Daily Investment Brief — May 13, 2026

May 13, 2026 — Today's analysis of the global cloud GPU market, derived from Signwl's proprietary pricing data, news intelligence, and SEC filings.

Executive Summary

• AWS Is Hitting a Structural Trainium Capacity Ceiling, Not a Temporary Spike
• H100 On-Demand Is Price-Anchored by Enterprise Reserved Deals — The Spot Market Has Already Bifurcated
• Broadcom/Custom ASIC "XPU Substitution" Is the Structural Long-Term Threat to the Entire H100 Spot Market

1. Price Movers Summary

AWS ASIC Spot Surge — Inferentia & Trainium on Fire

The dominant signal across 24h and 7d is a broad-based AWS ASIC spot price spike:

Interpretation: These are exclusively AWS ASICs (Inferentia, Trainium) and one L4. Single-provider (n_providers=1), zero spread — meaning this is raw AWS spot market movement, not inter-provider competition. The breadth across geographies (Ohio, Oregon, Tokyo, Mumbai, Sydney, Stockholm) suggests a genuine demand surge for inference/training capacity rather than a single-region event. Trainium's sustained 7d move (~330%) in us-ohio is particularly notable — consistent with the AI agent CPU supercycle narrative in news feeds.

SSD Provisioned IOPS — Synchronized Global Collapse (~-92%)

A simultaneous ~92% drop across 9+ regions (Paris, Stockholm, London, Cape Town, São Paulo, Tokyo, Virginia, Dublin, Seoul) in both 24h and 7d, while the 30d delta is flat — this happened within the last 7 days. With n_providers=3 and massive spread_pct (2,400–3,900%), this is almost certainly a provider repricing event — one major provider slashing IOPS pricing to match competitors, dragging down the median. The 30d view shows the inverse: select regions (Illinois, Oslo, Doha, Auckland) saw IOPS spike +1,200–2,350% last month, suggesting a chaotic repricing normalization cycle in storage. Not a GPU signal, but a significant cloud storage market event.

30d Notable Movers

Investable Hypotheses

All data validated. Here is the complete hypothesis report:

Hypothesis 1: AWS Is Hitting a Structural Trainium Capacity Ceiling, Not a Temporary Spike

Thesis

The Trainium spot surge in us-ohio (+332% in 7 days) is not a noise event — it represents a structural capacity ceiling being breached in real-time as Amazon's own internal workload migration to Trainium outpaces available spot inventory. With $225B in Trainium backlog (confirmed by Amazon CEO Andy Jassy) and Amazon already running "more Trainium than NVIDIA GPUs" on Bedrock, spot pools are being drained by committed capacity allocations, leaving zero slack for spot buyers.

Supporting Signals

• Pricing data (CONFIRMED): The 90-day Trainium (us-ohio) history reveals a critical structural break. Prices were elevated Jan 25–Mar 15 (~$0.25–$0.49/hr), then collapsed to near-zero ($0.024–$0.025/hr) from May 3–11 — this is the signature of a sustained over-provisioning phase ending, followed by an overnight snap to $0.107/hr on May 12 (the date's current observation). The pattern is demand-driven inventory exhaustion, not repricing.
• News (CONFIRMED): Amazon CEO Andy Jassy (AOL.com, May 12): "Amazon has a $225 billion backlog for its custom Trainium AI accelerator chips alone, with total AWS backlog reaching $364 billion. Amazon is deploying more of its custom Trainium AI accelerator chips than Nvidia GPUs." Amazon's custom silicon business has surpassed $20B annual revenue run rate.
• ASIC feed: About Amazon blog confirms Trainium delivers more performance per dollar and is the default runtime for Bedrock workloads.
• Multi-region corroboration: us-oregon also spiked +184% (7d), consistent with capacity exhaustion across all major Trainium clusters.

What Would Confirm/Deny

• Confirm: Spot prices remain above $0.08/hr for 5+ consecutive days (no mean-reversion); INFERENTIA spot in us-ohio also rises (cross-ASIC congestion signal)
• Deny: Price collapses back to $0.024/hr within 72 hours (one-time AWS internal scheduling artifact)
• Key ticker to watch: TRAINIUM|SPOT|us-ohio and TRAINIUM|SPOT|us-oregon
• SEC trigger: CoreWeave 10-Q (CRWV, May 8) — if they report AWS as a counterparty or discuss Trainium in competitive context, this validates market tightening

Implied Trade or Risk

• LONG AWS (AMZN): Trainium tightness directly translates to higher cloud margin — custom chips replace ~$10–15/hr Nvidia GPU spend with internally-manufactured chips at a fraction of the cost. This is a margin expansion story, not just a revenue story. The $225B backlog is the most concrete forward indicator in cloud compute today.
• SHORT/UNDERWEIGHT NVIDIA (NVDA) selectively: Not a structural short, but AWS systematically substituting Trainium for H100/H200 in inference pipelines removes a meaningful demand vector. Jassy explicitly stated AWS "saves tens of billions" by using custom chips.
• WATCH: AMD MI350P (launched this week per Manila Bulletin) targets exactly the inference-at-enterprise use case that Trainium dominates in cloud. AMD may capture the on-premises equivalent of the Trainium trade.

Hypothesis 2: H100 On-Demand Is Price-Anchored by Enterprise Reserved Deals — The Spot Market Has Already Bifurcated

Thesis

The H100 ON_DEMAND (us-virginia) 90-day history reveals a perfectly flat price ceiling at ~$8.69/hr held for 4+ months, with a brief spike to $11.47 in mid-January before snapping back. This is not natural market pricing — it is a reserved contract anchor. Meanwhile, global H100 pricing shows a 32x price dispersion ($0.43/hr in Amsterdam to $13.88/hr in Arizona), meaning a structural two-tier market has formed: hyperscaler-anchored core regions vs. undersupplied fringe markets. The spot market (where it exists) has completely diverged from on-demand.

Supporting Signals

• H100 us-virginia 90-day history (CONFIRMED): Iron-flat median at $8.69/hr with n_providers=2 (AWS + one other). The stability is extraordinary — across tariff announcements, DeepSeek panic, and macro volatility, this price didn't budge. This is characteristic of a bilateral contract floor, not open-market pricing.
• Global dispersion data (CONFIRMED): H100 ON_DEMAND in nl-amsterdam is $0.43/hr (n=1, single provider slash pricing) vs. $13.88/hr in us-arizona. Ireland is at $5.77/hr, Oregon at $5.96/hr, Iowa at $5.98/hr — a cluster of "spillover" providers (likely CoreWeave / independent GPU clouds) pricing below the AWS floor.
• Institutional confirmation: Argentum AI (scan step) is actively pushing clients toward long-duration GPU cluster deals vs. spot — consistent with bifurcation thesis.
• CoreWeave (CRWV) 10-Q: Customer concentration risk with 2 major clients. CoreWeave is one of the "sub-AWS" providers in the H100 spread — their business model depends on the price gap between their $5–6/hr and AWS's $8.69/hr floor.

What Would Confirm/Deny

• Confirm: H100 spot (SPOT pricing type) shows meaningful positive discount to ON_DEMAND in us-virginia; RESERVED_1YR H100 shows further compression below $8.69/hr (proving the OD price is the anchor, not the floor)
• Deny: H100 us-virginia breaks below $8.00/hr OD or above $10/hr — either would suggest the bilateral floor is collapsing or demand is spiking past reserved allocation
• Key ticker to watch: H100|SPOT|us-virginia, H100|RESERVED_1YR|us-virginia

Implied Trade or Risk

• LONG CoreWeave (CRWV) on spread compression thesis: If the H100 OD floor holds at $8.69/hr, CoreWeave (pricing at $5.77–$5.96/hr in Dublin/Oregon) can simultaneously undercut AWS AND maintain margin vs. hardware cost. Their risk is the floor collapsing — watch the nl-amsterdam anomaly ($0.43/hr) which may signal one provider (likely a European GPU cloud) distress-pricing.
• SHORT fringe H100 providers: The Amsterdam $0.43/hr outlier is priced at 1/20th of Arizona pricing. This is economically unsustainable unless subsidized or in distress — watch for provider exits in this region.
• FERC RISK: The Maryland/PJM complaint could force Virginia cost internalization. AWS us-east-1 (us-virginia) is the highest n_observations market in the dataset. A retroactive transmission cost allocation could move that $8.69/hr floor upward, which would actually benefit CoreWeave's relative position.

Hypothesis 3: Broadcom/Custom ASIC "XPU Substitution" Is the Structural Long-Term Threat to the Entire H100 Spot Market

Thesis

The current H100 spot market operates on the assumption that GPU demand is GPU-class demand. But across three hyperscalers simultaneously — Amazon (Trainium), Google (TPU v5), and now Alphabet/Meta/Anthropic/OpenAI (all partnering with Broadcom for XPU custom silicon) — a massive diversion of training and inference workloads away from NVIDIA GPUs is underway. If Broadcom's CEO projection of $100B+ in AI revenue by end-2026 materializes (from $20B in 2025), the serviceable market for GPU spot is being systematically hollowed out. Spot prices could collapse in 12–18 months as H100 clusters are stranded by custom silicon.

Supporting Signals

• News (CONFIRMED): Broadcom has custom AI chip deals with Alphabet, Meta, Anthropic, and OpenAI (AOL.com, May 12). Broadcom generated $20B AI revenue in 2025, projects $100B+ by end-2026 — a 5x increase. Broadcom-Meta 3-year partnership extended (MSN, May 12) covering chip design, packaging, and networking.
• AWS custom silicon confirmation (CONFIRMED): Amazon explicitly using more Trainium than NVIDIA GPUs on Bedrock; $225B Trainium backlog vs. "1 million NVIDIA GPUs" deal — the ratio is flipping toward custom silicon.
• GCP TPU v5: 30d spike in TPU_V5LITEPOD ON_DEMAND in Dublin (+381%) confirms GCP expanding its own non-GPU compute footprint.
• Depreciation data (CONFIRMED): Inference chip depreciation curve shows 13.05% annual decay rate (CROSS_SECTIONAL, R²=0.47), implied useful life 7.7 years. But the GENERATION_GAP data shows L4→L40S transition destroyed L4 value 80.52% in just 0.57 years. If custom silicon does to GPUs what L40S did to L4, the depreciation curve accelerates to near-total within 2 years.
• Catalog survival: GAUDI2 (Intel's custom AI chip) is already inactive across AWS, GCP, Azure — a cautionary example of how fast non-NVIDIA custom chips can be abandoned. The risk cuts both ways.

What Would Confirm/Deny

• Confirm: H100 30d prices show broad decline across 3+ regions (not just node-level repricing); AMD MI350P gains >10% of inference benchmark citations within 90 days (proxy: cloud provider SKU launches)
• Deny: Broadcom misses $100B projection; custom chip yields at advanced packaging nodes remain constrained (CoWoS bottleneck); Meta's MTIA custom chip underperforms (historical pattern)
• Key signal: Watch for AWS to add a "Trainium3" spot listing — that would signal generation transition and accelerate H100 spot depreciation

Implied Trade or Risk

• LONG Broadcom (AVGO): P/E of 82, PEG of 2.0, 41% expected earnings growth. If $100B AI revenue materializes, current valuation may still be cheap on a forward basis. Broadcom is the pick-and-shovel play for the entire hyperscaler custom silicon wave.
• LONG Korean power/electrical equipment stocks: The GPU article explicitly notes Blackwell racks require 3.3x more power than Hopper; custom XPU chips are equally power-hungry. Transformer lead times have increased to 4 years. This is a multi-year duration trade.
• SHORT legacy GPU clouds: Any GPU cloud running pure H100 spot without long-term contracted revenue faces accelerating depreciation risk as custom silicon substitution erodes the spot price floor. Depreciation curve implies 13%/year on inference class GPUs even before XPU substitution — add substitution pressure and real depreciation could exceed 25–30%/year.

Hypothesis 4: The Nordic/EU AI Infrastructure Wave Is Real and Underpriced — Inferentia2 Stockholm Is the Signal

Thesis

The Inferentia2 Stockholm spot surge (+153% in 7 days, +153% in 24 hours — both identical, meaning this happened in a single day) occurring simultaneously with the Nscale $790M Norway financing and Denmark's 60 GW grid queue oversubscription is not coincidence. AWS is expanding Nordic inference capacity at exactly the moment when European AI infrastructure is hitting a structural supply shortage. The Nordic region (hydropower, cool climate, available grid capacity) is becoming the EU's preferred AI datacenter zone, but capacity is being consumed faster than it is being built. This creates a geographic pricing asymmetry: EU Inferentia2 spot prices are about to mean-revert upward to APAC parity.

Supporting Signals

• Pricing data (CONFIRMED): Inferentia2 Stockholm: $0.083/hr (7d +153.3%). Inferentia2 Sydney: $0.642/hr (7d +65%, 24h +183%). Inferentia2 Tokyo: $0.613/hr. Stockholm is still priced at 1/7th of Sydney/Tokyo levels — a massive geographic discount that may be artificial (new cluster pricing in) or structural (EU regulatory friction).
• News (CONFIRMED): Nscale $790M financing from Nordic banks (DNB, Nordea, SEB) for Narvik, Norway campus — utility-style accordion facility. Denmark 60 GW grid queue vs. 7 GW peak demand = 8.6x oversubscription. Nebius (NVIDIA-backed) breaking ground in Missouri but previously Finland.
• Regulatory confirmation: Denmark's oversubscription signals demand is real but physically constrained. NERC's new "emerging large loads" guidelines may further slow EU-adjacent grid interconnection.
• Crypto-to-AI conversion: IREN Ltd signed a $3.4B 5-year AI cloud contract with NVIDIA deploying up to 5 GW of infrastructure. The Crypto-to-AI infrastructure shift (Data Center Knowledge, May 12) confirms power-rich brownfield sites in the Nordics are being rapidly converted.

What Would Confirm/Deny

• Confirm: Inferentia2 Stockholm price sustains above $0.08/hr for 2+ weeks; EU regulatory filings show AWS data center applications in Sweden/Norway/Denmark within 60 days
• Deny: Stockholm price collapses back to $0.033/hr (7d-ago level) — meaning the spike was a transient scheduling artifact, not structural demand
• Key tickers: INFERENTIA2|SPOT|se-stockholm, H100|ON_DEMAND|se-stockholm ($9.45/hr, 2 providers — relatively expensive vs. Dublin at $5.77/hr, suggesting supply tightness)
• Intel feed signal: Denmark/Norway power grid permit applications for AWS/hyperscalers would be the ultimate confirmation

Implied Trade or Risk

• LONG Nscale (private, pre-IPO): $790M Nordic financing at utility-grade terms suggests this company is being structured for an eventual infrastructure-bond or REIT-style exit. Nordic banks treating this as industrial infrastructure = creditworthy collateral. Watch for IPO filing.
• LONG Equinix (EQIX) EU campuses: EU datacenter REITs with existing grid connections in Nordic markets have the highest strategic optionality. NERC's new interconnection guidelines make brownfield grid connections worth a premium. NRC (US nuclear regulator) going colo with Equinix is a stamp of institutional validation.
• LONG Swedish/Norwegian grid equipment companies: The Denmark 60 GW queue at 8.6x oversubscription will force utility-scale transformer/switchgear procurement. Korean heavy electrical equipment companies (cited in GPU news) are already seeing 238% trading volume increase — the EU equivalent has not yet moved.
• RISK: 14-state legislation to ban new data centers could spread to EU member states. The Box Elder County backlash is a leading indicator for community resistance — Norway and Denmark currently lack this dynamic but it could emerge.

Hypothesis 5: The Crypto-to-AI Power Conversion Trade Is Now in Its Institutional Phase — Not Priced In Yet

Thesis

The conversion of Bitcoin mining infrastructure to AI/HPC compute — already confirmed across IREN, Applied Digital, TeraWulf, Core Scientific, CoreWeave, Hut 8, and Crusoe Energy — has moved from opportunistic to institutional and contracted. IREN's $3.4B 5-year NVIDIA contract and Hut 8's $9.8B 15-year lease represent utility-grade commitments. The key insight: these companies have pre-permitted, grid-connected power in markets where new interconnection takes 4+ years. The Trainium and Inferentia spot spikes may be partially driven by demand migrating to these converted facilities as AWS/Azure spot pools tighten. The market is valuing these companies as speculative tech, but they should be valued as infrastructure REITs with AI upside optionality.

Supporting Signals

• SEC filings (CONFIRMED): IREN 8-K (May 11), CoreWeave 10-Q (May 8), HIVE Digital 8-K (May 12) — three converted miners filing simultaneously in a single day. Institutional coordination signal.
• News (CONFIRMED): IREN $3.4B NVIDIA deal for 5 GW infrastructure. Hut 8: $9.8B, 15-year lease at Beacon Point Texas (investment-grade tenant — this is essentially a corporate bond). BlockchAIn hired former Amazon infrastructure exec to lead conversions. Crusoe Energy's methane-powered datacenter now hosts Oracle/OpenAI facility in Abilene.
• Power scarcity confirmation (CONFIRMED): NERC reliability guidelines for "emerging large loads"; NV Energy rerouting residential power to data centers; FERC/Maryland complaint; Box Elder County Utah 40,000-acre approval. Pre-permitted power is a scarcity asset.
• Government EDA: Utah PSC docket 26-057-02 shows "Millard County approved a multi-billion dollar AI data center" — consistent with mining→AI conversion geography (Utah has significant mining infrastructure).
• Regulatory structural support: Vermont legislature considering sales/use tax exemptions for data centers (EDA feed, May 5) — states are actively competing for converted mining facilities.

What Would Confirm/Deny

• Confirm: IREN files follow-on capital raise within 60 days; Hut 8 or Applied Digital announces investment-grade credit rating or IG-rated tenant; CoreWeave discloses mining-converted facility as revenue contributor in next 10-Q
• Deny: GPU spot prices collapse broadly (reduces urgency for alternative supply); FERC tightens large-load interconnection rules retroactively, stranding brownfield conversions; major converted miner fails to execute (hardware delivery delays, power quality issues at converted sites)
• Key metric: Watch the Trainium (us-ohio) spot price — Ohio has significant former mining infrastructure. If IREN/Applied Digital is running Trainium workloads out of converted Ohio facilities, the spot surge is directly traceable.

Implied Trade or Risk

• LONG IREN (ASX: IXI / NASDAQ: IREN): $3.4B 5-year contract with NVIDIA = $680M/year contracted revenue floor. Market cap needs to be compared against this annuity. If this is valued as a power utility with GPU upside, it is structurally mispriced vs. pure AI/tech comps.
• LONG Hut 8 (HUT): $9.8B 15-year lease with investment-grade tenant = ~$653M/year recurring infrastructure revenue. This is REIT-grade cash flow. The HUT market cap should be benchmarked against industrial REITs, not crypto companies.
• INFRASTRUCTURE ARBITRAGE: The core insight is that pre-permitted power in constrained grids is worth 50–100x more than it was 18 months ago. The conversion premium — avoided interconnection queue cost (up to 4 years × carrying cost) plus existing permits — is not being captured in current valuations of smaller converted miners.
• RISK: AMD MI350P launches this week targeting existing datacenter infrastructure (air-cooled, PCIe, no power upgrades required) — if enterprise compute shifts on-premises, the HPC cloud rental model these converted miners depend on faces structural headwinds. Also: SMCI ($10.2B Q3, $43B backlog) is the picks-and-shovels play on the server build-out these conversions require — potentially more liquid and less binary than individual mining companies.

Hypothesis Conviction Matrix

Immediate Action Items for Next Briefing Step

1. Monitor TRAINIUM|SPOT|us-ohio daily for 5-day sustainment above $0.08/hr (confirms H1)
2. Pull H100|SPOT|us-virginia — if spot exists and prices meaningfully below OD, confirms bifurcation (H2)
3. Track Broadcom Q2 earnings guidance vs. $100B projection (H3 confirmation event)
4. Watch INFERENTIA2|SPOT|se-stockholm for 2-week price sustainment (H4)
5. Monitor IREN and Hut 8 follow-on capital raises (H5 acceleration signal)

Market Pulse

AWS's custom ASIC spot markets ignited across six geographies simultaneously — Trainium (Ohio, Oregon) and Inferentia/Inferentia2 (Tokyo, Mumbai, Sydney, Stockholm) posted 7-day gains of 65–443%, driven by a genuine demand-velocity event where new AWS capacity provisioned in early May was absorbed within days. Simultaneously, the H100 market revealed a starkly bifurcated five-tier structure — $2.48/hr spot to $8.69/hr OD list, a 3.5x spread — with reserved contract prices locked immovably at $6.13/hr for 30+ consecutive days, confirming bilateral contract anchoring at scale. The undercurrent across all signals: demand for AI inference and training is outrunning capacity build rates, while pre-permitted power and grid interconnection are becoming the binding constraint, not silicon.

Key Movers

Note on ASIC data quality: All ASIC spot moves show n_providers=1 (AWS-only), zero spread — raw AWS spot auction prices, not inter-provider medians. Signal fidelity is high but represents a single counterparty.

Investable Insights

H1 — AWS Trainium: The Re-Provisioning Trap (Conviction: 5/5)

Thesis: AWS's Trainium spot market is exhibiting a demand-velocity signature that exceeds build velocity. The canonical pattern: spot exhaustion → AWS re-provisions new capacity → that capacity is consumed within days → repeat. The May 3–11 collapse to $0.024/hr was not mean-reversion; it was a new cluster coming online. The snap-back to $0.107 on May 12 means that entire new capacity tranche was absorbed in under nine days. This pattern is the most bullish possible cloud compute signal — demand outrunning the world's largest cloud provider's ability to add supply.

Key Evidence:

• 90-day Trainium Ohio history shows two full crunch cycles: Jan 25 – Feb 21 peak ($0.490/hr), normalization through May 2, re-provisioning May 3–11 ($0.024), snap-back May 12 ($0.107)
• Oregon corroborates: independently trending from $0.079 (Mar 11) → $0.305 (May 9), spot discount compressed from 90% → 71.6%
• Andy Jassy (AOL.com, May 12): $225B Trainium backlog; Amazon deploying "more Trainium servers than NVIDIA servers" on Bedrock; custom chip revenue exceeds $20B annual run rate, growing 40% QoQ
• Trainium access confirmed "fully booked" by third-party customers: Anthropic, OpenAI, Uber, Meta

Key Risk: AWS provisions a large new Trainium2 cluster and price collapses back to $0.024 for 30+ days. Watch for 5-day sustainment above $0.08/hr as the confirmation threshold.

Implied Action:

• LONG AMZN: Trainium tightness = hyperscaler margin expansion. Custom chips replacing $10–15/hr NVIDIA spend at fraction of cost. $225B backlog is the most concrete cloud forward indicator available.
• WATCH CoreWeave (CRWV): 10-Q (May 8) flagged customer concentration; if Trainium displaces H100 in AWS's own workloads, CoreWeave's AWS-adjacent H100 rental business faces substitution pressure in the 12–18 month window.

H2 — H100 ON_DEMAND: Five-Tier Market with a Bilateral Contract Floor (Conviction: 5/5)

Thesis: The H100 OD market in us-virginia is not freely traded. A bilateral contract floor has frozen the reserved price at exactly $6.13/hr with 0.0% delta over 30 days. The AWS OD list price ($8.69/hr) has held across DeepSeek panic, tariff announcements, and macro volatility — extraordinary stability characteristic of a rack-rate ceiling above a negotiated floor. Meanwhile, the spot market ($2.48/hr) represents AWS dumping surplus H100 capacity at 28 cents on the OD dollar — the clearest available proof that Trainium substitution is already moving committed H100 workloads off-GPU. The result is a 3.5x spread from spot to OD, the most extreme bifurcation seen in cloud compute.

Key Evidence:

• 90-day H100 OD us-virginia: median $8.45–$8.74/hr with one brief anomaly (Jan 20–24: $11.22–$11.47/hr, 4 days only), then snap-back — floor behavior, not trend
• H100 SPOT us-virginia: $2.48/hr, spot_discount_pct = 63.2%, trending slightly up (+2.9% 24h, +0.7% 7d) — not collapsing, but well below OD
• H100 RESERVED_1YR us-virginia: $6.13/hr, 30d delta = exactly 0.0%, spread 75.1% ($4.46–$7.80), volume discount 31.9%
• Global dispersion confirms bifurcation: nl-amsterdam $0.43/hr (n=1, likely distress-pricing) vs. us-arizona $13.88/hr — 32x spread across global H100 OD markets
• Sub-AWS OD floor held at $5.96/hr (CoreWeave/independents) for full 90 days

Key Risk: FERC grants Maryland's complaint (MD vs. PJM, filed May 7), forcing Virginia data center operators to internalize transmission upgrade costs — could shift the $8.69/hr OD floor upward and paradoxically benefit lower-cost providers.

Implied Action:

• LONG CRWV on spread compression: If AWS OD holds at $8.69, CoreWeave at $5.77–$5.96/hr undercuts AWS AND maintains hardware margin. Their risk is the Amsterdam $0.43/hr outlier — watch for EU provider distress or exit.
• SHORT fringe H100 providers: Amsterdam at $0.43/hr (1/20th of Arizona pricing) is economically unsustainable without subsidization. This is a distress signal, not a competitive price.
• WATCH H100 spot discount: If discount compresses below 60%, it signals H100 spot pools are tightening — which would strengthen CoreWeave's reserved-vs-spot value proposition.

H3 — Broadcom/XPU Substitution: 24-Month Structural Bear Case for GPU Spot (Conviction: 3/5)

Thesis: All four major hyperscaler AI workloads are in active migration away from NVIDIA GPUs toward custom silicon: Amazon (Trainium), Google (TPU v5), and Alphabet/Meta/Anthropic/OpenAI (all signed with Broadcom for custom XPU design). Broadcom's AI revenue grew from $20B (2025) with a projection to $100B+ by end-2026 — a 5x jump. If this materializes, the addressable market for third-party GPU spot is being systematically hollowed out. This is not a 2026 event — H100 prices are still flat — but the depreciation trajectory is accelerating: inference chip cross-sectional depreciation at 13%/year, with the L4→L40S GENERATION_GAP precedent showing 80% value destruction in 0.57 years if a generation transition is violent.

Key Evidence:

• Broadcom has custom AI chip deals with Alphabet, Meta, Anthropic, and OpenAI simultaneously (AOL.com, May 12)
• Amazon: $20B+ custom chip run rate, Trainium "fully booked" — substitution already occurring at scale
• GCP: TPU_V5LITEPOD OD Dublin +381% (30d) — GCP expanding non-GPU footprint in EU
• H100 spot at $2.48/hr vs. $8.69 OD is itself a substitution artifact — AWS dumping H100 surplus freed by Trainium migration
• Depreciation data: inference class 13.05%/yr (R²=0.47); GENERATION_GAP L4→L40S showed 80% destruction in 0.57 years as a worst-case template
• Counterbalance: Amazon committed to 1 million NVIDIA GPUs by end-2027; Wall Street expects Nvidia $79B FY2027 Q1 sales — coexistence, not replacement, near-term

Key Risk (to thesis): Custom chip yields at advanced packaging nodes (CoWoS) remain constrained; Intel GAUDI2 is already inactive across all three hyperscalers — a cautionary precedent for how fast non-NVIDIA chips can be abandoned.

Implied Action:

• LONG AVGO: Broadcom is the hyperscaler-agnostic pick-and-shovel for the custom silicon wave. P/E of 82, PEG ~2.0, 41% expected earnings growth — if $100B AI revenue materializes, current valuation may still be cheap on forward basis. Q2 earnings guidance is the key confirmation event.
• RISK: Pure H100 spot cloud operators face 13%+ annual GPU depreciation before XPU substitution pressure is added. Real effective depreciation could approach 25–30%/yr — structurally uneconomic for operators without long-term contracted revenue.

H4 — Nordic/EU Inferentia2: EU-Specific Demand Surge, Not Global (Conviction: 3/5)

Thesis: The Inferentia2 Stockholm spike (+153% in a single day, after a 73-day gap in listings) followed by instant demand absorption reflects EU-specific AI inference demand driven by GDPR/EU AI Act compliance requirements forcing workloads into EU-region infrastructure. Stockholm at $0.083/hr vs. Sydney at $0.642/hr ($7.7x geographic discount) suggests new Nordic cluster capacity is seeding at below-market prices — but the near-instant spike to $0.083 from $0.033 shows demand velocity is high. Simultaneously, the Nscale $790M Norway financing (Nordic banks treating AI DCs as utility infrastructure) and Denmark's 60 GW grid queue at 8.6x oversubscription confirm EU AI infrastructure is a real capital deployment theme.

Key Evidence:

• Inferentia2 Stockholm: 73-day listing gap (Feb 21–May 2), re-entry at $0.0330 on May 3, spike to $0.0835 on May 12 (+153%)
• Critical contradictory signal: Inferentia2 us-virginia SPOT fell -49% (7d) to $0.026/hr — weakens global shortage narrative but supports geographic migration thesis (EU regulatory pull)
• Nscale $790M accordion facility from DNB/Nordea/SEB — Nordic banks treating AI DCs as bond-grade infrastructure (Data Center Knowledge, May 12)
• Denmark: 60 GW grid queue vs. 7 GW peak demand = 8.6x oversubscription — physical capacity constraint confirmed
• H100 us-stockholm: $9.45/hr OD with 2 providers — more expensive than Dublin ($5.77), suggesting supply tightness at general GPU level too

Key Risk: Stockholm spike is a scheduling artifact from the 73-day re-entry, not genuine demand. Virginia's -49% drop is the real signal. Confirmation threshold: Stockholm Inferentia2 sustains above $0.08/hr for 14+ days.

Implied Action:

• LONG EQIX EU campuses: Equinix EU DCs with existing grid connections in Nordic markets carry the highest optionality as NERC-style guidelines make brownfield grid connections a premium asset. NRC (US nuclear regulator) selecting Equinix colo validates institutional demand.
• WATCH Nscale (private): $790M Nordic financing structured as utility-grade accordion facility is pre-IPO architecture. Nordic banks treating this as industrial infrastructure = creditworthy collateral. Monitor for S-1 filing.

H5 — Crypto-to-AI Power Conversion: Infrastructure REIT Mispricing (Conviction: 3/5)

Thesis: Bitcoin mining infrastructure with pre-permitted, grid-connected power is the scarcest asset in AI infrastructure today, where new interconnection timelines have extended to 4+ years. IREN's $3.4B 5-year NVIDIA contract ($680M/year floor) and Hut 8's $9.8B 15-year lease (~$653M/year recurring) should be benchmarked against industrial REIT comps, not crypto companies. The market is applying a speculative-tech multiple to what are now investment-grade infrastructure annuities. The Trainium Ohio surge may be partially traceable to demand absorbing at converted mining facilities — Ohio has significant former mining footprint.

Key Evidence:

• IREN 8-K (May 11, SEC), CoreWeave 10-Q (May 8), HIVE Digital 8-K (May 12) — three converted miners filing simultaneously
• IREN: $3.4B, 5-year NVIDIA contract. Hut 8: $9.8B, 15-year lease at Beacon Point Texas with investment-grade tenant — corporate bond economics
• NERC "emerging large loads" guidelines (regulatory intel feed) create friction for NEW interconnections while protecting EXISTING connections — asymmetric regulatory moat for converted miners
• Utah PSC docket 26-057-02: Millard County approving multi-billion dollar AI DC — consistent with mining→AI geography
• Vermont considering sales/use tax exemptions for data centers — state-level competition for converted facilities

Key Risk: AMD MI350P (launched this week) targets air-cooled, PCIe-compatible enterprise deployments — if compute shifts on-premises, the HPC cloud rental model these converted miners depend on faces structural headwinds. Also: H100 SPOT us-virginia at $2.48/hr (63% discount) suggests ample H100 spot availability, implying converted miners have not yet ramped GPU workloads fully.

Implied Action:

• LONG IREN: $680M/year contracted revenue floor against market cap — if valued as power utility with GPU upside optionality (vs. current crypto-tech multiple), structural mispricing is significant. Next catalyst: follow-on capital raise (expected within 60 days based on growth trajectory).
• LONG HUT: ~$653M/year recurring infrastructure revenue on 15-year lease = REIT-grade cash flow. Benchmark against industrial REITs, not crypto.
• LONG SMCI (picks-and-shovels): $10.2B Q3 revenue (+2x YoY), $43B order backlog. Liquid cooling moat directly serves converted mining facilities requiring dense GPU server buildouts. More liquid and less binary than individual miner names.

Risk Flags

FERC/PJM Transmission Cost Socialization — Virginia Data Center Liability

Maryland filed a complaint with FERC (reported May 7) seeking to reverse the allocation of $2B in transmission upgrade costs to Maryland ratepayers — costs incurred to serve Virginia data centers. If FERC rules in Maryland's favor, the precedent forces data center operators to internalize transmission upgrade costs previously socialized across ratepayers. Primary exposure: AWS us-east-1, Equinix Ashburn, Digital Realty Northern Virginia — the highest-density data center market in the world. A retroactive cost allocation could add $0.50–1.50/hr to Virginia compute costs. Watch FERC docket resolution; monitor for 8-K disclosures from EQIX and DLR.

14-State Data Center Legislation / Grid Displacement Risk

CNBC (May 9) reported 14 states considering legislation to ban or pause new data center construction. Maine passed a ban (subsequently vetoed). NV Energy is actively rerouting power from 49,000 Lake Tahoe residential customers to serve Google/Apple/Microsoft data centers near Tahoe-Reno (Fortune, May 12). This residential displacement dynamic is escalating community opposition and providing political fuel for restrictive legislation. Northern Nevada alone projects 5,900 MW of new data center demand by 2033. The regulatory risk is asymmetric: it does not affect existing permitted facilities (moat for H5 plays) but severely constrains greenfield builds, reinforcing the scarcity premium on pre-permitted brownfield power.

Amsterdam H100 Outlier — Potential EU Provider Distress

H100 OD in nl-amsterdam is $0.43/hr from a single provider — 1/20th of US-Arizona pricing ($13.88/hr), 1/20th of AWS US OD ($8.69/hr). This is economically unsustainable at any reasonable hardware cost basis unless: (a) the provider is subsidized, (b) the hardware is fully amortized, or (c) the provider is in financial distress and liquidating capacity. If (c), a sudden exit of this provider could cause a discrete upward repricing across EU H100 markets. Monitor for any EU GPU cloud shutdown announcements or restructuring filings.

SSD Provisioned IOPS Repricing Cascade — Storage Market Instability

A simultaneous ~92% collapse in SSD IOPS pricing across 9+ global regions (Paris, Stockholm, London, Cape Town, São Paulo, Tokyo, Virginia, Dublin, Seoul) in the past 7 days — while 30-day data shows select regions (Illinois, Oslo, Doha, Auckland) spiked +1,200–2,350% last month — indicates a chaotic repricing normalization cycle. With n_providers=3 and spread_pct of 2,400–3,900%, at least one major provider aggressively slashed IOPS pricing to match or undercut competitors. While not a GPU signal, this is relevant for any operator running inference pipelines with high storage I/O requirements — storage costs may be materially below recent budget assumptions.

Custom Silicon Yield / CoWoS Packaging Bottleneck

The core risk to H3: CoWoS advanced packaging capacity at TSMC remains constrained, limiting the rate at which Broadcom and Amazon can scale custom ASIC production. Intel GAUDI2 has already gone inactive across all three major hyperscalers (AWS, GCP, Azure) — a live example of how quickly a custom chip can be abandoned if yields or performance disappoint. The 5x Broadcom AI revenue projection ($20B → $100B) is contingent on packaging capacity expanding on schedule. Any TSMC CoWoS allocation news is a key H3 confirmation/denial input.

NERC "Emerging Large Loads" Guidelines — Interconnection Timeline Risk

NERC has formalized reliability guidelines for data center interconnection (regulatory intel feed), treating AI infrastructure as utility-scale loads. This formalizes and potentially extends interconnection review timelines for greenfield builds — directly extending the effective scarcity of pre-permitted power assets (tailwind for H5) while raising the regulatory risk profile for any operator planning new grid connections in constrained markets (headwind for H4 Nordic expansion).

Watch List for Next Briefing

Brief compiled May 12, 2026. Pricing data: cross-provider median model (AWS, Azure, GCP, OCI). ASIC spot prices are single-provider (AWS-only). All price levels in USD/hr per unit. This brief is for analytical purposes; nothing herein constitutes investment advice.