Daily Investment Brief — May 12, 2026

May 12, 2026 — Today's analysis of the global cloud GPU market, derived from Signwl's proprietary pricing data, news intelligence, and SEC filings.

Executive Summary

• ** Investable Hypothesis Report — May 11, 2026**
• HYPOTHESIS 1 — HIGH CONVICTION
• HYPOTHESIS 2 — HIGH CONVICTION
• HYPOTHESIS 3 — MEDIUM CONVICTION

I. PRICE MOVERS SUMMARY

24-Hour Horizon

Signal: Flat. Zero movers above the 0.5% threshold in the last 24 hours. Markets are in a holding pattern — no intraday volatility detected.

7-Day Horizon — Notable Swings (±28% to ±108%)

These are the most actionable price signals in the current cycle:

Key pattern: Legacy/older-gen GPUs (M60, MI25) showing wild bi-directional volatility across regions — classic thin-market behavior. The RTX PRO 6000 spike in California suggests early developer demand for Blackwell-class hardware.

30-Day Horizon — Structural Moves

Investable Hypothesis Report — May 11, 2026

Five grounded, data-backed hypotheses derived from pricing signals, depreciation curves, SEC filings, and regulatory intelligence.

HYPOTHESIS 1 — HIGH CONVICTION

"The A10 Inference GPU Is in a Structural Price Decline — But the Pacific Northwest Has Hit a Tactical Floor"

Thesis: The A10 is 5.1 years into a generational cycle where the depreciation model implies ~13% annual price decay and a 7.7-year useful life. At $0.73/hr in us-oregon — the cheapest on-demand A10 in the world — it has now hit a regional floor not seen anywhere else. The -4.0% 7-day move and -3.99% 30-day move in Oregon suggest an active supply flush, while globally prices are drifting down 1–7% per week. This is the last window before the A10 begins exiting catalogs.

Supporting Signals:

• Depreciation model: Cross-sectional regression → 13.05% annual decay, implied useful life 7.7 years (R²=0.47, p=0.059, borderline significant) — inference GPU class
• A10 is already delisted from AWS and GCP; only Azure remains active — catalog survival is in terminal phase
• us-oregon ON_DEMAND: $0.73/hr — vs. $2.69/hr in in-hyderabad (3.7× premium) and $1.88/hr global average
• 7-day delta: -4.0% in Oregon; -6.1% in Frankfurt; -6.7% in Tokyo — all pointing the same direction
• Generation Gap analysis: L4→L40S shows 80.5% implied annual decay when next-gen launches — A10's successor (L40S at $1.88/hr) is already 2.6 years old and widening the price gap
• A10G performance-normalized price ($0.031/TFLOP) is higher than T4 ($0.027), meaning A10G is not even the best value per compute unit in its class

What Would Confirm It:

• Check A10G|ON_DEMAND|US_EAST_VIRGINIA for AWS delisting confirmation
• Watch for Azure removing A10 from its Standard NC-series catalog (CATALOG_SURVIVAL|A10|azure = currently is_active: true)
• If A10 on-demand in Oregon drops below $0.65/hr in the next 14 days, that confirms the supply flush is accelerating

What Would Deny It:

• A10 price stabilizing at $0.73/hr for 2+ weeks → floor is real and durable
• New A10-specific inference workloads (e.g., small LLM serving at <7B parameters) creating renewed demand

Implied Trade:

• Short-duration compute buyers: Immediately lock in A10 on-demand in us-oregon or us-ohio ($0.76/hr) for inference workloads that don't require Hopper-class memory bandwidth. Capture 40–50% discount vs. global average before catalog exit
• Infrastructure investors: Azure's continued active listing of the A10 while AWS/GCP have exited is an anomaly — monitor for Azure repricing this upward (supply squeeze) or downward (clearance). Either move is a signal
• Short A10-heavy cloud providers: Any mid-tier cloud heavily weighted toward A10 inventory faces margin compression as spot pricing compresses toward $0.60/hr floor

HYPOTHESIS 2 — HIGH CONVICTION

"BTM Power Is No Longer Niche — It Will Become the Mandatory Architecture for New US AI Compute, Repricing Stranded Grid-Dependent Assets"

Thesis: PJM's proposed curtailment mechanism, ERCOT's 198 GW queue overflow, Florida's SB 484, and NERC's Level 3 Alert have converged in a single week. The regulatory framework for grid-dependent data centers is shifting from permissive to adversarial. Behind-the-meter natural gas generation is now the only architecture that guarantees uninterruptible compute power — and the market hasn't fully priced this bifurcation between BTM-enabled and grid-only facilities.

Supporting Signals:

• PJM white paper (May 10): Three proposed paths all involve either curtailment, differential reliability standards, or demand pricing for new data center loads — every option imposes a cost on grid-only facilities (AOL, May 10)
• ERCOT queue (Ascend Analytics, May 5): 198 GW of new large load applications in Q1 2026 alone = ERCOT's current entire peak load. SB6 in Texas requires large loads to curtail during grid emergencies — effectively mandating BTM for any facility that cannot tolerate interruption
• NERC Level 3 Alert (May 5): Real-world grid frequency instability from data center load spikes — this is the observable event that triggers the policy
• Florida SB 484 (May 7): Template legislation — "data centers pay their own infrastructure costs." 50 states are watching
• IREN/Nvidia deal ($3.4B, 5yr): Nvidia is specifically using a former Bitcoin miner with BTM power infrastructure, not a hyperscaler colo. The choice is not accidental
• AZIO AI + EVTV South Texas (May 11): 6MW BTM gas + 5MW AI compute — the model is replicating
• Williams Companies (WMB) Socrates project, Ohio H2 2026: $7B+ power innovation capital, Adjusted EBITDA guidance raised (AOL, May 10)

What Would Confirm It:

• Query query_intel_feed(category="building_permits", keyword="behind-the-meter") or keyword="natural gas" for permit surge in Texas/Ohio
• PJM formally adopting any of its three curtailment proposals = immediate headline trigger
• 2+ additional states introducing Florida SB 484 analog legislation within 60 days
• SEC filings from CoreWeave (CRWV), Crusoe, or Lambda Labs disclosing BTM power procurement

What Would Deny It:

• Congress passes federal preemption of state data center utility laws (unlikely in current environment)
• Grid operators extend interconnection queues by 3+ years, effectively delaying the curtailment crisis
• Rapid deployment of utility-scale storage (>20 GW) buffers peak demand — not visible in current PJM queue (gas=105 GW, storage=minimal)

Implied Trade:

• Long WMB, GEV, ETN, VST: Power infrastructure and natural gas pipeline operators with explicit data center exposure. WMB has 13.9% upside to $85.87 target per 24/7 Wall St. (AOL, May 10); GEV had $2.4B in data center equipment orders in a single quarter
• Long IREN: $3.4B Nvidia contract locked in; the BTM Bitcoin miner-to-AI pivot is complete and validated. Warrant coverage at $70 means Nvidia has skin in the game
• Risk flag for CRWV: CoreWeave's 10-Q (May 8) confirms customer concentration risk AND does not disclose BTM power arrangements — if it is grid-dependent in PJM territory, that's an existential operational risk that is not priced into the equity

HYPOTHESIS 3 — MEDIUM CONVICTION

"AMD's Cloud GPU Strategy Is Structurally Failing — The MI25 Price Collapse Is Not Temporary, It's Catalog Exit in Progress"

Thesis: The AMD MI25 (Instinct) is exhibiting the worst asymmetric price behavior in the dataset: -86% SPOT_DEV in London, -68% ON_DEMAND in Oregon — but simultaneously +83% ON_DEMAND in Texas. This is not normal supply/demand; it's the fingerprint of a dying product being cleared from inventory in high-competition markets while a single provider stubbornly holds price in a thin market. The broader AMD training GPU catalog shows a structural weakness: MI100, MI210, MI250X are all inactive across all three major clouds despite being only 3–5 years old — faster delisting than comparable Nvidia generations.

Supporting Signals:

• Catalog Survival data (training class): MI100 (5.5yr), MI210 (4.5yr), MI250X (3.9yr) — ALL inactive on AWS, Azure, GCP. Only MI300X (2.4yr) remains active on Azure only
• By comparison, A100_40GB (6yr) and A100_80GB (5.5yr) are still active on AWS + Azure + GCP — Nvidia chips survive 2–3x longer in cloud catalogs at similar ages
• MI25 in ie-dublin: -49.3% in 7 days on ON_DEMAND. Only 4 total observations = this provider is quietly clearing inventory
• MI25 Texas ON_DEMAND: +83% 7d = single provider spike in thin market, not organic demand
• Rackspace/AMD MoU announced May 11: Stock +103.5%, but management gave unchanged full-year guidance and "revenues expected to begin in 2027" — the market is pricing in a turnaround that is 18+ months away with no current pricing signal
• Intel's Gaudi2/Gaudi3 both inactive across all providers despite being 2–4 years old — the non-Nvidia training ecosystem is broadly being abandoned at the cloud provider level
• AMD's MI355X (Jan 2025, 1.4yr): also inactive everywhere — the newest generation hasn't even achieved initial cloud deployment

What Would Confirm It:

• CATALOG_SURVIVAL|MI300X|azure flips to is_active: false — would signal the last AMD training foothold in major clouds is gone
• MI25 disappears from all ON_DEMAND listings in the next 30 days (currently active in es-madrid, in-mumbai, se-stockholm, ie-dublin)
• Check query_tickers(component="MI300X") for any active pricing — if none found, AMD's cloud GPU presence is effectively zero

What Would Deny It:

• Rackspace/AMD partnership launches ahead of 2027 with real workloads and pricing traction
• Azure announces MI300X expansion to additional regions (Microsoft has a heavy AMD equity/partnership relationship)
• AMD MI400X (next-gen) achieves cloud deployment before MI300X is fully delisted

Implied Trade:

• Tactical short AMD (AMD): The cloud GPU catalog data suggests AMD's AI accelerator revenue story has no near-term pricing validation. The stock surged on a 2027 revenue story. The infrastructure evidence (catalog exits, price dumps, thin market) tells a different story. (Note: AMD has offsetting EPYC CPU business — this is a sector-specific risk, not a whole-company short)
• Arbitrage window closing: Any buyer currently running AMD MI25 workloads on legacy contracts should evaluate migrating before the catalog exits and pricing spikes (as thin-market M60 demonstrated +108% this week)
• Watch for Rackspace delivery risk: RACK stock +103% on an MoU with 2027 revenues and unchanged guidance = classic hype vs. substance divergence

HYPOTHESIS 4 — MEDIUM CONVICTION

"The Custom ASIC Wave Is Transitioning from Threat to Reality — Broadcom (AVGO) Is the Picks-and-Shovels Winner Regardless of Whose ASIC Wins"

Thesis: Broadcom's custom XPU division grew 106% YoY to $8.4B in Q1 and management is projecting $100B+ AI chip revenue by end of 2026. Amazon's Trainium hit $20B annualized with $225B in purchase commitments — fully booked, customers reserving Trainium4 (18 months out). Google launched two new TPU generations at Google Cloud Next 2026 (TPU 8t/8i). This is no longer a threat to Nvidia — it is a confirmed, growing parallel market. Broadcom's position is structurally superior to any single ASIC: it designs chips for Google, Meta, Anthropic, and OpenAI simultaneously. It wins regardless of which hyperscaler's model dominates.

Supporting Signals:

• Broadcom Q1 2026: $8.4B custom AI chip revenue (+106% YoY), CEO projecting $100B+ end of 2026 vs. $63.8B total 2025 revenue — this would represent >150% of current total company revenue from AI alone (AOL, May 11)
• Broadcom expanded Meta MTIA partnership (May 11) — now covering Meta's Training and Inference Accelerator, adding to existing Google TPU relationship (MSN, May 11)
• Amazon Trainium: $225B committed backlog, 40% QoQ growth, Trainium3 (30-40% better perf/price) essentially sold out, Marvell (MRVL) as the networking/design partner with 42% YoY revenue growth (Foreign Policy Journal, May 11)
• Google TPU 8t/8i just launched: separate training and inference chips signal hyperscalers are willing to invest in purpose-built silicon at every generation — this is not a one-off (MSN, May 11)
• TPU_V5LITEPOD pricing surge in us-nevada (+85.9% in 7 days) and new geographies (ie-dublin, ca-montreal +217-381% over 30 days) — GCP is actively expanding TPU availability, not contracting it
• Our depreciation/catalog data shows that ASIC-class accelerators (Trainium, Inferentia) have longer catalog survival than AMD chips and comparable to Nvidia at AWS — Inferentia v1 is still active at 6.4 years

What Would Confirm It:

• query_ticker_history("TPU_V5LITEPOD|ASIC|ON_DEMAND|us-nevada") showing sustained price stabilization after initial spike = real market, not artifact
• Broadcom Q2 2026 earnings (expected ~September) confirming >$10B quarterly XPU revenue run rate
• Any of Anthropic, OpenAI disclosing internal Broadcom XPU deployment at scale (they're currently customers of both Trainium and Broadcom per news)
• SEC filings from Broadcom showing capex expansion for advanced packaging (CoWoS/SoIC) — the production constraint is packaging, not design

What Would Deny It:

• Broadcom's $100B projection relies on 3 hyperscaler customers; if Google or Meta delays their next-gen XPU deployment, revenue timeline slips 12–18 months
• Nvidia's GB300/Vera Rubin NVL-scale systems are so dominant at frontier model training that XPUs remain limited to inference/CapEx reduction use cases

Implied Trade:

• Long AVGO: The 106% YoY custom chip growth at $8.4B quarterly run rate is not priced in conservatively. The $100B projection implies >$16B quarterly — extreme but the trajectory is real. Broadcom is the only company simultaneously designing chips for Google, Meta, Anthropic, and OpenAI
• Long MRVL: Marvell is Broadcom's networking analog — designing the interconnect fabric for Trainium at Amazon. $225B backlog needs networking silicon. Fiscal 2026 +42% YoY
• Pair trade: Long AVGO/MRVL, short Intel (INTC). Intel's Gaudi2/3 are already inactive in all cloud catalogs despite being 2–4 years old. CEO Lip-Bu Tan has acknowledged the failures. Intel's AI chip strategy is in "coming into shape" mode — 15 years of failed accelerator efforts (CRN, May 11)

HYPOTHESIS 5 — MEDIUM CONVICTION

"CoreWeave's Customer Concentration Risk Is the Most Underappreciated Structural Risk in AI Infrastructure Equities"

Thesis: CoreWeave's freshly filed 10-Q (May 8, 2026) explicitly discloses "significant customer concentration risk" with Customer A and Customer B representing "material portions of revenue in both Q1 2025 and Q1 2026." These are widely believed to be Microsoft and Meta. CoreWeave is deeply leveraged — it borrowed heavily to build GPU clusters at scale. If either customer reduces commitments (whether due to in-house ASIC buildout, renegotiation, or slower-than-expected AI ROI), CoreWeave has limited revenue diversification to absorb the shock. The pricing environment for the GPUs CoreWeave operates is also showing signs of structural softening.

Supporting Signals:

• CoreWeave 10-Q (CRWV, May 8): Explicitly names "significant customer concentration risk" — Customer A and Customer B are "material" in both Q1 2025 and Q1 2026, meaning this is not improving (SEC EDGAR)
• Hypothesis 4 above: Both Customer A (Microsoft) and Customer B (Meta) are simultaneously building custom ASIC capacity (Azure Maia/GB200, Meta MTIA via Broadcom). Every dollar spent on in-house chips is a dollar not spent renting CoreWeave GPUs
• CoreWeave's primary inventory (H100s): H100 at $2.49/hr median on 3 providers, launched 4.14 years ago. The annual decay model projects H100 prices compressing as GB200/GB300 Blackwell supply ramps through 2026. Falling H100 spot prices = falling CoreWeave revenue per GPU-hour
• B200 catalog survival (2.15yr): AWS already lists B200 as active while Azure/GCP do not — if Blackwell adoption accelerates, H100 fleet becomes rapidly commoditized
• Power risk: No BTM power disclosure in the 10-Q. If CoreWeave is grid-dependent in PJM territory, it faces potential curtailment under the proposed new framework — with a leveraged balance sheet, that's an existential event

What Would Confirm It:

• query_tickers(component="H100", pricing_type="ON_DEMAND") showing sustained price compression over 30-day horizon as B200 supply ramps
• Either Customer A or Customer B reduces forward commitments — would likely surface in a subsequent 8-K or 10-Q disclosure
• Microsoft or Meta announces expansion of in-house AI chip capacity (displacing third-party GPU rental)
• CoreWeave files a covenant waiver or credit facility amendment (would appear in 8-K)

What Would Deny It:

• AI demand growth outpaces ASIC supply — CoreWeave fills excess H100 capacity with new enterprise/mid-market customers, reducing concentration
• CoreWeave lands a 3rd major customer (e.g., Apple, Salesforce, government) at material scale before FY2026 year-end
• H100 prices stabilize as Microsoft/Meta lock in long-term reserved contracts at current rates

Implied Trade:

• Short CRWV or buy protective puts if: (a) H100 spot prices decline >10% over next 30 days, OR (b) either major customer discloses ASIC expansion plans
• Risk-adjusted sizing: This is not a slam-dunk short — CoreWeave has $225B+ in committed backlog echoing Amazon's Trainium model. But the equity may be pricing in a customer diversification that has not yet happened
• Relative value: Short CRWV vs. Long IREN — IREN has a named Nvidia contract, BTM power, and a 5-year locked revenue stream. CoreWeave has concentration risk, no confirmed BTM power, and generational GPU risk

Hypothesis Scorecard

Data sourced: Cross-provider compute pricing, SEC EDGAR (CRWV 10-Q May 8, IREN 8-K May 11, KEEL 10-Q May 11), NERC Level 3 Alert, Ascend Analytics ERCOT queue data, AOL Finance, MSN, Analytics India Magazine, CRN | Briefing: May 11, 2026

Market Pulse

Spot and on-demand GPU markets are largely flat at the 24-hour horizon — zero tickers moved more than 0.5% intraday. The action is at the 7–30 day scale: legacy GPUs (M60, MI25) are exhibiting violent bi-directional swings of ±70–108%, a textbook thin-market fragmentation pattern. Meanwhile, the structural story is dominated by regulatory inflection: PJM's proposed curtailment mechanism for new data center loads, ERCOT's 198 GW queue overflow, and Florida's SB 484 all crystallized in the same week, shifting the power access question from engineering risk to policy risk. The custom ASIC buildout (Broadcom +106% YoY, Trainium $20B annualized) is confirming that hyperscalers are structurally motivated to exit the GPU rental market — on a 12–24 month horizon.

Key Movers

7-Day Horizon — Biggest Swings

30-Day Structural Moves (Selected)

Investable Insights

H2 — BTM Power Is No Longer a Niche Strategy; It Is Becoming the Minimum Viable Architecture for New US AI Compute

Confidence: 5/5 Highest conviction

Thesis: Grid-dependent data centers are entering a hostile regulatory environment simultaneously across multiple US jurisdictions. Behind-the-meter natural gas generation has crossed from a competitive differentiator into an operational prerequisite for uninterruptible AI compute. The market has not yet repriced the divergence between BTM-enabled and grid-only facilities.

Key Evidence:

• PJM white paper (AOL, May 10): Three proposed curtailment frameworks — every option imposes material cost on new grid-connected loads. 67M consumers served; the grid operator has no choice but to act.
• ERCOT queue (Ascend Analytics, May 5): 198 GW of new large-load applications in Q1 2026 alone — equal to ERCOT's entire current peak load. SB6 requires large loads to curtail during emergencies.
• NERC Level 3 Alert (May 5): Real-world frequency instability from data center load spikes. The physical event that drives the policy.
• Florida SB 484 (signed, May 7): Data centers pay their own infrastructure costs; municipalities can block projects. A legislative template that is spreading.
• BaRupOn case study (Data Center Knowledge, May 8): Named operator. $35M grid interconnection cost. 2029 utility power availability date. Secured ~60 MW on-site via Kinder Morgan. "Hyperscalers have reversed their stance on BTM from disinterest to active reconsideration."
• IREN/Nvidia $3.4B, 5-year deal (IREN 8-K, May 11): Nvidia chose a former Bitcoin miner with BTM infrastructure, not a hyperscaler colo. That selection is not accidental.
• AZIO AI + EVTV South Texas (Stock Titan, May 11): 6MW BTM gas + 5MW AI compute. The model is actively replicating.

Implied Action:

• Long WMB, KMI: Both are named natural gas pipeline suppliers to AI data centers with direct BTM revenue. WMB has a $85.87 analyst target (24/7 Wall St., May 10); KMI confirmed in Liberty County, TX.
• Long GEV, ETN: GE Vernova reported $2.4B in data center equipment orders in a single quarter (+71% organic). Eaton at record margins. These are multi-year capex cycles.
• Long IREN: $3.4B Nvidia contract locked; BTM pivot from Bitcoin mining is validated and de-risked by the Nvidia warrant structure.
• Risk flag on CRWV: CoreWeave's 10-Q discloses no BTM power arrangements. If it is grid-dependent in PJM territory, curtailment is an unpriced existential risk on a leveraged balance sheet.

H5 — CoreWeave's Customer Concentration Risk Is Compounding With H100 Commodity Pricing Invasion

**Confidence: 3.5/5 **

Thesis: CoreWeave's CRWV 10-Q (May 8) explicitly names "significant customer concentration risk" — Customer A and B (widely identified as Microsoft and Meta) are material in both Q1 2025 and Q1 2026, meaning the concentration is structural, not transitional. This was already a known risk. What the H100 pricing data adds is a second, independent pressure vector: European H100 markets are showing commodity-floor dynamics ($0.43/hr in Amsterdam; −60% crash in Frankfurt in April) that could propagate to US enterprise pricing within 3–6 months.

Key Evidence:

• CRWV 10-Q (SEC EDGAR, May 8): Explicit "significant customer concentration" language; unchanged in both years of disclosure.
• H100 Frankfurt (90-day ticker history): Collapsed from $4.40 to $1.78/hr between January and April 16 — a −59.5% move in a single month driven by one new entrant. This is the US template risk.
• H100 Amsterdam: $0.43/hr median — a single provider pricing at ~17% of US enterprise rates. Loss-leader or distressed seller; either scenario signals oversupply building in the global H100 pool.
• H100 global spread: $0.43/hr (Amsterdam) to $13.20/hr (Madrid) — a 31× range on identical hardware. Markets with competitive entry collapse rapidly.
• H4 (Broadcom/Trainium) context: Both Customer A (Microsoft, Azure Maia) and Customer B (Meta, MTIA via Broadcom) are simultaneously building custom ASIC capacity. Every dollar of internal chip capacity reduces CoreWeave GPU-hour rentals.
• B200 pricing (ticker data): $12.83–14.60/hr, active in only 3 US regions with 1–2 observations each. CoreWeave can upgrade to Blackwell, but doing so on a leveraged balance sheet requires fresh capex.

Implied Action:

• Short CRWV or protective puts contingent on: (a) H100 spot prices in US declining >10% over 30 days, OR (b) either major customer disclosing ASIC expansion plans in a public filing.
• Relative value pair: Short CRWV / Long IREN — IREN has a named Nvidia contract, BTM power, and a 5-year locked revenue stream. CoreWeave has concentration risk, no confirmed BTM power, and generational GPU obsolescence risk.
• Not a slam-dunk short: CoreWeave has a substantial committed backlog. Size accordingly; use the H100 price compression as the trigger, not the thesis alone.

H4 — Broadcom Is the Custom ASIC Picks-and-Shovels Winner, Regardless of Which Hyperscaler's Model Dominates

**Confidence: 3.5/5 **

Thesis: The custom ASIC market is transitioning from threat-to-Nvidia to confirmed parallel market. Broadcom's structural position — simultaneously designing chips for Google (TPU), Meta (MTIA), Anthropic, and OpenAI — means it wins regardless of which customer's AI architecture prevails. This is the only risk-diversified position in the ASIC supply chain.

Key Evidence:

• Broadcom Q1 2026 (AOL, May 11): $8.4B custom XPU revenue, +106% YoY. CEO projecting $100B+ AI chip revenue by end of 2026. Current total company revenue is ~$64B annually — this projection is aggressive but the trajectory is real.
• Broadcom + Meta MTIA expansion (MSN, May 11): Broadcom now covers both Meta's training and inference accelerators, adding to the existing Google TPU relationship.
• Amazon Trainium (AOL, May 11): $20B+ annualized run rate, +40% QoQ, $225B in committed backlog. Trainium3 sold out. Trainium4 reservations open 18 months out.
• Google TPU 8t/8i launched (MSN, May 11): Separate training and inference chips at the same generation — confirms hyperscalers are investing in purpose-built silicon generationally, not as a one-off.
• H100 commodity collapse as structural motivation: The $4.40→$1.78/hr Frankfurt crash shows GPU market pricing is subject to competitive disruption. Custom ASICs at internal cost structure avoid this volatility entirely — Trainium/TPU don't appear in our public ticker feed, confirming captive pricing with zero competition.
• TPU_V5LITEPOD pricing surge (+86–381% in new regions): GCP is expanding TPU availability geographically, not contracting — active deployment, not a paper product.

Implied Action:

• Long AVGO: $8.4B quarterly custom chip run rate with two of the top-three AI spenders as customers. The $100B guidance is speculative, but the direction is confirmed.
• Long MRVL: Marvell is the networking/interconnect partner for Trainium at Amazon. A $225B backlog requires networking silicon. +42% fiscal 2026 YoY.
• Pair trade: Long AVGO/MRVL, short INTC. Intel's Gaudi2/Gaudi3 are inactive across all cloud catalogs despite being 2–4 years old. Intel has failed the AI accelerator market across three product generations.
• Key risk: Broadcom's $100B projection depends on 3 hyperscaler customers. A single delayed XPU deployment slips the timeline 12–18 months.

H1 — A10 Inference GPU Is in Structural Price Decline; Pacific Northwest Has Hit a Tactical Floor

**Confidence: 4/5 **

Thesis: The A10 is 5.1 years into its cycle, already delisted from AWS and GCP (Azure-only survivor), with 14 of 19 tracked regions declining week-over-week and zero regions posting positive 7-day momentum. The structural bear case is unambiguous. The tactical question is whether the current Oregon ($0.73/hr) and Frankfurt ($1.006/hr) levels represent a short-term floor before the next leg lower — or are already clearing prices.

Key Evidence:

• Global consensus: 14/19 A10 ON_DEMAND regions down 7d. Zero regions up. Tokyo −6.7%, Frankfurt −6.1%, Dublin −3.6%. This is not noise.
• Frankfurt 90-day history: Began 2026 at $0.89/hr, peaked at $1.29 on March 1 (failed bounce), grinding toward $1.00 support — held for 5 consecutive days as of May 11.
• Catalog exit underway: AWS and GCP have delisted A10. Azure maintains active listings in 19 regions — anomalous survival that likely ends within 6–9 months per depreciation model trajectory.
• Depreciation model: Cross-sectional regression yields 13.05% annual price decay; implied useful life 7.7 years. A10 is in its penultimate catalog phase.
• Mexico floor: $0.84/hr — flat for 30 days. The true equilibrium floor may be $0.84, not $0.73.
• Generation Gap: L4→L40S showed 80.5% implied annual decay when next-gen launched. A10's successor (L40S at $1.88/hr) is 2.6 years old and widening the gap.

Implied Action:

• Short-duration compute buyers: Lock in A10 on-demand in us-oregon ($0.73/hr) or de-frankfurt ($1.006/hr) for inference workloads not requiring Hopper-class memory bandwidth. This is a 40–50% discount to global average before catalog exit.
• Revised time horizon: This is a 6–9 month structural short, not a 2–4 week tactical trade. The $1.00 floor in Frankfurt has held; premature to call it broken.
• Monitor trigger: Azure removing A10 from Standard NC-series is the high-conviction confirmation. Any drop below $0.65/hr in Oregon signals the supply flush has accelerated.

H3 — AMD Cloud GPU Presence Is Chronically Fragmented; Catalog Exit Is Slower and Messier Than a Clean Narrative

Confidence: 3/5 (downgraded)

Thesis: AMD's training GPU catalog (MI100, MI210, MI250X, MI355X) is inactive across all three major clouds. The MI300X survives only on Azure. By contrast, comparable-age Nvidia chips remain active on 2–3 providers simultaneously. However, the MI25 price collapse in London (−86%) is not a clean exit signal — it is thin-market automated repricing noise from a single provider, oscillating 5–10× in single weeks.

Key Evidence:

• Catalog survival (depreciation data): MI100 (5.5yr), MI210 (4.5yr), MI250X (3.9yr), MI355X (1.4yr) — all inactive across AWS, Azure, GCP. Only MI300X (2.4yr) survives on Azure.
• Nvidia comparison: A100_40GB (6yr) and A100_80GB (5.5yr) remain active on AWS+Azure+GCP simultaneously.
• MI25 Dublin 90-day history: $0.017–$0.392/hr range over 90 days — 5–10× swings within weeks from a single provider. This is automated repricing, not strategic exit.
• Rackspace/AMD MoU (AOL, May 11): Stock +103% on the day, but management gave unchanged full-year guidance and "revenues expected to begin in 2027." Classic hype-vs-substance divergence.
• Niche demand exists: Cisco/AMD benchmark work with MI300X at 128-GPU enterprise scale — AMD has pockets of demand in regulated industries (Rackspace target market) that won't clear cleanly.

Implied Action:

• Tactical short AMD (AMD): Valid but lower conviction. The cloud GPU catalog data contradicts the stock's AI narrative, but AMD's EPYC CPU business offsets the accelerator weakness. Size as a sector-specific, not whole-company, short.
• Hard trigger before acting: Wait for MI300X|Azure to flip to inactive. That is the last AMD training foothold in major clouds — its delisting is the high-conviction confirmation signal.
• Rackspace (RACK) watch: +103% on an MoU with 2027 revenue start and unchanged guidance. Monitor for fundamental follow-through.

Risk Flags

Immediate / Acute

Emerging / Medium-Term

Watch List (Lower Urgency)

Charts

Chart 1: A10 Frankfurt 90-Day Price Trend — Confirms the structural decline from $1.29 peak (March 1) toward the $1.00 psychological support level. The orange dashed line marks the support zone currently being tested.

Chart 2: H100 Regional Price Dispersion — Illustrates the full fragmentation of H100 pricing: $0.43/hr (Amsterdam, distressed) to $13.20/hr (Madrid, monopoly). Red bars indicate anomalous pricing. The dashed line shows the US average ($2.49). This is the commodity risk vector for CoreWeave.

Chart 3: 7-Day Price Movers — The full asymmetric picture: legacy GPU spikes (M60 +108%, MI25 +83%) driven by thin-market illiquidity vs. A10/MI25 sell-offs in competitive regions. Green = price declines (buying opportunities); red = price spikes (thin-market noise).

Chart 4: GPU Catalog Survival by Vendor and Age — The most important structural chart in this brief. AMD has zero active GPU generations on major clouds beyond the MI300X (Azure-only). Intel has zero. Nvidia maintains multi-cloud active status across all generations back to 6 years. AWS ASICs (Trainium/Inferentia) show the longest survival at category.

Briefing compiled: May 11, 2026 | Data: Cross-provider compute pricing (AWS, Azure, GCP, Oracle), 90-day ticker histories, SEC EDGAR (CRWV 10-Q May 8, IREN 8-K May 11), NERC Level 3 Alert (May 5), Ascend Analytics ERCOT queue (May 5), Data Center Knowledge (May 8–11), AOL Finance (May 10–11), MSN (May 11), POWER Magazine (May 11)