Daily Investment Brief — June 7, 2026

GPU & Cloud Compute Daily Intelligence Brief

June 7, 2026 | For Pricing Analysts & Infrastructure Investors

Market Pulse

The compute market is entering a regime transition. Training GPU spot markets are bifurcating sharply by region — H100 SPOT in Montreal has rallied +140% over 30 days to $4.71/hr, compressing the spot-to-on-demand discount from 74% to 39% in a single month against a flat $7.68/hr on-demand ceiling, while Frankfurt has collapsed -27% to $1.79/hr (below on-demand, an inversion) and Dublin has shed -50% to $1.28/hr. This is not a global squeeze — it is a geographic demand concentration that is structurally more durable than a uniform tightening, because data-residency and latency constraints prevent customers from arbitraging across regions. Simultaneously, AWS is executing a deliberate floor reset on its proprietary ASIC stack (Inferentia SPOT +928% in Mumbai, +518% in Tokyo over 30 days), while inference GPU prices soften (A10 SPOT Ohio -44% in 7 days). The convergence of these three forces — training GPU regional tightening, AWS ASIC repricing, and energy/permitting becoming the binding supply constraint — marks the end of the GPU-supply-constrained era and the beginning of a power-access and proprietary-silicon era. Operators and investors still positioned for the prior regime face the most immediate risk.

Key Movers

• = thin-market noise; flag noted but not actionable. All prices are $/hr per GPU or per ASIC unit.*

Investable Insights

H1 — H100 Spot Discount Structurally Compressing: Neocloud Margin Squeeze With 45-Day Fuse Confidence: 4.5 / 5

Thesis: The H100 spot market in GCP/neocloud-concentrated regions is approaching a price-clearing event that could eliminate the spot arbitrage undergirding CoreWeave's infrastructure economics within the next 45–60 days. The mechanism is straightforward: on-demand prices are anchored flat (Montreal OD has shown 0% change over 30 days at $7.68/hr), while spot is racing toward that ceiling at a pace that implies convergence before Vera Rubin can deliver meaningful supply relief. At $4.71/hr today versus $1.96/hr 30 days ago, Montreal has consumed 35 of the 38 percentage-point gap remaining before full spot-to-OD convergence. The spread_pct explosion within Montreal spot listings (from 79% to 176%) signals that the two active providers are diverging sharply — a textbook pre-clearing fragmentation. Once spot exceeds ~$6.00/hr (78% of OD), the economics of contracting GPU capacity on the open market become indistinguishable from on-demand pricing. For CoreWeave, which finances its GPU fleet at market rates and contracts it out to hyperscalers at locked-in rates, this is a gross-margin vice. The only supply-side relief is Vera Rubin — confirmed for Q3 2026 delivery to all four hyperscalers with SK Hynix holding ~60-70% of HBM4 allocation (Tech Times, June 6) — but that timeline provides no H100 relief before August earnings. CoreWeave's Q1 2026 already showed the early stress: EPS of -$1.40 vs. estimate of -$1.20 (a -16% miss), gross margin at 65.5% (the floor of its guided range), and FCF burning at -$4.7B per quarter — accelerating, not stabilizing, even as the cost environment worsens. The stock at $100 is -27% from its May 6 close of $138 and -46% from its 52-week high of $187; the 52-week low of $63.80 is 36% below today.

Key Evidence:

1. Montreal H100 SPOT +140% in 30 days ($1.96→$4.71/hr), spot-to-OD discount compressed from 74% to 38.5% — mathematically 38 percentage points remain before full convergence (live ticker data, June 7)
2. H100 ON_DEMAND flat in every region — Montreal, Virginia, Ohio, London all showing ≤0.5% 30-day change — the OD ceiling is anchored, amplifying the spot squeeze (live ticker data, June 7)
3. Stockholm SPOT +66%, Ohio SPOT +40%, Madrid SPOT +31% over 30 days — multi-provider corroboration across 5 distinct markets confirms this is structural, not regional (live ticker data, June 7)
4. CoreWeave Q1 2026: EPS -$1.40 vs. -$1.20 estimate (-16% miss), gross margin 65.5%, FCF -$4.7B/qtr, total debt $35.1B vs. $2.3B cash (equities fundamentals data, June 7)
5. Vera Rubin delivery confirmed Q3 2026 (SK Hynix HBM4 certified, ~60-70% allocation) — the only supply catalyst, but provides zero relief before August 12 CRWV earnings (Tech Times, June 6)
6. Anthropic diversifying compute across CoreWeave, Google Cloud, and SpaceX Colossus campuses — CoreWeave's largest contracted customer is reducing concentration risk (Foreign Policy Journal, June 4)
7. Qualifying evidence: Frankfurt SPOT -27% and Dublin SPOT -50% over 30 days confirm the squeeze is NOT global — regional arbitrage exists for non-latency-constrained workloads, which limits the severity of full OD convergence in aggregate

Implied Action:

• Tactical short CRWV ahead of August 12 earnings, with $6.00/hr Montreal H100 SPOT as the trigger signal. Any Q3 guidance language citing "cost environment" or "spot market pricing" is the catalyst for a retest of the $63.80 52-week low
• Operators: lock in 1-year reserved H100 contracts immediately. Reserved OD is flat while spot is racing toward it — the window to capture the remaining arbitrage is 30-60 days at current trajectory
• Avoid: Pure-play neocloud positions with concentrated customer exposure (CoreWeave's APLD/CORZ tenant chain creates contagion risk — see Risk Flags)
• Monitor: H100|SPOT|ca-montreal weekly; Vera Rubin delivery confirmation or slip from NVIDIA/SK Hynix

H2 — AWS Executing a Two-Phase ASIC Pricing Reset — Inference GPU Margins Are Collateral Damage Confidence: 3.0 / 5

Thesis: AWS appears to have concluded the seeding phase of its Inferentia/Trainium adoption strategy and is resetting spot floor prices across multiple geographies in a coordinated 30-day window. The scale of the moves — Inferentia SPOT from sub-cent levels to $0.101/hr in Mumbai (+928%), $0.316/hr in Tokyo (+518%), and $0.095/hr in Ohio (+286%) — is inconsistent with organic demand discovery. These are deliberate floor resets by an operator with pricing authority over its own proprietary silicon. The strategic logic is clear: AWS spent 18-24 months using deeply discounted Inferentia/Trainium spot pricing to migrate workloads away from GPU instances; once that migration is achieved, the customer is locked in to the AWS-specific compilation toolchain (Neuron SDK) and can no longer trivially move back. The repricing then begins. However, confidence is held at 3.0 because (1) no corroborating AWS pricing page changes or customer announcements have been found in the news feed, (2) the A10 inference GPU market is not in freefall — it is frozen, with most regions showing zero change and only Ohio (-44%, 7d) and a handful of European markets showing deliberate markdowns, suggesting inference GPU commoditization is operating through the performance-normalization channel (L4→L40S generation gap implies 77.5% annual performance-normalized decay) rather than spot price collapse. The real and more immediate threat to GPU inference economics is the Google-Blackstone $5B TPU joint venture (Rolling Out, May 25), targeting 500 MW of TPU capacity within one year — Google's TPU is market-available inference silicon at hyperscaler scale, a more credible near-term competitor than Inferentia alone. The timeline for this hypothesis extends to 12-18 months, not 60-120 days as initially framed.

Key Evidence:

1. Inferentia SPOT: Mumbai +928% ($0.01→$0.101/hr), Tokyo +518% ($0.051→$0.316/hr), Ohio +286% ($0.025→$0.095/hr) — all within a single 30-day window (live ticker data, June 7)
2. A10 SPOT Ohio -44% in 7 days (now $0.111/hr); Paris -19% over 30 days — inference GPU softening in select markets (live ticker data, June 7)
3. Cross-sectional depreciation model: inference-class hardware at 12.6%/year (7.9-year useful life); performance-normalized decay for A10 at 77.5%/year — effective margin compression is through the denominator (depreciation data, June 7)
4. Google-Blackstone $5B TPU JV, targeting 500 MW within 12 months — the more credible near-term inference ASIC threat (Rolling Out, May 25)
5. Qualifying evidence: No corroborating news on AWS pricing page changes or customer migration announcements found; A10 market is frozen (not declining) in 25 of 28 regions — weakens the "coordinated trap" narrative

Implied Action:

• Underweight pure-GPU inference operators (Bit Digital BTBT -11% 5d, HIVE Digital HIVE -17% 5d) — market is reacting but may be only partially pricing the 12-18 month structural shift
• HIVE August 13 earnings: forward guidance on inference GPU utilization rates is the key disclosure — watch for language on inference vs. training revenue mix
• Long inference software/tooling that is ASIC-agnostic — companies building the MLOps and optimization stack that works across Inferentia, TPU, and GPU benefit from AWS's and Google's ecosystem lock-in plays without bearing the hardware commodity risk
• Monitor: AWS Neuron SDK release cadence as a proxy for Inferentia adoption depth; Google-Blackstone TPU site/power contract announcements as the 90-day confirmation trigger

H3 — Power Access Has Replaced GPU Access as the Alpha-Generating Scarcity Variable Confidence: 4.5 / 5

Thesis: The binding constraint for new AI compute capacity has silently shifted from GPU supply to power grid access, and the equity market has only partially priced this. The convergence of signals is the strongest cross-feed pattern in the dataset: NERC's new reliability guideline on "Risk Mitigation for Emerging Large Loads" (filed May 6, 2026) establishes the first formal compliance framework for datacenter-scale power draws, raising the barrier and cost of new grid interconnection. FERC's active working group on "Powering AI and Data Center Infrastructure" signals regulatory intervention in how datacenter power is priced is no longer hypothetical. Fourteen states are considering construction bans. Northern Virginia interconnection queues now run 14 years. ComEd is raising residential bills 12% citing datacenter load in Illinois. The IEA projects global datacenter consumption doubling to 945 TWh by 2030 — exceeding Japan's total consumption. The arithmetic is brutal: no combination of renewable build rates, grid interconnection timelines, and demand response programs can accommodate this ramp on the proposed timeline. Interconnection queues filed today determine 2028-2029 capacity — the scarcity is already locked in. Operators who hold secured, permitted power interconnections outside the most congested markets are sitting on an increasingly irreplaceable asset. The equity expression requires care: Applied Digital (APLD) is the most direct play but carries a critical concentration risk — its CoreWeave lease (150 MW at Ellendale, ND) means APLD is a derivative of CRWV financial health. The cleaner expression is Core Scientific (CORZ), which disclosed 590 MW of leased power capacity across 11 US facilities in its May 6 10-Q, is actively pursuing new site acquisitions to convert mining capacity to AI hosting, and has a more diversified tenant base that is not wholly dependent on CoreWeave's continued solvency.

Key Evidence:

1. NERC "Risk Mitigation for Emerging Large Loads" guideline — formal compliance requirements for datacenter power draws for the first time; filed May 6, 2026 (NERC, confirmed in intel feed)
2. Northern Virginia interconnection queue: 14 years — power access is the hard ceiling for the most critical US datacenter market (Forbes, June 2026)
3. IEA: datacenter consumption doubling to 945 TWh by 2030 — exceeds Japan's total national consumption (IEA, 2026)
4. ComEd raising residential rates 12% citing 80+ operating datacenters in northern Illinois; Florida SB 484 blocking residential cost-shifting — regulatory fragmentation of datacenter operating costs by state is accelerating (news feed, June 7)
5. CAISO interconnection queue report confirmed live (CAISO, May 7, 2026); Evergy Missouri IRP revision explicitly identifying datacenter demand as requiring forecast revisions (intel feed, May 8)
6. 14 states considering datacenter construction bans or pauses — leading edge of a permitting backlash that will restrict new supply geographically (news feed, June 7)
7. CORZ: 590 MW leased power, 11 US facilities — mining-to-AI-hosting conversion pipeline creates rapidly deployable power-permitted capacity (CORZ 10-Q, May 6, 2026)
8. APLD: 150 MW CoreWeave lease at Ellendale, ND — creates direct contagion exposure to CRWV financial health; stock -16% in 5 days, -10.3% intraday on June 7 (equities data, June 7)

Implied Action:

• Long CORZ as the cleaner power-scarcity expression: conversion economics, diversified tenant pipeline, active site acquisition strategy; watch the July/August earnings for new power contract announcements
• APLD re-examine on July 29 earnings — if CRWV lease is flagged as a risk or revenue decelerates, the -16% 5-day move may be just the beginning; if management reaffirms tenant diversification, the $66.77 analyst target vs. $39.62 current implies 68% upside. Do NOT chase before earnings clarity on CRWV exposure
• Datacenter REITs with multi-tenant, non-Virginia power positions are the structural long — NERC large-load compliance effectively raises the moat cost for new entrants, creating durable margin expansion for existing permitted facilities
• Monitor: New ISO interconnection queue reports (CAISO, PJM) for queue growth or timeline extensions; any M&A targeting permitted 500+ MW sites outside Virginia/Texas will signal institutional recognition of the scarcity

H4 — SpaceX as $24B/Year Neocloud Operator: Premium Compression Is a 12-24 Month Story Confidence: 4.0 / 5

Thesis: SpaceX's entry as a committed neocloud operator — Colossus-1 (220,000+ chips) entirely captive to Anthropic at ~$1.25B/month, Colossus-2 (110,000 GPUs) committed to Google at $920M/month through June 2029 — represents a $2B+/month ($24B+ annualized) run-rate that makes SpaceX effectively the second-largest neocloud behind CoreWeave virtually overnight. The mechanism that makes SpaceX structurally dangerous is not GPU count — it is Starlink's $4.5B in annual operating profit subsidizing $13B in 2025 datacenter capex plus $7.5B in Q1 2026 alone. No pure-play neocloud can match a capital structure where the AI infrastructure buildout is funded by a captive satellite internet cash flow. This creates a cost-of-capital advantage that will compress the premium neoclouds charge over hyperscaler on-demand rates over a 12-24 month horizon. Critically, the current SpaceX GPU fleet is NOT adding open-market supply — both Colossus deployments are take-or-pay captive arrangements. What it does is pull Google and Anthropic demand off the open market, reducing the demand pool competing for spot GPUs. The market confirmed the signal: SpaceX-Google disclosure drove CRWV -7.1% to $100.39 and NVDA -6.2% on 216M shares in a single session (MLQ.ai, June 6). The regional spot data contains a bearish divergence that confirms the supply fragmentation is already operating: Frankfurt H100 SPOT -27% and Dublin -50% over 30 days, while Montreal tightens. European regions where non-CoreWeave supply is more available are clearing at steep discounts — demand is already routing to the cheapest available option.

Key Evidence:

1. SpaceX Colossus-1: 220,000+ chips committed to Anthropic at ~$1.25B/month; Colossus-2: 110,000 GPUs committed to Google at $920M/month through June 2029 — combined $2B+/month run-rate (MLQ.ai, June 6)
2. SpaceX datacenter capex: $13B in 2025 + $7.5B in Q1 2026 alone, subsidized by Starlink's $4.5B annual operating profit — structural cost-of-capital advantage over pure-play neoclouds (MLQ.ai, June 6)
3. CRWV -7.1%, NVDA -6.2% on the SpaceX-Google disclosure — single-session equity confirmation (MLQ.ai, June 6)
4. Frankfurt H100 SPOT -27%, Dublin SPOT -50% over 30 days — European GPU oversupply in non-CoreWeave-concentrated markets confirms demand is routing to cheaper supply (live ticker data, June 7)
5. CRWV FY2025 revenue $5.13B vs FY2024 $1.92B (+116% YoY), but FCF -$8.56B TTM, debt/equity 738x — premium pricing is required to service this leverage; supply fragmentation is existential at this balance sheet (equities fundamentals, June 7)
6. Qualifying evidence: Both SpaceX GPU fleets are captive take-or-pay — they do NOT add open-market spot supply, limiting the direct pricing impact on the spot market we track

Implied Action:

• CRWV bear case on 12-24 month horizon: EV/Revenue 14x and EV/EBITDA 29x with $35B in debt and -$8.6B FCF requires sustained pricing premium that a Starlink-subsidized competitor structurally undercuts; analyst consensus at $140 mean (33 analysts) appears materially stale
• The bull-case hedge to size around: If SpaceX capacity remains entirely captive to Google/Anthropic with no open-market listings, the H1 tightening thesis (Montreal SPOT racing toward OD) can coexist with the H4 bear thesis — they operate on different timescales
• Monitor: SpaceX IPO filing — compute revenue disclosure would be the single most important new data point for pricing the neocloud sector's long-run competitive dynamics; CoreWeave Q2 revenue growth rate (August 12) — slowdown below 80% QoQ with margin compression would confirm thesis

H5 — Taiwan Concentration Risk: Unpriced Tail With 5-15% Probability Confidence: 3.5 / 5

Thesis: The AI GPU supply chain carries a geopolitical tail risk of unprecedented concentration that equity markets are pricing at approximately zero. Nvidia's Taiwan spending has surged from $10-15B annually five years ago to $150B/year today, with AMD adding $10B — $160B+ of AI GPU investment now runs through TSMC's ~90% monopoly on advanced AI silicon nodes. SK Hynix holding 60-70% of HBM4 allocation for Vera Rubin means even the memory stack for next-generation GPUs is concentrated in Northeast Asia. The intensity of black-market demand for this silicon is now confirmed: Taiwan authorities detained three Super Micro Computer executives for alleged document forgery on AI chip exports (TradingView, May 26), demonstrating that the scarcity premium is so severe that criminal channels have emerged. On-demand H100 prices are flat globally — if geopolitical risk were pricing in, you would see a risk premium in OD rates or reserved contract yields. It isn't there. The depreciation data confirms the duration of exposure: H100 remains active across AWS, Azure, and GCP at 4.21 years post-launch; A100 at 5.56-6.06 years. In a supply disruption, there is no domestic US substitution for 24-36 months minimum — Intel's "year-end AI datacenter chip launch" (MSN, June 6) is the only partial hedge, and it runs through TSMC's advanced nodes too. This is a 12-36 month structural tail, not a near-term trade.

Key Evidence:

1. Nvidia Taiwan spending: $150B/year (up from $10-15B five years ago); AMD committed $10B — $160B+ flowing through a single geopolitical chokepoint (Crypto Briefing, June 6)
2. TSMC ~90% monopoly on advanced AI silicon nodes — no domestic US equivalent at scale; TSMC Arizona N3 fab not expected until 2026-2027
3. Super Micro executives detained in Taiwan for alleged AI chip export document forgery — black market intensity confirms scarcity premium (TradingView, May 26)
4. Nvidia Vera Rubin introduced at GPU Technology Conference in Taipei — product launch geography deepens the Taiwan dependency narrative (multiple sources, June 2026)
5. H100 ON_DEMAND flat globally over 30 days — zero geopolitical risk premium embedded in current market pricing (live ticker data, June 7)
6. H100 catalog survival: 4.21 years post-launch, still active at AWS/Azure/GCP — duration of exposure is multi-year with no substitution option (depreciation data, June 7)
7. Qualifying evidence: "Silicon shield" deterrence theory is widely discussed in institutional media — awareness is high, but pricing still zero. Could remain unpriced indefinitely unless a trigger event occurs

Implied Action:

• Long NVDA puts with 18-24 month duration — the asymmetry is favorable: NVDA benefits from business-as-usual (monopoly deepens), loses catastrophically in disruption. Long-dated puts are the cleanest expression of tail risk that doesn't require calling timing
• Long physical GPU inventory operators (CoreWeave, APLD, HIVE) as a tail hedge only — their existing H100/A100 fleets become extraordinarily scarce assets in a supply disruption scenario, providing asymmetric upside that partially offsets their H1/H4 bear exposure in normal conditions
• Long TSMC (TSM) — benefits from both the silicon shield deterrence narrative and compounding demand; the risk is disruption of its own facilities, but TSM's capex moat deepens as every rival falls further behind
• Avoid: Portfolio construction with 100% positive correlation to a Taiwan event — GPU chips, hyperscalers, AI neoclouds, and datacenter REITs all move together in disruption. Diversification across this space provides no Taiwan hedge

Risk Flags

⏱ Immediate (0-30 Days)

R1 — Montreal H100 SPOT Approaching Price-Clearing Event The 38.5% remaining spread between Montreal SPOT ($4.71/hr) and the OD ceiling ($7.68/hr) is closing at approximately 35 percentage points per month. If the current rate holds, full convergence is 33-45 days away. At $6.00/hr (78% of OD), the neocloud GPU arbitrage model mathematically breaks — operators contracting GPU capacity at near-OD spot prices cannot profitably resell at contracted rates. Watch this ticker weekly. A CRWV pre-announcement or earnings pull-forward would be the equity signal; spot breaching $6.00/hr is the market signal.

R2 — Frankfurt H100 SPOT at/Below On-Demand: Market Inversion Frankfurt H100 SPOT at $1.79/hr vs. on-demand of $1.78/hr is a market inversion — spot pricing has breached the OD floor. This is either a single provider clearing below cost (unsustainable) or the OD price in Frankfurt is itself anomalous. Either way, it signals a distorted market where normal arbitrage logic has broken down. Do not use Frankfurt as a reference price for European GPU cost modeling without understanding the provider identity behind the anomaly.

R3 — Thin-Market Artifacts Masking Real Signals Multiple movers in the 24-hour scan (L4 Virginia SPOT +101%, AMD MI25 Texas +86%) are single-provider, thin-inventory events with no demand signal content. The risk is that automated systems or models treating these as demand signals generate false positive positions. Explicitly filter n_providers=1 and volume < 5 observations before acting on any 24-hour mover.

Near-Term (30-90 Days)

R4 — CRWV-APLD-CORZ Contagion Chain A contagion pathway discovered in the stress test: CoreWeave leases 150 MW from APLD (Ellendale, ND) and 590 MW from CORZ across 11 facilities. If H100 SPOT convergence compresses CoreWeave gross margin below 60%, CoreWeave's debt service capacity ($35.1B in debt, $2.3B cash, FCF -$4.7B/qtr) comes under stress. The second-order effect is potential lease renegotiation pressure on APLD and CORZ — which the equity market has not priced as correlated. APLD's -16% 5-day move and -10.3% intraday on June 7 suggests the market is beginning to price it; CORZ has not moved as sharply. Watch APLD July 29 earnings for any CRWV-related revenue guidance revision — that's the trigger for the contagion read-through.

R5 — Vera Rubin Delivery Slip Is a Double-Edged Catalyst Vera Rubin delivery to AWS, Azure, GCP, and Oracle is confirmed to begin Q3 2026 with SK Hynix HBM4 certified (Tech Times, June 6). A slip in this timeline — from HBM4 yield problems, TSMC N3 capacity constraints, or SK Hynix allocation reductions — would be simultaneously bullish for H100 spot prices (no relief valve) and bearish for every company that has made capex commitments contingent on Vera Rubin throughput. The HBM4 supply chain is the single most critical component to monitor: SK Hynix holds 60-70% of allocation. Any yield warning from SK Hynix on HBM4 should be treated as an immediate H100 SPOT bull catalyst and a CRWV Q3 cost risk.

R6 — Google-Blackstone TPU JV: 500 MW in 12 Months Is Aggressive Google's $5B joint venture with Blackstone targeting 500 MW of TPU capacity within one year (Rolling Out, May 25) is the most underappreciated near-term threat to GPU inference economics. If achieved, this introduces a hyperscaler-quality inference alternative at Blackstone-financed scale into the market in 2027 — before most GPU inference operators have depreciated their A10/T4 fleets. Watch for TPU JV site selection and power contract announcements as the 90-day confirmation signal. The absence of any announced site implies aggressive timeline risk; a site announcement would accelerate the inference GPU commoditization timeline from H2's thesis.

Structural (6-24 Months)

R7 — Energy as a Hard Supply Ceiling: The 2028 Capacity Lock-In Interconnection queues filed in 2026 determine 2028-2029 capacity. Northern Virginia's 14-year queue is the extreme case, but CAISO, PJM, and MISO queues are all extending as datacenter demand accelerates. The IEA's 945 TWh by 2030 projection cannot be met on current grid build rates. The risk is not that demand exceeds supply in 2030 — it is that the capacity constraints are already locked in and the market has not priced the operating cost implications for new datacenter entrants. States passing residential cost-shifting bans (Florida SB 484) and construction pause legislation will create differentiated operating cost structures that are not reflected in current cloud pricing, which is set nationally or globally. Watch for NERC large-load compliance deadlines materializing in interconnection filings — these will be the first observable market signal of the constraint becoming binding.

R8 — Taiwan Concentration Deepening as Demand Accelerates The $160B+ annual investment concentration in Taiwan is not stable — it is growing. Nvidia's Taiwan spend is up 10x in five years; AMD's $10B commitment is new. Every dollar of AI GPU capex that flows to TSMC deepens the concentration. The TSMC Arizona N3 fab (targeting 2026-2027) is the only credible partial hedge, but it will serve a small fraction of global demand. Intel's year-end AI chip target (MSN, June 6) runs through TSMC's advanced nodes. The risk compounds over the 12-36 month window: as more AI workloads move onto next-generation silicon (Vera Rubin, Blackwell Ultra), the fraction of global AI compute dependent on TSMC's specific facilities increases rather than decreases. Portfolio managers with AI infrastructure sector exposure should stress-test their Taiwan scenario assumptions against a 30-day, 90-day, and 12-month supply disruption horizon.