Daily Investment Brief — June 3, 2026

Market Pulse

The AI compute market is bifurcating in real time: premium GPU capacity is tightening from the bottom up while legacy storage and low-tier inference hardware deflates simultaneously. H100 SPOT prices have completed a confirmed multi-region trough — Ohio's floor of $1.52/hr on April 23 now looks clearly established, with 13 consecutive days of recovery to $2.27/hr — yet ON_DEMAND at $7.73/hr hasn't moved in 30 days, leaving a 3.4x spread that becomes increasingly unstable as spot demand builds. Simultaneously, AWS's proprietary ASIC spot market (Inferentia, Trainium) recorded the most extreme price surges in the entire dataset over 30 days — +856% in Mumbai, +306% in Ohio — signaling genuine absorption of idle ASIC spot capacity, not a data artifact. The macro backdrop is unambiguous demand confirmation: Alphabet's $80B equity raise (citing supply lagging demand), a $16B OpenAI/Oracle Michigan groundbreaking, and the Anthropic/SpaceX Colossus deal locking 220,000 H100/H200/GB200 GPUs into a private bilateral agreement — all removing chips from the public spot market precisely as spot prices are recovering. The binding long-run constraint is now visibly shifting from silicon to power: 14-year grid interconnection queues in Northern Virginia, NERC's first draft reliability guideline targeting AI data center loads, and 14 states weighing moratoriums define a regulatory environment that creates durable pricing power for operators sitting on existing interconnected, permitted power.

Key Movers

Noise filter: ON_DEMAND_DEV tier moves (A100 Iowa +678%, M60 Zurich +730%, V100 Oregon +401%) reflect single-provider catalog recalibrations with near-zero observation counts. Do not trade on these. The GB200 Seoul 24h spike (+319%) reverted within the measurement window. Similarly, the Amsterdam H100 SPOT at $0.071/hr (near-zero) is a single-provider deeply discounted SKU, not a market-clearing price.

Investable Insights

H1 — H100 SPOT Reflation: The Floor Is In, Spread Compression is the Trade Confidence: 4 / 5

Thesis: H100 SPOT prices have completed a confirmed multi-region trough and entered a momentum-driven recovery regime. The Ohio floor of $1.52/hr on April 23, 2026, was followed by 13 consecutive daily closes moving higher — an unbroken linear ascent to $2.27/hr today, a +49% move in six weeks. Montreal provides the more dramatic corroboration: from a $7.46/hr peak in early January, prices cascaded to a floor of $1.85/hr on May 11 as a second, cheaper provider entered and disrupted the market — then recovered +122% to $4.10/hr in just 23 days, with spread widening to 127% confirming the cheaper provider is still present but being outbid. Stockholm reinforces the pattern independently, recovering +57% over 30 days to $3.29/hr from a February trough around $1.72/hr — notable because European capacity was the anticipated supply overhang that was supposed to suppress prices everywhere. It isn't.

The mechanism is clear: Alphabet's $80B equity raise explicitly cited demand exceeding supply (Data Center Knowledge, June 2, 2026). The Anthropic/SpaceX Colossus 1 deal locked 220,000 H100/H200/GB200 GPUs into a private bilateral contract, removing them from public spot markets (IndexBox, June 2, 2026). The OpenAI/Oracle Michigan $16B groundbreaking adds another large private cluster (MLive.com, June 2, 2026). These events represent a structural withdrawal of chips from public markets precisely as enterprise adoption is accelerating — the classical supply-demand squeeze that precedes spot-to-ON_DEMAND spread compression.

The ON_DEMAND "wall" is the trade's crux: Ohio ON_DEMAND is frozen at $7.73/hr with zero 30-day delta, creating a 3.41x spot-to-ON_DEMAND spread. Historically, wide spreads of this kind resolve in one of two directions: spot rises toward ON_DEMAND (demand tightening), or spot falls back (oversupply). With every regional signal pointing upward except Frankfurt (where a single-provider discount is anchoring EU prices lower at $1.79/hr, a -27% 30d move), the probability distribution strongly favors spread compression from below.

Key Evidence:

1. Ohio H100 SPOT: confirmed floor $1.52/hr on April 23, recovery to $2.27/hr (+49%) in 13 unbroken daily closes (live ticker data, June 3, 2026)
2. Montreal H100 SPOT: $1.85/hr floor May 11 → $4.10/hr today (+122% in 23 days), spread at 127% — demand outpacing cheapest provider's capacity (live ticker data, June 3, 2026)
3. Stockholm H100 SPOT: +57% 30d to $3.29/hr — EU supply overhang narrative is not manifesting in prices (live ticker data, June 3, 2026)
4. H100 ON_DEMAND Ohio: $7.73/hr, 0% 30d delta, 2 providers, range $5.97–$9.50/hr — static wall awaiting spot compression (live ticker data, June 3, 2026)
5. Alphabet $80B equity raise citing demand exceeding supply for AI services (Data Center Knowledge, June 2, 2026)
6. Anthropic/SpaceX Colossus: 300MW, 220K GPUs locked bilateral — effective supply withdrawal from public spot market (IndexBox, June 2, 2026)
7. Qualifying evidence — Frankfurt SPOT at $1.79/hr (-27% 30d): Single-provider discount acting as a regional gravity anchor. If European workloads route through Frankfurt to seek lowest cost, the EU reflation thesis partially stalls. Does not invalidate US/Canada/Nordic recovery but caps the breadth.

Implied Action:

• Primary trade: Long CRWV (CoreWeave, $119) — H100 SPOT reflation directly expands margin on spot-priced capacity. Revenue +111% YoY to $6.2B; analyst high target of $303 implies near-3x. August earnings is the repricing catalyst if the spot trend holds. Risk: -$8.6B TTM FCF, 0.315 current ratio — balance sheet requires continuous capital access.
• Confirmation gate: H100 SPOT Ohio sustaining >$2.50/hr on the next 7-day observation locks in the trend. Watch Dublin and Oregon spreads (currently 3,700–18,000%): when the cheapest floor provider lifts, the market median jumps.
• Deny signal: Frankfurt SPOT recovering to >$2.00/hr (currently the exception) would actually be bullish — its persistence at $1.79/hr is the primary risk to the thesis.

H2 — Broadcom ASIC Earnings: Structural Rotation Catalyst, Today Confidence: 4 / 5

Thesis: The +856% Inferentia SPOT surge in Mumbai and +306% Trainium surge in Ohio are not pricing algorithm artifacts — they are market-level confirmation that AWS's own ASIC utilization has risen to the point where less capacity spills into spot. This is the behavioral fingerprint of structural ASIC adoption acceleration, and Broadcom's Q2 FY2027 earnings (reporting today, June 3) is the single highest-time-sensitivity catalyst in this cycle for validating or disrupting that thesis. Goldman Sachs has explicitly forecast ASIC demand will surpass GPU demand in the coming years; today's print is the first live stress test of that forecast against reported numbers.

The fundamental setup is exceptional. AVGO has delivered ten consecutive EPS beats without exception, with the most recent (Q1 FY2026, March 4, 2026) posting a +29.8% positive surprise ($1.95 reported vs. $1.50 estimated). Quarterly revenue has accelerated from $14.9B to $19.3B in the most recent print (+29.5% YoY), gross margin expanded to 68.1%, and free cash flow hit $8.0B — the highest in company history. Today's EPS consensus of $2.40 represents a 23% step-up from the last reported quarter — a high bar, but one set by a business whose AI custom silicon revenue is now material enough that Marvell's 10-Q explicitly flags its own competing custom ASIC (XPU) segment as a material revenue concentration risk (Marvell 10-Q, Q1 FY2027, June 2026). The ASIC shift is visible across multiple independent data sources: AWS spot pricing, Marvell's disclosure, and Broadcom's own beat cadence.

Broadcom's key hyperscaler customers — Google (TPU), Meta (MTIA) — are simultaneously signaling accelerating ASIC deployment: Google is doubling CapEx this year (embedded in Alphabet's $80B raise) and broke ground on a new Sweden AI data center with a €5M local skills fund to expedite permitting. Each new Google facility is a marginal unit of custom silicon demand that flows directly to AVGO's P&L. The outstanding question — whether a third major hyperscaler (Apple or Microsoft, both rumored) has signed a custom ASIC program — would be the single-sentence catalyst for a meaningful re-rate beyond the $486 mean analyst target.

Key Evidence:

1. AVGO: ten consecutive EPS beats, most recent +29.8% surprise ($1.95 vs. $1.50 est.) on March 4, 2026 — zero misses in the trailing 10-quarter dataset (equities fundamentals data)
2. AVGO Q2 FY2026 revenue $19.3B, gross margin 68.1%, FCF $8.0B — highest in company history (equities fundamentals data)
3. Today's EPS estimate $2.40 at 25.96x forward PE, 1.0x PEG — "fair value for growth," not speculative (equities fundamentals data)
4. AWS Inferentia SPOT Mumbai: +856% 30d to $0.105/hr; Trainium SPOT Ohio: +306% 30d to $0.102/hr — simultaneous absorption across two chip families and three geographies confirms demand event, not algorithm noise (live ticker data, June 3, 2026)
5. Marvell 10-Q Q1 FY2027: custom ASIC (XPU) hyperscaler segment flagged as material revenue concentration — independent corroboration of structural ASIC shift (SEC filing, June 2026)
6. Google Sweden AI data center groundbreaking + €5M local skills fund — marginal hyperscaler ASIC demand signal (Intel feed, June 2026)
7. Qualifying evidence: AVGO is up +14% in a single week into the print, trading at $481.57 within 1.5% of its 52-week high of $488.82. Expectations are fully loaded; any guidance caution triggers a reversal on reported beats.

Implied Action:

• Primary trade: Long AVGO through today's earnings. At $481 with a $630 analyst high target, the structural ASIC thesis has 31% further upside on earnings delivery — without needing multiple expansion, only EPS compounding. Position sizing should reflect the 14% pre-run risk.
• Re-rate trigger: Any commentary confirming a third major hyperscaler custom ASIC design win (beyond Google/Meta). This would be a step-change catalyst, not a marginal beat.
• Post-earnings structural expression (lower time-sensitivity): Long MRVL as the competing ASIC custom silicon play — its 10-Q disclosure of hyperscaler concentration signals it is a Broadcom revenue share beneficiary regardless of who wins specific programs.
• Deny signal: AVGO guiding Q3 AI revenue below the implied $12–13B run-rate, or any commentary about a hyperscaler "pause" in custom silicon ordering, would disrupt both the ASIC pricing thesis and the Inferentia/Trainium spot surge interpretation simultaneously.

H3 — The Converted-Miner Pipeline: Fastest Path to New Compute Capacity Confidence: 3 / 5

Thesis: The fastest way to add AI compute capacity in a market defined by 14-year grid interconnection queues, NERC reliability rulemaking targeting AI data center loads, and 14-state legislative pressure is to repurpose infrastructure that already has grid connections. Cipher Digital's disclosure of long-term anchor contracts with both AWS and Google (SEC 8-K, June 2026) is the proof-of-concept: a converted cryptocurrency miner with existing high-power infrastructure becomes a hyperscaler colocation facility faster and cheaper than any greenfield alternative. This is not an anomaly — it is a repeatable playbook, and the capital flowing into this category (CIFR +54% YTD, BTDR +148% over 3 months, HIVE +116% over 3 months) reflects the market beginning to price it.

The structural case is simple: a Bitcoin mining facility with 50–200MW of existing grid access, cooling infrastructure, and power purchase agreements is worth dramatically more as AI colocation than as a miner — and hyperscalers, facing internal build timelines of 3–5 years per greenfield site, are paying the premium to prove it. Microsoft's own disclosures confirm it added more cloud capacity in the last 18 months than in Azure's first decade. BlockchAIn Digital Infrastructure's S-1 ($60M raise, filed June 2, 2026) signals new capital is explicitly organizing around this theme. The HIVE Digital simultaneous 10-K and 8-K filing (FY ending March 31, 2026) is the most immediate unresolved catalyst: an 8-K filed alongside a 10-K almost always contains a material event, and in HIVE's context, an anchor tenant announcement is the most likely disclosure.

The basket differentiates on quality: APLD ($47.86, strong_buy, $64.59 analyst target, +83% YTD) is the highest-quality expression — larger, longer-dated contracts, next earnings July 29. HIVE ($4.54, +116% 3m, 6.9x forward PE) is the asymmetric speculative play on the unresolved 8-K. CIFR ($26.29, +54% YTD, confirmed hyperscaler contracts) has the least upside remaining — only 20% to mean target, with sub-1-year cash runway as a binary risk. BTDR ($18.76, +148% 3m) is deeper on the pivot but loss-making.

Key Evidence:

1. Cipher Digital (CIFR): confirmed long-term data center anchor contracts with AWS and Google; stock +54% YTD, +9.5% on the disclosure day (SEC 8-K, June 2026)
2. HIVE Digital: simultaneous 10-K + 8-K filing (FY ended March 31, 2026) — material event pending; stock at $4.54, 6.9x forward PE, +116% over 3 months (equities data, June 3, 2026)
3. APLD: $47.86, $64.59 analyst target (+35%), strong_buy consensus, +83% YTD; North Dakota/Texas panhandle siting in power-abundant states (equities data, June 3, 2026)
4. PJM Northern Virginia: 14-year interconnection queue (Forbes, via Intel feed); ComEd +12% residential bills from data center load — structural barrier advantaging existing connected infrastructure
5. 14 states weighing data center moratoriums; Maine legislature passed a ban (governor vetoed) — regulatory fragmentation compresses greenfield site supply (CNBC, May 9, 2026; Intel feed)
6. BlockchAIn Digital Infrastructure S-1: $60M raise (Intel feed/SEC, June 2, 2026) — new capital explicitly organizing around miner-to-datacenter thesis
7. Qualifying evidence: CIFR sub-1-year cash runway confirmed in SEC filing — binary capital markets risk. Alphabet's $80B raise suggests hyperscalers prefer ownership over leasing at scale; the addressable market for third-party converted miners may compress over a 12-month horizon.

Implied Action:

• Basket construction: APLD (largest weight — quality + July 29 earnings catalyst), HIVE (speculative weight — 8-K binary), CIFR (small weight — upside capped, cash risk real).
• Primary catalyst to watch: HIVE 8-K content. If it discloses an anchor tenant (hyperscaler or frontier AI lab), it re-rates from 6.9x to at least 15–20x forward PE, implying a $12–15 price target vs. $4.54 today.
• Hedge: Short legacy colo/hosting names without AI transition traction (e.g., secondary Equinix/Digital Realty assets in Northern Virginia submarkets facing the 14-year queue ceiling).
• Avoid: CIFR as a primary position — the confirmed hyperscaler contract is already priced at 39x forward PE, and the cash runway risk creates a binary downside that the 20% upside-to-target doesn't compensate.

H4 — Power as Durable Alpha: Grid-Constrained Supply is the Structural Moat Confidence: 4 / 5

Thesis: The binding constraint on AI compute capacity has shifted from silicon to power, and this transition is structural rather than cyclical. The evidence is convergent across pricing data, regulatory filings, and equity markets. The H100 SPOT geographic cross-section is the cleanest single proof: Dublin at $1.27/hr and Iowa at $1.24/hr (wind-dominated, hyperscaler mega-campus states) are the cheapest SPOT markets globally, while Mumbai at $5.35/hr, São Paulo at $9.30/hr, and Dubai at $2.88/hr — power-constrained or grid-uncertain regions — command sustained premiums. This is not latency or provider competition explaining the spread: it is power cost and reliability being mechanically priced into GPU compute through the decomposition model's base cost subtraction.

The regulatory acceleration makes this durable. NERC's draft Reliability Guideline on Risk Mitigation for Emerging Large Loads (NERC, May 2026) is the first federal-level acknowledgment that AI data center loads are a grid reliability threat — if enacted, it increases interconnection compliance costs and slows approval timelines further. Northern Virginia's 14-year queue is the most extreme outcome to date. Florida SB 484 (already enacted) bars utilities from passing data center costs to residential customers and empowers local governments to deny projects — a direct mechanism for compressing new capacity in a key US market. The geographic implication is a structural wedge: operators in Iowa, Oregon, Ohio, and Nordic Europe have a durable cost and permitting advantage that cannot be replicated in Virginia, California, or APAC without multi-year lead times.

Nebius Group (NBIS, $260.58, +200% over 3 months) is the single equity name most directly monetizing this premium. Building pan-European AI infrastructure in Nordic markets with abundant hydro/wind power and permissive zoning, NBIS trades at a 0.63 PEG ratio — rare for a hypergrowth name — suggesting that despite the 3-month rally, the market has not fully priced the European power-availability premium against North American grid constraints. Applied Digital (APLD) captures the same dynamic domestically: sited in North Dakota and Texas panhandle, +83% YTD, with a $64.59 analyst target implying 35% further upside.

Key Evidence:

1. H100 SPOT power-cost correlation: Iowa $1.24/hr, Dublin $1.27/hr, Oregon $1.70/hr (abundant power) vs. Mumbai $5.35/hr, São Paulo $9.30/hr (constrained) — near-perfect geographic correlation (live ticker data, June 3, 2026)
2. NERC draft Reliability Guideline targeting AI data center interconnection behavior — frequency/voltage trip requirements could impose retrofit costs and slow queue approvals (NERC, May 2026)
3. Northern Virginia grid interconnection queue: 14 years — the most extreme bottleneck in the dataset (Forbes, via Intel feed)
4. ComEd Illinois: +12% residential electricity bills attributed to data center load — demand-driven rate pressure in Midwest (Intel feed)
5. 14 states considering data center legislation; Florida SB 484 enacted; Maine ban passed (governor vetoed) — regulatory fragmentation actively compressing new capacity geographies (CNBC, May 9, 2026)
6. NBIS (Nebius Group): $260.58, +200% 3m, 0.63 PEG, "buy" consensus — most direct equity expression of EMEA power-availability premium (equities data, June 3, 2026)
7. APLD: +83% YTD, $64.59 target (+35% from $47.86), power-advantaged siting in ND/TX (equities data, June 3, 2026)
8. Qualifying evidence: Faster-than-expected SMR/modular nuclear permitting (NRC) could partially break the bottleneck by 2028. Nvidia's DSX software platform promises 40% density improvements from Vera Rubin, which increases effective compute output per watt — reducing (but not eliminating) the power constraint.

Implied Action:

• Primary long: NBIS — the European power-availability premium play at 0.63 PEG with a 200% 3-month run that still has structural tailwinds. Position sizing should reflect the post-run valuation: confirm contracts of Colossus/Anthropic magnitude are in the pipeline before sizing up.
• Domestic long: APLD — highest quality power-advantaged neocloud, next earnings July 29, clear analyst target at $64.59.
• Regional pricing trade: Any neocloud operator building in Iowa, Oregon, or Ohio is structurally advantaged over Virginia/Texas peers; track ON_DEMAND pricing spreads between US_EAST_VIRGINIA and US_MIDWEST_IOWA — if Virginia develops a sustained premium, it confirms the bottleneck thesis in pricing data.
• Structural short: Datacenter REITs with concentrated Northern Virginia exposure — the 14-year queue caps incremental growth ceiling and new supply timelines. No near-term equity names specified (not in dataset), but the thesis is monitorable via public REIT disclosures.
• Deny signal: NRC granting a first commercial SMR operating license in 2026–2027 would be a significant near-term negative for the power scarcity thesis.

Risk Flags

Immediate (0–30 Days)

R1 — Frankfurt H100 SPOT Divergence: The European Floor Anchor Frankfurt SPOT at $1.79/hr is down -27% over 30 days — a single provider cutting prices to fill capacity in a market where every other European region is reflating. The risk is behavioral: if enterprise workloads routing through Europe begin treating Frankfurt as the cost-minimizing on-ramp for H100 capacity, it creates a gravity anchor that slows the Ohio/Montreal/Stockholm reflation. Monitor Frankfurt n_providers: if it stays at 1, the discount is idiosyncratic and contained. If a second provider enters at the Frankfurt price level, the European reflation thesis stalls. Threshold to watch: Frankfurt SPOT crossing back above $2.00/hr (reversal) or dropping below $1.50/hr (capitulation that draws arbitrage demand).

R2 — AVGO Pre-Run Risk: 14% in a Week Into a Print Near 52-Week High Broadcom enters its Q2 FY2027 earnings at $481.57 — within 1.5% of the 52-week high of $488.82, up 14% in a single week. The structural ASIC thesis is sound, but the near-term risk is asymmetric in the negative direction: even a modest in-line beat (vs. positive surprise) could trigger profit-taking. Any guidance softness on AI custom silicon revenue trajectory — order pause, customer timeline slip, VMware integration drag — would create a sharp reversal despite reported numbers that would look strong in isolation. The 93x trailing PE leaves no room for negative surprise. Position sizing on the earnings trade should be calibrated to the pre-run, not just the conviction in the structural thesis.

R3 — CIFR Cash Runway: Binary Capital Markets Risk Cipher Digital's sub-1-year cash runway (confirmed in the June 2026 SEC filing) at $26.29/share — only 20% below its $31.60 analyst mean target — is a classic distress/upside binary. If capital markets tighten (risk-off event, credit spread widening), CIFR cannot execute its hyperscaler colocation contracts and the equity goes to near-zero irrespective of the contract value. The +9.5% one-day move suggests the market is not pricing this risk adequately. Avoid large positions; use as a confirmation vehicle only.

Near-Term (30–90 Days)

R4 — Vera Rubin Production Ramp: H100 Demand Displacement Horizon Nvidia has confirmed Vera Rubin mass production is underway (Jensen Huang's visit to SK hynix for HBM4 supply confirmation, Ad Hoc News, June 3, 2026), with the chip promising 40% more GPU density and 5x inference throughput over H100. The depreciation clock for H100 started ticking at production ramp, not at retail availability — hyperscalers place orders and reroute capex now. The 6–9 month window to Vera Rubin commercial availability is the period of maximum H100 pricing power, but within that window, enterprise buyers who can wait will pause H100 ON_DEMAND commitments and pull forward GB200 planning. Watch for any cloud provider blog post announcing Vera Rubin/GB200 NVL availability dates — that announcement, not actual delivery, will be the H100 ON_DEMAND pricing inflection point.

R5 — GB200 Single-Provider Catalog Pricing: Competition Could Shock Any Region Every GB200 SPOT market globally (18 regions, zero exceptions) shows a single provider, zero spread, and near-zero 30-day delta. This creates concentrated risk in both directions: the moment AWS, Azure, or GCP announces competing GB200 pricing in any of these regions, it triggers a 20–40% price correction in that region's catalog as competitive pricing discovery begins. The Seoul SPOT at $12.26/hr and Ohio SPOT at $13.54/hr are the most likely first-mover markets given existing provider infrastructure. Monitor n_providers weekly on GB200|SPOT|us-ohio and GB200|SPOT|kr-seoul — a change from 1 to 2 is the regime-change signal.

R6 — A10 Inference Tier Softening: Leading Indicator of Mid-Tier Margin Compression The A10 Oregon SPOT at $0.085/hr, down -24% over 30 days, may be the earliest signal of structural inference-tier commoditization. The A10 has been the workhorse for mid-market AI serving (image generation, LLM inference at moderate scale). If Intel's Crescent Island (LPDDR5, air-cooled, lower power envelope, Times of India, June 2, 2026) begins qualifying at cloud providers over the next 12 months, it adds a third competitive lane (alongside Nvidia and AMD) specifically targeting this price/performance segment. Neocloud operators with heavy A10 exposure and near-term contract renewals should be monitoring this carefully — a further -15–20% move would compress margins materially before a renewal window.

Structural (6–24 Months)

R7 — Power Grid Bottleneck: Risk of Legislative Crystallization The current regulatory environment (14 states weighing moratoriums, NERC rulemaking, 14-year Virginia queue) is at an inflection point: it could crystallize into permanent supply constraints (beneficial for existing operators, as analyzed in H4) or trigger a federal pre-emption response that fast-tracks interconnections and eliminates the moat. The NERC guideline is still in draft — if the final rule is diluted under industry lobbying (the data center sector has significant political weight), the compliance burden decreases and the constraint loosens. Conversely, if any state moratorium is enacted and survives judicial review (Maine's was governor-vetoed; a different state's governor might sign), it creates a legal template that spreads rapidly. The bifurcation is binary: structural moat or regulatory resolution. Track NERC comment period close and any state legislative session conclusions through August 2026.

R8 — Hyperscaler Insourcing Compresses Neocloud Addressable Market Alphabet's $80B equity raise signals a preference for owning compute rather than leasing it from neocloud operators. If the three major hyperscalers (Google, Amazon, Microsoft) collectively bring 5–10 GW of new owned capacity online by 2027 (consistent with announced capex trajectories), the addressable market for neocloud/converted-miner colocation shrinks proportionally. The Cipher Digital and HIVE models depend on hyperscalers choosing to lease rather than build — an assumption that is structurally correct in the near term (build timelines too long) but becomes less defensible as hyperscaler-owned capacity reaches operating scale. The APLD/NBIS equity thesis is more durable (direct enterprise customer relationships, not hyperscaler-dependent) and should be weighted accordingly.