Daily Investment Brief — September 14, 2026

Signwl Compute Intelligence Briefing — September 14, 2026

Final Edition | Data as of 07:31 UTC | Pipeline anomaly rate: 0.17% (clean)

Market Pulse

The GPU compute market is running two simultaneous dynamics that are fundamentally at odds with the dominant consensus narrative. The "universal GPU scarcity" story, driven by Oracle's 97.9% utilization headline and $30B AI backlog, is real at the frontier — but structurally false in the commodity H100 pool. Multi-provider Western US H100 markets are actively deflating: Montreal H100 ON-DEMAND fell 25.8% in seven days to $7.65/hr, Frankfurt collapsed 27% to $1.76, Tokyo dropped 31% to $7.27 — while Africa South and São Paulo single-provider H100 markets sit unmoved at $13.53 and $9.90/hr respectively with a 75% per-TFLOPS cost premium and zero competitive pressure to close the gap. Against that deflationary backdrop, the clearest genuine demand signal in the entire dataset is Trainium SPOT in Virginia: up 92.1% over 30 days, spot discount compressed from 87% to 57%, with a $21.50/hr max-price AZ outlier confirming that the cheapest availability zones are already exhausted. The market is bifurcating along two axes simultaneously — competitive vs. monopoly geography, and commodity H100 vs. proprietary ASIC — and most investors are reading only the aggregate utilization number.

Key Movers

Investable Insights

H1 — AWS Trainium SPOT Compression Is the Earliest Leading Indicator of a Genuine ASIC Utilization Inflection Confidence: 5 / 5

Thesis: The Trainium SPOT market in us-virginia is the single highest-conviction signal in today's dataset, and it is almost entirely invisible in public market commentary because Trainium is not a GPU. At 57% spot discount today — down from an estimated 87% at the start of August — the trn1 pool in Virginia is approaching the point where spot and reserved pricing converge, which historically precedes either a list-price increase or a full inventory pull from spot availability. The signature of genuine capacity absorption is textbook-clear: the median fleet price has moved monotonically from $3.21/hr on August 19 to $2.82/hr today (median falling as cheap AZs are consumed), while the max price has escalated from $9.50 to $21.50/hr over the same period — a single availability zone clearing at 1.34x the median chip rate because every other AZ at that region is at capacity. This is not noise. The pipeline's own anomaly diagnostics flagged today's +50.4% 24-hour move as a genuine delta event (confirmed non-artifact alongside Ohio's -49.4% unloading event on August 23 — together they paint a pool in dynamic flux, not steady decline).

Three demand vectors are pulling Trainium toward full utilization simultaneously: AWS's AI backlog growth (Oracle's $30B print is a market-wide indicator, not solely Oracle-specific demand); the GPT-Live-1 and long-context inference workload class that requires low-latency, always-on ASIC infrastructure resistant to spot preemption; and the Trainium 4 announcement (September 11, confirmed by Next Platform) with Broadcom providing packaging integration, which signals AWS is scaling the platform faster than public roadmap implied — existing-generation demand will not be relieved by new supply imminently.

Key Evidence:

1. Trainium SPOT Virginia up +92.1% over 30 days, +50.4% in the last 24 hours — largest sustained move across the entire ASIC sub-universe (live ticker, Sep 14 2026)
2. Spot discount compressed 87% → 57%; at 45% the spot/reserved economics invert and reserved looks better priced (live ticker, Sep 14)
3. Max observed price for trn1.32xlarge (16-chip fleet, 128 vCPU / 512 GB RAM): $21.50/hr, up from $4.10/hr on August 15 — a 424% max-price surge while median rose only 46% (instance price history, AWS, Aug 15–Sep 14)
4. Ohio Trainium SPOT still at 90% discount — corroborates that the Virginia signal is regional pool exhaustion, not a catalog artifact (live ticker, Sep 14)
5. Trainium 4 announced September 11 (Broadcom integration, status: ANNOUNCED) — next-generation supply not arriving before mid-2027 at earliest, leaving current-gen pool supply-constrained through the cycle peak (component tracker, confidence: medium)
6. No UPCOMING_PRICE_CHANGE alert for any Trainium SKU — AWS has not telegraphed a list-price increase, meaning spot compression will do the price discovery work that a list-price raise typically accomplishes (price alert scan, Sep 14)

Implied Action: Lock trn1.32xlarge or trn1n.32xlarge 1-year reserved capacity in Virginia immediately. The $0.17/chip-hr SPOT rate still represents a compelling deal vs. on-demand implied (~$0.50/chip-hr), but the pool is visibly draining. For investors: AWS's AI product revenue in Q4 2026/Q1 2027 will be the first financial disclosure to reflect this utilization — AMZN does not break out Trainium revenue explicitly, but cloud segment operating margin should widen if ASIC utilization is compressing spot discounts. The Trainium signal is also a directional read for Google's TPU utilization (never reported publicly); if the ASIC absorption trend is AWS-wide, a parallel signal in GCP's TPU SPOT discount would be the confirmation that proprietary AI accelerators are structurally tightening ahead of H100.

H2 — The H100 Market Has Bifurcated Permanently: Monopoly Markets Carry a 75% Per-TFLOPS Premium That Will Compress as Competition Enters Confidence: 4 / 5

Thesis: The aggregate "GPU shortage" narrative obscures a market structure that is bifurcating in real time. In the competitive Western US (Oregon: 3 providers, $0.10/TFLOPS-hr; Ohio: 2 providers, $0.115/TFLOPS-hr), H100 pricing is already mid-cycle deflation — neoclouds clearing below $0.50/hr SPOT in Oregon represent distressed inventory, not tight supply. In the single-provider monopoly markets — Africa South at $0.202/TFLOPS-hr (Azure sole), São Paulo at an implied ~$0.235 (single neocloud), Warsaw at $0.182 (single provider, $12.17/hr) — the same silicon commands a 75-100% per-TFLOPS premium with zero competitive pressure and zero spread. This is not a shortage premium; it's a structural absence of competition that will correct as neocloud operators expand geographic coverage.

The deflation in multi-provider markets is already accelerating: Montreal -25.8% over 7 days, Frankfurt -27%, Tokyo -31% — all multi-provider locations where the entry of a second or third operator has compressed margins. The path from monopoly to competitive equilibration is visible in the Korea Seoul ticker: Seoul H100 sits at $8.63/hr on-demand but already has a 9,111% min-to-max spread with GCP's aggressive SPOT pricing ($0.16/hr min) providing the competitive anchor. That is what early-stage competitive entry looks like in this market.

Key Evidence:

1. Africa South H100 ON-DEMAND: $13.53/hr, $0.202/TFLOPS-hr (Azure sole provider, 4 observations). US-West Oregon: $6.67/hr, $0.0995/TFLOPS-hr (3 providers, 40 observations). 75% premium per unit of compute for identical silicon (price-performance table, Sep 14)
2. Montreal H100 ON-DEMAND: $10.31/hr → $7.65/hr in 7 days (-25.8%). Frankfurt: $2.42 → $1.76 (-27%). Tokyo: flat → -31%. All multi-provider markets actively deflating (ticker history, Sep 7–Sep 14)
3. Africa South, São Paulo, Warsaw, Osaka H100 ON-DEMAND: 0% change across 24h and 7d, spread_pct = 0. No competitive discipline (live tickers, Sep 14)
4. Korea Seoul H100: $8.63/hr median, min $0.16/hr, spread 9,111% — illustrates the deflation path once GCP enters as a second provider (live ticker, Sep 14)
5. Pipeline anomaly rate for AWS/Azure H100 tickers: effectively 0% (Stage 1 overall 1.87%, but GCP GPU_type anomaly is a known stable artifact). The directional moves above are validated data (anomaly diagnostics, Sep 14)
6. H100 catalog survival: still active at all three hyperscalers after 4.48 years, longer than A10/A10G (now inactive on Azure and GCP). Price decay has not yet begun at list-price level; the market is doing the price discovery through competitive entry, not hyperscaler list cuts (depreciation tracker, Sep 14)

Implied Action: Any buyer currently paying $13.85/hr (Arizona single-provider) or $13.53/hr (Africa South) for H100 on-demand should restructure procurement toward Oregon or Ohio neoclouds immediately — the per-TFLOPS cost savings are 50-70% for equivalent compute. For investors: Applied Digital (APLD, -33.8% / 3 months) and Cipher Dynamic (CIFR, -24% / 3 months) operate H100 capacity in non-competitive geographies. Their margin compression will accelerate as neocloud operators expand into APAC and LATAM — the equity market has partially priced this, but the pricing data suggests the deflation is still early. The more durable equity expression is operators with the most diverse geographic footprint (CoreWeave, CRWV, -6.8% / 3 months) rather than single-geography names.

H3 — B200 "SPOT" Markets Are Isolated Catalog Pricing, Not Market Discovery: The São Paulo 33% Premium Will Collapse Within 60 Days Confidence: 4 / 5

Thesis: Every single one of the eleven live B200 SPOT tickers carries n_providers = 1, spread_pct = 0.0, and null natex data. There is no B200 spot market — there are B200 catalog prices published by individual neocloud operators in geographic isolation. The São Paulo B200 "SPOT" at $6.41/hr, up 55.1% over seven days and 6.5% in a single day, is not demand-driven price discovery; it is a single operator launching a new geography and repricing experimentally without any competitive reference. The fundamental tell: São Paulo's B200 SPOT has a null spot_discount_pct because there is no B200 on-demand offering in São Paulo against which to compute a discount — the operator is pricing in a vacuum. Compare this to the Virginia B200 SPOT ($4.84/hr), which represents a genuine 63.1% discount to Virginia's B200 on-demand at $12.82/hr. The São Paulo operator is not constrained by either a competitive floor or an on-demand ceiling.

The regional dispersion is dramatic and entirely explained by this isolation effect: São Paulo at $6.41 is 33% above Virginia ($4.84), 87% above Jakarta ($3.43), and $3.00 above Seoul. Jakarta and Hyderabad meanwhile are collapsing — -30% and -34.3% over seven days respectively — because those are also single operators cutting unilaterally after overpricing their launch. The B200 market has the price structure of four separate closed auctions, not one global GPU market. This will normalize as the second operator activates B200 capacity in each region, which is a when-not-if question given the 115 active datacenter construction projects in the tracker.

Key Evidence:

1. All 11 B200 SPOT tickers: n_providers = 1, spread_pct = 0.0, natex_median = null. Zero market structure (live tickers, Sep 14)
2. São Paulo B200 SPOT: $6.41/hr, +55.1% / 7d, +6.5% / 24h. spot_discount_pct = null (no on-demand comparator). $1.57/hr above Virginia reference (live ticker, Sep 14)
3. Jakarta B200 SPOT: $3.43/hr, -30.0% / 7d. Hyderabad: $4.59/hr, -34.3% / 7d. Independent operators cutting without market signal (live tickers, Sep 14)
4. Virginia B200 ON-DEMAND at $12.82/hr — the only competitive reference price; São Paulo operator has ignored it entirely in their SPOT pricing (live ticker, Sep 14)
5. Microsoft datacenter query returns zero records for Brazil — the São Paulo B200 SPOT is definitively not an Azure deployment; it is a neocloud operator without hyperscaler pricing discipline (datacenter tracker, Sep 14)
6. The GB200 South America premium ($26.99/hr SPOT, $0.300/TFLOPS-hr) is standard Azure geographic multiplier applied uniformly across APAC/LatAm — not a strategic sovereign play — confirmed by the same premium structure in Azure Seoul and Azure Africa South H100 (price-performance table, Sep 14)

Implied Action: Buyers requiring B200 capacity should default to Virginia ($4.84/hr SPOT) or Ohio ($4.91/hr) and absolutely not commit to São Paulo contracts at current prices — the premium has no structural basis and collapses the moment a second provider enters. Watch for spread_pct > 0 or n_providers > 1 in the B200|SPOT|br-saopaulo ticker as the entry signal; it will precede a rapid convergence toward $4.80-5.00. For investors: the first neocloud to achieve multi-region B200 SPOT coverage (Virginia + Ohio + 2+ international) will establish itself as the price-setter for Blackwell-class inference globally. Currently no public neocloud company has confirmed multi-region B200; this is the correct operational milestone to track in Q4 2026 earnings calls.

H4 — The Open-Weights MoE Pricing Wedge Is Real, But the Demand Route Is Through API Adoption, Not GPU Rental Self-Hosting Confidence: 3 / 5

Thesis: DeepSeek V4 Flash Vision (released September 10) benchmarks at 90.9 on GPQA Diamond — above Claude Opus 5 on several evaluations — and prices at $0.15/M input tokens, a 67x discount to GPT-6 Astra at $10/M. The economic incentive for enterprises to switch is real: a production workload running 100M tokens/day costs $15/day on DeepSeek vs. $1,000/day on GPT-6 Astra. However, the hypothesis that this drives meaningful self-hosted GPU rental demand is incorrect on the unit economics: serving the full 763B-parameter MoE model — even with only 8-16B parameters active per forward pass — requires the complete weight matrix resident in memory, approximately 500-600GB of VRAM, or a full 8x H100 server. At US-West on-demand prices ($53/hr for an 8x H100 node), serving DeepSeek at $0.60/M output tokens generates roughly $0.43/hr in API revenue against $53/hr cost. The GPU rental self-hosting margin is deeply negative at hyperscaler list prices.

The correct economic beneficiary of the MoE pricing wedge is not the GPU spot market — it is the DeepSeek API itself (capturing enterprise migration from OpenAI), and secondarily the server hardware layer (Dell, HPE) capturing the on-prem buildout by enterprises running internal inference at fixed cost amortized across continuous utilization. An enterprise replacing $5,000/day in GPT-6 Astra spend with a $500/day reserved GPU server running DeepSeek is the plausible scenario; renting spot GPUs at $0.30-0.50/hr SPOT and building an API business on top of them is not. The trade flows corroborate the inference infrastructure build: $7.08B in flash/SSD imports to the US in July 2026 (+499% YoY, +75.3% three-month momentum) directly supports the model weight caching, KV-cache offloading, and embedding storage stacks required for large-context inference at enterprise scale. Chips (+40% YoY but -13% MoM in the same period) confirm the cycle has shifted from training cluster accumulation toward inference storage buildout.

Key Evidence:

1. DeepSeek V4 Flash Vision: 763B parameters total, 8-16B active, MIT license, GPQA Diamond 90.9, available HuggingFace. Pricing: $0.15/M input, $0.60/M output (model tracker, confirmed Sep 10)
2. GPT-6 Astra: $10/M input, $50/M output (GA Sep 3). Claude Fable 5.1: same tier. 67x pricing gap confirmed (model tracker, Sep 3–10)
3. DeepSeek V4 Flash hardware requirement: 500-600GB VRAM for full model serving (8x H100 minimum). H100 8x server US-West: ~$53/hr on-demand. Revenue potential at $0.60/M output: ~$0.43/hr. Unit economics inverted at list prices (model tracker sourcing, yottalabs.ai, Sep 10)
4. US flash/SSD imports (HS 852351): $7.08B in July 2026, +499% YoY, +75.3% three-month momentum (RISING). IC imports (HS 8542): $4.37B, +40% YoY but -13% MoM (FLAT/decelerating). Import divergence confirms infrastructure cycle has rotated to storage (trade flow data, July 2026)
5. H100 US-West Oregon SPOT min prices: sub-$0.50/hr (neocloud clearing). Even at $4/hr for 8x server SPOT, $0.43/hr revenue doesn't cover costs. Self-hosting revenue model doesn't pencil (live ticker, Sep 14)
6. Dell Technologies +37% over 3 months at 19.8x forward P/E with 0.6x PEG — the equity market is pricing the server hardware layer correctly, not the rental layer (equities feed, Sep 14)

Implied Action: Long Dell Technologies (DELL) as the cleanest expression of inference infrastructure capex regardless of which GPU rental operator or model API wins. Long AMD (AMD, +10.2% / 5 days, $516, 0.51x PEG) on the thesis that MI300X picks up enterprise on-prem inference server sales as a hyperscaler-independent option, though our pricing data carries no AMD GPU tickers to corroborate GPU rental demand. Avoid the "GPU rental self-hosting of open-weights models" trade — the unit economics don't work at current H100 prices. Monitor H100 US-West spread compression (currently 94% min-to-max) as the signal that reserved GPU demand for inference workloads is absorbing neocloud inventory; that would be the first data point that validates the self-hosting thesis getting closer to penciling.

Inventory & Channel Health

The composite inventory regime is MIXED / NEUTRAL — the Firestore DIO collections have not populated this cycle, leaving the hyperscaler and distributor layers silent. The absence of data is not a bullish or bearish read; it is a gap that the next ARW/SNX/AVT filing will resolve. However, the trade flow layer tells enough of the story independently.

The most analytically important single number in today's briefing is the flash/SSD import signal: $7.08B in US imports in July 2026 against roughly $1.2B in July 2025 — a 499% YoY surge with a +75.3% three-month momentum reading classified RISING by the pipeline. This is not a glut signal in isolation, because there is no corresponding DIO data showing inventory piling up at distributors. The benign interpretation — which the IC import data partially corroborates — is that flash/SSD pull-through demand is real and being absorbed immediately into inference infrastructure (model weight storage, KV-cache offloading, embedding databases for RAG). The bearish interpretation — which the silent DIO layer cannot yet rule out — is that storage is being over-ordered relative to actual inference deployment pace, which would show up as rising ARW/SNX DIO in the next quarterly snapshot.

The watch signal is explicit: If the next DIO update for Arrow Electronics (ARW) or SYNNEX/TD Synnex (SNX) shows DIO rising more than 10 days QoQ while flash import momentum remains above 50% three-month, that is a CHANNEL_STUFFING alert in formation and a bearish signal for storage-adjacent equipment capex (NetApp, Pure Storage, Western Digital). Today, no such alert exists.

Supply & Demand Pipeline

Supply — the physical build is genuinely large-scale and geographically concentrated. The construction tracker's top-of-funnel is dominated by multi-gigawatt centralized campuses: AWS New Carlisle Indiana (2.25 GW planned, active construction), Oracle-OpenAI Stargate in Saline Township Michigan ($56B, 1.4 GW, building), xAI Shelby County Tennessee (1.56 GW, natural gas-powered), and Meta Project Anthem in Tulsa (1.5 GW). The power source pattern across these builds — natural gas dominant, with Monarch Compute Campus in West Virginia explicitly going off-grid to escape Northern Virginia's capacity-constrained grid — confirms that grid access rather than land or capital is the binding constraint on 2027-2028 supply additions. TVA's new rate structure with a 10% increase for facilities over 5 MW is a direct energy-cost headwind for Tennessee/Southeast datacenter economics, and one that will filter into GPU rental pricing only with a 12-18 month construction-to-deployment lag.

Political friction is materializing at scale, not at the margin. Project Sail in Newnan, Georgia (936 MW planned) faces an active 8,000-member opposition coalition and litigation. CyrusOne Waverly Illinois (634 MW) is in court over noise violations. The Stargate Michigan facility faces state-level proposals for water use caps. This is not NIMBY noise — it is organized, legally active opposition at projects representing over 2.5 GW of planned capacity. The construction-to-operational timeline for any project currently in permitting dispute should be extended by 12-24 months in any supply model.

Demand — the model release calendar is driving qualitatively new inference workload classes. GPT-Live-1 (full-duplex voice at $0.05/minute, released September) represents an always-on, latency-critical inference workload that is fundamentally incompatible with spot preemption — this demand class will migrate toward reserved Trainium and TPU capacity, directly supporting Hypothesis 1's ASIC utilization thesis. The 1M-context-window models (GPT-6 Astra, Claude Fable 5.1, DeepSeek V4 Flash) at scale impose enormous KV-cache memory requirements per active session, which drives per-GPU utilization up even at flat request volume. The effective GPU demand per user-session for a 1M-context inference call is estimated at 10-20x higher than for a standard 8K-context call — this is the structural driver of the Oracle 97.9% utilization figure, not simply more users.

Component — Trainium 4's Broadcom packaging integration (announced September 11, component tracker confidence: medium) suggests AWS is scaling Trainium manufacturing capacity significantly, with Broadcom's advanced packaging capability enabling higher chip density per server. This is directionally bullish for AWS's long-run Trainium economics but introduces a Broadcom supply dependency that is worth monitoring for potential packaging capacity constraints in 2027.

Pre-Announced Price Changes

No UPCOMING_PRICE_CHANGE alerts are currently active across AWS, GCP, Azure, or Oracle for any GPU, CPU, RAM, or storage SKU. This is the cleanest "no imminent hyperscaler repricing" signal the system can produce. All price movement visible in the data today is either market-driven (Trainium SPOT compression, multi-provider H100 deflation) or single-operator catalog updates (B200 regional repricing). No hyperscaler has telegraphed a future-dated price change.

Risk Flags

Immediate (0–30 Days)

R1 — B200 São Paulo Overpayment Risk Any enterprise that signed a B200 capacity contract in Brazil at or near the current $6.41/hr SPOT price in the last seven days has locked in a rate that is 33% above the Virginia reference with zero structural justification. The price moved +55.1% in seven days by a single operator with no competitive constraint; it can move -33% just as quickly when a second provider enters or when the current operator re-anchors to the Virginia benchmark. No long-term B200 capacity commitments should be made in São Paulo, Seoul, or Hyderabad until n_providers ≥ 2. Monitoring trigger: B200|SPOT|br-saopaulo spread_pct > 0.

R2 — Pipeline Artifact Contamination in the 7-Day Mover List The vCPU UNMAPPED tickers showing ~12,000,000% deltas from Ohio, Seoul, and Stockholm are confirmed pipeline artifacts (unit-mapping failures in the Stage 1 decomposition, not price events). Any automated alert system or quant model consuming the 7-day mover list without filtering these tickers is generating false signals. This is not a new problem — the anomaly diagnostics show it as a persistent structural issue — but it remains unresolved and will continue contaminating derivative analyses until the mapping layer is patched.

R3 — Trainium SPOT Availability Evaporation Window The $21.50/hr max-price observation for trn1.32xlarge in Virginia indicates at least one availability zone is already at or near full capacity. Buyers who intended to run long-horizon workloads on Trainium SPOT face increasing preemption risk in the Virginia region. The window for locking reserved capacity at favorable rates is measurably narrow: if spot discount reaches sub-40% and holds for five consecutive days, the reserved economics will have inverted and reservations will clear at a premium to current SPOT pricing.

Near-Term (30–90 Days)

R4 — H100 Monopoly Market Repricing: Forced Procurement Restructuring The 75% per-TFLOPS premium in single-provider H100 markets (Africa South, Warsaw, São Paulo, Osaka) will attract competitive entry as neocloud operators expand globally. When the second provider enters any of these markets — most likely Hyderabad (Microsoft's largest India region, active neocloud expansion) — the price correction could be rapid and violent, consistent with Montreal (-25.8% / 7 days after competitive entry). Any enterprise with a long-term H100 contract in a monopoly market should be building optionality to shift workloads to competitive regions. Investors long APLD or CIFR in concentrated geographies should model a 25-40% revenue-per-GPU reduction over the next two quarters as the leading scenario.

R5 — DeepSeek V4 Flash Open-Weights Model Export Risk DeepSeek V4 Flash Vision is MIT-licensed and available on HuggingFace. If US export control policy extends from GPU hardware to model weights — a regulatory risk that has been discussed in policy circles but not enacted — a large fraction of the open-weights inference thesis (H4) would be disrupted overnight. The model tracker shows DeepSeek available globally with no restricted regions currently. This is a tail risk with a non-trivial and rising probability given the trajectory of China AI policy and the November 2026 US election cycle.

R6 — Flash/SSD Import Surge Without DIO Corroboration $7.08B/month in US flash imports is an extraordinary volume — nearly 4.4x the simultaneous IC import figure. If the DIO data for distributors (ARW, SNX) when it populates shows DIO rising QoQ, this import surge will be reclassified from "inference infrastructure pull-through" to "channel stuffing in formation." That reclassification would be bearish for storage-adjacent infrastructure names (NetApp, NTAP; Pure Storage, PSTG; Western Digital, WDC) that are currently trading on the inference infrastructure capex theme. The next ARW/SNX quarterly DIO read is the single most important incoming datapoint for this risk.

Structural (6–24 Months)

R7 — Power Grid Constraint Becomes a 2027 Supply Ceiling The physical evidence for a structural grid constraint is unambiguous: the Monarch Compute Campus went off-grid specifically because Northern Virginia's transmission capacity is exhausted; TVA is repricing for large loads; Pennsylvania is imposing new data center restrictions. The construction tracker's top 10 builds represent over 15 GW of planned capacity against US grid additions of roughly 50-60 GW annually (all uses). New AI datacenter builds competing for interconnection queue slots against solar and wind additions will face 18-36 month interconnection delays in constrained regions. The power bottleneck is not a 2026 pricing driver — GPU rental prices today show no power-cost premium in constrained markets — but it will suppress new supply additions in 2027-2028, which is exactly the period when B200 and GB200 demand will peak. Projects currently UNDER_CONSTRUCTION but facing permitting or grid delays (Sail/Georgia, Waverly/Illinois) should be probability-weighted accordingly in any supply model.

R8 — The China Domestic Chip Ecosystem Is Maturing Faster Than NVIDIA's Revenue Model Assumes Cambricon's reported +108% growth, Moore Threads +147%, and Biren's +1,998% growth (from a small base) in the Chinese domestic AI accelerator market — sourced from Q3 2026 news analysis — signal that CUDA ecosystem lock-in is not permanent in China. NVIDIA's China datacenter revenue has already fallen from $25B (FY2025) to $19.7B (FY2026), a 21% decline with Q3 FY2027 guidance assuming zero China datacenter compute revenue. At 86x trailing P/E, NVIDIA's equity is pricing in significant growth from non-China markets absorbing more than 100% of the lost China volume. If Chinese domestic chips achieve 60-70% parity on CUDA-compatible workloads within 18 months — a plausible trajectory given the Biren/Moore Threads roadmaps — the global AI chip addressable market could contract meaningfully from its current consensus estimate.