Daily Investment Brief — September 25, 2026

Signwl Compute Intelligence Brief

September 25, 2026 | Final Investment Edition

Market Pulse

The dominant narrative today is SPOT/ON_DEMAND inversion at the inference tier — a structural event, not a data artifact. Across six L40S regions simultaneously, SPOT prices have surged 56–102% over 30 days while ON_DEMAND list prices remain frozen, compressing the hyperscaler SPOT discount from historical norms of 50–65% to sub-5% in Frankfurt and negative in Oregon. Oregon's SPOT ticker has already crossed above its ON_DEMAND list price — the first formal inversion in the inference tier this cycle. The mechanism is clear: a sustained wave of open-weight model releases (Xiaomi MiMo-V2.6-Pro at 1.02T parameters, Qwen multimodal stack) is absorbing inference SPOT capacity faster than providers are releasing it, and the AWS catalog reflects pricing last updated when the discount was 54% wide, not the 16.5% that prevails today. Simultaneously, GCP is executing what appears to be a deliberate, step-function markdown of its V5/V6E ASIC SPOT fleet — not demand weakness, but supply reclamation ahead of a V8 fleet transition — a directional decoupling between the GPU inference market (tightening) and the GCP ASIC market (intentionally flooded) that creates distinct and opposing trade setups.

Key Movers

Note: A100_40GB Iowa is a thin-market anomaly (315% bid/ask spread, 2 providers) and should be disregarded as a structural signal. B200 and related Blackwell figures are neocloud market reads, not unified hyperscaler model outputs.

Investable Insights

H1 — L40S SPOT/OD Inversion: The List-Price Increase Is Now a Matter of Days, Not Weeks Confidence: 5 / 5

Thesis: The L40S inference GPU SPOT market has progressed past the warning stage and into outright ON_DEMAND inversion in at least one major region, with a second at a 4.8% residual discount — below any rational threshold at which providers would maintain a meaningful SPOT program. The AWS ON_DEMAND list price for L40S has been frozen across all regions for 30+ days, while the real market — priced by the SPOT mechanism — has repriced upward 56–102% in the same window. In Oregon, the decomposed SPOT ticker now sits above the ON_DEMAND list (a negative volume discount, i.e., a market inversion). In Frankfurt, the SPOT sits at 95.3% of list — historically a boundary condition that precedes a list-price update within 1–2 billing cycles. What makes this analytically unusual is not that any single region has inverted; it is that eight or more regions are simultaneously approaching or at the trip-wire, a multi-region synchrony inconsistent with isolated demand shocks and consistent with global inference demand having structurally re-rated upward relative to the supply level at which AWS last priced. The underlying AWS instance evidence is unambiguous: the g7e.48xlarge (8× L40S, Nitro v5) is trading in the $16.55–16.84/hr spot range, and the g6e.48xlarge in specific AZs is intermittently trading above the $16.48 implied ON_DEMAND rate — meaning the real-world auction market has already crossed the list price ceiling before the ticker formally registers it. The model quality is impeccable here: Stage 2 decomposition anomaly rate for this run was 0.17%, with AWS Stage 1 at exactly 0.0%, eliminating any concern that the inversion is a modeling artifact.

Key Evidence:

1. Oregon L40S SPOT ticker at $1.62/hr, ON_DEMAND at $1.623/hr — ticker shows negative volume_discount_pct, confirming formal inversion (live ticker, September 25, 2026).
2. Frankfurt SPOT at $1.95/hr vs. $2.049/hr ON_DEMAND — 4.8% residual discount, below the 5% trip-wire threshold (live ticker, September 25, 2026).
3. Ohio SPOT has risen monotonically from $0.30/GPU/hr in late May to $1.36/hr today — no reversals in 120 days. The 90-day history confirms this is structural, not transient (ticker history, September 25, 2026).
4. Madrid SPOT stable at $0.45–0.55/hr January through May, then a step-change in late June that has since doubled; 7-day rate is +4.8%, confirming acceleration (ticker history, September 25, 2026).
5. g7e.48xlarge SPOT trading at $16.55–16.84/hr range across AZs; g6e.48xlarge intermittently at $17.05–22.50/hr — the upper bound is ~36% above implied ON_DEMAND (underlying instance query, September 25, 2026).
6. Zero UPCOMING_PRICE_CHANGE alerts in the database — the list-price increase has not been pre-announced and will arrive without warning when AWS publishes it (price alerts, September 25, 2026).
7. AWS 1-week price-change scan returned zero changes — ON_DEMAND list prices have not moved at all in the past 7 days, confirming the lag (price changes table, September 25, 2026).
8. Qualifying evidence: B200 SPOT in Brazil (-25% over 7d) and Hyderabad (-28% over 7d) show cooling at the frontier tier — the inference repricing is L40S/L4-specific, not a broad GPU bull market, which is consistent with demand concentrating at mid-tier inference workloads where open-weight model self-hosting is cost-optimal.

Implied Action:

• Buy L40S 1-year reserved capacity immediately. The reserved rate is locked at current ON_DEMAND list, and any list-price increase will be accompanied by a proportional reserved-rate increase. The window to lock pre-increase rates closes the moment AWS publishes the change — and with zero pre-announcement, that window could close with the next daily pricing scrape.
• For neocloud operators with L40S inventory (CoreWeave, Lambda): the SPOT data validates pricing power at materially higher ASPs. Any operator still pricing L40S at or near the old SPOT level is leaving direct margin on the table. The hyperscaler list-price freeze gives cover for a unilateral ASP increase now, before the hyperscaler formally reprices and normalizes the new level.
• Monitor daily: Watch the Oregon and Frankfurt tickers for the SPOT discount to reach 0% or below in both regions simultaneously — that dual-region inversion is the highest-probability trigger for an AWS price update publication. Also monitor for a PRICE_CHANGE or UPCOMING_PRICE_CHANGE alert on g6e.48xlarge or g7e.48xlarge specifically.
• NVDA equity second-order: Rising L40S utilization near full-rate on a 3-year-old GPU validates NVIDIA's inference silicon demand narrative ahead of Blackwell Ultra and Rubin inference variants, supporting the forward multiple.

H2 — GCP Is Executing a Structured V5/V6E ASIC Fleet Transition; Sub-$0.06 TPU Spot Window Is Time-Limited Confidence: 4 / 5

Thesis: GCP's ASIC SPOT pricing is not responding to weak demand — it is the product of a deliberate, operations-driven markdown sequence designed to flush workloads off V5-generation TPU capacity ahead of a V8 volume deployment. The diagnostic fingerprint is specific: discrete step-down cuts at irregular but scheduled intervals (not continuous drift), a constant catalog size (the pods haven't disappeared from the SPOT pool — n_observations is steady at 8 throughout), and a regional pattern that rotates US capacity out while seeding Europe (Amsterdam receiving new V5LitePod capacity at $0.027/hr). This is exactly what a datacenter operations team does when they need to reclaim a physical rack footprint — they don't pull the listing, they price it to zero. The V6E Texas collapse from $0.23/hr to $0.051/hr (-77% over 30 days, -44% in the last 7 days) and the V5LitePod Virginia collapse from $0.127/hr to $0.019/hr (-85% since August 25) are occurring simultaneously with the L40S SPOT surge — ruling out a general compute demand collapse as the driver. The news layer supports this with GCP disclosing TPU 8i production shipments via Broadcom in Q3 2026 and Morgan Stanley projecting $84B in Google TPU revenue by 2027 — a scale that requires power footprint currently occupied by V5 fleets.

Key Evidence:

1. V5LitePod Virginia SPOT history: flat at $0.127/hr from January 8 to August 24 (8 months, zero variance), then discrete step-downs: $0.054 (Aug 25), $0.039 (Aug 27), $0.023 (Sep 10), $0.019 (Sep 25) — step-function, not drift (ticker history, September 25, 2026).
2. n_observations for V5LitePod Virginia SPOT: constant at 8 throughout the markdown — catalog is intact, capacity is available, pricing is being deliberately lowered (ticker history).
3. TPU V6E Texas: $0.051/hr, down -77% over 30 days and -44% in 7 days; Tokyo V6E simultaneously up +30% in 7 days ($0.266/hr) — US-specific markdown, Asia equilibrium maintained (ticker data, September 25, 2026).
4. Amsterdam V5LitePod: newly listed, +54% over 7 days as a new regional entry at $0.027/hr — consistent with geographic rotation of retiring US capacity into lower-priority EU markets (ticker data, September 25, 2026).
5. TPU V6E Texas at $0.051/hr against a V6E performance spec of 918 TFLOPS (FP32 equivalent) implies an extraordinary $/TFLOPS cost — the cheapest available ASIC compute in the dataset for any JAX-portable workload (current ticker).
6. Qualifying evidence: No confirming news articles on "TPU V8 Trillium transition" were returned in the news feed — the thesis rests on pricing signals plus prior-reported intelligence (Broadcom TPU 8i production, Morgan Stanley projections). If V8 ramp slips, GCP has no reason to dump V5; recovery would be a denial signal.

Implied Action:

• Opportunistic workload operators: TPU V6E SPOT in Texas at $0.051/hr is a historically anomalous price for 918 TFLOPS of ASIC compute. Firms with JAX-portable training workloads have a narrow window — likely 4–6 weeks before either (a) the capacity disappears from the SPOT pool as GCP reclaims it, or (b) prices stabilize at a new floor once fleet transition is complete. The sub-$0.10/hr threshold for V6E is unlikely to persist past Q4 2026.
• Monitor: Watch V5LitePod Virginia SPOT for n_observations dropping from 8 toward 0 — that is the definitive signal that GCP has retired the pod from the SPOT pool. When that happens, the cheap window is formally closed.
• AVGO equity: If the V8 fleet transition thesis is correct, Broadcom's datacenter chip revenue should show acceleration tied to TPU 8i volume shipments in Q3/Q4 2026 earnings. This is a secondary catalyst worth cross-checking against AVGO's next quarterly disclosure.
• What to avoid: Don't build a long-term workload dependency on V5LitePod or V6E SPOT at current prices. The price signal is a liquidation incentive, not a stable cost structure. Migrate workloads opportunistically but have a Blackwell/GPU fallback.

H3 — NAND Import Surge: $7B/Month With Flat Exports Is an Unresolved Diagnostic Requiring Active Monitoring Confidence: 3 / 5

Thesis: The US flash/SSD import series (HS 852351) has risen from $1.18B in July 2025 to $7.08B in July 2026 — a +499% YoY acceleration that is idiosyncratic relative to the broader chip trade. Total IC imports (HS 8542) are up only +40% YoY with -13% MoM deceleration in July. The bilateral asymmetry — $7B imports, $1.62B exports, flat +1.6% export momentum — is the classic pre-positioning signature: goods arriving faster than they can leave, consistent with tariff front-running rather than organic demand. The complicating factor is that genuine demand is also real: the Monarch campus (8 GW full build, Phase 1 targeting late 2027 operations) and broader under-construction pipeline would require SSD procurement 12–18 months ahead of rack installation, placing legitimate procurement directly in the 2026 window. The hypothesis cannot yet distinguish between these two explanations; that distinction resolves with the August 2026 import data due approximately in October. A sustained plateau or sequential decline from the July $7.08B peak would confirm tariff front-running; continuation at or above that level would shift the weight toward a genuine AI storage procurement cycle.

Key Evidence:

1. US HS 852351 flash/SSD imports: $7.08B in July 2026 vs. $1.18B in July 2025, +499% YoY; 3-month momentum +75%; signal: RISING with near-monotonic monthly acceleration (trade flow data, September 25, 2026).
2. US flash/SSD exports: $1.62B in July 2026, +1.6% 3-month momentum — flat, with no acceleration mirroring the import surge (trade flow data, September 25, 2026).
3. Broad IC imports (HS 8542): +40% YoY but -13% MoM in July, flat 3-month momentum — NAND surge is idiosyncratic, not part of a general chip procurement wave (trade flow data).
4. Monarch campus: $71B total investment at 8 GW planned capacity by 2031; Phase 1 (460 MW, Anthropic contract) targeting late 2027 operational — legitimate procurement could begin 12–18 months prior, i.e., in 2026 (datacenter tracker).
5. No distributor DIO data (ARW/SNX/AVT) available in the inventory health layer — the standard channel-stuffing confirmation signal is absent, leaving the diagnostic incomplete (inventory health query, September 25, 2026).
6. Qualifying evidence: The Yongin Semiconductor Complex (Samsung/SK hynix, 15 GW planned) won't have KEPCO grid transmission until 2036, making NAND supply expansion from that facility a decade-distant event — medium-term supply scarcity is a plausible genuine demand driver for near-term NAND procurement.

Implied Action:

• Tactical: Do not act on this signal in either direction until August import data is published (~October). The two explanations have opposing implications: tariff front-running implies a Q4 NAND inventory overhang and ASP softening for NAND-exposed equities (Western Digital, Kioxia/Seagate); genuine AI demand implies continued tightness and supports SK Hynix and Micron NAND segments.
• Watch: Set an alert for August 2026 US Census trade data release. A sequential drop >20% from $7.08B is the front-running confirmation signal. Simultaneously, watch for ARW/SNX distributor DIO data to populate — if it surfaces with RED signals for NAND categories, the hypothesis upgrades to 4/5 conviction immediately.
• HBM is insulated from this risk. HBM is not tracked under HS 852351. SK Hynix's HBM3e franchise for H100/B200 and the Yongin supply-constraint dynamic for HBM remain structurally separate from the NAND overhang risk. Don't conflate the two memory stories.

H5 — Anthropic's Monarch Commitment: The Rubin Supply Chain Is Carrying an Unpriced Concentration Risk Confidence: 4 / 5

Thesis: Anthropic's $45 billion, 6-year compute contract at the Nscale Monarch campus (460 MW, Mason County WV, Vera Rubin platform) is the largest single-lab compute contract ever disclosed — and the market is not pricing the supply-chain risk embedded within it. Three compounding factors have emerged since the initial hypothesis: (1) the full Monarch campus targets 8 GW and $71B total investment, making it far larger than any single Anthropic contract implies; (2) Microsoft has exited a non-binding LOI for 1.35 GW, leaving a significant vacant anchor position that Nscale must fill before its planned IPO; and (3) Nscale's IPO is explicitly anchored on the Anthropic contract's creditworthiness. The Rubin hardware dependency is the central risk: CoWoS-L advanced packaging at TSMC remains a cited bottleneck for Blackwell Ultra (Rubin's predecessor), and Rubin is sequentially dependent on Blackwell Ultra's supply ramp. AWS has publicly committed to 2 million Blackwell Ultra and Rubin GPU deployments across 2027–2028, and Meta has a separate multi-year contract for millions more — these commitments compete with Anthropic's Monarch requirements for the same TSMC packaging capacity. If NVIDIA must allocate CoWoS-L capacity across AWS, Meta, Google, and Anthropic simultaneously, Monarch's late-2027 operational target carries non-trivial schedule risk. Anthropic's Claude inference pricing — Opus 5 at $50/MTok output — provides the revenue anchor against which this cost structure must deliver returns; a Rubin delay that extends Blackwell-era pricing into 2028 would pressure those unit economics directly.

Key Evidence:

1. Monarch campus: 8 GW planned by 2031, $71B total investment, Phase 1 at 2 GW current operational; Anthropic's contracted Phase 1 is 460 MW at $45B over 6 years (datacenter tracker, September 25, 2026).
2. Microsoft exited a non-binding LOI for 1.35 GW at Monarch — creating a major unfilled capacity position critical to campus economics and Nscale IPO viability (datacenter tracker, September 25, 2026).
3. Nscale IPO explicitly anchored on Anthropic contract: "After Microsoft Exited, Anthropic Signed $45B Deal Anchoring Nscale's IPO" (Tech Times, cited in datacenter tracker).
4. NVIDIA Q2 FY2027 data center revenue: $89B (+117% YoY), Q3 guidance $108B; AWS committing 2M Blackwell Ultra + Rubin GPUs 2027–2028; Meta multi-year millions of B/R GPUs — all competing for CoWoS-L capacity (news feed, September 24, 2026).
5. CoWoS-L advanced packaging cited as critical bottleneck in current supply chain intelligence (Seoul Economic Daily, September 24, 2026, cited in prior scan).
6. Stormwater incident (May 2026) and West Virginia state-of-emergency declaration (September 2026) introduce community/regulatory friction at site level; Nscale committed to covering cleanup costs (datacenter tracker).
7. Anthropic model portfolio: Claude Opus 5 at $10/MTok input, $50/MTok output (1M context); Claude Fable 5.1 at same tier for restricted verticals; Claude Science — multi-product inference stack that validates demand for the Monarch compute base (AI model tracker, September 25, 2026).
8. Qualifying evidence: The West Virginia microgrid model (first state-certified AI microgrid in the US) provides power independence from grid constraints — a genuine risk mitigation for the energy uncertainty that plagues other projects.

Implied Action:

• NVDA equity: The Monarch contract, alongside AWS and Meta commitments, establishes a committed demand floor for Rubin that is structurally bullish for FY2028 revenue. The risk is allocation — NVIDIA choosing who gets priority — not demand collapse. The supply constraint is a pricing-power argument for NVIDIA, not a revenue-miss argument.
• Nscale IPO watch: The Microsoft exit is a material undisclosed risk for any IPO investor. The gap between Anthropic's 460 MW contracted and the 2 GW Phase 1 target is real. Scrutinize Nscale's pre-IPO disclosures for Phase 2+ tenant commitments; a campus with one anchor tenant and a 1.35 GW vacant anchor slot is a concentration risk, regardless of Anthropic's creditworthiness.
• Anthropic token pricing: If Rubin deliveries slip, Anthropic's inference cost structure remains tied to Blackwell-era pricing longer than modeled. Watch for any Q4 2026/Q1 2027 Claude pricing updates — an output price increase from the current $50/MTok would be a signal of cost-structure pressure upstream.
• Broader inference pricing: If Anthropic's cost anchor rises due to Rubin delays, the entire mid-market inference API pricing stack tends to follow. This would actually accelerate the L40S SPOT/OD convergence thesis (H1) — high frontier API prices push more workloads toward self-hosted open-weight models on rented L40S capacity.

Inventory & Channel Health

The formal inventory composite is not populated in today's run — the DIO layer across distributors (ARW/SNX/AVT) and buyers carries no current records, so no composite regime (GLUT_FORMING / ABSORBING / MIXED / NEUTRAL) can be calculated, and no channel-stuffing alert has fired. This is a data gap, not a clean bill of health.

The single most alarming channel-health signal in the dataset is operating outside the formal DIO framework and must be read on its own terms: US flash/SSD imports at $7.08B/month (July 2026) against $1.62B in exports — a 4.4:1 import-to-export ratio with flat export momentum — is an extreme bilateral asymmetry. In prior hardware inventory cycles (the 2021–2022 NAND/DRAM glut, the 2015–2016 SSD channel stuffing episode), this kind of asymmetric import acceleration without export counterpart was followed within 2–3 quarters by distributor inventory build-up becoming visible in DIO. The fact that distributor DIO records are currently unpopulated should not be read as confirmation that the channel is healthy — it is a monitoring gap. If ARW or SNX DIO surfaces in the next data collection with a notable QoQ increase for storage categories, the hypothesis upgrades materially.

Broad chip imports (HS 8542) are actually decelerating — +40% YoY but -13% MoM in July — which makes the NAND surge structurally idiosyncratic rather than part of a general semiconductor demand surge. That idiosyncrasy is itself a diagnostic: if AI datacenter demand were the primary driver, you would expect broad chip imports to be equally elevated. The NAND-specific concentration of the import surge is consistent with tariff front-running, which tends to be product-specific (buy what you're most worried about being hit) rather than broad.

Watch trigger: August 2026 US Census flash/SSD import data (due ~October). A sequential decline >20% from $7.08B would be the highest-priority confirming signal for the inventory correction risk. A continuation above $7B would shift the balance toward genuine demand.

Supply & Demand Pipeline

Monarch Campus (Nscale, Mason County WV) — Updated Scale. The datacenter tracker now shows 8 GW planned capacity at $71B total investment by 2031, with 2 GW as the current Phase 1 operational target — materially larger than the $45B / 460 MW figures in prior reporting. Anthropic's contracted Phase 1 is the anchor; the remaining 1.54 GW of Phase 1 capacity (and 6 GW beyond that) has no confirmed tenants following Microsoft's departure. This gap is the most material single supply-pipeline development this week: a campus of this size with one anchor tenant and an IPO dependent on filling the vacant anchor slot represents concentrated execution risk for the 2027 inference capacity pipeline. Power source is the West Virginia state-certified microgrid, providing grid independence. Rubin hardware dependency links this directly to the H1/H5 pricing theses.

Oracle Project Jupiter / Stargate (New Mexico, 4 GW, $165B). Currently ~26% complete, with a natural gas pipeline delayed to February 2027, 103M gallons drawn from a single well since April, and the New Mexico Supreme Court having paused water-permit proceedings following the state Attorney General's challenge. Oracle confirmed publicly on September 24 that these are "commonplace at this scale" — the diplomatic language of a company managing active litigation. With 3,600 workers on site and the project targeting a flagship Stargate campus, any further permitting hold would be a near-term demand-side signal for tighter GPU supply through 2027, directly supporting the inference pricing thesis.

The Barn Stargate (Saline Township MI, Oracle/OpenAI, 1.4 GW, $56B). Building 1 near completion, operations targeted Q1 2028, DTE Energy 19-year power contract in place. This is the more near-term and lower-risk Stargate addition to the supply curve — no legal challenges documented, utility committed.

Demand-Side: Open-Weight Model Wave. Xiaomi's MiMo-V2.6-Pro (1.02T parameter MoE, 42B active, MIT license, $0.435/MTok input) topped Artificial Analysis' Intelligence Index for open-weight models on September 22. The UltraSpeed variant claims 20× faster output throughput. This is precisely the category of model that drives self-hosted inference demand on rented L40S capacity: large enough to require serious GPU memory, open enough to deploy on rented hardware, and cheap enough at the API to compete with closed-source alternatives only if operators are running their own clusters. The timing is directly correlated with the L40S SPOT surge that began accelerating in late June — the open-weight release cadence of Q2/Q3 2026 is the demand engine behind the convergence trade. Alibaba's parallel release of Qwen multimodal variants (Omni-Flash, LiveTranslate, Image 2.1) between September 18–21 adds specialized inference demand across audio, vision, and image generation — the mid-tier L40S inference GPU is the natural deployment target for all of these models.

Component-Level Signal: Rubin/CoWoS-L Bottleneck. TSMC's CoWoS-L advanced packaging is identified as a critical supply-chain bottleneck for Blackwell Ultra and sequentially for Rubin. With AWS (2M GPU commitment), Meta (multi-year millions), and now Anthropic/Monarch (460 MW of Rubin-class density) all in the queue simultaneously, TSMC's CoWoS-L allocation becomes the single physical constraint governing whether the 2027–2028 datacenter capex wave delivers as committed. This is not a rumor — it is confirmed in multiple independent intelligence sources — and it means NVIDIA's forward revenue guidance is only achievable if TSMC executes on time.

Pre-Announced Price Changes

No UPCOMING_PRICE_CHANGE alerts exist in the database. Zero provider-disclosed future repricing events from AWS, GCP, Azure, or Oracle. Every price move documented in this brief — including the L40S SPOT/OD inversion — is live market action, not telegraphed. When AWS publishes the L40S ON_DEMAND list-price increase that the market is pricing in, it will arrive without advance notice. This makes the Oregon and Frankfurt SPOT discount thresholds the most useful real-time leading indicators available.

Risk Flags

Immediate (0–30 Days)

R1 — L40S Oregon SPOT Inversion Creates Forced Reconciliation. The Oregon SPOT ticker has crossed above ON_DEMAND — a market state that cannot persist indefinitely without provider action. AWS must either (a) increase the ON_DEMAND list price, bringing it above the current SPOT level and resetting the discount to a rational level, or (b) absorb the situation as a data artifact if the inversion is AZ-specific and averaging effects resolve it. The risk for buyers is simple: any L40S ON_DEMAND reserved contract signed today may be immediately underwater if AWS reprices ON_DEMAND upward by 15–25% — which is what the SPOT market is implying. Watch daily. A PRICE_CHANGE alert on g6e.48xlarge or g7e.48xlarge is the zero-latency signal.

R2 — Thin-Market A100 Iowa Signal Should Not Be Propagated. The A100_40GB Iowa SPOT at $1.09/hr with a 315% bid/ask spread across two providers is a thin-market distortion. It will self-correct as capacity returns. Do not interpret it as a structural A100 re-rate — the simultaneous -53% move in A100 Montreal confirms regional idiosyncrasy. Any analysis citing Iowa A100 as a training GPU demand signal is working from noise.

Near-Term (30–90 Days)

R3 — Regulatory Encirclement of Power Is Becoming an Opex Line Item. Maryland's Executive Order on Responsible Data Centers (Utility RELIEF Act of 2026) mandates data center customers bear their own grid upgrade costs. Virginia's Data Center Accountability Framework (Executive Order 22, 2026) adds a "Data Center Fairness Fee Act" with EPA oversight provisions. TVA has enacted a 10% average rate increase for loads consuming over 5 MW — directly hitting xAI's Shelby County colossus operations. These are not proposals; they are enacted or near-enacted policies. For speculative datacenter builds (Holly Ridge 2 GW, Joliet 1.8 GW, Project Antelope 1.5 GW) without committed tenants, this regulatory shift materially alters the underwriting assumption. The opex headwind is real and accelerating multi-state.

R4 — GCP ASIC Markdown Overshoot Risk. If GCP's V5/V6E SPOT markdown has brought prices below the genuine clearing level — i.e., if workloads absorb all available capacity at $0.051/hr V6E — GCP faces a situation where it has given away margin on capacity that would have cleared anyway at a higher price. The risk for GOOG equity is not large (ASIC SPOT is a small revenue line), but for ASIC-dependent workload operators, the risk is that the sub-$0.10/hr window closes faster than expected as demand absorbs the discounted capacity.