Daily Investment Brief — June 1, 2026

Market Pulse

The GPU rental market is undergoing a generational transition in real time: H200 on-demand pricing has detonated in peripheral markets (São Paulo -76.4%, Sydney -58.9% over 7 days) while core EU/ME prices hold at $12.83–$14.46/hr — a 3.5x gap that is structurally unsustainable and will close within 60–90 days as Blackwell rack deployments complete. Simultaneously, AWS's B200 spot market has deepened to three regions at a consistent $3.32–$4.57/hr — a 67–74% discount to a perfectly frozen on-demand list price of $12.82/hr — establishing the Blackwell generation's effective market-clearing price well below its nominal rack rate. The dominant narrative is hardware-generation rotation compressing legacy GPU economics while energy and interconnection constraints emerge as the only non-compressible cost in the stack. The Frankfurt H100 on-demand market — which has fragmented to a $0.40–$5.49/hr intra-market spread (1,281%) — is the clearest available signal that the secondary H100 market is in price-discovery breakdown, with reserved-rate holders still locked at $9.51/hr against a floor that has already fallen below $0.50/hr.

Key Movers

Noise alert: MI25 and Intel V710 "surges" are first-ever catalog listings from a single provider (OCI), not genuine demand spikes. Deltas are meaningless; the signal is the activation event itself and the $0.07–$1.38/hr price tier being established.

Investable Insights

H1 — The H200 Cliff Is 30–60 Days From Hitting Core Markets: Front-Run CRWV Margin Compression Confidence: 4 / 5

Thesis: H200 on-demand pricing is executing a geographic cascade — periphery collapses first, core follows within one to two billing cycles. São Paulo dropped from $12.74 to $3.01/hr and Sydney from $9.23 to $3.79/hr — both in a single observation step at month-end, consistent with hyperscalers executing a deliberate Blackwell-era inventory clearance. The "core firewall" is now showing first cracks: Paris and Madrid are both -2.0% on the 7-day, Milan is -1.8%, while Frankfurt (+0.31%) and London (+0.23%) haven't moved yet. When — not if — those premium markets follow, the H200 premium that neoclouds like CoreWeave built their $6–12/hr realized revenue model on will evaporate. The Ad Hoc News (May 31) reported hyperscalers dumped H200 rates ~40% in three weeks as B200/GB200 rollouts accelerated; that's a secondary-market confirmation of what the ticker data shows directly. Jensen Huang's Computex confirmation that AI token generation is now profitable explains why the rotation is happening now: inference economics have cleared the bar, accelerating Blackwell rack deployment.

CoreWeave's balance sheet makes this thesis not merely a revenue story but a potential credit event. Q1 FY2026 (filed May 8) showed $2.08B in revenue against $7.7B in capex — a 3.7x ratio — with free cash flow of -$4.71B. This follows a -30% Q4 EPS miss and a -22% Q1 EPS miss, two consecutive quarters of worsening execution. At $35.1B in total debt against $2.27B cash and a current ratio of 0.315, CoreWeave has less than 32 cents of current assets for every dollar of current liabilities. A 30–40% decline in realized revenue per GPU-hour — which is what the H200 periphery collapse mathematically implies for a global rate card — would not leave the debt stack serviceable at anything close to current interest expense.

Key Evidence:

1. H200 OD São Paulo: $3.01/hr, 7d delta -76.4%; Sydney: $3.79/hr, 7d delta -58.9%. Both moves occurred in a single step, confirming hyperscaler repricing event rather than gradual erosion. (Cross-provider pricing tickers, June 1, 2026)
2. H200 OD Paris: -2.0% over 7d; Madrid: -2.0%; Milan: -1.8% — three wobbling core markets that are the leading edge of the cascade. (Cross-provider pricing tickers, June 1, 2026)
3. H200 OD Frankfurt ($13.08/hr) and London ($12.93/hr): single-provider tickers with no verified 90-day history — the "stability" reading is a point-in-time snapshot, not a confirmed trend. (Ticker history query, June 1, 2026)
4. B200 on-demand (Virginia) at $12.82/hr vs. H200 OD London at $12.93/hr — near-parity pricing eliminates any remaining premium justification for H200 in the Blackwell era. (Query tickers, June 1, 2026)
5. Ad Hoc News (May 31): "H200 rental rates collapsed ~40% in three weeks from $7/hr to $4/hr as hyperscalers rotate to B200/GB200" — independent secondary-market validation of the ticker signal.
6. CRWV Q1 FY2026 10-Q (filed May 8): Revenue $2.08B, FCF -$4.71B, total debt $35.1B, current ratio 0.315, two consecutive EPS misses (-30% in Q4, -22% in Q1). (SEC filing, equity fundamentals)
7. Qualifying caveat: Frankfurt and London H200 prices are single-provider, lack 90-day history, and are still flat as of today. The cascade thesis is not yet confirmed in core markets — it remains a 30–60-day forward call.

Implied Action: Short CRWV or buy puts with August/November expiry. August 12 is the Q2 2026 earnings date — squarely within the window when core-market H200 repricing should be visible in realized revenue. The -22% Q1 EPS miss pattern and 0.315 current ratio mean CRWV cannot absorb a 30–40% revenue/GPU-hr hit from a position of financial strength; the debt stack amplifies the pricing signal into a potential credit event. Stop-out trigger: Frankfurt and London H200 prices remain above $12/hr through June 15 with no follow-through in Paris/Madrid/Milan. Relative long: Nebius (NBIS, $231, +153% in 3 months) — more geographically diversified, faster growth, lower debt concentration — as a pair trade against CRWV short.

H2 — B200 Spot Is the Effective Training Price Floor: The On-Demand Rate Is a Fiction Confidence: 5 / 5

Thesis: AWS has established a structurally bifurcated B200 market: on-demand frozen at $12.82–$14.60/hr across all three active regions (zero delta over 24h, 7d, and 30d — behavior consistent only with a deliberately held list price), while spot clears at $3.32–$4.57/hr — a 62–74% discount. This is not a supply imbalance; it is textbook price discrimination. AWS is protecting the OD "rack rate" for enterprise procurement contracts while making the newest, most capable chip available at spot for any workload tolerant of interruption. The consequence is that the effective market price for B200 compute is already $3.32/hr in Virginia — and the three-region geographic expansion (Ohio at $4.57/hr, Oregon at $4.35/hr, with a 30d +16.2% drift in Oregon confirming workload migration) means the AWS spot B200 market is deepening, not thinning. Virginia's 30-day -25.4% move to $3.32/hr signals AWS is flooding capacity as Blackwell racks come online, not restricting supply.

The cross-architecture comparison reveals the training economics shift is already live: B200 spot at $3.32/hr (Virginia) vs. H100 on-demand at $2.94/hr median (Frankfurt, 8 observations, 2 providers) — near-parity on a $/hr basis, but B200 delivers approximately 82% higher BF16 throughput than H100. On a cost-per-TFLOP basis, B200 spot is already cheaper than H100 on-demand in Frankfurt, and dramatically cheaper than H100 on-demand in US markets ($8.91/hr median in Virginia). Any model developer with checkpoint capability — which encompasses essentially all serious training workloads — has already entered a new training economics regime without necessarily knowing it. Inferentia SPOT in Mumbai at $0.111/hr (+922% over 30d) and Trainium SPOT in Ohio at $0.108/hr (+328% over 30d) corroborate the broader AWS proprietary silicon inference bid that Jensen Huang's "tokens are profitable" Computex comment predicted.

Key Evidence:

1. B200 SPOT us-virginia: $3.32/hr, 30d delta -25.4% (deepening supply); B200 OD us-virginia: $12.82/hr, 30d delta 0.0% (frozen list price). Spot discount: 74%. (Query tickers + ticker history, June 1, 2026)
2. B200 SPOT three-region expansion confirmed: us-ohio ($4.57/hr, +9.1% 30d), us-oregon ($4.35/hr, +16.2% 30d) — geographic deepening is the key new signal from stress testing. (Query tickers, June 1, 2026)
3. B200 SPOT 90-day Virginia history: started $3.31/hr in January, troughed at $2.71/hr in February, mini-surge to $4.45/hr on May 2–3, settled back to $3.32/hr today. 90 days of continuous price discovery on a new architecture is remarkable. (Ticker history, June 1, 2026)
4. H100 OD Frankfurt median: $2.94/hr — B200 spot at $3.32/hr is within 13% on a $/hr basis but delivers ~82% more compute throughput (B200: ~1,800 BF16 TFLOPS vs H100: ~989 BF16 TFLOPS). B200 spot is now cheaper on a cost-per-TFLOP basis than the cheapest H100 OD market. (Cross-provider tickers, June 1, 2026)
5. B200 OD on-demand price shows 0.0% delta across 24h, 7d, and 30d in all three regions — this is the pricing fingerprint of a deliberately held list price, not a market-clearing price. (Query tickers, June 1, 2026)
6. Jensen Huang at Computex (May 2026): "AI token generation is now profitable" — economic validation that inference-scale Blackwell deployment is accelerating, which drives the spot supply surge. (News feed, May 2026)
7. Qualifying caveat: AWS is the only provider with real spot market mechanics; GCP/Azure "spot" discounts are fixed estimates. The B200 spot market is currently AWS-only and AWS-region-constrained (US only). EU/APAC B200 spot expansion would be the key confirmation of a fully global liquid market.

Implied Action: Operational: shift interruptible training jobs to B200|SPOT|us-virginia at $3.32/hr immediately. This is a 63% cost reduction vs. H100 OD in Virginia ($8.91/hr) for superior hardware. On a cost-per-TFLOP basis it beats every H100 market globally except Frankfurt distressed-floor ($0.40/hr). For investors: the $3.32–4.57/hr effective B200 price is structurally deflationary for any provider selling H100 OD at $6–9/hr on multi-month contracts. The spread is untenable beyond 90 days. Watch for AWS re:Invent announcements confirming B200 spot market depth, and for B200 SPOT first listings in EU regions as a confirmation catalyst.

H3 — H100 Reserved Contracts Have Inverted: A Hidden Credit Event Is Building Confidence: 4 / 5

Thesis: The Frankfurt H100 on-demand market has fractured in a way that reveals the secondary market is in price-discovery breakdown — and reserved-rate holders have no escape. The intra-market spread is 1,281%: one provider is dumping H100 at $0.40/hr while another holds at $5.49/hr, with an 8-observation median of $2.94/hr and a -42% volume discount (meaning bulk buyers in this market are getting H100 below the median). Against this backdrop, a single provider simultaneously doubled its H100 RESERVED_1YR Frankfurt rate from $4.76/hr to $9.51/hr in 7 days — a transparent attempt to lock in long-term commitments at elevated rates as the spot floor collapses beneath the buyer's feet. The reserved rate is now 24x the cheapest available OD instance in the same city. In Dublin, H100 SPOT is clearing at $1.28/hr (with instances visible at $0.014/hr at the extreme), Amsterdam at $0.071/hr. The ratio of committed reserved rate to spot floor has moved from an uncomfortable 2–3x to an economically indefensible 50–100x in low-demand European markets.

This is not just an arbitrage curiosity — it is the mechanics of how a credit event builds. CoreWeave's $35.1B debt stack was financed on the assumption that GPU revenue/hr stays elevated. Every dollar of reserved-rate compression flows directly to the top of the income statement. APLD's (Applied Digital, $47.28) master-lease structure with CoreWeave (150MW, Polaris Forge 1, Ellendale ND) provides partial insulation because APLD's revenue is power-and-space, not GPU-hours — CoreWeave bears the rate compression risk. But if CoreWeave's debt becomes stressed, APLD loses its anchor tenant, which represents a different but equally serious risk. Core Scientific's 590MW CoreWeave contract across 11 facilities (10-Q, May 6) is the same exposure, at three times the scale.

Key Evidence:

1. H100 OD Frankfurt: median $2.94/hr, min $0.40/hr, max $5.49/hr — 1,281% spread, 2 providers, 8 observations, -42% volume discount. This is the most structurally distressed market in the dataset. (Query tickers, June 1, 2026)
2. H100 RESERVED_1YR Frankfurt: $9.51/hr, +100% over 7 days (from $4.76/hr) — provider attempting to lock in buyers as spot floor collapses. (Top movers + query tickers, June 1, 2026)
3. H100 RESERVED_1YR premium EU/ME: London $7.13/hr, Dublin $8.90/hr, Zurich $9.99/hr, Dubai $10.06/hr — all flat for 30 days; committed contracts not repricing despite deteriorating spot/OD markets. (Query tickers, June 1, 2026)
4. H100 SPOT Dublin: $1.28/hr median (with visible floor near $0.014/hr); Amsterdam: $0.071/hr — reserved/spot ratios of 100x+ in low-demand European markets. (Query tickers, June 1, 2026)
5. CRWV Q1 FY2026: total debt $35.1B, cash $2.27B, current ratio 0.315, FCF -$4.71B/quarter on $2.08B revenue. Two consecutive EPS misses (-30% Q4, -22% Q1). (10-Q/equity fundamentals, May 8 + June 1, 2026)
6. APLD Q1 FY2026: CoreWeave 150MW master lease (Polaris Forge 1); APLD is isolated from GPU rate compression by lease structure but exposed to CoreWeave anchor-tenant credit risk. FCF -$720M on $127M revenue — 5.7x capex/revenue ratio justified only by infrastructure scarcity premium. (APLD fundamentals, June 1, 2026)
7. Qualifying caveat: The Frankfurt H100 OD fragmentation (1,281% spread) may partly reflect thin-market auction noise on a 2-provider, 8-observation ticker. The reserved rate doubling is from a single provider and could be a data artifact. Verify with H100|RESERVED_1YR|us-virginia as a deeper-liquidity cross-check.

Implied Action: If you hold H100 reserved capacity at $7–10/hr, initiate contract renegotiation conversations now — before providers see the OD market deterioration fully in their own revenue data. The $9.51/hr Frankfurt RESERVED ask is a last-gasp attempt to lock in buyers; reject it. For investors: monitor CRWV bond credit spreads (company has $35B of publicly filed debt obligations) for any rating action as the August 12 earnings approach. The Frankfurt fragmentation is the canary; the core-market H200 price cascade is the mechanism; CoreWeave's 0.315 current ratio is the fuse.

H4 — The Budget Inference Tier Is Real: $0.07–$1.38/hr Legacy GPUs Signal Market Segmentation Confidence: 3 / 5

Thesis: Oracle Cloud Infrastructure has simultaneously listed AMD MI25 (Vega 20, ~2018) GPUs across 15+ global regions and Intel V710 (2017 virtualization GPU) in two US markets — all first-ever catalog listings showing identical 24h/7d/30d deltas, confirming these are activation events rather than price moves. At $0.065/hr in Frankfurt (MI25) to $1.38/hr in Madrid, these SKUs are positioned 10–50x below H100 pricing in the same markets. The underlying business logic is sound: as frontier inference (GPT-5 class, Claude 4 class) migrates to B200-class hardware, a large residual market exists for cost-sensitive workloads — embeddings, RAG retrieval, classification, small-model fine-tuning — where 16–24GB of Vega VRAM is fully adequate. Intel's Computex launch of the Crescent Island GPU with 480GB VRAM validates that the inference memory wall is real for frontier models, but the inverse is also true: for sub-frontier workloads, legacy VRAM is more than sufficient at a fraction of the cost.

The thesis is real but the market is embryonic. Stockholm is the only MI25 market with 12 observations and a stable price (~$0.70/hr, +9% over 7d) — every other region has 1–3 observations and zero spread, meaning prices are list-set by OCI without competitive validation. The 25x price dispersion across markets ($0.065/hr Frankfurt to $1.38/hr Madrid) reflects opportunistic regional pricing, not market clearing. The budget inference tier exists; it is not yet investable as a market signal.

Key Evidence:

1. MI25 catalog activation: 15+ regions, all showing delta-24h = delta-7d = delta-30d, confirming first-ever listings within a ~30-day window. Price range: $0.065/hr (Frankfurt) to $1.38/hr (Madrid) — 21x dispersion on identical hardware. (Query tickers, June 1, 2026)
2. MI25 Stockholm: n_observations = 12 (highest in MI25 dataset), $0.70/hr, +9% 7d — only market with genuine pricing depth. All others: n_observations ≤ 3, zero spread. (Query tickers, June 1, 2026)
3. Intel V710 US listings: Texas $0.87/hr (+1,840% delta), Virginia $0.44/hr (+308%) — identical catalog activation pattern, single provider. (Top movers, June 1, 2026)
4. Intel Computex 2026: Crescent Island GPU at 480GB VRAM — validates inference memory-wall segmentation that creates demand for the budget tier at the opposite end. (News feed, May 2026)
5. H100 Frankfurt floor: $0.40/hr OD (distressed) vs. MI25 Frankfurt: $0.065/hr — H100 floor is now only 6x the MI25 legacy price, a much narrower gap than the 127x nominal gap. (Cross-ticker comparison, June 1, 2026)
6. Qualifying caveat: Single-provider market with no verified price history means these prices have no competitive anchor and could be revised without warning. Stockholm's 12-observation depth is encouraging but still insufficient for trend analysis; no 90-day history was returned for any MI25 ticker.

Implied Action: This is a watch-and-test play, not a trade today. The operational opportunity — testing MI25 workloads for batch embedding, small-model serving, and classification at $0.07–$1.38/hr — is real and accessible now. The investment signal becomes tradeable when n_providers moves from 1 to 2+ in at least 3 markets, confirming competitive pricing rather than unilateral catalog maintenance. Watch for A10, T4, or L4 ON_DEMAND_DEV listings at sub-$0.50/hr globally (fresh activations already visible in Texas, Ohio, Singapore) as the next phase of budget-tier buildout. The larger implication — further commoditization pressure on mid-tier A100-class capacity at $2–4/hr — is a second-order bear case for any neocloud with significant A100 inventory.

H5 — Power Is the Non-Fungible Asset: Infrastructure Scarcity Will Outperform GPU Operations Confidence: 5 / 5

Thesis: Every data stream in this analysis — pricing, news, regulatory intel, SEC filings — converges on a single bottleneck: guaranteed megawatts at a predictable price are the irreplaceable input that determines who survives the GPU price war. GPUs are becoming commoditized (this brief documents H100 floors at $0.40/hr and H200 at $3.01/hr); energy and interconnection are not. Northern Virginia's grid interconnection queue now stretches 14 years (Forbes / Data Center Dynamics), meaning any project initiated today in the world's densest AI datacenter market cannot receive power until 2040. ComEd (northern Illinois, 80+ operating datacenters) issued a +12% residential rate hike in June explicitly attributed to datacenter load growth — the externalities of AI infrastructure are now flowing directly into utility rates and, consequently, into community and legislative opposition. Fourteen US states are actively considering datacenter bans or pauses; Maine passed one (vetoed by governor); Florida's SB 484 prevents utilities from passing datacenter costs to residential ratepayers.

The NERC reliability guidelines update for "emerging large loads" (intel_feed, May 6) is the formal regulatory acknowledgment that gigawatt-scale datacenter interconnections are a new class of grid risk — the kind of acknowledgment that generates years of compliance process for new entrants. SoftBank's €75B / 5GW commitment in France is fundamentally a European grid capacity land-grab before the interconnection queue closes, not merely a capex announcement. China's offshore wind-powered undersea datacenter in Shanghai Lingang (24MW live, 22.8% lower PUE) is a proof-of-concept for a structurally cheaper energy model that, if scalable, establishes a permanent ~20–25% compute cost advantage for Chinese AI operators.

The public equity plays that best express this thesis — APLD ($47.28, analyst target $64.59, strong buy, 11 analysts) and IREN ($63.54, analyst target $79.84) — both have multi-year power contracts and are spending 5–6x revenue on infrastructure capex that makes sense only if power-secured sites command a durable scarcity premium. APLD's recent +143% EPS beat (most recent quarter: +$0.09 vs. -$0.21 estimate) confirms the model is operationally improving even through the investment phase. Core Scientific's 590MW CoreWeave contract (10-Q, May 6) across 11 geographically distributed facilities is the best publicly documented example of the model: Core Scientific owns the power asset; CoreWeave buys the access; Core Scientific captures the scarcity rent.

Key Evidence:

1. Northern Virginia grid interconnection queue: 14-year wait for new large-load connections. (Forbes / Data Center Dynamics, cited in news feed, June 2026) — the dominant US AI datacenter market is effectively closed to new power supply.
2. NERC reliability guidelines update for "emerging large loads" (intel_feed, power_grid category, May 6, 2026) — formal regulatory acknowledgment that gigawatt-scale datacenter interconnections require new reliability frameworks, creating multi-year compliance complexity.
3. ComEd +12% residential rate increase, June 2026, explicitly attributed to northern Illinois datacenter load growth — political blowback is becoming quantifiable and legislative. (News feed, June 2026)
4. 14 US states considering datacenter bans/pauses; Maine passed legislation (vetoed); Florida SB 484 enacted. Virginia resident comfort with new datacenters dropped 34 percentage points since 2023. (News feed, regulatory intel, June 2026)
5. SoftBank €75B / 5GW France commitment — at 5GW, represents ~5% of France's total grid capacity to a single operator. Confirms hyperscaler-scale operators are pre-committing European grid capacity aggressively. (News feed, May/June 2026)
6. APLD Q1 FY2026: $775M capex vs. $127M revenue (6x ratio), CoreWeave 150MW master lease, 52-week range $6.68–$50.73, current price near top with analyst consensus $64.59. Most recent quarter: EPS +$0.09 vs. -$0.21 estimate (+143% beat). (Equity fundamentals, June 1, 2026)
7. Core Scientific 10-Q (May 6, 2026): 590MW contracted CoreWeave capacity, 11 facilities across 7 states — geographic diversification away from the constrained Virginia market is itself a power-scarcity hedge. "Pursuing acquisition of new sites including land and power capacity" — explicit land-banking behavior. (SEC filing, May 6, 2026)
8. JLL projection: ~100GW of new datacenters 2026–2030; IEA projects datacenter electricity demand hits 945 TWh by 2030 — exceeds Japan's total national consumption. The physical infrastructure requirement is a structural multi-year demand pull. (News feed, 2026)
9. Qualifying caveat: Power-infrastructure equities (APLD, IREN) have already run significantly (+73% and +55% over 1 month respectively). The scarcity thesis may be partially priced in at current levels. New site permitting delays could be asymmetric — they hurt new entrants but also cap existing operators' expansion optionality.

Implied Action: Long APLD (analyst target $64.59, +37% upside from current $47.28) and IREN (analyst target $79.84, +26% upside from current $63.54) as the primary power-scarcity thesis expressions. Monitor HIVE Digital ($4.52, +103% in 1 month, renewable-energy focus) as a higher-beta/higher-risk variant. Short or underweight pure GPU-rental plays without owned power assets — they face margin compression from both directions simultaneously: falling GPU rental rates (H100 floor now $0.40/hr in Frankfurt) and non-compressible or rising energy costs. The clearest short pair against the power-infrastructure long is CRWV: it rents power from Core Scientific and APLD (bearing the cost), rents GPUs to customers (with the rate compression risk), and carries a 738x debt-to-equity ratio as the cushion. Key catalysts: New power purchase agreements (PPAs) from hyperscalers for 10+ years at below-market rates; any additional state-level datacenter legislation following the Florida SB 484 model; NERC formal rule-making on large-load interconnection.

Risk Flags

Immediate (0–30 Days)

R1 — H100 Frankfurt Secondary Market Breakdown The 1,281% intra-market spread ($0.40–$5.49/hr, 2 providers, 8 observations) with a -42% volume discount is the most acute immediate risk signal in the dataset. This is not normal price discovery — it is a market where one provider is dumping H100 capacity at distressed prices while another holds out. When a market fragments this severely, one of two things follows: the high-end provider capitulates and prices collapse to the low-end floor, or the low-end provider exits and the floor disappears. Either outcome is negative for anyone holding H100 OD capacity at $3–6/hr. Watch for the Frankfurt median to break below $2/hr as the confirmation signal.

R2 — H100 RESERVED_1YR Frankfurt: Possible Pricing Trap A provider doubled its H100 RESERVED_1YR Frankfurt rate from $4.76 to $9.51/hr in 7 days — while spot in the same market trades at $0.40/hr. This is either a data artifact (single-provider, possible entry error) or a deliberate lock-in attempt. Either way, any enterprise buyer signing a 1-year H100 reserved contract at $9.51/hr today is entering a 24x premium-to-spot commitment in a market where the secondary floor has already collapsed. Do not sign new H100 reserved contracts above $3/hr in EU markets without a market-exit clause.

R3 — MI300X Dev-Tier Global Reprice: AMD Compliance Exposure The simultaneous -77.5% (Seoul), -39.4% (Paris), -39.3% (London), -19.7% (Frankfurt) reprice of AMD MI300X ON_DEMAND_DEV tickers — against a counter-move of +38.3% in São Paulo and +37.7% in Melbourne — is too coordinated to be random. Combined with BIS export guidance issued May 31 updating compliance requirements for MI350-class chips reaching Chinese-affiliated entities in any jurisdiction, this looks like AMD (or a partner cloud) restructuring its dev-tier access. Any enterprise customer with MI300X dev-tier capacity in Seoul or EU markets should verify that their contracts are not subject to renegotiation or termination in the next 30-day compliance window.