Daily Investment Brief — July 31, 2026

GPU & Compute Market Intelligence Brief

July 31, 2026 | Finalized Analysis

Market Pulse

The compute market is in simultaneous training scarcity and inference expansion — two dynamics that rarely coexist at this intensity. The clearest signal is a 200-day price peg on H100 ON_DEMAND in Virginia at $8.69/hr: not a ceiling, not a floor in the conventional sense, but an administered price held steady across a two-provider market while demand continues to absorb every available unit. That stability is now transmitting upstream into predecessor hardware — A100_80GB spot in Oregon and Virginia is up +36% and +44% over 30 days — and downstream into geographic expansion, where inference-class L4 and T4 instances are appearing in new regions (Dubai, Stockholm, Milan, São Paulo) at prices that carry 2–3x premiums over established US markets. The structural narrative is a supply-demand mismatch across every tier of the stack, held in place by two slow-moving constraints: the AMD MI450/Helios delivery gap (H1 2027) and a power interconnection queue that regulators themselves acknowledge is being systematically underestimated.

Key Movers

Investable Insights

H1 — H100 Price Floor: The Most Durable Peg in Cloud History

**Confidence: 5 / 5 **

Thesis: The H100 on-demand market in Virginia has operated as a price peg — not a free market — since January 12, 2026. On that date, the median price stepped from $6.47/hr to $8.72/hr in a single session (+34.8%), a brief thin-market anomaly pushed it to $11.47 during January 20–24, and then the two-provider market reconstituted at $8.69–$8.72 and has held that level for approximately 200 consecutive days through today, July 31. The 30-day delta is +0.95% — effectively zero. This is not an equilibrium reached organically; this is a market where demand continuously absorbs every unit offered at the prevailing price, and providers have no incentive to cut. The mechanism sustaining the floor is twofold: first, the AMD MI450/Helios supply gap (full production shipments begin H1 2027, with CoreWeave's Vera Rubin deployments still ramping) means there is no near-term substitute for large-scale H100 training capacity; second, B200 — the natural displacement threat — is available from exactly one provider in three US regions at $12.82/hr (Virginia), $12.83/hr (Ohio), and $14.60/hr (Oregon), with zero international presence. At 47% above H100 on-demand, B200 is only rational for Vera Rubin-specific workloads where the 246kW rack density and 10x token/megawatt efficiency advantage justify the premium. For cost-sensitive buyers, H100 remains the dominant choice — and the market is pricing that precisely. The scarcity premium is also appearing in non-competitive markets: Osaka reached $10.93/hr (+12.8% over 30d), Zurich $9.91/hr (+20.3%), and São Paulo $9.90/hr — all single-provider markets with no competitive pressure. Rather than the ~17%/year decay implied by the cross-sectional depreciation model, H100 is actively appreciating in thin-market geographies.

Key Evidence:

1. H100 Virginia ON_DEMAND median: $8.69/hr, 30d delta +0.95%, held within ±1% range for ~200 days (live ticker data, July 31, 2026).
2. Single-session step-jump from $6.47 → $8.72 on January 12, 2026 — a 34.8% re-rate that never reversed (90-day ticker history, Virginia).
3. B200 ON_DEMAND: single provider, 3 US regions only, $12.82–$14.60/hr — zero international catalog presence confirming generation supply constraint (live ticker data, July 31).
4. H100 Osaka +12.8%, Zurich +20.3%, Warsaw +11.5% over 30 days — appreciation, not depreciation, in non-competitive markets (live ticker data, July 31).
5. A100_80GB SPOT in Oregon +36% and Virginia +44% over 30 days — predecessor generation absorbing demand overflow from H100 scarcity (live ticker data, July 31).
6. Depreciation catalog shows H100 active simultaneously on AWS, Azure, and GCP — no provider has delisted, confirming the asset is demand-constrained (query_depreciation, July 31).
7. Cross-sectional model implies ~17%/yr decay for training-class GPUs — directly contradicted by 200 days of flat-to-rising price action (query_depreciation, July 31).

Qualifying Evidence:

• Frankfurt H100 shows a -32% apparent 30d decline — but this is a 687% spread distortion driven by a low-price outlier, with the max price still at $3.13/hr. This is a thin-market artifact, not a structural break.
• Virginia H100 SPOT discount widened to 68.65% (spot now ~$2.33/hr), up from ~41–42% earlier in the year. This could reflect AWS releasing reserved-capacity pools — worth monitoring as a leading indicator of supply increase.

Implied Action:

• Lock in 1-year reserved capacity at ~$3.30–$3.70/hr effective rate — the arbitrage vs. on-demand ($8.69/hr) or spot sublicensing (~$5–6/hr gross spread per GPU) is the widest sustained gap in the dataset.
• Long CRWV and APLD on H100 floor durability — both companies' contract revenue is directly anchored to this price level. A break below $8.00 is the single stop-loss trigger for both positions.
• Monitor: H100|SPOT|us-virginia discount rate weekly. If spot discount narrows back below 50% (spot rising above ~$4.35), it signals pool tightening — bullish for reserved arb. If spot discount expands above 75% (spot falling below $2.17), it signals AWS releasing new supply into spot — bearish leading indicator for on-demand.
• Monitor: B200 catalog entries. First B200 listing in EU or APAC would signal supply unlocking and is the primary threat to this thesis.

H2 — Non-US A100 Geographic Arbitrage: Monopoly Pricing in Thin Markets

**Confidence: 5 / 5 **

Thesis: The global A100_80GB SPOT market is not one market — it is a collection of structurally disconnected regional pools with price dispersions that would be immediately arbitraged away in a liquid market but persist indefinitely in physical infrastructure. The most extreme case: Singapore A100_80GB SPOT at $3.78/hr, a single-provider monopoly with 0% spot discount (meaning the provider could charge more by listing as on-demand and has chosen not to), trading at 3.7x the competitive US norm of $0.88–$1.05/hr for an identical GPU. This premium has not compressed in over 30 days — the 30d delta is +0.18%, essentially flat at the peak. No second provider has entered. The mechanism is straightforward: physical data center buildout in Singapore requires 2–3 years of construction and permitting lead time, and the Europe Hyperscale data center market growing at 18% CAGR (Arizton/Yahoo Finance, July 29) validates that capacity expansion is a multi-year process even in faster-growing markets. The Tel Aviv story is the directional complement: a market that went from initial pricing ($3.04/hr) through an extraordinary surplus flush (collapsing to $0.16/hr by mid-March — a 95% spot discount, meaning AWS was essentially clearing inventory at marginal cost) and has since staged a 550% recovery to $1.04/hr by July 31, still rising at +2.7%/day. That progression from surplus to scarcity in a single region over seven months is the clearest demonstration of the thin-market lifecycle in the dataset. Frankfurt provides the third data point: 30 days ago, Frankfurt had a 32.5% spread between its two providers (one charging a significant premium over the other). Today the spread has collapsed to 2.5% — a new entrant matched the incumbent almost exactly within weeks, compressing the premium completely. This is precisely the competitive entry model that Singapore has not yet experienced.

Key Evidence:

1. Singapore A100_80GB SPOT: $3.78/hr, 1 provider, 0% spot discount, 30d delta +0.18% — monopoly premium holding for 30+ days (live ticker data, July 31, 2026).
2. Singapore price vs. competitive norm: 3.7x Ohio ($0.88/hr), 3.6x Dublin ($0.98/hr), 3.6x London ($1.05/hr) — identical hardware, identical GPU spec (live ticker data, July 31).
3. Tel Aviv A100_80GB SPOT: $0.16/hr floor in mid-March → $0.45/hr by April 21 → $1.04/hr July 31 — a 550% recovery arc with +2.7% 24h momentum (90-day ticker history, July 31).
4. Frankfurt A100_80GB SPOT: 32.5% spread 30 days ago → 2.5% today — new entrant competitive entry compressed the premium in under 4 weeks (live ticker data, July 31).
5. Tokyo A100_80GB SPOT: -15.8% over 30 days, 2 providers, 51.5% spread — second provider entered and is aggressively undercutting, same mechanism as Frankfurt (live ticker data, July 31).
6. Oregon + Virginia A100_80GB SPOT both up 36–44% in 30 days despite being 2-provider markets — confirms aggregate supply shortage, not just thin-market effects (live ticker data, July 31).
7. Europe Hyperscale data center market growing at 18% CAGR; Switzerland/Norway at ~$14/watt buildout cost — marginal capacity in non-US markets will price at premium for years (Arizton/Yahoo Finance, July 29, 2026).

Qualifying Evidence:

• Frankfurt's apparent -32% 30d median decline is a spread distortion artifact (687% spread, low-price outlier); the max Frankfurt price is $3.13/hr. Not a genuine price decrease.
• Singapore's $3.78/hr could be underpriced relative to Singapore on-demand benchmarks (Zurich H100 ON_DEMAND is $9.91/hr). The 0% spot discount means there's potentially further upside if re-listed as on-demand.

Implied Action:

• Capacity holders with A100_80GB in Singapore should eliminate spot discounts entirely — the market is clearing at on-demand parity already. Any discount is leaving revenue on the table.
• Equity preference for APAC/LATAM-exposed operators: Oracle Cloud (strong Asia-Pacific build), Equinix (Singapore LD5/LD6 footprint), over pure-US neoclouds. Oracle's APAC presence gives it direct exposure to the Singapore premium without a single-tenant dependency.
• Monitor: Any new data center permitting in Singapore (intel_feed building_permits), or a second cloud provider announcing Singapore A100 capacity. The Frankfurt timeline suggests premium compression happens within 4–6 weeks of a second provider entering — the signal will be visible in spread data before it hits prices.
• Avoid: Locking in long-term reserved contracts for A100 capacity in thin-market regions as a buyer — the expected value is negative given the monopoly premium. Use spot or on-demand and accept the premium.

H3 — CoreWeave Pre-Earnings: Revenue Acceleration vs. Debt Load Binary

**Confidence: 3.5 / 5 **

Thesis: CoreWeave's August 11 earnings represent the most significant single catalyst event for the GPU compute market in Q3 2026. The company has built a revenue progression that is not just growing but accelerating quarter over quarter: $982M → $1.21B → $1.36B → $1.57B → $2.08B in Q1 2026. Critically, Q1 2026 operating cash flow was $2.98B — solidly positive — meaning the business is genuinely cash-generative at the operating level. The $35B debt load and $7.7B of Q1 capex mean free cash flow is deeply negative (-$4.7B in Q1 alone), but this is the consequence of building a $99.4B contracted backlog, not of operational failure. The H100 Virginia price floor is the single most important external validator for CoreWeave's thesis: with on-demand stable at $8.69/hr for 200 days, there is no evidence that Microsoft (62% of 2025 revenue) or any other customer is renegotiating contracts at lower rates. Google's announcement of $195–205B capex for 2026 and its explicit admission that it needs third-party compute in Q3 (Memeburn, July 26) is a direct demand signal for CoreWeave capacity — Alphabet's need to source external GPU time almost certainly channels through CRWV's infrastructure. Vera Rubin production being confirmed at CoreWeave facilities (Indiatimes, July 28) adds a next-generation revenue overlay: at 246kW per Vera Rubin rack versus ~30kW for H100, per-rack revenue can expand by an order of magnitude if buildout is managed. The bear case is real: $35B of debt against a $40B market cap means equity holders are effectively leveraged on a thin margin of safety, GAAP EPS has missed consensus in four of five recent quarters (driven by interest expense, not operations), and the stock is down 35.3% over 3 months and 51.8% from its 52-week high. The $73.90 stock price versus $138 analyst consensus implies 87% upside — a dispersion that signals genuine uncertainty, not mispricing.

Key Evidence:

1. Q1 2026 revenue: $2.08B, up from $1.57B in Q4 2025 — sequential acceleration confirmed (CRWV fundamentals, July 31).
2. Q1 2026 operating cash flow: $2.98B — positive and substantial, proving underlying cash economics (CRWV fundamentals, July 31).
3. Q1 2026 gross margin: 65.6% (vs. 69.4% TTM) — slight compression, directionally worth monitoring (CRWV fundamentals, July 31).
4. Revenue backlog: $99.4B contracted — provides multi-year revenue visibility at current GPU pricing (CRWV fundamentals, July 31).
5. H100 Virginia ON_DEMAND: $8.69/hr, stable 200 days — no contract re-pricing pressure visible externally (live ticker data, July 31).
6. Google $195–205B 2026 capex + admission of need for third-party compute in Q3 — direct demand signal for CRWV capacity (Memeburn, July 26, 2026).
7. Vera Rubin systems confirmed in production at CoreWeave facilities — next-gen revenue opportunity (Indiatimes, July 28, 2026).
8. EPS misses in 4 of 5 recent quarters — all driven by interest expense, not gross margin (CRWV fundamentals, July 31). Operational fundamentals are intact; debt service is the GAAP drag.
9. $35B total debt vs. $40B market cap — equity cushion is thin; any operational miss amplifies downside non-linearly.

Qualifying Evidence:

• CRWV -35.3% over 3 months, -51.8% from 52-week high suggests the market has already priced in substantial risk. The pre-earnings setup has asymmetric upside if Q2 revenue beats consensus.
• August 11 EPS consensus: -$1.45 — the market is expecting a GAAP loss. The question is by how much, and whether guidance for the backlog delivery timeline satisfies.

Implied Action:

• Event-driven pre-earnings long with defined stop at the $60.55 52-week low. The combination of accelerating revenue, positive operating cash flow, Google demand signal, and H100 floor stability creates a bullish setup for the print.
• Stop triggers: Revenue below $2.0B, gross margin below 60%, or any commentary suggesting contract renegotiation or customer concentration risk increasing. A Microsoft Azure capex reduction (watch MSFT 8-K filings) is the highest-stakes read-through signal.
• Key question to watch in the transcript: Guidance on what fraction of the $99.4B backlog is H100 vs. B200/Vera Rubin contracts, and the re-pricing mechanism for contract renewals. This will determine whether H100 floor durability translates directly to revenue durability.
• Do not hold through a miss — $35B debt means a downside scenario is asymmetric. This is an event trade, not a core position.

H4 — CIAC Power Self-Funding: Hyperscaler Grid Moat in Formation

**Confidence: 3 / 5 **

Thesis: Amazon's filing with Virginia's Dominion Energy seeking CIAC authorization — the ability for hyperscalers to voluntarily fund their own transmission infrastructure and bypass the standard 3–4 year utility interconnection queue — is a regulatory inflection point that has not yet been priced into the equity spread between power-secured and power-insecure compute operators. If approved, CIAC becomes a replicable template: Amazon and other Tier 1 hyperscalers would effectively pay to privatize their grid access, compounding their capacity advantage over every neocloud and colocation operator that must wait in the standard queue. The mechanism self-reinforces: faster grid access → faster GPU deployment → faster revenue generation → greater capital availability to fund the next CIAC filing. The enabling conditions are all visible simultaneously. FERC's own Q3 2026 CAISO Forecasting Forum explicitly acknowledged that grid planning models are becoming structurally low-biased — AI demand is outrunning the historical calibration baseline, meaning the grid bottleneck is worse than regulators realize. CAISO's load forecast models are built on historical consumption patterns that predate hyperscale AI training demand; the implication is that regulatory projections for power availability will continue to prove optimistic, sustaining the queue backlog longer than official timelines suggest. The Vera Rubin rack density milestone — 246kW per rack vs. 16kW legacy average — means operators still in the interconnection queue when Vera Rubin reaches full deployment cannot compete. The NextEra/Brookfield $100B Paducah campus announcement (AJOT, July 29), chosen for its former DOE uranium enrichment site's existing high-voltage transmission infrastructure, is the strategic template for the next 5 years: you don't build the power, you find existing power. Core Scientific's $14B AMD MI450 deal deploying 1.1GW across 5 pre-powered campuses demonstrates the same logic at enterprise scale. Regulatory uncertainty remains the primary risk: Dominion Energy has raised legitimate FERC jurisdictional objections to the CIAC mechanism, and FERC's own show-cause order on co-located loads could supersede state-level CIAC approvals.

Key Evidence:

1. Amazon CIAC filing at Dominion Virginia Rider T1 proceeding — live regulatory action targeting 3–4 year queue bypass (Data Center Knowledge, July 29, 2026).
2. FERC CAISO Q3 2026 Forecasting Forum: explicit acknowledgment that load growth models are structurally low-biased, AI demand outrunning historical baseline (intel_feed, federal_regulatory).
3. FERC show-cause order on "new transmission services for flexible large loads" and "co-located loads" — direct federal regulatory attention on the exact mechanism Amazon is pursuing (intel_feed, federal_regulatory).
4. Utility data center load forecast potentially exceeding 90GW over 5 years per FERC 2024 State of Markets — described as "potentially understated" (intel_feed, federal_regulatory).
5. NextEra/Brookfield $100B Paducah campus (2GW gas + 2.6GW battery) sited at former DOE uranium enrichment site — existing high-voltage infrastructure was the decisive site-selection criterion (AJOT, July 29, 2026).
6. Core Scientific $14B AMD deal: 1.1GW across 5 pre-powered campuses, MI450 deployments H1 2027 (Indiatimes, July 29, 2026).
7. Rack density jump: 16kW average (2025) → 27kW average (2026) → 246kW for Vera Rubin — operators without secured power are locked out of the highest-revenue hardware (Data Center Dynamics, July 29, 2026).
8. Electrical estimator job postings in Louisville KY and Tempe AZ for $25M+ project portfolios — mission-critical electrical construction supply chain is at capacity (intel_feed, job_postings).

Qualifying Evidence:

• Dominion Energy's FERC jurisdictional objection is substantive — federal preemption could block state-level CIAC mechanisms entirely.
• No FERC docket activity confirming or denying the filing has appeared in regulatory databases yet — the process is early-stage.
• EPA pollution rule relaxation enabling gas-fired datacenter power (Impakter, July 29) is recent and not legally settled; a legal challenge could delay the gas-fired power backstop.

Implied Action:

• Long power-secured operators: EQIX and DLR as landlords with existing, multi-region, pre-contracted power portfolios. Both benefit from the bifurcation without single-event regulatory risk.
• Underweight power-insecure neoclouds: APLD and BTDR have project-by-project power exposure — each campus is an individual queue dependency. The CIAC moat widens the structural gap against them.
• Watch for second-mover CIAC filings: If Google or Microsoft files a similar request in a different state PUC within 60 days, that is industry-wide template adoption — a significant re-rating event for the entire power-secured operator category.
• Asymmetric macro hedge: Long transmission infrastructure ETFs (e.g., IIF or similar mid-continent grid plays) as a hedge against the scenario where FERC rejects CIAC but tightens grid access rules, creating an even more acute bottleneck.
• Regulatory trigger: Watch FERC docket numbers related to Virginia/PJM CIAC filings. A staff recommendation — positive or negative — will reach the equities market before it reaches mainstream financial media.

Risk Flags

Immediate (0–30 Days)

R1 — CRWV August 11 Binary: Debt Optics Could Overwhelm Operational Beat CoreWeave's $35B debt load generates GAAP interest expense that has driven EPS misses in 4 of the last 5 quarters. Even if Q2 2026 revenue lands at $2.4B+ and gross margin holds above 65%, a GAAP EPS miss versus the -$1.45 consensus (if interest expense is higher than modeled) could trigger algorithmic selling in a stock already down 51.8% from its high. The operational business is not the risk — the optics of a deep GAAP loss in the same quarter that Google admits it needs third-party compute is the communication risk. Watch for management commentary on the debt refinancing timeline and capex deceleration signals as the primary earnings-call variables.

R2 — Thin-Market Noise Masking Real Signal in V520 and Inferentia The 24-hour top movers include V520 Montreal SPOT (+130%) and Inferentia Stockholm SPOT (+108% over 7 days) — both reversals of prior collapses in single-provider, zero-spread pools. These are AWS spot micro-pool fluctuations with no directional validity. The danger is treating them as demand signals for ASIC inference hardware when they are inventory-clearing artifacts. Disregard both for structural analysis.

R3 — H100 Virginia Spot Discount Widening: Leading Indicator Watch The Virginia H100 spot discount has expanded from ~41–42% to 68.65%, pushing spot prices from ~$5.10/hr down to ~$2.33/hr while on-demand holds at $8.69. This divergence can mean either (a) AWS releasing reserved capacity into spot pools — a supply increase that would eventually pressure on-demand — or (b) demand rotating from spot into reserved contracts (bullish, as buyers lock in forward pricing). The resolution of this ambiguity within the next 30 days will determine whether H1 (the H100 floor thesis) needs reassessment.

Near-Term (30–90 Days)

R4 — AMD Helios Bridging Demand Pulls Forward: Trainium Spot Surge Is Reflexive Trainium SPOT surging +286–393% across three regions in 30 days reflects real demand — but that demand is partly a bridge to AMD MI450 deliveries beginning H1 2027. When MI450 units start shipping to Core Scientific, Anthropic, and the broader customer base under the $5B+ AMD deals, some of the Trainium bridging demand unwinding is plausible. The 7-day momentum (+37–46% weekly across Oregon, Melbourne, São Paulo) suggests the surge is still in early innings, but the H1 2027 delivery timeline creates a natural ceiling for the bridging trade. Monitor Trainium SPOT monthly for signs of a plateau.

R5 — APLD Capital Structure: Preferred Dividend and Covenant Pressure Window Applied Digital's multi-tranche preferred stock (Series E, E1, G) issued to fund expansion carries dividend obligations that compound against the deeply negative free cash flow position. The cloud-services-held-for-sale classification means a monetization event has not yet occurred, delaying balance sheet repair. With capex running at $500M–$775M per quarter against ~$250M annualized gross profit, the capital structure depends on continued equity market access. A risk-off environment, a CoreWeave earnings disappointment, or a GPU price correction could close that window precisely when APLD needs the next equity offering. October 8 earnings are the confirmation gate — watch for any pre-announcement or delay as an early warning.

R6 — Singapore A100 Premium: First Entrant Risk Is Binary The Singapore A100_80GB SPOT monopoly at $3.78/hr is the highest-confidence structural opportunity in the brief — but it resolves violently in one direction. The Frankfurt case study shows that a second provider compressed a 32.5% spread to 2.5% within weeks of entry. Singapore's premium is 3.7x competitive norm, meaning a second entrant has enormous incentive to undercut. Any cloud provider announcing Singapore GPU capacity (Oracle, Azure, Google all have Singapore regions or expansion plans) triggers immediate premium compression. The opportunity is real but has a non-zero near-term termination risk.

Structural (6–24 Months)

R7 — Vera Rubin Rack Density Creates a Facility Obsolescence Cliff NVIDIA's Vera Rubin platform requires 246kW per rack — against a 2025 industry average of 16kW and a 2026 average of 27kW (Data Center Dynamics, July 29). Only 1 in 5 operators is prepared for 50–70kW AI racks today. At 246kW, the relevant fraction drops to near zero for legacy facilities. This creates a bifurcation: operators who have invested in liquid cooling, upgraded power delivery, and secured high-capacity interconnects will capture a disproportionate share of Vera Rubin revenue; operators in legacy facilities will be physically unable to deploy the hardware regardless of capital availability. The implications for APLD, BTDR, and other neoclouds that have built facilities but not necessarily to Vera Rubin specifications are significant — a stranded-asset risk that is slow-moving but asymmetric. This risk will not manifest in Q3 or Q4 2026, but by H2 2027, facility-spec differentiation will be visible in utilization rates.

R8 — Grid Planning Model Lag: The Capacity Gap Is Worse Than Projected FERC's own CAISO Forecasting Forum acknowledged that utility load growth models are becoming low-biased because AI training demand is structurally outside the historical calibration range. The FERC 2024 State of Markets report flagged that data center load forecasts "could exceed 90GW over 5 years" and described this as potentially understated. The implication is that every infrastructure investment decision being made today using official utility load forecasts is being made with systematically optimistic power availability assumptions. For operators with 2025–2026 permit filings, this means actual power delivery dates will be later than permitting timelines suggest. The compounding effect: operators who modeled 18-month grid connection timelines may face 24–36 months in practice, stranding GPU capex in facilities that can't be powered on schedule. This is a tail risk for every neocloud with open-ended permitting in mid-continent markets (Texas, Midwest, PJM).

R9 — B200/GB200 Supply Unlock: The H100 Floor Has a Known Destroyer The H100 price floor hypothesis has a single identifiable defeat condition: a B200 supply surge that gives Microsoft, Google, and other large-volume customers the ability to shift H100 workloads onto next-generation hardware at a lower effective cost per token. B200 is currently available from one provider in three US regions at $12.82–$14.60/hr. That's too expensive to displace H100 for general workloads — but Vera Rubin's 10x token/megawatt efficiency changes the calculus at scale. If NVIDIA resolves its CoWoS/advanced packaging bottlenecks and B200 supply expands to 3+ providers by Q4 2026, the competitive pressure on H100 pricing begins. This is a known risk with a known trigger condition (B200 catalog expansion), making it one of the most monitorable structural risks in the dataset. Check B200 provider count and regional coverage monthly — the transition from 1-provider to 2-provider in Virginia will be the first detectable signal.

Data sourced from cross-provider GPU pricing tickers, 90-day price history, SEC EDGAR filings, regulatory intel feeds, and news feeds. All prices as of July 31, 2026. This brief is for analytical and informational purposes. Not investment advice.