and the APAC-specific revival of AWS Inferentia spot demand paint a picture of a GPU market that is structurally segmenting across three independent axes — generation, geography, and customer type — faster than any single provider's pricing desk is tracking.

Key Movers

Hyderabad -52% is extrapolated from 14-day history; instance snapshot unavailable today — neocloud may have pulled listing.

Investable Insights

H2 — Azure's Global H200 Monopoly Is Extracting 57% Scarcity Rent Outside the US, with a 3.1x Hidden Arbitrage Sitting Beneath It Confidence: 5 / 5

Thesis: Azure is the sole provider in every H200 ON_DEMAND market on earth — confirmed by the ticker_prices_v2 decomposition table across all 18 canonical regions. The pricing consequences of this monopoly are most visible outside the US: Stockholm H200 ON_DEMAND sits at $0.197/TFLOPS, London at $0.192/TFLOPS, and Tokyo at $0.210/TFLOPS. Compare these to US-Ohio at $0.093/TFLOPS — Azure's own cheapest H200 market, where it faces head-on AWS H100 competition in the p5 corridor. The 57% $/TFLOPS premium EU/APAC customers pay over US-Ohio customers for the same GPU generation is not a performance differential; it is uncontested geographic pricing power. The mechanism is simple and structural: outside the US, no hyperscaler competes on H200, and no neocloud has yet listed H200 at scale through the hyperscaler pricing model. What makes today's read urgent is the discovery that Nebius — a neocloud, therefore not wired into the unified model — is offering H200 SXM in Stockholm at approximately $34.40/node (8 GPUs), implying a decomposed $4.30/GPU/hr: a 3.1x discount to Azure's $13.18/hr ticker price in the same region, and invisible to every procurement system reading the cross-provider ticker. The EU H200 market has three tiers that the ticker collapses into one: Azure ON_DEMAND at $13.18 (the model's anchor), Nebius at $4.30 (out-of-model), and H100 SPOT at $2.74 (the performance-adjusted alternative). The scarcity rent Azure is extracting is real — and so is the arbitrage sitting directly beneath it.

Key Evidence:

1. ticker_prices_v2 (unified model, Sep 18): Azure confirmed as sole provider in all 18 H200 ON_DEMAND canonical regions globally — zero competitive entries.
2. H200 ON_DEMAND $/TFLOPS: Stockholm $0.197, London $0.192, Tokyo $0.210 vs US-Ohio $0.093 — a 57%–126% premium across EU/APAC markets vs the most competitive US region (live ticker, Sep 18).
3. H200 ON_DEMAND 30-day delta: effectively 0% in all EU/APAC markets — scarcity rent is locked in, not converging toward US prices.
4. query_underlying_instances H200 Stockholm (Sep 18): Five rows returned — four Azure Standard_ND96isr_H200_v5 at $101.76–$114.69/full node, plus one Nebius gpu-h200-sxm at $34.40/node = $4.30/GPU/hr (not in ticker).
5. US-Ohio H200 at $0.093/TFLOPS vs H100 at $0.115/TFLOPS — H200 is 19% cheaper per TFLOPS where competition exists, confirming the EU premium is not fundamental to H200 economics.
6. H100 SPOT Stockholm $/TFLOPS: $0.041/hr (price_performance table, Sep 18) — the alternative tier for EU workloads that can tolerate spot interruption.
7. B200 SPOT Stockholm entry (AWS, first listed Sep 14): $1.67/hr decomposed, rising from $1.79/hr naive at first collection to $1.91–$1.95/hr naive today — new supply entering the Nordic market directly below the H200 ON_DEMAND band.

Implied Action:

• EU compute buyers: Do not procure H200 ON_DEMAND through the hyperscaler ticker without first checking Nebius. At $4.30/hr vs $13.18/hr for the same chip, the delta is $8.88/hr per GPU — on an 8-GPU job running 2,000 hours, that is a $142,000 savings from reading one additional price source. The Nebius tier is the correct H200 ON_DEMAND reference price for sophisticated EU buyers today.
• Long EU GPU infrastructure (non-Azure): Nebius, Pantheon AI (1 GW approved Croatia campus, construction January 2027), and any operator colocating H200 DGX directly in European data centers faces a market paying $0.197/TFLOPS to Azure — compelling unit economics for 2026–2027 commissioning windows. The infrastructure return on EU H200 deployment is materially higher than US markets where competition has compressed margins.
• Avoid EU H200 ON_DEMAND lock-in: The Pantheon AI Croatia campus (Q1 2029 operational) and the B200 SPOT Stockholm entry signal that the EU H200 scarcity premium has a defined expiration date in the 18–30 month range. Reserved/committed H200 capacity at Azure EU rates initiated today would be repriced by a competitive market before the contract expires.
• Monitor: Croatian power grid interconnection filings (Pantheon AI timeline trigger), any second H200 provider appearing in ticker_prices_v2, and Nebius expanding H200 node count in Stockholm or Frankfurt.

H3 — H100 Stockholm SPOT Has Completed a Textbook Market Cycle and Is at Structural Buy Entry, Now Capped by a B200 Ceiling Confidence: 5 / 5

Thesis: The H100 SPOT Stockholm ticker has traced a complete market cycle in nine months with enough structural clarity to constitute a repeatable price pattern. From a single-provider $3.75/hr regime in December, through a price war floor of $1.72/hr in February, to a tight $2.57/hr equilibrium in May, a demand-surge peak of $4.20/hr on July 15, and now a controlled descent to $2.74/hr — the spread metric has been the leading indicator throughout. Spread peaked before price peaked and compressed before price bottomed. Today's 13.6% spread is the fastest compression rate seen since the original equilibrium formation in May, when both providers locked in at 2.9% spread and $2.57/hr. That equilibrium zone — $2.50 to $2.65/hr — is now approximately 5–10% below the current price, and the instance-level data confirms the mechanism: AWS's p4d/p5 spot floor in eu-north-1 has declined continuously from $27.8/hr (8-GPU node) on August 31 to $11.0/hr today, converging from below toward Azure's fixed-rate floor of ~$21.26/hr. The critical new development that changes the medium-term calculus entirely is the B200 SPOT Stockholm entry: AWS's p6-b200.48xlarge, first listed September 14 at $14.32/hr (8 GPUs = $1.79/GPU naive), has been live for five consecutive collection days with increasing intraday observations (4 → 25 → 25 → 25 → 12). At $1.67/hr decomposed, B200 SPOT Stockholm delivers frontier-generation silicon at a price that creates a performance-adjusted ceiling on what H100 can ever command in this market again. A chip with ~2x H100's memory bandwidth is now available on AWS spot for 39% less than current H100 spot. This is not a temporary listing — it is a structural shift in the Nordic supply profile.

Key Evidence:

1. 90-day H100 SPOT Stockholm history (Sep 18): Complete V-shape confirmed — $3.75/hr single-provider (Dec), $1.72/hr floor (Feb 21), $2.57/hr equilibrium (May 15, spread 2.9%), $4.20/hr peak (Jul 15), $2.74/hr today (spread 13.6%).
2. query_instance_price_history 45-day series (AWS eu-north-1, Sep 18): Floor min_price declined from $27.8/hr (8-GPU node) on Aug 31 to $0.74/hr today — 18 consecutive days of monotonic descent. Not noise; genuine real-time spot clearing.
3. Azure H100 SPOT in Stockholm: Flat at ~$21.26/hr node median throughout the 45-day series — confirming spread compression is AWS converging upward toward Azure's fixed floor, not Azure chasing down.
4. B200 p6-b200.48xlarge Stockholm first listed September 14 (5 consecutive collection days, observation count rising). Decomposed: $1.67/hr. Naive: $1.91–$1.95/GPU/hr. Stable and rising slightly (+7% over 5 days) — not a catalog artifact.
5. $/TFLOPS price performance (Sep 18): Stockholm H100 SPOT at $0.041/TFLOPS — cheapest H100 SPOT among all two-provider markets globally; -18.7% in 24h.
6. H100 ON_DEMAND Stockholm: $9.44/hr, 2 providers, 0% 30-day delta — stable anchor confirms the spot market's $2.74 is ~71% below on-demand, within historical norms (not an inverted market).
7. Model anomaly diagnostics (Sep 18): H100 Stockholm delta flagged as an anomaly in GCP Europe-North SPOT at -30.2% — but the AWS instance price history independently confirms the floor decline is real, not a modeling artifact.

Implied Action:

• Spot buyers with EU inference or fine-tuning workloads: Entry at $2.74/hr is 9% above the identified equilibrium floor of ~$2.50–$2.57/hr and 35% below the recent peak. The downside is bounded by a historically confirmed price floor; the upside is hard-capped by B200 SPOT at $1.67/hr. This is one of the cleaner asymmetric buy setups in the dataset — maximum loss from the entry point is ~10%, maximum gain vs the August peak is ~35%.
• Sophisticated spot bidders on AWS EC2 eu-north-1 specifically: The ticker median of $2.74/hr is a cross-provider construct. AWS's actual spot floor at the node level is now at $0.74/hr for p4d instances — sophisticated bidders clearing through the real auction may be achieving prices dramatically below the ticker. This is a structural advantage for any procurement team bidding directly into AWS spot vs using broker intermediaries that price to the ticker.
• Avoid H100 RESERVED or ON_DEMAND in EU Stockholm: At $9.44/hr, any reservation locked in today will look expensive by 2027 as B200 SPOT availability in the Nordics expands. The B200's September 14 entry is almost certainly a soft launch — AWS typically broadens SPOT availability as supply normalizes. H100 reserved EU pricing initiated today carries 12–24 month obsolescence risk.
• Trigger to close the position: Spread re-widening above 30% would signal one provider breaking upward again (the July pattern). Set an alert on the H100|SPOT|se-stockholm spread field.

H4 — The Inferentia SPOT Revival Is a Real Utilization Signal with a Distinctive APAC Signature, Not a Repricing Event Confidence: 3 / 5

Thesis: AWS Inferentia Gen 1 — the oldest custom ASIC in any hyperscaler catalog, launched December 2019 — is experiencing a genuine secondary-market utilization surge that is geographically concentrated in APAC and the Americas, and structurally absent in Europe. This is not AWS formally repricing the chip: the ON_DEMAND anchor in Virginia sits at exactly $0.521/hr with zero movement over any time horizon. What is happening is spot pool demand absorbing previously idle capacity. The geographic signature is the most analytically valuable data point: Virginia SPOT is up +101% over 30 days to $0.0805/hr, Tokyo is up +38.1% to $0.37/hr, Mumbai up +16.6% to $0.12/hr, Brazil up +10.2%. Meanwhile Frankfurt is down -23.3% and London down -12.9%. The Tokyo Inferentia spot discount has compressed to just 41% versus Virginia's 73% — meaning Japanese enterprises are bidding meaningfully more aggressively for Inferentia capacity than US customers. This APAC-specific demand concentration is consistent with a single mechanism: the deployment of open-weights MoE inference models with small active parameter counts (8–16B active) onto hardware originally optimized for static graph neural networks. The AWS Neuron SDK has accumulated MoE support updates since 2024; DeepSeek V4 Flash Vision Exp, released September 10 with MIT license and only 8B active parameters during prefill, is precisely the model profile that runs efficiently on Inf1's constrained memory with high throughput. Japanese enterprise adoption of Chinese open-weights models for cost-optimized inference — given the MoE architecture's efficiency on older ASICs — is the most plausible demand mechanism available that fits both the timing and the geographic signature.

Key Evidence:

1. Inferentia SPOT Virginia: $0.0805/hr, +101% over 30 days, from a trough near $0.001/hr in February 2026. Describes a complete V-shape recovery (live ticker, Sep 18).
2. Inferentia ON_DEMAND Virginia: $0.521/hr, 0.0% delta across all horizons (24h, 7d, 30d) — confirms this is a utilization event in the spot pool, not a formal AWS pricing action.
3. APAC geography: Tokyo Inferentia SPOT $0.37/hr (+38.1% 30d); Mumbai $0.12/hr (+16.6%); Brazil $0.06/hr (+10.2%). European markets: Frankfurt -23.3%, London -12.9%. Directional split is unambiguous (cross-provider ticker, Sep 18).
4. Tokyo spot discount: 40.8% vs Virginia's 72.9% — Japanese customers bidding materially above global average for the same ASIC, suggesting localized supply constraint or inelastic demand.
5. CATALOG_SURVIVAL depreciation data: Inferentia Gen 1 (Dec 2019, 6.8 years old) remains active in AWS catalog — the longest-lived hyperscaler custom ASIC. Active listing plus rising spot price = platform not in end-of-life mode.
6. DeepSeek V4 Flash Vision Exp (model tracker, released Sep 10): 763B total parameters, 8B active during prefill / 16B during decode via MoE. MIT license. $0.15/M input on OpenRouter. This model profile fits within Inf1 memory constraints with high inference throughput — the structural mechanism for ASIC demand.
7. What it is NOT: Montreal +1,155% is a near-zero base artifact from a single catalog repricing. Virginia ON_DEMAND flat rules out formal repricing. European declines rule out global surge. This is a geographically specific utilization signal.

Implied Action:

• Inference platform operators running open-weights MoE models under 20B active parameters: The cost differential between AWS Inferentia spot ($0.08/hr in Virginia, $0.37/hr in Tokyo) and H100 SPOT ($2.74/hr Stockholm, ~$8.69/hr US on-demand) is a 30–100x factor at current prices. For high-throughput, latency-tolerant inference on well-quantized MoE models, the Inferentia ASIC deserves formal benchmarking against the assumed H100 default. Any inference startup not evaluating this stack is burning unnecessary GPU budget.
• Watch Inf2 SPOT as the structural confirming signal: Inferentia Gen 2 (Inf2) carries 4x the accelerator memory, supports larger model weight files, and covers the 70B+ active-parameter range that Gen 1 cannot. If Inf2 SPOT activates in APAC with the same geographic signature — rising in Tokyo and Mumbai while European markets lag — the MoE-to-ASIC thesis graduates from a utilization observation to a structural hardware shift.
• Caution on scale: Inferentia's absolute price levels ($0.08–$0.37/hr) mean the dollar-value signal is tiny relative to H100 or B200 markets. This is a leading indicator, not a macro GPU trade. The inference hardware bifurcation thesis (ASIC for commodity MoE inference, frontier GPU for training and frontier serving) is directionally right but measured in quarters, not months.
• Deny signal to watch: If Virginia Inferentia SPOT falls back below $0.02/hr within 30 days, the utilization surge was a single large customer terminating a batch job. A sustained floor above $0.05/hr through October would confirm secular demand.

H5 — B200 SPOT Is Bifurcating into Two Structurally Different Markets: AWS-Anchored Hyperscaler Stability vs Neocloud Emerging-Market Collapse Confidence: 3 / 5

Thesis: The B200 SPOT market has emerged globally across nine regions, but the pricing trajectories of those markets are diverging along a single structural fault line: whether the market is anchored by AWS's genuine real-time spot auction or by a standalone neocloud operator setting list prices on web infrastructure. AWS-anchored B200 SPOT markets — Ohio ($4.91/hr), Oregon ($5.34/hr), Virginia ($4.84/hr), and now Stockholm ($1.67/hr) — show modest, stable movement of 1–2% over 30 days, with confirmed spot discounts of 63–64% off on-demand references. The neocloud markets are a different animal entirely. Hyderabad, where a single operator's B200 SPOT price declined in a perfectly linear fashion from approximately $7.40/hr on September 4 to $3.58/hr on September 17 before the listing vanished from today's snapshot entirely, represents the most acute example of frontier GPU oversupply at the neocloud layer. Jakarta B200 SPOT at $2.87/hr is down -24% over 7 days. The Hyderabad collapse — -52% in 14 days before the listing was pulled — is the fastest frontier GPU price collapse in the dataset and carries the hallmark of an operator that over-provisioned capacity in anticipation of India-market demand that has not materialized at the expected price point. Stockholm's simultaneous entry at $1.67/hr (AWS, five collection days, rising) is the structural inverse: a hyperscaler bringing B200 SPOT online as genuine auction capacity, immediately establishing a lower price point than any other market in the world — not because demand is weak, but because AWS's spot pricing mechanism clears at the marginal utility value of the capacity, which for brand-new hardware is substantially below the neocloud markup. Note that Hyderabad's instance-level data is absent from today's snapshot — the neocloud has likely pulled or reset its listing — so the specific endpoint of the Hyderabad leg requires reconfirmation before trading on it.

Key Evidence:

1. B200 SPOT Stockholm (AWS p6-b200.48xlarge): First listed September 14, five consecutive collection days, observation count rising (4→25→25→25→12). Price $14.32/hr at first collection, $15.3–$15.6/hr range today ($1.67/hr decomposed, $1.91–$1.95 naive). Rising, not falling — genuine spot demand absorption (instance price history, Sep 18).
2. US B200 SPOT markets (Ohio $4.91/hr, Oregon $5.34/hr, Virginia $4.84/hr): 30-day change +1–2%, spot discounts 63–64% confirmed. AWS hyperscaler-anchored pricing regime, stable (live tickers, Sep 18).
3. Hyderabad B200 SPOT: 14-day history showed a perfectly linear daily decline from $7.40/hr (Sep 4) to $3.58/hr (Sep 17). Instance snapshot absent on Sep 18 — neocloud listing pulled or collection gap. Endpoint requires next-day reconfirmation.
4. Jakarta B200 SPOT: $2.87/hr, -24% over 7 days, -31% over 30 days. Single provider, no hyperscaler involvement — independent market read only.
5. B200 Hyderabad/Stockholm inversion: $3.58 (neocloud, falling, now absent) vs $1.67 (AWS, rising) for the same GPU generation — 115% gap with directionally opposite momentum. No NatEx pairs to validate.
6. GCP Hong Kong RAM SPOT: +79% delta in the anomaly diagnostics — independent signal of APAC compute market pricing stress consistent with a supply/demand dislocation in the region.
7. US NAND/SSD imports (HS 852351, Jul 2026): +499% YoY, 3-month momentum +75.3% — storage pre-positioning consistent with datacenter wave arriving 2027–2028, suggesting B200 supply is loosening ahead of that wave.

Implied Action:

• Immediate buy for memory-intensive workloads: Stockholm B200 SPOT at $1.67/hr decomposed delivers 76x more HBM per dollar than H100 SPOT Stockholm at $2.74/hr (8TB B200 HBM vs 80GB H100 HBM). For memory-bound training workloads that can tolerate interruption, Stockholm B200 SPOT is currently the best unit-economics computing surface on earth by this metric. Monitor for daily before it widens to other EU regions.
• Neocloud APAC B200 investors: risk flag. The Hyderabad collapse pattern — a single operator setting and then abandoning a pricing curve that lost 52% of its value in 14 days — is the textbook signal of overprovisioned capacity meeting demand shortfall. Neoclouds with heavy APAC/South Asian B200 deployment who are not yet at the AWS-anchored price tier are exposed to further capitulation as AWS B200 SPOT expands geographically.
• NVDA ASP risk (long-horizon): B200 scarcity, assumed to sustain through end-2027, appears to be eroding ahead of schedule. If AWS can offer B200 SPOT in Stockholm at $1.67/hr — implying on-demand reference of roughly $5/hr — then the Blackwell ASP premium built into sell-side NVDA models needs revisiting. Not a 2026 earnings event, but a 2027 forward estimate risk.
• Reconfirmation needed: Pull query_instance_price_history(accelerator='B200', region='IN_HYDERABAD', days=30) on September 19 to determine whether the Hyderabad listing reappears at a reset price or has been permanently withdrawn. The answer changes the interpretation from "finding a floor" to "operator capitulation."

Inventory & Channel Health

The full DIO/inventory stack returned no data for this cycle, and purchase commitment tables are similarly empty — leaving the channel health analysis to be inferred from trade flow and macro indicators alone, which limits precision. That said, the trade signal that does exist is directionally clear and consequential: US NAND/SSD imports (HS 852351) are running at +499% year-over-year for the July 2026 read, with a 3-month momentum of +75.3% (RISING signal). Broad integrated circuit imports (HS 8542) are essentially flat at -13% month-over-month, decelerating. This divergence — storage surging while chips flatten — is the fingerprint of operators pre-staging storage infrastructure ahead of datacenter wave completions rather than absorbing a new chip shipment cycle. It is storage-side channel fill, not GPU glut. The implication for the B200 neocloud collapse observed in Hyderabad is subtle but important: if operators are pre-staging SSDs for 2027–2028 rack deployments, the B200 units being over-provisioned by neoclouds today were likely sourced on the expectation of colocation demand that won't materialize until datacenter builds complete. The flash import surge is corroborating evidence that the datacenter pipeline is real, but the timing mismatch between hardware procurement and facility readiness is creating inventory pressure at the operator level — visible in B200 neocloud pricing behavior — even as no formal DIO signal exists in the dataset today.

Supply & Demand Pipeline

Supply — The Nordic Market Is Opening Up Fast The most immediate supply-side shift is AWS's B200 SPOT entry in Stockholm on September 14. This is not just a price event — it is the first evidence that AWS's Blackwell (p6 family) deployment has reached European availability with genuine spot auction mechanics. Five consecutive collection days with rising observation counts (4→25) indicates the pool is growing, not contracting. At the same time, Project Jupiter (Oracle/OpenAI, Santa Teresa NM, 2,450 MW under construction) carries an active legal risk: the New Mexico Supreme Court paused permit proceedings over water usage after the project well pumped 103 million gallons, and the natural gas pipeline is delayed to February 2027. A construction stoppage at the largest single AI campus in the dataset would directly impact CoreWeave's compute roadmap and OpenAI's capacity commitments — material to the trajectory of H100 and H200 demand in the US market.

Supply — EU Sovereign Capacity on the Horizon Pantheon AI's 1 GW Croatia campus (€50B, approved, construction starting January 2027, commercial Q1 2029) is the single most consequential EU supply event in the tracker. It has a defined timeline, specific geography, and would introduce non-Azure H200-class capacity into the EU market at scale. If delivered on schedule, it terminates the H2 hypothesis's medium-term thesis by 2029. The Stargate Michigan campus (Oracle/OpenAI, 1,400 MW, Building 1 nearly complete targeting early 2028, DTE Energy 19-year power contract approved) is the nearest-term large US completion signal — relevant to H100/H200 demand anchoring in the 2027–2028 window.

Demand — MoE Architecture Is Reshaping Inference Hardware Economics DeepSeek V4 Flash Vision Exp (released September 10, 763B total parameters, 8–16B active via MoE, MIT license, $0.15/M input) is the specific model driving the Inferentia SPOT revival hypothesis. Its MIT license makes it self-hostable on any hardware including Inf1; its sub-20B active parameter count fits within Inf1's constrained memory; and its APAC adoption explains the geographic signature (Tokyo +38%, Mumbai +17%) in the Inferentia SPOT data. GPT-5.6's simultaneous 20% token price cut (released September 1, OpenAI) confirms that API-level competition is forcing the entire inference stack to find cheaper silicon — Inferentia is the most visible hardware-side response. These two model events are working in the same direction: compressing the value of frontier GPU hours for inference workloads while expanding the utilization of sub-$0.10/hr ASIC capacity.

Risk Flags

Immediate (0–30 Days)

R1 — B200 Hyderabad Listing Absence Unresolved The most pressing data-integrity issue: B200 SPOT Hyderabad returned zero instance rows in today's snapshot after showing a perfectly linear -52% price decline over 14 days. The neocloud operator (likely Hyperstack based on India footprint) has either pulled the listing to reset pricing or the collection cycle missed it. Until tomorrow's snapshot, the endpoint of the Hyderabad collapse is unverified. A listing reappearing at a new, lower price floor (below $2/hr) would confirm genuine demand shortfall and neocloud capitulation — the most bearish read for India/APAC frontier GPU infrastructure investment. A listing not reappearing would indicate the operator has exited the India B200 spot market entirely, which is more bullish for remaining operators' pricing power.

R2 — Tokyo A100_80GB Repricing Has No SKU-Level Attribution The A100_80GB SPOT Tokyo ticker shows a 222% unilateral repricing from one of two providers since August 17 — but neither the AWS raw pricing tables nor the instance price history contain a matching SKU in Tokyo SPOT. The repricing is from a neocloud not covered by the hyperscaler SKU tables. Until the specific operator and instance can be identified, the hypothesis of "inelastic demand causing unilateral repricing" cannot be separated from "a catalog scraping artifact at a stale neocloud web listing." The spread widening (now 225%) rather than converging is the most important near-term signal — if the floor provider raises prices to follow within 10 days, the demand thesis holds. If it stays flat, the high-price observation was noise.