Daily Investment Brief — June 13, 2026

GPU & Cloud Compute Intelligence Brief

June 13, 2026 | Automated Briefing — Scan → Hypothesize → Corroborate → Final

Market Pulse

The GPU spot market is bifurcating along a clean geographic fault line: North American and European markets with genuine supply constraints — Montreal, Stockholm, Frankfurt — are printing multi-tier pricing breakouts simultaneously, while US domestic and Indian markets sit in oversupplied equilibrium. The most definitive signal in today's data is Montreal's H100 three-tier convergence: SPOT at $4.98/hr (+121% in 30 days), ON-DEMAND repricing +23.6% overnight to $7.67/hr, and — most strikingly — 1-year RESERVED contracts jumping +42.5% in a single 24-hour window to $7.01/hr with spread expanding 916%. Reserved pricing doesn't move 42% in a day on noise; it moves when a provider fundamentally reassesses forward capacity availability. Against a backdrop of $130B in North American data center projects halted in Q1 2026 alone, 14-year interconnection queues in Northern Virginia, and a New York permit moratorium on facilities above 20MW, this is regulatory friction manifesting directly in price structure — not a transient anomaly. The dominant narrative is this: announced capex is not the same as delivered capacity, and the spread between Oracle's $638B remaining performance obligations and the physical reality of constrained interconnection queues is the central tension in infrastructure pricing for the next 12–24 months.

Key Movers

• = disregard as thin-market artifact. = high conviction signal. = monitor. = confirmed reversal.*

Investable Insights

H1 — Montreal Is the Canary: Three-Tier H100 Breakout Confirms Structural Canadian Supply Floor Confidence: 5 / 5

Thesis: The H100 market in Montreal has undergone a textbook supply exhaustion event across all three pricing tiers simultaneously — a combination that cannot be explained by catalog errors or thin-market artifacts. Spot tightening typically precedes on-demand repricing by days to weeks as providers drain spot pools before raising catalog rates; that sequence is now complete. The definitive confirmation is the reserved tier: a +42.5% single-day move in 1-year committed pricing, with spread expanding from 7.4% to 75.3% in 24 hours, signals that at least one provider has fundamentally revised its forward capacity model. Reserved contracts are not adjusted lightly — they anchor enterprise sales negotiations, committed-use discount calculations, and balance-sheet capacity planning. A 42% revision implies the provider's internal forecast of available Montreal H100 capacity over the next 12 months has materially deteriorated. The mechanism is structural: Quebec and Ontario face the same regulatory and energy friction ($130B in North American projects halted Q1 2026, NERC "Emerging Large Loads" guidelines tightening interconnection, 14-year queues in comparable metro markets) that prevents rapid supply normalization. The contrast with US domestic markets is the tell — Arizona, Illinois, and Warsaw all sit at $2.40–2.60/hr SPOT with zero spread and flat 30-day deltas, reflecting a rationed but comfortable equilibrium. Montreal at $4.98/hr SPOT with a 197% spread is a different market entirely, and the spread is still widening, not compressing. Stockholm is the shadow confirmation: at $3.50/hr SPOT (+67.3% over 30 days) with spread expanding from 58.7% to 73.3% and two providers, it is running the same playbook four to six weeks behind Montreal — the Nordic hydropower advantage attracts new entrants who then bifurcate the market before one exits or capacity gets absorbed. The cleanest near-term watch: if the Montreal RI spread holds above 75% for five consecutive trading days, the structural re-underwriting is confirmed and on-demand pricing has room to move another 15–20% to close the SPOT/OD inversion that currently exists ($4.98/hr SPOT vs. $7.67/hr OD — an inverted market where spot is pricing below on-demand, the rarest signal of inventory stress).

Key Evidence:

1. H100|SPOT|ca-montreal: $4.98/hr, 30d delta +121%, spread expanding from 26% → 197% over 30 days; 2 providers (live ticker, Jun 13, 2026)
2. H100|ON_DEMAND|ca-montreal: $7.67/hr, +23.6% in 24 hours, spread spiked from 10% → 36% intraday (live ticker, Jun 13, 2026)
3. H100|RESERVED_1YR|ca-montreal: $7.01/hr, +42.5% in 24 hours, spread expanded from 7.4% → 75.3% — a 916% intraday spread expansion (live ticker, Jun 13, 2026)
4. A100_40GB|SPOT|ca-montreal: +117% over 30 days — older flagship tracking the same tightening, confirming it is a market-wide Canada phenomenon, not H100-specific (live ticker, Jun 13, 2026)
5. H100|SPOT|se-stockholm: $3.50/hr, +67.3% over 30 days, spread widening 58.7% → 73.3%; same two-provider bifurcation signature at a lower absolute level (live ticker, Jun 13, 2026)
6. $130B in North American data center projects halted in Q1 2026 alone; 48 projects blocked totaling $156B in all of 2025 (Startup Fortune, Jun 12, 2026)
7. New York one-year moratorium on permits for data centers >20MW enacted June 5, 2026; 71% of Americans oppose local data center construction (Gallup, reported Jun 12, 2026)
8. Northern Virginia grid interconnection queue now stretching 14 years (Forbes, May 31, 2026) — benchmark for how far regulatory friction can delay supply

Implied Action:

• Long: Neocloud operators with existing Canadian capacity requiring no new permits — they are earning a structural rent that is widening, not compressing. The key screen is balance-sheet stability (see H5 risk); avoid highly leveraged names even with the right geographic exposure.
• Monitor (trigger): H100|RESERVED_1YR|ca-montreal spread holding above 75% for 5 consecutive days = full structural confirmation. H100|SPOT|se-stockholm reaching $4.00/hr = Stockholm breakout initiating.
• Avoid: Any neocloud announcing Canadian expansion contingent on new permitting or grid interconnection — they will miss timelines by 18–36 months and the rents will have compressed by the time they go live.
• Avoid: US domestic H100 spot positions (Arizona, Illinois, Warsaw) — these are efficiently priced single-provider markets with no near-term catalyst.

H2 — H100 Decay Is Already Priced Aggressively; GB300 Will Accelerate the Inevitable Confidence: 4 / 5

Thesis: The depreciation data reveals a structural divergence between AWS and Azure on H100 useful-life assumptions that is the largest matched-pair gap in the entire dataset — and the direction of the disagreement has direct implications for multi-year reservation decisions. AWS encodes a 22.3% annual decay rate for H100, implying a 4.5-year useful economic life and a 1-year reserved price of $41.28/hr vs. $63.30/hr on-demand. Azure, by contrast, prices H100 at only 16.1% annual decay (6.2-year life, 1yr RI at $62.92/hr — barely below OD), which is functionally a statement that Azure does not believe H100 will depreciate meaningfully within 12 months. That is a 40% divergence in useful-life assumptions between the two largest hyperscalers for the identical chip, and one of them is wrong. The generational escalation in the RI-implied data makes AWS's more aggressive assumption look correct: A100_40GB decays at 19.1% (AWS) → A100_80GB at 20.2% → H100 at 22.3% → H200 at 24.5% (Azure) → GB200 at 24.5% (Azure). Each successive generation enters the market with a higher baked-in obsolescence premium, because buyers increasingly price NVIDIA's own release cadence back into the asset. The GB300 is the catalyst that converts this from an abstract depreciation argument to a concrete capital allocation decision: NVIDIA's own June 12 benchmarks confirm GB300 NVL72 delivers 20x better agentic coding performance per megawatt versus H200. Extrapolating conservatively, this implies a 5–7x improvement in per-token economics for inference workloads versus H100. Any enterprise committing to H100 RESERVED_3YR today is locking in at a rate that embeds a chip becoming economically obsolete mid-contract. The cross-sectional regression provides the base rate: 17.5% annual decay across all accelerators (p=0.016, R²=0.23), with inference chips slower at 12.3% — but INFERENTIA's RI-implied decay of 20.8% and L40S at 25.3% both exceed the cross-sectional inference average, suggesting the inference market is not as structurally stable as the regression implies.

Key Evidence:

1. AWS H100 RI-implied decay: 22.27% annually — OD $63.30/hr, 1yr RI $41.28/hr; implied useful life 4.5 years (depreciation model, Jun 13, 2026)
2. Azure H100 RI-implied decay: 16.13% annually — OD $73.15/hr, 1yr RI $62.92/hr; implied useful life 6.2 years (depreciation model, Jun 13, 2026)
3. Generational escalation in RI-decay: A100_40GB 19.1% → H100 22.3% → H200/GB200 24.5% — each generation enters with faster expected obsolescence (depreciation model, Jun 13, 2026)
4. GB300 NVL72: 20x better agentic coding performance/MW vs H200 — NVIDIA Developer benchmark (NVIDIA Developer, Jun 12, 2026)
5. Cross-sectional decay regression: 17.5% annual across all chips (p=0.016, R²=0.23); inference slower at 12.3% but with high individual-chip variance (depreciation model, Jun 13, 2026)
6. AMD MI300X (Azure): only 14.6% annual decay — lowest of any training chip; either defensive Azure pricing or genuine inference flexibility premium (depreciation model, Jun 13, 2026)
7. GB300 launch 1.4 years ago; H200 2.6 years old already at 24.5% decay — NVIDIA's cadence is compressing chip lifetimes in real time

Implied Action:

• Avoid: H100 RESERVED_3YR commitments for training workloads. AWS's 22.3% decay rate already embeds significant risk; GB300 at scale within 18 months means the chip could be economically stranded mid-contract.
• Prefer: Shorter-duration commitments (SPOT or ON_DEMAND) on H100 for training, accepting the spot risk premium as insurance against obsolescence.
• Long inference infrastructure (qualified): The 12.3% cross-sectional inference decay supports longer commitments on dedicated inference silicon in principle — but validate at the chip level before committing. INFERENTIA's managed end-of-life and L40S's 25.3% decay rate mean not all "inference chips" are created equal.
• Watch: AMD MI300X pricing — if Azure's 14.6% decay assumption is a competitive subsidy rather than a genuine market view, it creates a long-duration inference arbitrage for buyers who can run on ROCm.
• Monitor: GB300 availability timelines — every quarter of delay in broad GB300 deployment extends H100's economic life and compresses the urgency of this thesis.

H3 — INFERENTIA EU: Custom Silicon Structural Lock-In, Not a Trading Signal — Yet Confidence: 3 / 5

Thesis: The AWS Inferentia2 market in Europe is exhibiting a specific pattern that is strategically interesting but not immediately actionable: completely frozen on-demand pricing across 18 global regions combined with a volatile spot signal in Frankfurt. The freeze is not a data artifact — Frankfurt at $0.434/hr, Paris at $0.422/hr, London at $0.451/hr, Tokyo at $0.425/hr, and all US regions in the $0.521–0.523/hr range show exactly 0.0% change across every time horizon. This is a deliberate AWS catalog management decision, consistent with a chip in a managed transition phase where AWS is extracting utilization from fully-deployed capacity without incentivizing new customer acquisition. The strategic logic is lock-in: enterprises that migrated workloads to Inferentia2 (bearing the Neuron SDK transition costs from CUDA) face significant switching costs, making them price-inelastic to modest OD rate changes. AWS has no incentive to cut prices (demand is captured) and no incentive to raise them (it would trigger re-evaluation of switching costs). The SPOT signal in Frankfurt (+317% over 30 days) is real but requires a thin-market qualifier: INFERENTIA SPOT appears to have near-zero consistent daily liquidity, meaning the +317% may represent a single day's observation reappearing after weeks of absence rather than sustained repricing. The macro context is unambiguously supportive of the broader custom silicon theme: the xPU market is projected at $77.7B in 2026 (+107.8% YoY, Futurum Group), MediaTek has doubled its AI ASIC revenue target to $2B, and Ubuntu's production-ready TPU VM images (Canonical, May 28) signal that custom silicon is reaching mainstream deployment tooling. INFERENTIA's 6.5-year active catalog life — while older chips like Gaudi2 (4.1 years) and A30 (5.2 years) have already exited — is the most operationally significant data point: AWS is keeping this chip in active global deployment well beyond what the RI-implied 4.8-year useful life would suggest, extracting rent from a fully-depreciated base.

Key Evidence:

1. INFERENTIA ON_DEMAND: $0.434/hr (Frankfurt), $0.422/hr (Paris), $0.451/hr (London), $0.425/hr (Tokyo) — all show 0.0% delta across 24h, 7d, and 30d horizons (live ticker, Jun 13, 2026)
2. INFERENTIA SPOT Frankfurt: +317% over 30 days — but zero consistent daily SPOT records; qualifies as thin-market observation (live ticker, Jun 13, 2026)
3. INFERENTIA RI-implied decay: 20.8% annually (AWS), implying 4.8-year useful life vs. 6.5-year active catalog life — AWS is running the chip beyond its own financial model's implied retirement date (depreciation model, Jun 13, 2026)
4. Gaudi2 (4.1yr old): is_active: false; A30 (5.2yr old): is_active: false; INFERENTIA (6.5yr old): is_active: true across 18+ regions — anomalous survival (depreciation/catalog model, Jun 13, 2026)
5. xPU market: $77.7B projected 2026, +107.8% YoY (Futurum Group, via news feed, Jun 13, 2026)
6. Ubuntu production-ready TPU VM images — Canonical announcement (May 28, 2026) — signals custom silicon reaching mainstream DevOps tooling
7. Google Gen-10 "Icefish" TPU: Samsung/TSMC split manufacturing confirmed; MediaTek AI ASIC revenue target doubled to $2B

Implied Action:

• Monitor, don't trade yet: Watch INFERENTIA|ON_DEMAND|de-frankfurt (currently $0.434/hr, flat) for the first positive delta — that is the signal AWS is ready to monetize the EU demand it is currently absorbing at frozen prices. First positive OD delta in Frankfurt = actionable entry for the custom silicon theme.
• Long Broadcom (AVGO): Dominant ASIC co-design partner for hyperscalers; the xPU adoption wave benefits AVGO regardless of which chip wins. This is the cleanest liquid proxy for the custom silicon trend.
• Watch MediaTek: Doubling ASIC revenue target to $2B positions it as the challenger in the ASIC design space; less liquid but higher beta to the theme than AVGO.
• Avoid: Treating the +317% SPOT move as a liquid signal — INFERENTIA SPOT is a sub-institutional thin market. The OD freeze is the real data.

H4 — APAC Fragmentation: Japan/Korea Demand-Pull vs. India Supply-Push — Four Distinct Markets, Not One Story Confidence: 3 / 5

Thesis: The APAC GPU market is not a single zone — it is four countries running four different market dynamics simultaneously, driven by sovereign AI policy investment rather than unified price competition. Japan and Korea exhibit demand-pull tightening (prices rising while spreads compress, meaning existing providers converge on a market-clearing price under genuine demand pressure). India exhibits supply-push loosening (prices falling while spreads also compress, new entrants establishing floor prices in an oversupplied market). Southeast Asia remains in early price discovery. The Japan signal is validated across multiple chip classes — V100_32GB SPOT +365% over 30 days, TPU_V6E SPOT +144% over 30 days, T4G entering the Tokyo spot market — a multi-chip, multi-vintage demand signal that points to Japan's AI infrastructure buildout absorbing compute across the stack, not just premium chips. Korea is the sharpest single signal in APAC: Seoul H100 SPOT has compressed from a 45.2% spread 30 days ago to 7.0% today, implying two providers are converging on a market-clearing price, and that spread is still tightening. When Seoul reaches zero or near-zero spread, one provider will effectively set the market price unilaterally — a near-monopoly condition that historically precedes sharp OD repricing. India inverts the thesis in instructive ways: Mumbai H100 SPOT at $4.66/hr, down 33% over 30 days from $6.92/hr, with a 2,712% spread (one provider quoting near $0.19/hr) confirms new entrants are dumping spare capacity. The Meta/Reliance Jamnagar announcement — 168MW Phase 1, backed by ~1GW of secured clean energy — explains the dynamic: infrastructure investment announcements attract new neocloud entrants before the demand is mature enough to absorb them. India spot will remain soft for 12–24 months until the Jamnagar campus comes online.

Key Evidence:

1. H100|SPOT|kr-seoul: $2.64/hr, +14.5% over 30 days, spread collapsed from 45.2% → 7.0% — near-monopoly dynamics forming (live ticker, Jun 13, 2026)
2. H100|SPOT|jp-tokyo: $2.65/hr, +11.4% over 30 days, spread compressing from 59.6% → 23.1% — genuine demand-pull tightening (live ticker, Jun 13, 2026)
3. V100_32GB|SPOT|jp-tokyo: +365% over 30 days — Japan absorbing aging compute across vintage tiers (live ticker, Jun 13, 2026)
4. TPU_V6E|SPOT|jp-tokyo: +144% over 30 days — GCP latest TPU hitting APAC capacity constraints (live ticker, Jun 13, 2026)
5. H100|SPOT|in-mumbai: $4.66/hr, -33% over 30 days from $6.92/hr; 2,712% spread with floor near $0.19/hr (live ticker, Jun 13, 2026)
6. Meta/Reliance Jamnagar: 168MW Phase 1, ~1GW clean energy secured — India AI campus announcement (CarbonCredits.com, Jun 12, 2026)
7. H100|SPOT|id-jakarta: $1.56/hr, +5.1% over 30 days, 100% spread — nascent price discovery, two providers establishing positions (live ticker, Jun 13, 2026)

Implied Action:

• Monitor Korea as leading APAC indicator: H100|SPOT|kr-seoul spread reaching zero (from current 7%) is a 2–4 week trajectory at current compression pace. When it does, the effective monopoly pricing in Korean SPOT creates upward OD pressure. Watch weekly.
• Underweight India spot: Mumbai H100 SPOT will remain soft for 12–24 months as Jamnagar supply comes online. Any neocloud pricing India capacity as a premium market in current presentations is misrepresenting the current dynamic.
• Japan as quality signal: The multi-chip, multi-vintage demand signal (V100 aging compute + TPU latest gen simultaneously tightening) is more credible than a single-chip move. Japan infrastructure investment is broad-based, not a FOMO bid on premium chips.
• Watch Jakarta: +5.1% with 100% spread and two providers is early-stage Montreal — much earlier on the trajectory, but the same two-provider bifurcation dynamic is present.

H5 — Neocloud Equity Froth vs. Spot Economics: CRWV's 738x Debt/Equity and the Beta Trap Confidence: 5 / 5

Thesis: The neocloud equity complex is running a leveraged carry trade on GPU spot rents that is structurally fragile in three specific ways: (1) the rent premium they capture is geographically concentrated and contingent on regulatory constraints that can normalize without warning; (2) the capex machine required to maintain growth is consuming capital at a rate that makes the business model a zero-FCF enterprise at scale; and (3) the SPAC/shell proliferation in the sector confirms late-cycle capital formation dynamics where institutional players have already priced the primary opportunity and retail is reaching for the next vehicle. CoreWeave's Q1 2026 fundamentals crystallize the contradiction: revenue of $2.08B in a single quarter (implying $8B+ annualized run rate), gross margins of 69.4% — genuinely world-class — yet burning -$4.71B in free cash flow in that same quarter against $7.70B in quarterly capex. The company is spending $3.70 in capex for every $1.00 of revenue. Debt stands at $35.1B against $2.3B in cash — a 738x debt-to-equity ratio. The Hut 8 Beacon Point deal (Jun 6) is the archetype: $4.25B borrowed at 6.129% for 352MW in Texas, backed by a $9.8B 15-year triple-net lease with an AA-rated tenant, generating $261M/year in interest before the facility even opens in Q3 2027. The entire model is a leveraged bet on H100/H200 spot rents remaining elevated — and our pricing data shows what happens when that bet goes wrong: Frankfurt H100 OD clears at $1.76/hr in a multi-provider efficient market vs. Montreal's $7.67/hr in a constrained two-provider market. The 4x spread between these two prices for the identical chip is a pure regulatory rent, not a productivity premium. When any one of the following occurs — a third provider enters Montreal, Stockholm normalizes, or GB300 at scale reduces the effective H100 utilization rate — that rent compresses. The high-beta neocloud equities (HUT beta 6.0, APLD beta 5.6, AIB beta 8.3, IREN beta 4.2) mean a 20% spot price compression translates to 100–160% equity drawdown for the most leveraged names. The five consecutive SPAC 8-K filings in four days (SharonAI, Spark I, WhiteFiber, Axiom, Axe Compute) — zero operating substance, zero successful scrapes — are the sentiment indicator confirming the cycle is still inflating at the retail margin even as CRWV trades 46% below its 52-week high of $187.

Key Evidence:

1. CRWV Q1 2026: Revenue $2.08B (run-rate $8B+), gross margin 69.4%, FCF -$4.71B, capex $7.70B — $3.70 capex per $1.00 revenue (equities fundamentals, Jun 13, 2026)
2. CRWV balance sheet: $35.1B total debt vs. $2.3B cash — debt/equity ratio 738x; EV/revenue ~26.4x on negative FCF (equities fundamentals, Jun 13, 2026)
3. CRWV EPS misses: -22% Q1 2026, -30% Q4 2025 — two consecutive quarters; analyst target range $36–$303 (a 740% spread) signals zero consensus on fair value (equities/SEC feed, Jun 13, 2026)
4. CRWV trading at $100.55 vs. 52-week high of $187 — already 46% below peak despite record revenue (equities feed, Jun 13, 2026)
5. Frankfurt H100 OD: $1.76/hr (multi-provider, efficient) vs. Montreal H100 OD: $7.67/hr (two-provider, constrained) — 4x spread for identical chip; the entire neocloud margin is this geographic rent (live tickers, Jun 13, 2026)
6. Hut 8 Beacon Point: $4.25B debt at 6.129% for 352MW Texas facility; $261M/yr interest expense before Q3 2027 opening; $9.8B 15-year triple-net with AA-rated tenant (intel feed, Jun 6, 2026)
7. Five consecutive SPAC/shell 8-K filings in 4 days: SharonAI (SHAZ), Spark I (SPKL), WhiteFiber (WYFI), Axiom Intelligence (AXIN), Axe Compute (AGPU) — zero scrape-able substance (SEC feed, Jun 13, 2026)
8. Neocloud beta profile: HUT 6.0, APLD 5.6, AIB 8.3, IREN 4.2 — 20% spot compression = 100–160% equity drawdown at these betas (equities feed, Jun 13, 2026)

Implied Action:

• Structural short thesis on high-beta neoclouds (HUT, AIB, IREN): These names are priced for sustained spot rent premium that is geographically contingent and supply-constrained. Any supply normalization event — third provider in Montreal, GB300 mass deployment, regulatory relaxation — collapses the thesis. Use defined-risk structures (puts, collars) given the beta profile.
• CRWV: Nuanced — not a simple short. The revenue trajectory and gross margin are real; the risk is the capex machine running out of road. Q2 earnings (Aug 12, consensus -$1.25 EPS) are the next binary. A third consecutive miss combined with guidance for accelerating capex is the stress trigger. The bear case analyst target of $36 is the scenario where spot prices compress before Beacon Point-style facilities finish construction.
• Long infrastructure picks-and-shovels: Vertiv (VRT, power and cooling), Corning (GLW, fiber), Celestica (CLS, networking switches, +345% in the period) — these capture the capex wave without the spot rent concentration risk.
• Avoid all micro-cap SPACs (AGPU, SHAZ, AXIN, SPKL, WYFI) — no revenue, no filings that scrape, betas of 3–8x, and the AGPU 52-week high of $32.10 vs. current $7.11 shows the retail momentum cycle has already peaked and reversed.
• Monitor: CRWV Q2 earnings August 12 as sector-wide catalyst.

Risk Flags

Immediate (0–30 Days)

R1 — Montreal Reserved Pricing Inversion Creates Enterprise Confusion The current state — SPOT at $4.98/hr, OD at $7.67/hr, and 1yr RESERVED also at $7.01/hr — is an unusual inverted structure where reserved pricing offers almost no discount to OD (7.67 vs. 7.01 = 8.6% discount vs. the typical 25–35%). This means enterprise customers comparing commitment tiers are receiving misleading pricing signals. If the RI spread (currently 75.3%) compresses as one provider corrects its pricing, the 24-hour RI move could partially reverse, creating a false normalization signal. Watch the RI spread daily; a reversal below 50% within 72 hours would indicate a catalog error rather than a structural repricing.

R2 — DEV/Thin-Market Noise Masking Real Signals The 24-hour mover list is dominated by single-provider DEV-tier listings with n_providers=1 and spread=0% — RTX Pro 6000 (+1,428%), M60 (+442%), V100_32GB Stockholm (+409% 24h, but this is the DEV tier distinct from the SPOT tier). These can create false urgency or mask real signals. The ALVEO_U30 Oregon +32,972% (30-day) and T4G Tokyo +9,927% (7-day) are in the same category. Analysts should filter any mover with n_providers=1 and spread=0% as a first-pass noise rejection.