Daily Investment Brief — September 26, 2026

Signwl Daily Briefing — September 26, 2026

Mode: NEWS INTERROGATION — Headlines stress-tested against live pricing, capacity, and inventory data.

Market Pulse

The dominant signal in today's data is a SPOT/on-demand inversion in US inference-class GPUs — a structural market breakdown, not a spike. L40S SPOT prices in Ohio and Oregon have blown through both the on-demand list rate ($1.62/hr) and the 1-year reserved rate ($1.09/hr), with raw AWS instance clearing prices implying $2.04–$2.31/GPU-hr on the g6e/g7e families — 27–58% above a frozen reserved rate that AWS has not touched in over 30 days. The spot floor across 14 stable L40S SKUs has risen 120% since August 22, with zero catalog churn, confirming this is genuine demand absorption, not accounting noise. Pipeline data explains the demand side: four major open-weight model releases in a single week (Xiaomi MiMo-V2.6-Pro at 1.02T parameters, Alibaba Qwen-Image-2.1, MiMo-V2.6-Flash, MoGe-3) are flooding inference-class spot pools with new self-hosted workloads. Meanwhile, neocloud equity (APLD -33.9%, CIFR -29.5%, AIB -31.1% over 3 months) is diverging sharply from the GPU demand signal it should benefit from — a coherent flag that surging rental revenue is being consumed by escalating power and regulatory costs that operators cannot hedge. The binding constraint on the supply side is increasingly not silicon but MW of licensed, grid-connected data center power, most acutely in Northern Virginia, where a 2-year substation delay at the Plaza 500 campus and Dominion's Golden Substation serving existing load — not new projects — capture the bottleneck precisely.

Key Movers

TPU Amsterdam (+831% 7d) and T4G Korea (+11,317% 30d) should be disregarded as percentage moves — both start from near-zero bases with n=1 provider and reflect new regional listings, not demand. The L40S Singapore +127% 24h figure was explicitly flagged by the pipeline anomaly diagnostics but is corroborated by the 30-day trend.

Investable Insights

H1 — The L40S On-Demand List-Price Hike: Window Open, Clock Running Confidence: 4.5 / 5

Thesis: AWS's L40S on-demand list price is frozen at $1.6237–$1.6254/hr across every US region — a manually-set catalog rate that has not moved by a single cent in over 30 days, confirmed by zero rows in the price changes table for the full trailing month. The spot market has meanwhile sprinted to a point that makes this frozen rate economically incoherent. In Ohio, SPOT now clears at $1.392/hr implied at the ticker level — and at the raw instance level, the g7e.48xlarge SPOT is clearing at $16.34–$16.88/hr for an 8-GPU node, implying $2.04–$2.11/GPU-hr: spot has already exceeded on-demand. In Oregon and Virginia, the implied SPOT rate is $1.718–$1.726/hr, 6% above the on-demand list. The mean spot discount in us-west-2 has compressed to just 25.5% (IQR: 17–31%), and in Frankfurt, the p25 of the spot discount distribution touches 8.2% — some AWS instance families are effectively trading at par with on-demand in the spot market. Critically, the volatility flag is false across all nine measured regions: this is a sustained drift, not a spike, and the 45-day AWS instance price history shows the spot floor rising from $0.27/hr on August 22 to $0.59/hr today across 14 stable, unchurned SKUs — a 120% floor increase with no catalog changes.

The demand mechanism is concrete. Four major open-weight model releases landed in a single week: Xiaomi's MiMo-V2.6-Pro (1.02T parameters, MIT license, 1M context window), MiMo-V2.6-Flash (310B), Alibaba's Qwen-Image-2.1 (7B, Apache 2.0), and MoGe-3 (vision). Each is globally available and self-hostable, creating a new category of inference workload that competes directly for spot pools outside the API economy. The L40S (48GB GDDR6, 91.6 TFLOPS FP32) is the only hyperscaler-available GPU with sufficient VRAM to serve 100B+ parameter models cost-effectively at inference scale — above the underpowered T4/A10G, below the over-expensive H100 — giving it a structural moat in the inference market that is not replicable by the generation above or below it.

The competitive floor is set by RunPod, which lists L40S ON_DEMAND at $0.79/GPU-hr (independent market read, not in the unified model) — a stable rate with no movement across today's collection window. This creates a ceiling on how aggressively AWS can reprice: enterprise buyers with SLA requirements will accept a premium over RunPod, but not an unlimited one. The $0.83 gap between RunPod and AWS on-demand is the price of reliability. AWS has room to raise list prices toward $1.90–$2.10/hr (the Frankfurt/Seoul analog, and the level at which intraday g6e SPOT is already clearing) while still commanding a ~60% premium over the neocloud floor.

Key Evidence:

1. AWS g7e.48xlarge SPOT clearing at $16.34–$16.88/hr (8 GPUs → $2.04–$2.11/GPU-hr), confirmed by raw instance data today — spot exceeds on-demand ($1.62/hr) at the instance level.
2. g6e.48xlarge SPOT in Ohio at $18.44–$18.52/hr (8 GPUs → $2.31/GPU-hr) — a second independent instance confirming the inversion.
3. AWS spot floor on all L40S SKUs: $0.27/hr (Aug 22) → $0.59/hr (Sep 26), +120% in 45 days across 14 unchanged SKUs — no churn.
4. Spot discount mean: us-west-2 at 25.5% (IQR 17–31%); us-east-1 at 33.5% (p25 = 23%); Frankfurt p25 = 8.2%. All volatility_flag = false.
5. Zero rows in the price changes table for L40S over the past 30 days — on-demand frozen.
6. Zero UPCOMING_PRICE_CHANGE alerts for any L40S SKU — the window is definitively open today.
7. RunPod L40S ON_DEMAND: $0.79/GPU-hr, static across all collection timestamps — competitive floor confirmed but not AWS-equivalent product.
8. Pipeline anomaly diagnostics: AWS Stage 1 anomaly rate 0.0%, Stage 2 0.17% — today's data is clean.
9. Qualifying caveat: No natex cross-check exists for L40S (component_price_estimates returned zero rows) — the decomposed price rests entirely on the two-stage residual model without independent method validation.

Implied Action: Lock L40S 1-year reserved contracts in Ohio and Oregon immediately at $1.09/hr — before an UPCOMING_PRICE_CHANGE alert makes this the consensus trade. If on-demand reprices to $1.90–$2.10 (the instance-level clearing range), the reserved rate has implied value of $1.27–$1.41/hr, an 17–29% paper gain on the committed rate. At 32–128 GPU deployment scale, this translates to $1,700–$6,700/month in pure cost savings per cluster. Monitor trigger: any PRICE_CHANGE or UPCOMING_PRICE_CHANGE alert on AWS for the g6e or g7e instance families. Secondary trigger: spot discount mean in us-west-2 breaking below 20% (currently 25.5%). The position closes (adversely) only if new L40S capacity floods the market and discounts re-widen above 40%, or if Azure/GCP re-list L40S as active (currently marked inactive on both platforms).

H2 — Neocloud Underperformance Is Structural, Not a Contrarian Buy Confidence: 3.5 / 5

Thesis: The most analytically striking divergence in today's dataset is not in the GPU pricing tables — it is in the neocloud equity cohort. APLD (-33.9%), CIFR (-29.5%), AIB (-31.1%), and BTDR (-26.0%) are all down 26–34% over three months at exactly the moment when L40S spot prices — the primary revenue driver for GPU cloud operators — are up 83–135% over the same period. If these companies were efficiently passing through GPU demand to their equity holders, their stock prices should be rising, not collapsing. The market is saying, rationally, that the incremental rental revenue is not reaching the equity holder. The mechanism is the cost side, and the regulatory data over the past week makes it concrete.

In seven days, California eliminated environmental exemptions for data centers and mandated they pay for their own grid upgrades (Sep 21, Governor Newsom), Maryland created a new written-determination approval gate for any project ≥25 MW (Sep 23, Governor Moore), Texas issued a halt on TCEQ data center permits (Sep 21, Governor Abbott), and New Jersey issued its largest-ever enforcement fine — $1.07M for 62 unpermitted backup generators (Sep 22). These are not symbolic regulations. They directly raise the capital cost and operating cost of the GPU cloud operators who are expanding aggressively precisely when the permitting environment is tightening. At Lysander, NY, advisory committee filings documented 450,000 gallons of diesel per 24-hour outage for a 270 MW facility — backup power is a material EBITDA line item. The PJM-wide load projection (Pennsylvania PUC, Sep 14) shows 200+ TWh of new data center load between 2029–2030 across mid-Atlantic markets where APLD, CIFR, and AIB are concentrated — and the regulatory trajectory says that load will be charged for its own grid infrastructure at commercially significant rates.

The cohort, however, is not monolithic. CoreWeave (CRWV, -8.7%) and DigitalOcean (DOCN, +25.8% over 1 month) are behaving differently. CRWV has $48.3B market cap, a $141 analyst mean target against $87.59 current price (61% upside), and a diversified customer base. DOCN's SMB-managed-cloud model is largely insulated from large-load permitting dynamics. The differentiation matters: the structural short is in undercapitalized, GPU-concentrated operators without pre-secured grid connections (APLD, CIFR, AIB) — not the entire neocloud category. Digital Realty (DLR, -5.9%), whose 288 MW Northern Virginia campus is fully pre-leased to investment-grade hyperscalers and targeting H1 2027, represents the opposite risk profile — owning pre-permitted, grid-connected capacity that no regulatory order can un-permit. At -5.9% over 3 months, it is being sold alongside the neoclouds despite holding the asset moat they are desperately trying to build.

Key Evidence:

1. APLD -33.9% / CIFR -29.5% / AIB -31.1% / BTDR -26.0% over 3 months, concurrent with L40S SPOT +83–135% in the same period — the GPU demand signal they sell is not translating to equity value (equity data, Sep 26).
2. APLD forward PE -104x; CIFR forward PE 134.5x; AIB forward PE -3.2x — deeply unprofitable, priced for execution in a deteriorating permitting environment.
3. California SB/AB grid-cost mandate + NJ enforcement fine + Texas TCEQ halt + Maryland ≥25 MW approval gate — four distinct state regulatory actions in seven days (intel feed, Sep 21–23).
4. PJM-wide 200+ TWh data center load projection 2029–2030 (PA PUC filing, Sep 14) — grid cost socialization is ending as demand scale becomes undeniable.
5. CRWV -8.7% vs. cohort -27–34%; DOCN +25.8% (1 month) — stratification within the cohort confirms it is business-model-specific, not sector-wide.
6. DLR 288 MW campus: UNDER_CONSTRUCTION, fully pre-leased, targeting H1 2027 (datacenter tracker) — the moat the undercapitalized neoclouds lack.
7. CoreWeave 8-K filed Sep 17 — convertible debt presentation. Lenders pricing execution risk explicitly.
8. APLD earnings due Oct 8 — the nearest binary event; gross margin vs. power cost per GPU-hour is the critical read.

Implied Action: Avoid APLD, CIFR, and AIB on a 3-month horizon ahead of the Oct 8 APLD earnings — the margin compression thesis gets its first public data point there. Watch the income statement for realized power cost per GPU-hour vs. rental revenue per GPU-hour; any compression confirms the structural thesis. DLR at -5.9% on fully-pre-leased, grid-connected assets is the relative value expression on the long side — the infrastructure moat is real and regulatory actions strengthen rather than threaten it. Do not apply a blanket neocloud short: CRWV and DOCN require individual analysis. The short thesis is specifically on operators without pre-secured grid connections who are expanding into the most restrictive regulatory environment the US data center industry has ever faced.

H3 — Northern Virginia Grid Constraint: The Forward Pricing Signal Not Yet in Tickers Confidence: 4 / 5

Thesis: Northern Virginia is the largest compute cluster in the world, and the regulatory and infrastructure evidence accumulated in the last week makes a compelling case that it is approaching the point where new grid capacity cannot enter the market at the pace required by demand. When that transition happens — and the datacenter tracker data suggests it is 12–18 months away, not 36 — the current pricing equality between L40S on-demand in us-east-1 (Virginia, $1.6237/hr) and us-east-2 (Ohio, $1.6254/hr) will break. A Virginia geographic premium will emerge, and the first indicator will be the spot discount compressing faster in us-east-1 than in alternative US regions — a signal that is already directionally visible today (Virginia spot discount mean 33.5% vs. Oregon 25.5%, with Virginia running hotter).

The datacenter evidence is the strongest it has been. The Plaza 500 campus (Fairfax County, 176 MW planned) has had its Dominion Energy substation hearing postponed for nearly two years — now targeting December 2024 rescheduling, meaning the project is under construction without finalized power delivery commitments. The Golden Substation (Dominion Energy, Loudoun County) — the largest grid investment in the NoVA cluster — is explicitly remediation infrastructure for existing load, not capacity for new projects. Dominion's interconnection queue is reportedly 5–7 years for large loads; the Golden Substation addresses the back of that queue. Governor Spanberger's "data center accountability framework" adds a new Virginia-specific approval layer. The CAISO FERC order (Sep 21, 196 FERC ¶61216) defines large-load thresholds and commercial customer interconnection requirements that apply nationally, adding federal process overhead to state permitting friction.

The most strategically important data point in the construction tracker is negative: the only projects that will actually come online in Northern Virginia within 18 months are those already under construction with pre-committed grid access. Digital Realty/Blackstone's 288 MW campus (Manassas/Sterling, fully pre-leased, H1 2027) is in that category — it is effectively the last project to reach financial close before the interconnection bottleneck became acute. Everything behind it in the queue faces a structurally different approval environment. The Stargate UAE ($30B, 5 GW, Abu Dhabi) is arguably the most important data point here: OpenAI has explicitly chosen to build outside the US grid because US grid access has become unreliable at the scale they need. That is not a marginal optimization — it is a strategic pivot driven by the constraint this hypothesis identifies.

Key Evidence:

1. Plaza 500 substation hearing postponed ~2 years (datacenter tracker); project is under construction without finalized power delivery — the textbook definition of a capacity project at regulatory risk.
2. Golden Substation (Dominion, Loudoun County): UNDER_CONSTRUCTION, serving existing data center load — remediation infrastructure, not expansion capacity.
3. CAISO FERC Order 196 FERC ¶61216 (Sep 21): new large-load thresholds for commercial customers — federal layer atop Virginia SCC process.
4. Maryland EO (Sep 23): written determinations required for projects ≥25 MW — a state adjacent to Northern Virginia's grid territory is adding approval gates, not removing them.
5. Virginia spot discount mean (33.5%) compressing faster than Oregon (25.5%) — the earliest pricing signal the constraint is tightening.
6. L40S on-demand: Virginia $1.6237/hr = Ohio $1.6254/hr — no regional premium exists today. This is the gap the thesis predicts will open.
7. Stargate UAE ($30B, 5 GW planned, confidence 0.95): OpenAI's explicit geopolitical choice to build compute offshore, citing US grid access constraints.
8. PJM 200+ TWh demand forecast 2029–2030 (PA PUC, Sep 14) — the grid is being asked to absorb a load equivalent to adding multiple new industrial states' demand in 3 years.

Implied Action: This is a forward thesis — the trade does not exist in today's tickers because the premium hasn't materialized yet. The monitoring trigger is the first observation of L40S|ON_DEMAND|US_EAST_VIRGINIA diverging above L40S|ON_DEMAND|US_EAST_OHIO. Watch the NoVA intel feed specifically for any Virginia SCC ruling on the Plaza 500 substation, or any Dominion rate case filing that explicitly charges data centers higher large-load tariffs. On the equity side, operators with pre-committed grid access in Northern Virginia (DLR, EQIX) hold an option value not yet reflected in -5% to -6% 3-month performance. The risk expression: any DCF model for Northern Virginia compute capacity that extrapolates current power availability for new projects is mispriced. Do not underwrite new NoVA capacity without confirmed substation agreements and queue positions.

H4 — Nuclear-Backed EU Compute Will Reprice the 2028–2030 Inference Floor Confidence: 3 / 5

Thesis: The most consequential long-run signal in today's dataset is not in the spot market — it is in the France datacenter tracker, where four confirmed projects backed by sovereign capital have aggregated to a 2027–2029 supply pipeline of 1.7–2.7+ GW at structurally lower power costs than any US grid-dependent facility. Campus IA Fouju (240 MW by end 2027, 700 MW by 2029, UAE sovereign $54B, Mistral AI operator, Microsoft co-tenant, RTE fast-track confirmed) and SoftBank France (1.7 GW Phase 1 by 2027 across Dunkirk/Bosquel/Bouchain, €75B commitment, EDF as direct power partner with a 585 MW combined cycle plant collocated at Bouchain) represent the first time credible-scale, sovereign-backed compute capacity is being built with nuclear baseload power at industrial electricity prices of roughly €40–60/MWh — against US constrained-market data center power costs of $100–120/MWh. The power cost math translates to a structural per-GPU-hour cost advantage of $0.15–$0.25/hr at typical utilization rates: at current EU L40S on-demand of ~€1.87–2.05/hr (Frankfurt), that's a 8–13% margin advantage that will eventually be priced to capture EU market share, not to maximize margin.

The timeline has been clarified by the corroboration step: this is a 2027–2030 story, not 2024-cycle. The Fouju full completion date has extended to 2038. The SoftBank grid connection status at Loon-Plage and Bouchain is unconfirmed by RTE, and the RTE fast-track program (confidence: 0.55) is not fully documented in the public record. The Data4 Escaudain campus (Brookfield, 700 MW by 2030, RTE fast-track confirmed, confidence 0.82) adds a third project with independent capital. The aggregate pipeline — 2+ GW by 2029 with nuclear-backed economics — is now the most material structural pricing overhang for EU inference compute over a 3-5 year horizon, and no current pricing model for EU GPU instances accounts for it.

Key Evidence:

1. Campus IA Fouju: 240 MW end-2027, 700 MW by 2029, $54B UAE sovereign capital, Mistral AI operator, Microsoft co-tenant, RTE fast-track grid confirmed, confidence 0.95 (datacenter tracker).
2. SoftBank France: 1.7 GW Phase 1 by 2027, €75B, EDF direct power partner, 585 MW combined cycle collocated at Bouchain — baseload power with no interconnection queue for that site (datacenter tracker).
3. Data4 Escaudain: 700 MW by 2030, Brookfield capital, RTE fast-track confirmed, confidence 0.82 (datacenter tracker).
4. French nuclear industrial electricity price: ~€40–60/MWh long-term contract vs. US constrained-market $100–120/MWh — a 50–67% power cost discount translating to $0.15–0.25/hr per GPU.
5. EU L40S on-demand (Frankfurt): ~€1.87–2.05/hr — the floor these projects will price against to capture EU market share.
6. Bruyères-le-Châtel MGX/Bpifrance campus: already operational Q2 2026, 13,800 NVIDIA chips — the proof-of-concept for French nuclear-backed GPU compute is live (datacenter tracker).
7. Qualifying caveats: SoftBank RTE fast-track unconfirmed at two of three sites; Fouju full completion 2038 (extended from prior 2029 estimate); EU regulatory overhead (GDPR, AI Act) partially offsets power cost advantage; RTE fast-track program confidence 0.55.

Implied Action: This is a 3-year structural input to DCF models, not a this-quarter trade. Any valuation model for EU inference compute (whether for GPU cloud operators, AI model companies pricing EU API tiers, or datacenter REITs with EU exposure) that uses current L40S on-demand rates (~€1.87–2.05/hr) as a permanent floor beyond 2028 needs a 10–15% haircut on the terminal rate assumption. Near-term investment expression: watch for Mistral AI or any Campus IA Fouju tenant to announce EU-specific pricing tiers below AWS Frankfurt on-demand, which would be the first commercial test of nuclear-backed pricing power. Add EDF and RTE-linked instruments to watchlist as infrastructure beneficiaries — not covered in current Signwl equity universe. Risk: EU AI Act compliance overhead, EDF financial restructuring risk, and RTE grid connection delays are the three variables that could extend the timeline further.

Inventory & Channel Health

The inventory health, purchase commitment, and DIO datasets returned empty across all layers today — no composite regime, no individual ticker signals, no channel-stuffing flag. This is a data gap, not a market signal. It leaves two important sub-theses unverifiable:

• HBM supply constraint (Samsung/SK Hynix allocating 37% of HBM supply to specific customers, per the news feed) cannot be validated against component or inventory data. The thesis is plausible — HBM3e is a genuine bottleneck for H100/H200 production — but absent DIO data, it remains a press narrative without inventory-chain corroboration.

• Flash import surge as glut signal vs. demand signal: US flash/SSD imports hit $7.08B in July 2026, +499% YoY, +75% 3-month momentum. Without distributor DIO data (ARW, SNX, AVT typically lead hyperscalers by 1–2 quarters), this cannot be resolved into either channel stuffing or genuine demand absorption. At this volume, a glut forming in NAND while GPU prices rise would be an important divergence — watch for DIO data to return to the dataset.

Supply & Demand Pipeline

Supply — Critical Projects:

The Oracle Project Jupiter (New Mexico, 4 GW planned, OpenAI tenant) is the most consequential supply-side risk event in the tracker. A force majeure notice has been issued citing Energy Transfer's Green Chile gas pipeline delay (pushed to February 2027) and water disputes that prompted the New Mexico Supreme Court to pause permit proceedings. At 26% completion, this is Stargate's flagship US campus — and it is effectively stalled. If the pipeline delay holds, Stargate's 2028 operational target for New Mexico slips by at least one year. At 4 GW, the capacity loss is not trivial for OpenAI's inference scaling plans, and it explains OpenAI's parallel commitment to the UAE Stargate ($30B, 5 GW, Barakah nuclear, confidence 0.95) as a geographical hedge.

The Monarch campus (West Virginia, 8 GW full build, Anthropic, Phase 1 460 MW targeting late 2027) and the Barn Stargate (Michigan, 1.4 GW, OpenAI, DTE 19-year contract, targeting early 2028) are the two most operationally de-risked near-term US capacity additions. The Barn's DTE pre-negotiated power contract is the model: the only US large-scale project in the tracker that has secured utility commitment before breaking ground. Anthropic's Monarch campus runs NVIDIA Vera Rubin (GB200), meaning any Vera Rubin supply constraint directly gates when Anthropic can serve inference from it — a supply chain dependency that connects the Monarch timeline to TSMC's CoWoS-L packaging throughput for GB200.

Demand — Model Releases and Token Pricing:

The inference workload surge is structurally driven by open-weight releases. Four significant open-weight models landed in a single week ending Sep 21: Xiaomi MiMo-V2.6-Pro (1.02T parameters, MIT license, 1M context window), MiMo-V2.6-Flash (310B, MIT), Alibaba Qwen-Image-2.1 (7B, Apache 2.0), MoGe-3 (vision, open-source). All four are globally available and self-hostable. MiMo-V2.6-Pro is priced at $0.435/MTok input by providers — below GPT-4o and Claude equivalents — continuing the frontier-class capability commoditization trend that increases throughput volume demand rather than reducing it. Qwen-Image-2.1 at 7B parameters runs on RTX 3090-class hardware, directly linking to the RTX_PRO_6000 SPOT surge in Oregon (+132% over 30 days). This is not coincidence — small open models are consuming the consumer-grade inference spot pool in US-West.

Component — Alibaba Zhenwu V900:

Alibaba's announced Zhenwu V900 (216GB memory, 1.2TB/s interconnect, 3x predecessor performance, Sep 25) has shipped 560,000 units historically as of May 2026. The cluster-of-500,000 claim for the V900 is aspirational; no independent delivery confirmation exists. If V900 adoption accelerates within Alibaba's China datacenter footprint (20 GW global target by 2032), it could displace NVIDIA GPU demand in CN-domestic inference workloads. There is no pricing evidence for this displacement in today's data — APAC H100/H200 spot data is absent from the alerts table — so this remains a 12–24 month structural watch item rather than a current-day signal.

Pre-Announced Price Changes

No UPCOMING_PRICE_CHANGE alerts were detected across any provider over the past 7 days. This is the most forward-looking signal in the dataset, and its absence is itself meaningful: no hyperscaler has formally telegraphed a list-price revision despite the L40S spot inversion documented in H1. The window for buyers to lock 1-year reserved contracts before a hike materializes is therefore definitively open as of September 26, 2026. When this section becomes populated — when AWS files an UPCOMING_PRICE_CHANGE for any g6e or g7e family instance — the actionable window described in H1 and H4 closes simultaneously. Monitor daily.

Risk Flags

Immediate (0–30 days)

R1 — APLD Earnings Binary Event (Oct 8) Applied Digital reports on October 8. The analytical question is not whether revenue is growing — with L40S spot up 89% in Ohio, it almost certainly is — but whether power cost per GPU-hour is growing faster than realized rental revenue per GPU-hour. At a forward PE of -104x and a -33.9% 3-month drawdown, the market has already priced significant margin compression. A gross margin line that confirms this thesis will accelerate equity underperformance across the cohort (CIFR, AIB, BTDR) by sympathy. A gross margin beat — which would require APLD to have locked power purchase agreements well before the current spot electricity surge — would be a genuine contrarian surprise. Watch the power cost disclosure specifically, not just the revenue line.

R2 — L40S Singapore Anomaly Not Fully Resolved The pipeline anomaly diagnostics explicitly flagged the L40S Singapore SPOT +126.7% 24h move as anomalous. While the 30-day trend across nine other regions validates that the broad L40S spot inflation is real, the Singapore 24h figure remains suspect and should not be used in isolation to size positions or trigger alerts. The n_observations_total for Singapore is thinner than for US/EU regions, and a single provider re-entering or repricing a regional tier can produce this kind of one-day read. Wait for the 7-day read from Singapore to establish whether the 24h move is sustained or mean-reverting.