Daily Investment Brief — June 2, 2026

Market Pulse

The dominant structural event in today's data is a geographic bifurcation in H100 on-demand and spot pricing driven by H200 supply routing: US core markets (Iowa, Ohio, California) are being actively undercut by H200 on-demand — Ohio H200 at $6.29/hr versus H100 at $7.94/hr, California H200 at $6.06/hr versus H100 at $8.82/hr — while European and select APAC markets, starved of competitive H200 supply, are repricing H100 sharply upward. Frankfurt H100 on-demand has surged from a $1.78/hr floor six days ago to $3.58/hr today on a 1,603% spread blowout; Stockholm H100 spot is up +83.7% from its February trough to $3.16/hr and accelerating +24.6% in the last seven days alone. These are not independent local market events — they are two expressions of the same supply-routing shock, and they are hardening. Layered beneath the GPU price action, power infrastructure is transitioning from soft constraint to structural ceiling: a confirmed NERC draft reliability guideline targeting data center load, ComEd's June +12% rate increase flowing directly into Illinois H100 pricing (+13.3% over 30 days), and a 14-year Northern Virginia interconnection queue collectively cap US supply growth in ways that will persist for years, not quarters.

Key Movers

Thin-market noise advisory: All AMD MI25 tickers (10 markets, Frankfurt to Santiago) are catalog re-entry artifacts from a single provider (likely Oracle Cloud) bulk-listing legacy 2017 inventory. Zero spread, zero prior history, zero market relevance. Do not interpret as demand signals. Similarly, all V710 tickers are n_providers=1, zero-spread developer-preview listings — they evidence Intel cloud channel activity but not commercial production traction.

Investable Insights

H1 — H100 EU/APAC Spot Premium: H200 Supply Routing Creates a Durable Regional Arbitrage

Confidence: 4 / 5

Thesis: H200 on-demand supply is landing preferentially in US core markets, where it is actively undercutting H100 on-demand pricing — a supply-routing dynamic that is simultaneously depressing US H100 prices and creating acute scarcity pressure in European and select APAC markets that lack competitive H200 inventory. The mechanism is structural, not cyclical: H200 CoWoS/HBM3e supply is physically concentrated in US-proximate cloud regions first, leaving EU operators unable to offer H200 as a demand pressure valve. With Frankfurt H200 on-demand locked at $13.09/hr by a single provider with zero spot discount and zero 30-day price movement, European buyers face a binary choice between ultra-cheap remnant H100 capacity ($0.40/hr floor from what appears to be AWS overflow) and expensive incumbent H100 pricing above $6.77/hr — creating the 1,603% spread visible today. London's market structure is equally revealing: H100 on-demand was frozen at exactly $8.21/hr for four months — a hallmark of single-operator reservation pools — before a structural event on June 1 broke the freeze, dragged the median to $10.06/hr, and blew the spread to 61.7% as a new expensive-tier provider entered. These are not random pricing events; they are the anatomy of markets transitioning from thin/single-provider to contested on rising demand.

The EU spot picture provides independent corroboration. Stockholm H100 spot has traced a precise V-shape: peak at $4.05/hr on January 11, relentless compression to a $1.72/hr trough by February 28 (supply normalization), then methodical re-expansion over 60 days to $3.16/hr today — with the final leg ($2.53/hr → $3.16/hr) occurring in just seven days as a new high-price-tier provider entered at $3.76/hr. Madrid H100 spot at $4.02/hr (+47.3% over 30 days) is the dataset's most expensive EU spot price, with a 109% spread confirming a two-tier market, not noise. Montreal SPOT at $3.97/hr (+77% over 30 days) while Montreal on-demand fell 14.9% is the definitive structural inversion signal: new on-demand supply entered (confirming supply exists) while spot markets tightened on real demand, suggesting spot buyers have different time-sensitivity profiles and are willing to pay above on-demand rates for immediate availability.

One important qualification: H200 cheap-floor spot pricing in Stockholm ($1.77/hr floor, n_providers=2, 634% spread) means technically substitutable H200 capacity exists in that market at near-H100 prices. Users who can tolerate H200 spot volatility could sidestep the H100 premium — this partially caps the arbitrage ceiling.

Key Evidence:

1. Frankfurt H100 ON_DEMAND median surged from $1.78/hr (Apr 16–May 31 plateau) to $3.58/hr on June 2, with spread expanding to 1,603%; H200 ON_DEMAND in same region held exactly flat at $13.09/hr with zero-spread, n_providers=1 — confirming H200 is not a competitive pressure valve in EU (live ticker data, June 2, 2026)
2. H200 ON_DEMAND is undercutting H100 ON_DEMAND in three US core markets simultaneously: Ohio ($6.29 vs $7.94), California ($6.06 vs $8.82), Iowa ($7.73 vs $4.66 [H100 already compressing]) — the supply routing inversion is measurable (live ticker data, June 2, 2026)
3. Stockholm H100 SPOT 90-day history confirms V-shape: $4.05/hr peak → $1.72/hr trough (Feb 28) → $3.16/hr today, +83.7% from trough; final 7-day acceleration (+24.6%) driven by new provider entering at elevated price tier (ticker history, June 2, 2026)
4. Madrid H100 SPOT at $4.02/hr (+47.3% 30d), Montreal SPOT at $3.97/hr (+77% 30d) — two independent EU/Americas markets exhibiting the same upward repricing pattern (live ticker data, June 2, 2026)
5. Qualifying factor: Singapore H100 ON_DEMAND at $1.86/hr (-43.3% 30d) shows APAC is not uniformly tightening — supply-adequate APAC markets are commoditizing. The EU premium is specifically about grid-stable regions with lower AWS/multi-provider saturation.

Implied Action:

• Long European neocloud operators with H100 spot inventory in EU markets (particularly operators visible in Frankfurt and London multi-provider spreads) — they are capturing a regional price premium unavailable to US peers
• For enterprise procurement: Lock in EU H100 on-demand commitments now, before the Frankfurt/London spread normalizes upward. H200 on-demand parity in European markets appears 60–90 days away at minimum
• Trigger to exit: H200 spot listings appearing in Frankfurt or London with n_providers ≥ 2 would signal supply normalization and collapse the H100 EU premium — watch these tickers weekly
• Monitor: H100|GPU|ON_DEMAND|gb-london spread trajectory over the next 7 days — spread compression downward (expensive provider winning) confirms premium is hardening; further blowout suggests the market is still contested

H2 — Broadcom AVGO: Binary Catalyst Tomorrow as Multi-Silicon Narrative Matures

Confidence: 4 / 5

Thesis: Intel's Computex Crescent Island announcement (480GB LPDDR5X, 350W air-cooled, targeting LLM inference KV-cache workloads) and the simultaneous catalog-seeding of Intel V710 GPU instances in Oracle Cloud (Texas ON_DEMAND_DEV at $1.28/hr, Virginia at $0.68/hr, both with zero spread and single provider) are structurally correct as directional signals but 12–18 months early as investment events. The V710 tickers are classified as ON_DEMAND_DEV — developer-preview instances, not production compute — and reserved 1YR/3YR V710 pricing across all regions is declining slightly (Texas 1YR: $0.98/hr, -0.6% over 30 days; Virginia 1YR: $0.70/hr, -0.3%), the definitive tell that customers are not yet willing to commit capital to Intel GPU workloads. Intel itself confirmed sampling begins H2 2026 with no mass production timeline; CUDA ecosystem lock-in is acknowledged in Intel's own materials as the primary barrier. The direct Intel trade has no near-term revenue catalyst.

The investable beneficiary is Broadcom (AVGO), which is positioned as the infrastructure layer that wins regardless of which non-Nvidia GPU architecture emerges. The thesis has three independent legs: (1) OpenAI signed a confirmed multi-year Broadcom deal for co-developed accelerators and networking (Crypto Briefing, June 1, 2026), (2) analyst models now project Broadcom AI revenue exceeding $30 billion annually by FY2027 — a number that is likely to receive guidance language tomorrow, (3) Susquehanna maintained a Positive rating ahead of earnings specifically citing "custom XPU momentum." Broadcom's 3.5D XDSiP packaging platform serves the inference-specific hardware shift that Intel's own Computex materials quantify as a transition from 4:1 GPU:CPU (training) to 1:1 (inference), a ratio change that expands the custom silicon addressable market. Marvell's 10-Q (May 28) flagged gross margin pressure at 58.9% from lower-margin custom AI silicon — positioning Broadcom as the premium custom ASIC alternative at higher margins.

Key Evidence:

1. AVGO trading at $459.97 (+2.95% today), PEG 0.97, forward P/E 24.8x — cheapest AI infrastructure pure-play relative to earnings growth in the sector (equities feed, June 2, 2026)
2. 29.3M shares traded vs. 23.8M 90-day average — elevated volume heading into earnings; 52-week high at $465.92 (equities feed, June 2, 2026)
3. OpenAI multi-year Broadcom deal for co-developed accelerators and networking confirmed (Crypto Briefing / news feed, June 1, 2026)
4. V710 reserved pricing declining (-0.3% to -0.6% 30d) while on-demand is up — zero customer commitment to Intel GPU workloads despite catalog seeding (live ticker data, June 2, 2026)
5. Intel Crescent Island: sampling H2 2026, no mass production timeline, no named cloud provider partnership announced — 12–18 month monetization gap (Computex press coverage, June 1–2, 2026)
6. Risk factor: AVGO at 52-week highs with strong positioning means a guidance miss on AI ASIC revenue would create a swift pullback from an already elevated entry point

Implied Action:

• Long AVGO into June 3 earnings — the PEG of 0.97 implies AI ASIC growth is not fully priced at current multiples; strong AI revenue guidance would push through the $465.92 52-week high
• Do not initiate Intel (INTC) long on the Crescent Island narrative — the V710 developer-preview data confirms the timeline is 2027, not 2026; the CUDA moat remains the structural barrier
• Post-earnings: If AVGO guides AI custom silicon revenue >$25B FY2027, treat as confirmation signal for the multi-silicon non-Nvidia infrastructure thesis and hold through Crescent Island's H2 2026 sampling announcements
• Exit signal: If Broadcom guidance explicitly avoids quantifying AI custom silicon revenue trajectory, the thesis requires reassessment

H3 — Grid Cap Becoming Hard Ceiling: Powered Capacity Compounds Pricing Power

Confidence: 4 / 5

Thesis: The data center power constraint has crossed a qualitative threshold in the past 30 days: it is no longer a warning about future scarcity, it is an active mechanism already transmitting into compute pricing. The confirmation is not theoretical — it is visible in the spread between H100 and H200 pricing in Illinois. US Illinois H100 on-demand has risen +13.3% over 30 days to $12.84/hr in the ComEd service territory, the same grid region where ComEd explicitly cited data center load as the driver of a +12% residential rate increase effective June. The H200 in Illinois rose only +0.2% in the same period — implying H200 operators in that region have locked long-term power purchase agreements or co-located generation while the H100 operator is exposed to marginal tariff increases. This is a natural experiment confirming that grid cost exposure is differentially transmitting to older hardware deployments first, accelerating the depreciation curve for H100 relative to H200 within the same grid territory.

The regulatory superstructure is hardening simultaneously. NERC's confirmed draft guideline ("Risk Mitigation for Emerging Large Loads," physically confirmed on nerc.com/globalassets) instructs transmission owners and planners to obtain frequency and voltage trip settings from large load entities — this is the grid regulator treating hyperscale compute as a systemic reliability risk, with binding rule-making next. FERC's 2025 State of the Markets Report explicitly tracks data center capacity share by region, signaling migration from passive monitoring to active framework-building. The 14-year Northern Virginia interconnection queue, Florida SB 484 (prohibiting utility cost pass-through to residential customers, granting local governments denial authority), and 14 states considering construction bans collectively constitute a structural cap on US data center supply that cannot be resolved in less than 5–7 years absent breakthrough SMR deployment.

The operational consequence — "AI workload flexibility," geographic shifting of training jobs, DC-connected microgrids — is not a future scenario; it is already appearing in enterprise procurement discussions, and it is the structural explanation for Stockholm (+49.7% spot over 30 days), Seoul (+11.1%), and Madrid (+47.3%) tightening simultaneously while US Iowa spot idles at $0.03–$1.24/hr with a 7,178% spread. US excess capacity exists but is grid-inconvenient; international capacity is scarce but grid-stable.

Key Evidence:

1. Illinois H100 ON_DEMAND +13.3% over 30 days to $12.84/hr vs. Illinois H200 +0.2% in the same period — differential power cost passthrough in ComEd territory confirmed as a natural experiment (live ticker data, June 2, 2026)
2. NERC draft guideline on "Risk Mitigation for Emerging Large Loads" physically confirmed at nerc.com/globalassets — binding reliability standard in development, not advisory (intel feed, June 2, 2026)
3. ComEd June +12% rate increase explicitly attributed to data center load from 80+ operating facilities in northern Illinois (energy news feed, June 2, 2026)
4. Northern Virginia interconnection queue: 14-year wait for new grid connections cited in Forbes (energy news feed, June 2, 2026)
5. Florida SB 484 signed into law — prohibits utility cost pass-through to residential customers, grants local governments data center denial authority; template legislation with 14 states watching (government/regulatory feed, June 2, 2026)
6. GE Vernova (GEV) at $950/sh, +7.9% over 3 months, buy consensus — market pricing grid equipment scarcity into the primary power infrastructure supplier (equities feed, June 2, 2026)
7. Critical risk factor: Dominion Energy (D) at $64.61/sh, -3.48% today, -4.52% over 5 days — potential NextEra/$66.8B acquisition rumor creating noise; sustained Dominion decline would signal deal cooling and grid-narrative re-evaluation

Implied Action:

• Long GE Vernova (GEV) as the grid buildout beneficiary with pricing power structurally insulated from GPU commoditization — GEV revenue is gated by regulatory approval timelines, not GPU generations
• Long FPS (Forgent Power Solutions) (+56.6% over 3 months, +17.6% over 5 days, PEG 0.61, strong_buy consensus) as a smaller-cap power equipment proxy whose data center power exposure remains underpriced relative to growth
• Data center REIT screen: Identify REITs with fully powered, grid-connected capacity in non-Virginia US markets (Arizona, Ohio, Chicago, Dallas) and grid-stable EU markets — these assets carry an embedded optionality premium the REIT market may not yet have fully priced
• Watch Dominion (D) as the real-time signal: Any sustained move above $68 should be treated as M&A confirmation; a break below $62 signals the NextEra deal is cooling and reduces the power-scarcity premium on the eastern seaboard
• Hedge: If aggressive SMR deployment announcements materialize in Northern Virginia within 12 months, the queue relief timeline shortens and the constraint thesis softens — currently the market assigns near-zero probability to this within 36 months

H4 — EU/APAC Spot Premium Durable: Grid Routing Creates 12–18 Month Structural Window

Confidence: 3.5 / 5

Thesis: The geographic spot price premium in EU and select APAC H100 markets is not a transient pricing anomaly — it reflects a structural demand-routing consequence of US grid congestion that will persist for 12–18 months minimum. The workload flexibility strategy explicitly named in energy-sector reporting ("geographic shifting of training jobs," "peak-period load reduction") creates a new pool of geographically mobile demand that will preferentially route to regions with grid-stable, low-latency capacity. Stockholm, Madrid, Montreal, and Seoul are the beneficiaries; Iowa, Oregon, and Dublin (which still has abundant ultra-cheap AWS remnant capacity) are the release valve.

The Stockholm 90-day history is the cleanest analytical instrument for this thesis. The compression phase (January peak $4.05/hr → February trough $1.72/hr) was supply normalization as new provider capacity came online; the re-expansion phase (February → June, +83.7% from trough) is the demand re-routing signal. Crucially, the final acceleration from $2.53/hr to $3.16/hr occurred in just seven days, driven by a new high-price-tier provider entering at $3.76/hr — a provider willing to bid $3.76/hr for Stockholm H100 spot capacity where spot was $1.72/hr ninety days ago. This is not opportunism; it is a deliberate commitment to a geographically constrained capacity position.

One important structural caveat requires honest acknowledgment: US Iowa H100 spot has a floor at $0.03/hr with a ceiling at $2.44/hr (7,178% spread). The constraint is not that cheap US H100 spot capacity is unavailable — it is that the grid-queue for new permanent capacity in constrained US markets creates a long-term supply ceiling that is distinct from day-to-day spot availability. Spot buyers can access Iowa's $0.03/hr floor today; but operators cannot build additional Iowa capacity within 7–14 years given current interconnection timelines. The EU premium is therefore most durable for capacity reservation and long-term contract vehicles, less so for pure spot arbitrage against Iowa's abundant existing surplus.

Key Evidence:

1. Stockholm H100 SPOT: $1.72/hr trough (Feb 28) → $3.16/hr (June 2), +83.7% in 60 days; 7-day acceleration +24.6% on new provider entry at $3.76/hr; spread 1,545% confirms two-tier market forming (90-day history + live ticker, June 2, 2026)
2. Madrid H100 SPOT $4.02/hr (+47.3% 30d), Seoul H100 SPOT $2.64/hr (+11.1% 30d, spread compressing from 35.7% to 20.7% — market maturing toward cohesion) (live ticker data, June 2, 2026)
3. US Iowa H100 SPOT floor at $0.03/hr, ceiling at $2.44/hr, 7,178% spread — abundant existing cheap US capacity coexisting with $3–4/hr EU premiums confirms the constraint is structural (new supply creation), not current availability (live ticker data, June 2, 2026)
4. Energy news confirms "AI workload flexibility" and "geographic shifting of training jobs" is an explicitly named enterprise operational strategy (energy news feed, June 2, 2026)
5. European Commission AI gigafactory initiative confirmed as active policy (energy news feed, June 2, 2026) — indigenous EU compute demand growing independently of US routing
6. Key risk factor: H200 SPOT in Stockholm shows a cheap-floor provider at ~$1.77/hr (n_providers=2, spread 634%) — users who can tolerate H200 spot volatility can substitute at near-H100 prices, which caps the H100 spot premium ceiling in Stockholm specifically

Implied Action:

• Long EU capacity reservation vehicles — the structural window is 12–18 months before US grid constraints begin resolving or H200 supply reaches EU spot markets competitively
• Buy EU H100 spot capacity forward while the grid narrative is still being absorbed by the enterprise market — Madrid and Montreal represent the highest conviction positions based on spread maturity (lower spread = more cohesive, more reliable market pricing)
• Avoid Dublin as a spot premium play — the 18,231% spread (floor $0.12/hr, ceiling $2.53/hr) reflects bimodal market structure that has not yet resolved; the cheap AWS tier will dominate Dublin volume and suppress the median
• Monitor: Centrica's DC-connected microgrid deployments (15–20% efficiency gain cited in energy news) as the leading indicator for grid-bypass technology scale — if microgrid deployments accelerate in EU markets, the supply bottleneck dissolves within 24 months and the EU premium collapses

Risk Flags

Immediate (0–30 Days)

R1 — Broadcom Earnings Miss (June 3): AVGO is at 52-week highs ($465.92) with elevated volume (29.3M vs 23.8M average) heading into tomorrow's earnings. A guidance shortfall on AI custom ASIC revenue — specifically any failure to confirm the OpenAI accelerator deal's revenue contribution timeline — would create a swift reversal from a crowded long position. The binary nature of this catalyst makes the pre-earnings long position high-risk/high-reward. Size accordingly.

R2 — Frankfurt/London Spread Collapse: The H100 ON_DEMAND Frankfurt spread at 1,603% and London at 61.7% could resolve either upward (expensive provider wins share, premium hardens) or downward (cheap-floor provider floods the market, premium collapses). If the cheap-floor provider (likely AWS overflow at $0.40/hr) aggressively scales volume in Frankfurt over the next 7 days, the $3.58/hr median collapses back toward $1.80/hr, invalidating the EU premium thesis. Watch the spread direction daily.

R3 — Constellation Energy (CEG) Selloff as Proxy for Power Narrative Re-rating: CEG declined -7.7% today and -15.1% over one month without an obvious catalyst — a significant move for a utility. If the market is beginning to reprice the "nuclear-for-AI" premium, it would be a leading indicator for broader power-scarcity infrastructure re-ratings, potentially suppressing GEV and FPS valuations and weakening the grid-constraint investment thesis. The lack of a specific catalyst for CEG's decline warrants close monitoring; if the -15% monthly trend persists into the next week, the power infrastructure long thesis requires reassessment.

Near-Term (30–90 Days)

R4 — H200 Spot Listings Appearing in EU Markets: The most direct killer of Hypothesis 1. If H200 spot capacity with n_providers ≥ 2 appears in Frankfurt, London, or Amsterdam within 60 days — particularly at prices competitive with H100 ($6–8/hr) — it removes the pressure differential that is driving H100 EU premiums. Currently H200 Frankfurt is locked at $13.09/hr by a single provider with zero spot discount; any crack in that control would be the signal. Set a monitoring alert on H200|GPU|SPOT|de-frankfurt spread expansion.

R5 — Intel Crescent Island Benchmark Underperformance at H2 2026 Sampling: Intel's 480GB LPDDR5X architecture makes a credible theoretical case against HBM-constrained Blackwell inference deployments. But if sampling-phase benchmarks reveal memory bandwidth or interconnect throughput bottlenecks that undermine the KV-cache superiority claim, the entire multi-silicon narrative deflates. This would negatively impact Broadcom's custom ASIC positioning indirectly (market confidence in non-Nvidia silicon broadly) and reinforce Nvidia's CUDA moat. CUDA ecosystem lock-in is already Intel's own stated primary barrier — sampling benchmarks are the moment of truth.

R6 — Dominion Energy (D) Deal Resolution: Dominion at $64.61/hr with -3.48% today is behaving like a stock under M&A uncertainty, not a stable utility. The rumored NextEra/$66.8B acquisition, if confirmed, would be transformative: it concentrates AI power infrastructure ownership in the most critical US data center market and validates the grid-scarcity monetization thesis. If the deal falls apart, it likely triggers a Dominion re-rating downward, removes the power-concentration risk premium from NoVA data center assets, and weakens the pricing power thesis for grid-connected operators east of the Mississippi. Treat any D volume spike >3x average as the binary signal.

Structural (6–24 Months)

R7 — H100 Depreciation Cascade as H200 Supply Normalizes: H200 is already undercutting H100 on-demand in three US markets. São Paulo H200 on-demand is down -33.4% over 30 days — H200 price normalization is not hypothetical, it is in progress. When H200 breaks below $10/hr in any major market (currently tracking toward this in California at $6.06/hr and Ohio at $6.29/hr), it will trigger a repricing cascade for H100 that has not been priced into neocloud operator valuations. APLD (+71.6% 3m), BTDR (+138% 3m), and other GPU cloud operators carrying H100 on-demand inventory are vulnerable to a -20 to -30% pricing haircut in their core product over the next 12–18 months. The H200 SPOT discount of -42.4% to -64.6% below on-demand in Iowa and California respectively signals that H200 owners are already desperate to move excess inventory — this is the early tremor before the earthquake.

R8 — Regulatory Fragmentation Across State Lines: Florida SB 484 is now signed law. Maine's legislature passed a construction ban (governor vetoed). Fourteen states are actively considering similar legislation. The policy risk is not that one state acts — it is that regulatory heterogeneity across states creates an unpredictable patchwork where capacity expansion planning becomes a legal minefield. Data center operators who have committed capex in states currently considering bans face potential stranded asset risk that cannot be hedged easily. The NERC reliability guideline, once finalized as a binding standard, will create the federal floor — but the ceiling of state-level restrictions remains uncapped.

R9 — CoWoS/HBM Packaging Bottleneck Resolution Timeline: Nvidia's TSMC partnership expansion (cuLitho, cuEST) and TSMC's confirmed N2P node capacity expansion are accelerating H200/B200 supply timelines. If CoWoS packaging constraints resolve faster than the market expects — potentially by Q1 2027 — H200 supply floods into European markets within that window, collapsing both the H100 EU premium (Hypothesis 1) and the US grid-routing scarcity thesis (Hypothesis 4). The packaging bottleneck is currently the structural governor of the bifurcated pricing regime; its resolution is the single highest-impact risk to the entire H1/H4 thesis complex. Monitor TSMC quarterly CoWoS capacity guidance as the leading indicator.

Brief generated June 2, 2026. All prices from live cross-provider ticker data. Prices are $/hr per GPU unless otherwise specified. This brief is for analytical and informational purposes.