Daily Investment Brief — September 22, 2026

Market Pulse

The US compute market is undergoing a sub-market decoupling between competitive and monopoly-provider regions that has not been this pronounced since the initial H100 launch cycle. The defining dynamic today is not a demand surge — it is competitive exit: in Iowa, Illinois, Nevada, and Texas, H100 on-demand markets have consolidated to a single provider at $10.43–$11.52/GPU-hr, while the last genuinely competitive US market, Ohio, holds a multi-provider anchor at $7.73/hr with a $5.97 AWS floor still accessible. That $3.55–$5.55/hr spread between the competitive anchor and monopoly-provider markets — confirmed by 15 observations in Ohio versus a single zero-spread observation in Iowa — is the single most actionable pricing signal in the dataset today. Layered on top, two exogenous shocks are compressing the future supply pipeline simultaneously: Texas has frozen new data center permitting across ~30% of US capacity applications, and California's SB 886 (signed September 21) adds structural cost floors for any interconnection executed after January 1, 2027. The compute market is not in a demand-pull inflation — it is in a supply-side fragmentation where geography and provider concentration are the dominant pricing variables.

Key Movers

Thin-market noise: All APAC L4/A100 SPOT moves are GCP administrative spot discount adjustments (fixed estimates, not genuine market bids — n_providers=1, no matched-pair NatEx). TPU moves are GCP catalog restructuring. Disregard for demand inference. The AWS Inferentia Oregon/Montreal and B200 Seoul signals are genuinely market-sourced.

Investable Insights

H1 — The Midwest H100 Monopoly-Provider Fracture: Ohio Is the Last Competitive Anchor Confidence: 4.5 / 5

Thesis: Iowa's H100 on-demand repricing to $11.52/GPU-hr (+94% over seven days, from a stable $5.98–$6.00 baseline maintained since January) was not a list-price increase — it was a provider exit event. The raw instance collection returned zero H100 Iowa records on September 22, meaning the ticker is being held from the last observation before a cheap provider — almost certainly GCP or a neocloud pricing at ~$6/GPU-hr — vacated the Iowa market. Simultaneously, Illinois has held at $11.28 for 30+ days (single provider, zero spread, no compression), and Nevada ($10.73) and Texas ($10.43) are similarly consolidated. The result is a Midwest H100 on-demand market that looks like a $10–12/hr commodity — except Ohio, where AWS and Azure together produce 15 observations at a $7.73 median with a $5.97 AWS floor and a 59% spread confirming genuine competition. The Iowa catalog exit is being physically reinforced by QTS Cedar Rapids (7-building Blackstone campus, 4,500 construction workers/day, wind-powered) and Google Cedar Rapids consuming available local grid interconnection capacity, creating a structural barrier to new providers entering Iowa H100. The mechanism for Ohio to follow the Iowa/Illinois pattern exists: if either AWS or Azure reduces their Ohio H100 SKU coverage, the ticker snaps upward by $3–4/hr with no demand change required.

Key Evidence:

1. H100 Iowa 90-day history: stable at $5.98–$6.00 from January 1 through September 15, then abrupt break to $11.52 — a clean regime change, not noise (Signwl ticker history, Sep 22, 2026)
2. Iowa ticker: n_providers=1, spread=0%, zero raw instances in the September 22 collection — confirms provider exit, not repricing (Signwl raw instance query, Sep 22, 2026)
3. Illinois H100 at $11.28 has been stable for 30+ days at single-provider, zero-spread — confirming monopoly pricing can persist indefinitely without competitive correction (Signwl ticker, Sep 22, 2026)
4. Ohio H100: 2 providers, 15 observations, 59% spread, $5.97 AWS floor vs $7.73 median — the last US Midwest competitive pricing market (Signwl price_performance table, Sep 22, 2026)
5. QTS Cedar Rapids and Google Cedar Rapids are active under-construction hyperscale campuses absorbing Iowa power interconnection capacity — structural barrier to new provider entry (Signwl datacenter tracker, Sep 22, 2026)
6. Texas permit freeze diverts future demand to Indiana/Illinois, increasing load on the remaining Midwest competitive markets (Governor Abbott freeze order, per news feed, Sep 22, 2026)
7. GCP anomaly diagnostics flag H100 us-west ON_DEMAND at -70% in 24h — same provider exit pattern in a different region, corroborating Iowa diagnosis (Signwl anomaly_diagnostics, Sep 22, 2026)

Implied Action:

• Immediate compute buyer action: Secure Ohio H100 on-demand or 1-year reserved capacity NOW. The AWS floor at ~$5.97/hr represents a $5.55/hr discount to Iowa's current monopoly price — the most asymmetric compute hedge available in the US market today. Every month Ohio holds competitive pricing is time to lock exposure.
• Monitor trigger: Any PRICE_CHANGE alert on AWS p5.48xlarge or Azure NDv5 in Ohio is the earliest signal of provider exit from the competitive anchor. Set a watch on the H100|ON_DEMAND|US_EAST_OHIO ticker for n_providers dropping to 1.
• Real estate / infrastructure: Datacenter operators with approved Indiana capacity (Amazon New Carlisle 2,250MW, Joliet Technology Center 1.8GW) are valuation beneficiaries if the $10–12/hr Midwest floor persists — those assets clear at materially higher GPU rents than the $7.50/hr mid-2026 base case embedded in development pro formas.
• Avoid anchoring off Iowa: $11.52 is a monopoly price from a thin market. For benchmarking purposes, Ohio ($7.73) is the correct US Midwest reference rate.

H2 — Texas Permit Freeze Creates a Two-Tier Intrastate Asset Market: Interconnection-Executed Projects Become Scarce Confidence: 4.0 / 5

Thesis: Texas's permit freeze — which the state processes roughly 30% of all US data center permit applications through — does not affect projects uniformly. The freeze creates a clean binary: projects with executed utility interconnection agreements and construction underway are legally immune; projects still in permitting are effectively frozen for 12–24 months given the political dynamics. The market has not yet priced in which Texas projects fall in which bucket, and the information asymmetry is significant. The immune cohort is well-documented in the tracker: Hut8's Beacon Point in Corpus Christi ($4.25B financing closed June 2026, AEP Texas interconnection explicitly executed, substation construction commenced, 352MW operational expanding to 700MW by Q1 2027), Meta El Paso ($14B, BlackRock JV, 2,300 workers on-site, El Paso Electric building a dedicated 366MW gas plant for 2027 delivery), Google's Meitner Energy Center in Pampa (840MW, $14B, off-ERCOT power-first design), and IREN Sweetwater 2 (600MW, Nvidia $2.1B commitment, Microsoft tenant). Together these represent approximately 3–3.5GW of freeze-immune capacity. The exposed cohort includes the Lancium/Crusoe Childress MegaCampus (1GW, Oracle/OpenAI anchor, construction "scheduled to start Q3 2026" — if construction had not commenced before the freeze, this is the most consequential single project at risk). Losing 1GW of Stargate-adjacent Texas capacity from the 2027–2028 delivery schedule tightens the national H100/B200 absorption pipeline and directly supports above-$8/hr on-demand pricing nationally.

Key Evidence:

1. Texas H100 ON_DEMAND ticker at $10.43/hr, 1 provider, flat over 30 days — existing supply is stable; the freeze's price impact arrives 12–18 months out as approved-but-not-energized projects fail to deliver (Signwl ticker, Sep 22, 2026)
2. Beacon Point/Hut8 Corpus Christi: AEP Texas interconnection executed, $4.25B financing closed June 2026, substation under construction — explicitly freeze-immune (Signwl datacenter tracker, confidence 0.85, Sep 22, 2026)
3. Meta El Paso: BlackRock JV, 2,300 workers on-site, El Paso Electric dedicated gas plant under construction for 2027 — freeze-immune (Signwl datacenter tracker, Sep 22, 2026)
4. Lancium/Crusoe Childress MegaCampus: 1GW, Oracle/OpenAI anchor, construction "scheduled to start Q3 2026" — ambiguous freeze exposure; confidence 0.75 (Signwl datacenter tracker, Sep 22, 2026)
5. Texas processes ~30% of US permit applications; freeze creates 12–24 month supply gap for projects not yet past interconnection stage (news feed, Sep 22, 2026)
6. California SB 886/AB 2383 (signed Sep 21, 2026) adds Jan 1, 2027 cutoff for new interconnection cost allocation — Georgia, Indiana, and Louisiana become the primary alternatives, not new Texas entrants (California legislative record, per intel feed, Sep 21, 2026)
7. ERCOT Large Load IROL documents (filed Sep 16–17) model mitigation plans for identified large-load groups — the grid operator is already gating new interconnection independently of the permit freeze (Signwl intel feed, Sep 16–17, 2026)

Implied Action:

• Long operational Texas data center operators with locked interconnection: IREN (Sweetwater 1+2), Hut8 (Beacon Point + River Bend), and Nscale (Ward County, already partially operational) hold scarcity option value that was not priced 30 days ago. Monitor Hut8's Q1 2027 energization announcement for the clearest catalyst.
• Long Indiana infrastructure beneficiaries: AEP (serving both New Carlisle and Texas camps) and Duke Energy Indiana benefit from demand re-routing. The utility with the signed interconnection agreement for Amazon's New Carlisle 2,250MW campus sees multi-year large-load revenue certainty without freeze risk.
• Avoid companies with unconfirmed Texas construction start: CloudBurst's 1.2GW San Marcos campus (approved but construction start unconfirmed) is the clearest avoid. Verify construction commencement status before taking any position.
• Watch for: Louisiana Holly Ridge 2GW (approved, operator undisclosed, no construction activity) accelerating — it becomes the highest-optionality approved project in the country if Texas and California simultaneously tighten.

H3 — AMD MI300X: The Hike Is Real, But the Compute-Market Story Is Slower and Narrower Than the Equity Signal Suggests Confidence: 3.0 / 5

Thesis: AMD's announced ~10% Q4 2026 AI accelerator price hike is confirmed by multiple sources and its equities impact is already priced in — AMD crossed the $1 trillion market cap threshold and sits at $615/share, +20% over the past month. But the compute-market propagation of this hike is both structurally slower and geometrically narrower than the headline implies. Three constraints limit the impact: First, the hike has not touched a single SKU — every MI300X ticker in the dataset shows exactly 0.0% delta across all regions and all time horizons through September 22. Second, the MI300X public cloud market is an Azure monopoly: every MI300X ON_DEMAND instance in the raw collection is a single Azure SKU (Standard_ND96is_MI300X_v5 or Standard_ND96isr_MI300X_v5), with no AWS or GCP equivalents and no neocloud MI300X catalog in 60 days of collection. Third, a 10% increase on Azure's $9.60/hr instance price moves the decomposed per-GPU price from ~$3.23 to ~$3.55 in Ohio — still 69% below the Iowa H100 price. The MI300X/H100 $/TFLOPS-hr gap closes from 5.7× to 5.3× — directionally correct, economically immaterial for decision-making. The genuine investable content of the AMD story is in equities, not compute pricing. However, the confirmation of AMD's 107% YoY data center revenue growth (to $6.7B, now 58% of total revenue) and the $60B Meta supply agreement mean that AMD's revenue visibility through 2027 is exceptionally high — the hike adds pricing power on top of already-strong volume.

Key Evidence:

1. AMD price hike: ~10% on AI accelerators, Radeon GPUs, and motherboard chipsets, Q4 2026 start — confirmed by multiple independent sources including AD HOC News (Sep 21) and Chosun (Sep 22, 2026)
2. All MI300X tickers: 0.0% delta over 24h, 7d, and 30d — hike has not propagated to any SKU as of Sep 22 (Signwl ticker, Sep 22, 2026)
3. Raw instance collection: 100% Azure (Standard_ND96is_MI300X_v5, 8× GPU, $9.60/hr OH/CA, $11.808/hr Iowa, $13.44/hr Stockholm) — no AWS or GCP MI300X instances exist (Signwl raw instances, Sep 22, 2026)
4. Hyperstack MI300X: zero 60-day price history — neoclouds have not deployed MI300X clusters at commercial scale (Signwl instance price history, Sep 22, 2026)
5. $/TFLOPS-hr gap: MI300X $0.0198–$0.0249 vs H100 $0.115–$0.172 — 4.6–8.7× performance-normalized discount that a 10% hike reduces by only ~5% in absolute terms (Signwl price_performance table, Sep 22, 2026)
6. AMD data center revenue: $6.7B in Q2 2026, +107% YoY, 58% of total revenue; $1T market cap milestone reached this week; analyst target $616 with PEG ratio 0.57 (Signwl equities feed, Sep 22, 2026)
7. Azure Stockholm MI300X at $13.44/hr (8 GPU) — more expensive than Ohio at $9.60/hr — disproves any European cost advantage at the hyperscaler layer (Signwl raw instances, Sep 22, 2026)

Implied Action:

• Long AMD equity into Q4 earnings (November 3): The hike signals pricing power; Meta $60B supply agreement provides revenue certainty; PEG of 0.57 remains attractive despite the $1T cap milestone. The equities market leads the compute market by 30–60 days here, and the equity thesis is well-supported.
• Compute buyers: no urgency on RI hedging. The RI premium over on-demand in some geographies (Stockholm ON_DEMAND at $4.28 vs RESERVED_1YR at $4.40) is slightly inverted — do not lock in reservations based on hike timing. Wait for the first Azure price alert on ND-v5 SKUs before acting.
• Watch: Azure price alerts on Standard_ND96is_MI300X_v5 or Standard_ND96isr_MI300X_v5 in centralus (Iowa) and northcentralus (Illinois) — those premium-priced regions are where the hike propagates first and most visibly (current: $11.808/hr for the 8-GPU instance in Iowa).
• Neocloud MI300X gap: The absence of any neocloud MI300X catalog after 60 days of collection means there is no competitive check on Azure's MI300X pricing. AMD's monopoly distribution through Azure on public cloud is a risk concentration that compute buyers should note.

H4 — APAC Blackwell: Seoul Premium and Jakarta Stability Confirm Regional Inference Demand Is Real Confidence: 3.5 / 5

Thesis: The APAC GPU pricing signals are a mixed bag — the L4 and A100 SPOT surges in Taipei, Seoul, and Tokyo are GCP administrative catalog adjustments (fixed spot discount estimates, not demand bids) and should be discarded. But two B200 SPOT data points survive that scrutiny and are genuinely market-sourced. AWS's p6-b300.48xlarge in Jakarta (ap-southeast-3) has been collected at a stable $23.09/hr for an 8-GPU instance ($2.89/GPU naive, $2.71/GPU decomposed) across multiple collection intervals — this is real hyperscaler spot pricing, and it is the only hyperscaler-sourced B200 SPOT listing in the entire global dataset. Simultaneously, the neocloud B200 SPOT in Seoul (kr-seoul) has repriced from $3.94 to $5.97/GPU-hr over 30 days (+51.5%) — the only B200 SPOT market priced above US neocloud spot levels ($4.86–$5.35). Together these two data points tell a coherent story: AWS launched Blackwell spot capacity in APAC first (Jakarta), at below-US pricing, while independent APAC neocloud operators in Korea are pricing B200 at a demand premium to the US. This is the Blackwell APAC inference infrastructure being monetized ahead of a structural demand shift driven by the real-time voice AI model wave. GPT-Live-1 ($0.05/min, full-duplex voice, OpenAI, September 10), Gemini 3.8 Live (97-language simultaneous speech, $0.75/M tokens, September 15), and Qwen3.8-LiveTranslate (60-language real-time interpretation, Alibaba, September 18) all impose hard latency requirements — sub-100ms round-trip at the model layer — that make US-to-APAC inference physically nonviable at scale. Seoul-to-US ping alone exceeds 130ms. Regional GPU capacity is not optional for these workloads; it is a physics constraint.

Key Evidence:

1. AWS p6-b300.48xlarge Jakarta SPOT: $23.093/hr for 8 GPUs ($2.886/GPU naive), stable across 10+ hourly collection intervals — the only hyperscaler B200 SPOT instance in the global dataset (Signwl raw instances, Sep 22, 2026)
2. B200 SPOT Seoul: $5.97/GPU-hr, +51.5% over 30 days from $3.94 — the only B200 SPOT market above US neocloud levels of $4.86–$5.35 (Signwl ticker, Sep 22, 2026)
3. B200 SPOT Jakarta decomposed ticker: $2.71/GPU-hr — 44–50% below US B200 SPOT; confirms AWS is pricing APAC Blackwell spot aggressively to seed the market (Signwl price_performance model, Sep 22, 2026)
4. AWS Inferentia SPOT: +124% over 30 days in Oregon, +122% in 7 days in Montreal — real-time voice inference is already absorbing specialized inference silicon in the US, with APAC migration lagging by one structural step (Signwl ticker, Sep 22, 2026)
5. A100-40GB Iowa SPOT: -60.6% over 30 days — legacy inference silicon vacating the same geography where H100 reprice upward; confirms generational rotation (Signwl ticker, Sep 22, 2026)
6. Alibaba announced global datacenter expansion and new AI chip development on September 22 (stock +3%) alongside Qwen3.8-LiveTranslate GA — the largest non-US hyperscaler is simultaneously building APAC inference supply and its own silicon to serve it (Signwl news feed, Sep 22, 2026)
7. L4/A100 APAC SPOT "surges" are GCP administrative repricings (n_providers=1, fixed spot discount, no matched-pair NatEx) — these do NOT corroborate demand thesis and are excluded (Signwl anomaly_diagnostics, Sep 22, 2026)

Implied Action:

• For hyperscaler investors: APAC segment revenue in AWS Q3 earnings (early November) should outperform consensus driven by Blackwell spot absorption — watch Jakarta ap-southeast-3 regional contribution specifically. If Jakarta B200 SPOT prices move above $28/hr ($3.50/GPU), absorption is confirmed.
• NVIDIA remains the primary beneficiary: Seoul neocloud buyers paying $5.97/GPU-hr for B200 at a premium to US spot confirms APAC willingness to pay above the US floor for Blackwell-generation inference. NVDA at $227 (analyst consensus target $328, +44% implied upside) is the highest-conviction beneficiary across every hypothesis in this brief.
• Monitor Jakarta B200 SPOT stability: The AWS listing has been flat at $23.09/hr across multiple collection intervals. Any move above $28/hr signals absorption; any disappearance signals catalog restructuring. This is the single cleanest real-time inference demand signal available.
• Watch Alibaba (BABA): The simultaneous September 22 announcements — Qwen3.8-LiveTranslate GA, new AI chip development, and global datacenter expansion — are the components of a vertical integration play for APAC inference that could shift GPU demand away from US-sourced hardware within 18–24 months.

Inventory & Channel Health

The composite inventory regime returned no DIO records this cycle, preventing a formal composite-regime read. However, the trade-flow data provides a standalone signal that is significant enough to warrant inclusion.

US flash/SSD imports (HS 852351) came in at $7.08B in July 2026 — a +499% year-over-year surge with +75% three-month momentum, rated RISING. This is running at a level inconsistent with routine consumer or enterprise SSD procurement. By contrast, the broad IC import signal (HS 8542) is flat, down 13% month-over-month despite +40% YoY — the flash surge is category-specific. The most plausible explanation is AI infrastructure procurement: petabyte-scale NVMe arrays for training checkpoint storage, vector database backing, and inference KV-cache at projects like Project Jupiter (Santa Teresa, NM, 4GW planned) and Amazon New Carlisle (Indiana, 2,250MW). A Stargate-scale training cluster consumes several petabytes of NVMe in its initial fit-out, and the July timing aligns with Q3 2026 infrastructure procurement cycles for projects targeting 2027 energization.

The channel health risk is the absence of distributor DIO reads. Without ARW/SNX/AVT DIO data, there is no early-warning confirmation that this flash import surge is landing into genuine consumption versus building distributor inventory. The next Arrow Electronics and Synnex earnings reports are the highest-priority data events for resolving this ambiguity. A rising distributor DIO in Q3 against accelerating flash imports would confirm channel stuffing risk; flat or falling DIO would confirm real demand absorption.

Supply & Demand Pipeline

Supply — Two Projects Shifting the National Pricing Floor

The Saline Township, Michigan Barn campus (The Stargate / Oracle / Blackstone structure, 1.4GW planned, $56B commitment, 19-year DTE contract, first building near complete) is the most consequential near-term supply event in the US Midwest. Targeting first operations in early 2028, with 700+ tradespeople already on site, this campus will deliver the first large-scale Blackwell training cluster in the Midwest. When it comes online, it directly adds supply pressure to the region currently priced at Iowa/Illinois monopoly rates of $11–12/hr. The DTE contract with a pre-January 2027 interconnection execution date makes this explicitly grandfathered under California SB 886 — the project holds structural cost immunity that newer competitors cannot replicate.

Project Jupiter in Santa Teresa, NM (4GW planned, $165B, Stargate/Oracle, 26% complete) remains the most consequential supply variable nationally, and it is now materially at risk. The New Mexico Supreme Court paused permit proceedings over water usage — 103 million gallons pumped in April–August — and the enabling natural gas pipeline has slipped to February 2027. A 6–12 month delay on the nation's largest planned AI training cluster is not a rounding error for national H100/B200 absorption curves.

Demand — The Real-Time Voice Architecture Shift

The September 10–22 model release window produced a coherent new inference workload archetype: full-duplex, sub-100ms voice inference at global distribution. GPT-Live-1 (OpenAI, September 10), Gemini 3.8 Live (Google DeepMind, September 15), and Qwen3.8-LiveTranslate (Alibaba, September 18) are not incremental — they represent a workload that was not served by any cloud GPU product twelve months ago. Latency-constrained, memory-bandwidth-limited, and geographically distributed by necessity, real-time voice inference drives demand for inference-class silicon (L4, A10G, Inferentia) at regional proximity rather than large H100 clusters. This architectural shift is the mechanism behind the AWS Inferentia SPOT repricing in Oregon (+124% over 30 days) and Montreal (+122%) — not a training demand surge.

DeepSeek V4 Flash Vision (MIT license, 763B parameter MoE, 8–16B active parameters, $0.15/M input tokens via OpenRouter and 13 channels) is the model most likely to generate neocloud H100 demand at current pricing. At $0.15/M tokens and a compute-efficient architecture, the API is cheaper than self-hosting on Iowa H100 at $11.52/hr for most inference volumes — but the model's openness means any operator with cheap spot capacity (Oregon H100 at $5.96, Ohio at $5.97 AWS floor) can undercut the API at scale.

Component — AMD's TSMC Wafer Cost Increase Enters the Supply Chain

AMD's 10% AI accelerator hike is explicitly attributed to higher TSMC wafer costs. TSMC's N4/N5 process node — shared between AMD's MI300X (CDNA 3, N5) and NVIDIA's H100 (Hopper, N4) — is facing pricing pressure from the prioritization of data center over consumer, confirmed by NVIDIA's RTX 60 "Rubin" gaming GPU delay to 2028. With NVIDIA targeting $20B in Vera Rubin data center revenue in Q3 2027, the N4/N5 wafer allocation is structurally tilted toward training silicon through at least 2027. AMD's hike is the first visible pass-through of this wafer cost increase into cloud pricing; NVIDIA's own Blackwell pricing pressure should follow, though NVIDIA's pricing power is structurally higher and the pass-through may be absorbed in margin rather than list price.

Pre-Announced Price Changes

No UPCOMING_PRICE_CHANGE alerts are in the AWS, GCP, Azure, or Oracle pipeline as of September 22, 2026. The absence of pre-announced changes means the Iowa H100 repricing was not a scheduled tariff move and is consistent with a catalog restructuring (provider exit), not a communicated list-price increase. The AMD Q4 hike, while publicly announced, has not been encoded as an UPCOMING_PRICE_CHANGE alert in any provider's API — meaning it has not yet translated into a specific instance-level SKU change notification. The next 7-day alert window is the first opportunity to see how Azure encodes the ND-v5 refresh into its pricing API.

Risk Flags

Immediate (0–30 Days)

R1 — Ohio H100 Provider Concentration Risk Ohio is the last genuinely competitive US Midwest H100 market and the primary hedge for compute buyers who need US-East capacity below $8/hr. The AWS floor at $5.97 is accessible today with 15 observations and a 59% spread. But the Iowa pattern — stable for nine months, then abrupt single-day exit to $11.52 — shows that provider exits happen without pre-announcement. An AWS p5.48xlarge price alert in Ohio or an Azure NDv5 Ohio withdrawal would give 24–48 hours of warning at most. Buyers who rely on Ohio spot without any reserved-instance anchor are fully exposed to this scenario.

R2 — GCP Catalog Churn Contaminating APAC Inference Reads Both L4 Taipei (+117.7% over 30 days) and L4 Seoul (+100.7%) are GCP administrative spot adjustments — fixed discount estimates with n_providers=1 and no demand-bid mechanism. The anomaly diagnostics independently flag GCP GPU-type as a 34.95% anomaly rate dimension. Any analysis that treats these moves as demand signals is building on false premises. The risk is that GCP catalog churn continues to generate headline-level APAC spot "surges" that attract capital to thin-market positions before genuine multi-provider coverage develops.

R3 — B200 Stockholm Single-Observation Trap The B200 Stockholm SPOT ticker ($2.86, n_observations=1, n_providers=1, 7-day-old market) has generated significant apparent analytical interest as a European compute signal. The Azure MI300X Stockholm data definitively disproves any structural European hyperscaler cost advantage — Stockholm MI300X is 40% MORE expensive than Ohio, not cheaper. The B200 Stockholm signal is one neocloud's catalog entry. It could reprice to zero within 14 days with no market mechanism to prevent it. Do not build European GPU cost-advantage trades off a single observation with a seven-day history.

Near-Term (30–90 Days)

R4 — AMD MI300X Hike Propagation Uncertainty AMD's 10% hike is real and scheduled for Q4, but the MI300X public cloud market is an Azure monopoly with no neocloud alternatives at scale (zero Hyperstack/Lambda MI300X catalog in 60 days of collection). Azure's actual SKU repricing timing is unknown and has not been pre-announced. If Azure absorbs the TSMC wafer cost increase internally rather than passing it to ND-v5 list prices, the hike never enters the tickers — and the compute-market thesis of Hypothesis H3 fails even as the AMD equity thesis succeeds. Watch for Azure price alerts specifically on Standard_ND96is_MI300X_v5 through October.