Daily Investment Brief — September 17, 2026

SIGNWL COMPUTE INTELLIGENCE BRIEF

September 17, 2026 — Final Edition

Market Pulse

The GPU market is executing a sub-market decoupling that the headline narrative — "demand is insatiable, supply cannot keep up" — completely misses. Three independent markets are moving in opposite directions simultaneously. AWS's custom silicon (Trainium SPOT, Virginia) is in genuine pool exhaustion, with max bids hitting $21.50/hr against an on-demand list of $0.937/hr decomposed — a 100% spot-to-OD convergence in premium AZs that is the precursor pattern to a list-price increase. Meanwhile, H100 on-demand rates are flat to the basis point in US East (-0.05% over 30 days) and spot discounts remain anchored at 65–77% across Dublin, Dubai, and Madrid — the unmistakable signature of a market where committed hyperscaler capacity is fully contracted but incremental spot pools are chronically oversupplied. And at the frontier, B200 SPOT is deepening its discount in India (Hyderabad -30% over 7 days to $3.83/hr) and Southeast Asia (Singapore $2.73/hr) precisely because Blackwell supply has been deployed globally faster than enterprise demand has geographically diffused. The dominant framing today is inference ASIC pull-through vs. GPU fleet stranding: every workload that routes to Trainium3, Google TPU ($50B revenue run rate), or Meta's MTIA is a workload permanently removed from the H100 addressable market — and the neocloud equity markets, which are down 23–49% from peak, are pricing that permanent workload shift while GPU spot tickers have not yet caught up.

Key Movers

*Stockholm B200 figure is a decomposition residual from a single p6-b200.48xlarge AWS SKU at $14.63–14.84/hr whole-instance. Not a directly purchasable per-GPU rate.

Investable Insights

H1 — AWS Custom Silicon Pool Exhaustion Precedes On-Demand Rate Hike

Confidence: 5 / 5

Thesis: The AWS Trainium SPOT market in Virginia is not exhibiting a statistical anomaly or a catalog update — it is exhibiting textbook spot-pool exhaustion in real time. The 60-day price history traces a smooth, continuously accelerating auction: the median bid rose from $2.48/hr on July 19 to $2.64/hr today, a methodical +6.5% net increase in the clearing median. More decisively, the max bid within the same pool climbed from $8.99/hr to $21.50/hr over that same period — a 139% increase in the price some bidders are willing to pay, now converging on the $0.937/hr decomposed on-demand rate per Trainium chip (which equates to roughly $29.98/hr for the full trn1.32xlarge instance). When SPOT max bids equal on-demand list prices, the spot pool is effectively exhausted for priority workloads: buyers who cannot tolerate interruption are paying 100% of on-demand just to guarantee capacity in the auction. This is not ambiguous. The pattern is corroborated across silicon families: INFERENTIA SPOT in Oregon is up +157% over 30 days, and an anomaly flag was raised on INFERENTIA SPOT Canada at +30% in the same pipeline run. Two different AWS silicon architectures tightening simultaneously, across three regions, over the same 30-day window, points to a structural demand surge rather than a regional catalog event. The demand-side context resolves the "why": AWS SVP DeSantis confirmed Trainium3 is "largely sold out" (Crypto Briefing, September 16), Anthropic has committed to 5 GW of Trainium capacity and OpenAI to 2 GW through AWS from 2027, and the Amazon-Qualcomm $60B decade-long chip deal extends the architecture runway. On-demand rates for trn1./trn2. have not moved in 30 days ($0.937/hr and $0.617/hr for 1YR reserved respectively) — but that's precisely the point. The list price has not yet adjusted to reflect the utilization reality the spot market is already pricing. When that adjustment comes, 1-year reserved buyers who secured inventory at $0.617/hr will be holding a structurally advantaged position.

Key Evidence:

1. Trainium Virginia SPOT median +6.5% over 60 days, max bid +139% (July 19 → September 17, live ticker and instance price history data)
2. Max SPOT bid of $21.50/hr for trn1.32xlarge observed in premium Virginia AZ — converging on the on-demand equivalent (instance price history, September 17, 2026)
3. Three simultaneous price points for trn1.32xlarge within a single hour in Virginia ($2.56, $2.67, $21.50) confirm a two-tier market structure, not a catalog error (underlying instance data, September 17)
4. AWS Virginia SPOT pool has a 48% discount to OD — vs. 82% in Oregon and 90% in Ohio. The regional divergence is utilization geography, not model noise (ticker data, cross-region comparison)
5. AWS SVP DeSantis: "Trainium3 largely sold out"; Trainium4 reservations accumulating for 2027 targeting 6x FP4 gains (Crypto Briefing, September 16, 2026)
6. Anthropic committed to 5 GW, OpenAI to 2 GW of Trainium capacity through AWS from 2027 (news feed, September 2026) — demand commitments that rationalize Virginia pool exhaustion
7. Zero AWS list-price changes detected across all SKUs in the past 14 days (SKU-level price change tracker, September 17) — the SPOT move is auction-clearing, not catalog repricing, which means the gap between SPOT behavior and OD list price is widening, not normalizing
8. Broadcom custom XPU shipments +3.5x YoY, representing 73% of $16.7B AI revenue (TipRanks, September 16) — corroborates broad enterprise demand for custom inference silicon, of which Trainium is the AWS-native expression

Implied Action:

• Buyers: Secure trn1.* 1-year reserved at $0.617/hr now. The window is closing as Virginia spot convergence to OD accelerates. Oregon spot (82% discount, largely unutilized) remains a secondary option for interruptible workloads, but Virginia is the production AZ.
• Investors: Trainium tightening is a direct revenue tailwind for AWS within AMZN's Q3/Q4 reporting. Monitor the UPCOMING_PRICE_CHANGE alert channel daily for trn1./trn2. instance families — this is the pre-announced signal that will arrive before any OD price hike posts publicly. It has not fired yet as of today, but the spot-to-OD convergence means it is structurally overdue.
• Neocloud implication: Every Trainium workload at $0.50–0.65/hr effective cost is a workload that does not pay $5.96/hr for H100 OD on CoreWeave or Applied Digital. This is a partial mechanism driving CRWV's -23% 1-month equity underperformance — the market is pricing the Trainium substitution before the GPU tickers reflect it.

H2 — The CRWV Equity/GPU Spot Divergence Will Force a Binary Resolution at November 11 Earnings

Confidence: 4 / 5

Thesis: CoreWeave is trading at $83.35 against a 35-analyst consensus target of $144.46 — a 73% implied upside that the equity market has been consistently rejecting for four weeks. The -23% 1-month selloff, -19% in 5 days, and -46% from the 52-week high of $153.20 are not noise; they're the market embedding a specific hypothesis about utilization trajectories that the GPU ticker market has not yet priced. The H100 SPOT discounts that CRWV's fleet competes in remain anchored at 65–77% globally — Dublin at $1.34/hr, Madrid at $2.52/hr — which implies the market clearing price for H100 spot capacity is structurally capped by neocloud competition. The tension is this: CRWV's Q2 2026 gross margin of 67.4% on $2.58B revenue confirms that current-period utilization is not broken. What the equity is pricing is the transition period — the 6-9 months during which Vera Rubin racks (currently seven in production = 504 GPUs, analytically invisible in the P&L) ramp to revenue-generating scale, while the H100 fleet faces a pincer from inference ASICs below and Blackwell above. The financial structure amplifies this risk: $51.6B in total debt at 1,027% debt-to-equity and -$9.09B FCF TTM means a utilization shortfall of even 10-15% on the H100 fleet could trigger covenant discussions before Vera Rubin revenue fully offsets. The 35-analyst consensus at $144 is modeling neither the utilization cliff nor the capital structure covenant sensitivity — it is extrapolating the revenue CAGR (370% over three years) into a world where the hardware cycle cooperates. The price-performance data makes the bear case specific: B200 SPOT in CRWV-accessible markets (EU North $0.017/TFLOPS, Singapore $0.030/TFLOPS) is already 5x more compute-efficient per dollar than H100 ON-DEMAND at $0.089/TFLOPS in US East. Rational tenants evaluating workload migrations have a compelling quantitative case to pause H100 commitments and negotiate Vera Rubin terms instead.

Key Evidence:

1. CRWV at $83.35 vs. analyst consensus $144.46 (+73% implied upside) vs. 52-wk high $153.20; -23% 1-month, -19% 5-day (equities feed, September 17, 2026)
2. CRWV Q2 2026: revenue $2.58B, gross margin 67.4%, capex -$6.42B, FCF -$5.74B (single quarter); TTM FCF -$9.09B vs. total debt $51.6B (fundamentals data, September 17)
3. EPS surprise pattern: Q2 2026 beat +22.5%, but prior two quarters -22.4% and -30.0% — volatile execution track record relative to consensus (fundamentals data)
4. H100 SPOT discounts globally stable at 65–77%: Dublin $1.34/hr (-77% vs. OD), Dubai 71.7%, US East ~65% (ticker data, September 17) — no compression = no spot market tightening beneath CRWV's fleet
5. B200 SPOT price-performance at $0.017/TFLOPS/hr (EU North) vs. H100 ON-DEMAND at $0.089/TFLOPS/hr US East — 5.2x efficiency gap that rational tenants can arbitrage (unified price-performance analytics, September 17)
6. Vera Rubin production deployment: 7 racks = 504 GPUs across 2 CoreWeave regions (news feed); SemiAnalysis claim: NVL72 delivers 42% more profit/GW and 39% more revenue/GW vs. Blackwell GB300 (September 16) — economically attractive but sub-scale vs. $45.9B market cap
7. APLD down -49% from 52-week high, -25% 1-month with volume 105% of average (equities feed, September 17) — corroborates H100 neocloud equity broadly pricing the same risk, not CRWV-specific noise
8. Google TPU business at $50B annual revenue with <1 year average payback vs. >2 years for general-purpose servers (Goldman Sachs Communacopia, September 10, via news feed) — quantifies the workload migration mechanism

Implied Action:

• Event-driven long: CRWV into November 11 earnings is the asymmetric bet if GPU utilization exceeds 92% and capex guidance for Q4 comes in below $5B. At 6.1x EV/Revenue and 67% gross margin, a clean quarter re-rates the equity sharply toward $120+. The Q2 EPS beat (+22.5%) sets a challenging but not impossible bar.
• Hedge: CRWV put spread (capturing the -$1.54 Q3 EPS consensus as floor, with downside to a utilization-miss scenario) is appropriate given the FCF burn rate and prior EPS miss pattern. The utilization cliff is real if H100 tenant migration to Trainium or B200 SPOT is already underway.
• Barometer role: CRWV earnings on November 11 is the highest-information event in the GPU market for Q4 2026. Revenue per rack per day and any disclosure of Vera Rubin revenue contribution will be the two metrics that resolve the H100-fleet-obsolescence vs. smooth-generational-transition debate. Mark your calendar.

H3 — The Custom Silicon Pincer Is Shortening H100's Economic Life Faster Than Any Depreciation Model Reflects

Confidence: 5 / 5

Thesis: This is the most structurally important hypothesis in the brief, and it is the one most thoroughly corroborated across independent datasets. Three hyperscaler-adjacent compute consumers — each operating at a scale measured in hundreds of thousands of chips — have simultaneously disclosed they are shipping proprietary inference silicon that directly substitutes for H100: AWS (1M+ Trainium deployed, $25B annual revenue run rate), Google (TPU at $50B annual revenue, <1 year payback vs. >2 years for GPU servers), and Meta (hundreds of thousands of MTIA chips in production, MTIA 450 doubling HBM bandwidth, MTIA 500 adding 50% more in H2 2027). These are not pilot programs or roadmap promises — they are production deployments at scale, and they are serving inference workloads that would otherwise run on H100 GPUs. The mechanism of H100 economic life compression is not a successor GPU (B200 exists and is visible in pricing) — it is a wave of workload-specialized ASICs that collectively remove the inference market from H100's addressable revenue base. Simultaneously, Vera Rubin and B200 are encircling H100 from above for training workloads. H100 is caught in a pincer with no remaining sole-source advantage in either training or inference. The price-performance data quantifies the generational squeeze: H100 ON-DEMAND at $0.089/TFLOPS/hr in US East competes against B200 SPOT at $0.033–0.042/TFLOPS/hr in US core markets — a 2–3x efficiency disadvantage at the spot level, which means any rational buyer who can tolerate interruption is already running on B200. The depreciation module returns only CATALOG_SURVIVAL data for H100 with no annual decay rate populated — a model gap that means there is no systematic quantification of how fast H100 on-demand rates are decaying. That absence is itself a risk flag: the market lacks a validated obsolescence clock for the asset underpinning $10B+ in neocloud debt.

Key Evidence:

1. AWS: 1M+ Trainium chips deployed, $25B annual revenue run rate; Trainium3 "largely sold out," Trainium4 reservations building (Analytics Insight, September 11; Crypto Briefing, September 16, 2026)
2. AWS-Cerebras inference split confirmed: "Amazon split AI inference workloads between Trainium (computationally complex tasks) and Cerebras wafer-scale engine (memory-intensive tasks)" (IEEE Spectrum, September 15, 2026) — GPU is not even the fallback option; custom silicon is bifurcated by task type
3. Google TPU at $50B annual revenue with <1 year payback vs. >2 years for general-purpose GPU servers (Thomas Kurian, Goldman Sachs Communacopia, September 10, 2026)
4. Meta MTIA: "hundreds of thousands" of MTIA chips in production; MTIA 450 doubles HBM bandwidth vs. MTIA 400; MTIA 500 adds another 50% for H1/H2 2027; 1+ GW custom silicon deployment target in 12 months (Analytics Insight, September 11, 2026)
5. Broadcom custom XPU shipments +3.5x YoY, 73% of $16.7B Q3 AI semiconductor revenue; new long-term Google TPU co-development deal running "multi-tens of billions annually" (TipRanks, September 16)
6. B200 SPOT EU North: $0.017/TFLOPS/hr vs. H100 ON-DEMAND US East: $0.089/TFLOPS/hr — 5.2x efficiency disadvantage for H100 at on-demand rates (price-performance analytics, September 17)
7. H100 launched March 2022 (4.49 years active in catalog); no GENERATION_GAP or PERF_NORMALIZED annual decay rate in the depreciation module — the model has no systematic obsolescence clock for the most widely-held neocloud asset (depreciation data, September 17)
8. H100 ON-DEMAND US East flat at $5.96/hr (-0.05% over 30 days) — the absence of rate cuts is not a positive signal; it reflects contracted pricing floors, not demand strength (ticker data, September 17)

Implied Action:

• Buyers: Avoid any H100 3-year reserved contract. The evidence for workload migration to inference ASICs is too strong to lock in a 3-year commitment at $5.96/hr when Trainium SPOT clears at $0.50/hr for equivalent inference throughput. Prefer SPOT or 1-year reserved at most.
• Investors: Reduce exposure to H100-heavy neoclouds (CRWV, APLD, CIFR) that lack a credible Vera Rubin migration timeline on their balance sheet. APLD at -49% from 52-week high with October 8 earnings is the most direct short-duration expression, but it is already heavily discounted. The cleaner trade is long AMZN/NVDA (custom silicon tightening benefits both) vs. short the neocloud complex.
• Watch: H100 ON-DEMAND rate cuts at any hyperscaler are the signal that the custom silicon pincer has forced active repricing — this has not yet occurred, but it is the event that will mark the transition from a "soft landing" H100 story to an explicit obsolescence repricing cycle. Monitor daily.

H4 — HBM Scarcity Is Real but Channel-Bifurcated: NVIDIA-Locked Supply vs. Contestable Secondary Market

Confidence: 4 / 5

Thesis: Intel CEO Lip-Bu Tan's September 15 statement — memory prices up "5–7x with no relief until 2028" — is being absorbed by the market as a uniform HBM scarcity narrative. The reality exposed by our component and trade-flow data is more precise and therefore more actionable: the scarcity is real, but it is concentrated in the NVIDIA supply channel and effectively structural through at least 2027. SK Hynix holds approximately 60–62% of global HBM market share, with roughly 90% of that output committed under a $500B long-term supply agreement directly to NVIDIA — an agreement that locks HBM4 supply through 2027. For non-NVIDIA buyers, SK Hynix HBM4 spot availability is effectively zero through the agreement's term. The broader IC import data tells a more nuanced story: US all-IC imports (HS 8542) are up only +40% YoY with a -13% MoM print and flat 3-month momentum — decelerating channel momentum that is inconsistent with a single unified HBM shortage narrative. Meanwhile, US solid-state storage imports (HS 852351) are surging at +499% YoY — a separate memory cycle (NAND/flash) that is absorbing enormous import volume and should not be conflated with HBM. The critical counterweight to the 2028 timeline: Samsung is scaling HBM4 wafer output from 170,000 to 250,000 wafers/month by end of 2026, with yields reaching 80% as of August 2026. The AMD-Samsung Korea AI Center of Excellence (announced September 16) accelerates Samsung's HBM4 ramp with AMD as a committed offtake customer — directly addressing Samsung's bottleneck of qualifying HBM4 with major accelerator manufacturers. If Samsung passes NVIDIA's HBM4 qualification before end-2026, the "no relief until 2028" narrative breaks in 2027, potentially 12 months earlier than consensus.

Key Evidence:

1. SK Hynix: ~60–62% HBM market share, ~90% of output to NVIDIA under $500B long-term supply agreement locking HBM4 through 2027 (component tracker, confidence 0.85, September 13, 2026)
2. "HBM chips commanding prices exceeding their weight in gold in 2026" — SK Hynix component record characterization (component tracker, September 13)
3. US all-IC imports (HS 8542): +40% YoY, -13% MoM, 3-month momentum flat/decelerating (trade flow data, July 2026 census) — inconsistent with uniform HBM surge narrative
4. US solid-state storage imports (HS 852351): +499% YoY, +75% 3-month momentum, RISING signal (trade flow data, July 2026) — a separate flash/NAND cycle being conflated with HBM in press coverage
5. Samsung HBM4: scaling wafer output 170,000 → 250,000/month by end-2026; yields at 80% as of August 2026 (component tracker, September 2026) — ramp faster than 2028 relief timeline implies
6. AMD-Samsung Korea AI Center of Excellence announced September 16, 2026: co-development of HBM4, with AMD MTIA 450 doubling HBM bandwidth as demand pull (news feed)
7. SK Hynix Indiana fab targeting mass HBM production for H2 2029 — the actual 2028+ relief source for the NVIDIA-channel HBM shortage (component tracker)
8. Intel CEO Lip-Bu Tan, AI Infrastructure Summit: "Memory prices up 5–7x, no relief until 2028" (shattered.io, September 15, 2026) — primary source, but represents NVIDIA-channel pricing, not general market

Implied Action:

• Long: SK Hynix (KRX: 000660) and AVGO (as primary NVIDIA XPU supplier dependent on HBM bandwidth) through Q1 2027. The NVIDIA-channel scarcity is structural through the $500B agreement; the 5-7x price premium is real in that channel.
• Monitor: Samsung HBM4 NVIDIA qualification timeline. A qualification announcement before end-2026 breaks the 2028 relief narrative and triggers a rotation from SK Hynix to Samsung as the marginal HBM supplier — meaningful for both equity prices.
• Fade the 2028 uniform narrative: The broad memory market (flash, DRAM) is not co-moving with HBM scarcity. Avoid conflating flash import surges with HBM supply tightness — they are different cycles serving different customers.

Inventory & Channel Health

The inventory health system returned no composite regime signal and no DIO flags for the distributor layer (ARW, SNX, AVT) in this run. This is a genuine analytical gap, not a clean negative. The absence of a GLUT_FORMING or ABSORBING regime read means the most reliable leading indicator (distributor DIO leads hyperscalers by 1–2 quarters) is dark for this cycle.

What the trade-flow data can partially substitute: US solid-state storage imports (HS 852351) at +499% YoY with +75% 3-month momentum are a RISING signal that the US channel is absorbing flash/NAND inventory at an extraordinary pace. Historically, surging flash imports concurrent with flat or rising distributor DIO would indicate channel stuffing forming — but without the DIO read, we can only say the import absorption is large and accelerating. Separately, all-IC imports (HS 8542) are decelerating at -13% MoM — if HBM were driving a broad IC surge, this channel should also be rising. The divergence between flash (+499% YoY) and all-ICs (+40% YoY, decelerating) implies the import surge is NAND/consumer SSD rather than AI accelerator memory, and the AI memory supply chain is tighter and more directly contracted than trade flows can measure.

Priority action: Query query_inventory_health_history for ARW and SNX directly on the next cycle. If ARW DIO is rising from its current level while revenue decelerates, the HBM glut-forming thesis gains credibility against the Intel CEO scarcity narrative.

Supply & Demand Pipeline

Supply — What's Actually Landing Near-Term:

The datacenter tracker shows 141 active UNDER_CONSTRUCTION projects, but near-term MW additions with firm tenant demand and utility contracts are concentrated in three sites. xAI Shelby County (1.56 GW, TVA utility, 2026 target) is the most imminent delivery — its landlord is already contracted and construction is advanced. Stargate Michigan ("The Barn") (1.4 GW, Building 1 "nearly complete," DTE Energy 19-year contract, OpenAI/Oracle tenant) is targeting January 2028, placing it inside the H100-to-Vera Rubin transition window — the GPU generation that fills it will depend on which hardware is deliverable at scale in late 2027. Google New Florence Missouri ($15B, 1.5 GW, Ameren contracted, Building 1 vertical construction begun) represents Google's own training-capacity buildout, which will predominantly consume TPU and Blackwell-generation hardware rather than H100.

Demand — Model Token Pricing in Free Fall:

The model tracker surfaces a 167x output-token pricing spread between Claude Fable 5.1 ($50/MTok output, Anthropic, September 2026) and DeepSeek V4.1 Flash Vision ($0.60/MTok, MIT license, 763B MoE with 16B active decode). OpenAI's GPT-5.6 cut prices 20% on September 1 as a direct competitive response to Chinese frontier model pricing. This token-price compression is the demand mechanism behind ASIC tightening: at $0.60/MTok output, running profitable inference requires volume throughput measured in trillions of tokens monthly, which flows naturally to cost-optimized ASICs (Trainium, TPU) rather than premium H100 capacity. The Claude Fable 5.1 / DeepSeek spread will not survive at scale — the convergence pressure will continue forcing token prices down, and the compute demand that results will continue flowing toward the lowest-cost-per-token architecture, which today is Trainium SPOT at ~$0.50/hr effective Virginia clearing rate.

Component — HBM4 Supply Chain:

The AMD-Samsung Korea AI Center of Excellence announced September 16 is the most material component-level development of the week. Samsung's HBM4 wafer ramp to 250,000/month by end-2026 at 80% yield creates a potential second qualified HBM4 supplier for the market — the first significant supply-side relief signal for non-NVIDIA accelerator developers. If Samsung's NVIDIA qualification clears before Q1 2027, the HBM4 supply constraint becomes significantly less binding for the AMD MI400X and GB300 ecosystems. Watch AMD and Samsung joint announcements as the leading indicator.

Pre-Announced Price Changes

None. Zero UPCOMING_PRICE_CHANGE alerts were published by any provider (AWS, GCP, Azure, Oracle) across the full lookback window. This is the absence that matters most in today's data. In prior repricing cycles, a batch of provider-published future effective dates typically preceded major market moves by 2–4 weeks. The current silence, combined with AWS Trainium Virginia spot-to-OD convergence at $21.50/hr max bids, means one of two things: the repricing decision has been made internally but not yet published (the "quiet before the announcement" pattern), or AWS is managing the tightness through allocation controls rather than list-price adjustments. The UPCOMING_PRICE_CHANGE alert channel for trn1./trn2. instance families is the highest-priority monitoring target in the system as of today.

Risk Flags

Immediate (0–30 Days)

R1 — Trainium OD Rate Hike Without Warning (High Impact) The Virginia SPOT market has already signaled exhaustion at $21.50/hr max bids converging to list-price equivalency. If AWS publishes an UPCOMING_PRICE_CHANGE on trn1./trn2. without the usual 30-day lead time (which it has done in prior ASIC repricing cycles), buyers relying on current 1-year reserved rates at $0.617/hr are properly protected — but spot users would face an immediate cost step-change. The risk is asymmetric: compute buyers on spot Trainium without a reserved fallback could see effective costs double overnight. Watch the alert channel daily.

R2 — Iowa GPU Slot Reallocation Creating Thin-Market Artifacts (Low Impact, High Confusion Risk) The simultaneous A100 SPOT +157% and H100 OD -48% in GCP Iowa on a 24-hour basis is a genuine catalog event but carries no investable signal — it is GCP clearing older inventory to allocate zone capacity for Blackwell deployments. The risk is analytical rather than financial: any automated system that picks up the Iowa A100 move as a "scarcity signal" and recommends locking A100 capacity is misreading the data. This is a sell signal for legacy GCP GPU capacity in Iowa, not a buy.

R3 — H100 NatEx Absence Inflates Azure/AWS Price Gap Artificially (Model Risk) Zero NatEx pairs exist for H100 across all regions. The AWS/Azure H100 spread ($5.96 vs. $11.05 in US East/Central) may partially reflect bundled networking, storage, or egress costs rather than pure GPU pricing — a model artifact from Stage 2 residual decomposition when no matched-pair validation is available. Any contract negotiation using the Azure H100 price as a market reference should be treated with skepticism until NatEx pairs confirm the decomposed rate.

Near-Term (30–90 Days)

R4 — CRWV Covenant Risk if Q3 Utilization Misses (High Impact) CRWV's $51.6B debt load at 1,027% debt-to-equity is structured around continuous high-utilization GPU revenue. The Q3 EPS consensus is -$1.54, and the prior two quarters produced EPS misses of -22.4% and -30.0% before the Q2 beat. If Q3 revenue growth decelerates materially from the 112.5% YoY rate — for instance, to 70–80% YoY as H100 tenant migrations begin — the FCF burn at $6–8B/quarter becomes covenant-sensitive.