Daily Investment Brief — June 22, 2026

Market Pulse

The compute market has bifurcated cleanly along geographic lines: European inference GPU capacity is entering a new structural pricing regime while US capacity remains in oversupply, and the data is unambiguous. L40S SPOT in Frankfurt has traded within pennies of $2.23/hr for 73 consecutive days — not a trend, a new equilibrium — while the equivalent Ohio L40S falls toward $0.37/hr at a 73% spot discount that keeps widening. This EU/US split is confirmed by five independent L4 SPOT series: Paris (+20.4% over 30 days), Stockholm (+17.4%), London (+8.9%) are all appreciating, while Ohio (-23.9%) and Virginia (-37.7%) decline. The catalyst is policy-driven and structural: the Anthropic Claude export block for non-US nationals, confirmed June 12–13, has made sovereign EU inference infrastructure an enterprise imperative rather than a procurement preference. Meanwhile the generational overlap between H100 and B200 is resolving into a 1.83× price ratio ($1.78 vs $3.25/hr SPOT in Ohio) that is consistent with relative performance value — neither in distress, both finding floors — while the energy grid, not NVIDIA's fab, emerges as the binding constraint on all future supply additions.

Key Movers

Noise flags: V100 Tokyo +489% and Inferentia Virginia -96% are both thin single-provider anomalies — end-of-life or decommissioning signals with zero forward-looking validity. H100 SPOT in US-Iowa and US-Ireland show spreads of 8,797% and 6,580% respectively, indicating stranded sub-penny instances skewing medians; treat those regions' H100 pricing as unreliable.

Investable Insights

H1 — EU Inference GPU Scarcity Is a Structural Trade, Not a Spike Confidence: 5 / 5

Thesis: European inference GPU spot capacity is supply-constrained at a structural level, and the Anthropic Claude export block has transformed a gradual tightening into a hard inflection. The mechanism is compounding: EU enterprises that built production Claude Fable 5 / Mythos 5 applications now face a US government export control wall — not a pricing issue or API outage that can be remediated in a sprint cycle. Their options are local deployment (requiring EU GPU capacity), migration to open-weight models like Zhipu AI's GLM-5.2 (753B parameters, MIT license, released June 17), or loss of their core AI capability entirely. The first two paths both add demand for EU-region GPU capacity, and the EU AI Act's data provenance requirements mean that demand cannot be satisfied by routing inference to US clusters. OVHcloud's CEO made this explicit at VivaTech (Startup Fortune, June 18): the export block "just made our pitch a lot easier," with EU sovereign cloud IaaS now forecast to surge +83% to $12.6B in 2026. That is enterprise procurement, not spot speculation — it creates durable on-demand demand that will absorb spot pools and compress the spot/OD discount further. The pricing data has been tracking this story with perfect fidelity for 73 days.

Key Evidence:

1. L40S SPOT Frankfurt pinned at $2.23/hr since April 10 — 73 consecutive days without reversion; spot discount compressing from ~55% to ~54% (live ticker, June 22)
2. L40S SPOT Madrid +43.6% over 30 days to $0.88/hr, spot discount compressing from ~72% to ~69% vs. $1.705/hr on-demand — a steady, non-volatile uptrend with no single-day reversal > 2% (live ticker, June 22)
3. L4 SPOT sweep: Paris +20.4%, Stockholm +17.4%, London +8.9%, Frankfurt +2.0% over 30 days — all EU L4 regions positive, all core US L4 regions negative; directional split is 100% consistent (live tickers, June 22)
4. Inferentia SPOT Frankfurt +113% vs. Virginia -96% over 30 days — the same EU/US demand divergence across a completely different SKU class (ASIC), confirming the effect is regional and not SKU-specific (live tickers, June 22)
5. Anthropic Claude Fable 5/Mythos 5 export block confirmed as policy-driven (US government letter), effective June 12–13 — not reversible via API update; EU enterprises are structurally disrupted (InDaily SA, June 22)
6. OVHcloud VivaTech announcement: EU sovereign cloud IaaS forecast +83% to $12.6B in 2026 explicitly linked to US export controls as demand driver (Startup Fortune, June 18)
7. Qualifying risk: If the Anthropic block is reversed within 30 days — which is possible but unlikely given it is a government export control letter — the catalyst dissipates; additionally, GLM-5.2 adoption in EU enterprise contexts may face EU AI Act scrutiny on Chinese-origin model provenance

Implied Action:

• Long L40S and L4 on-demand rate cards in EU regions. Operators with reserved EU L40S capacity at the $1.70–$2.05/hr on-demand rack rate who can flex into the spot market at converging spot rates have durable carry. The Frankfurt pool is at new-equilibrium pricing; Madrid is still discovering its new floor.
• Long CRWV ($117.95, analyst target $140, "buy") as the only large-cap neocloud with confirmed EU operational footprint directly benefiting from EU sovereign inference demand. At 37% below its 52-week high of $187, it has given back the AI hype premium but not the underlying demand thesis.
• Monitor trigger: L40S|SPOT|gb-london and L40S|SPOT|de-frankfurt spot discount compressing below 50% — that is the signal the on-demand rate card itself is being bid up, representing an additional positive catalyst for EU neocloud operators.
• Avoid: Short-duration L40S SPOT trading in EU regions — the Frankfurt pool's 73-day price lock has zero volatility to exploit; the trade is structural, not tactical.

H2 — EU/US Inference Bifurcation Validates the Training → Inference Demand Shift Confidence: 4 / 5

Thesis: The geographic split in inference GPU pricing is not merely a supply story — it is the clearest real-time validation of the enterprise transition from AI experimentation to always-on production inference. When GPU spot prices in EU inference regions (L4 Paris, L40S Frankfurt) move directionally opposite to US inference regions (L4 Virginia -37.7%, L40S Ohio -17%) over the same 30-day period, the explanation cannot be a global demand cycle. It must be a regional deployment wave: EU enterprises are going live with production inference workloads that consume available spot capacity without releasing it, while US enterprises have already over-provisioned spot capacity that is now bleeding price. The A10G Ohio story — two full boom-bust cycles in 90 days, crashing from $0.65/hr to $0.27/hr and back to $0.55/hr — represents the older pattern: experimental burst jobs consuming and releasing single-provider pools. The L40S Frankfurt pattern (rangebound $1.14/hr in January, parabolic ramp to $2.23/hr by April 10, then locked for 73 days) represents the new pattern: sustained production deployment. The Vultr/HPE GB300 NVL72 deployment announcement (Data Center Knowledge, June 17) named this shift explicitly — "production deployments tied to customer-facing applications" — and the pricing series shows it arrived in EU markets first, most likely because EU enterprises face sovereign AI mandates that US enterprises do not.

Key Evidence:

1. L40S Frankfurt ON_DEMAND context: $2.00–$2.05/hr across EU regions vs. $1.62/hr in US-Ohio — EU rate cards are already structurally higher, and spot is converging toward them rather than diverging (live tickers, June 22)
2. A10G Ohio 90-day cycle: $0.27/hr Jan→March, spike to $0.65/hr April 25, crash to $0.27/hr June 11, re-spike to $0.55/hr today — identical parabolic pattern twice in 90 days = experimental burst workload, not structural demand (live ticker history)
3. L40S Ohio SPOT -17% over 30 days to $0.37/hr vs. L40S Frankfurt 0% change at $2.23/hr — a 6× price divergence on identical hardware (live tickers, June 22)
4. Vultr/HPE GB300 NVL72 deployment: explicit validation of "training → production inference" transition from a hyperscaler-adjacent operator (Data Center Knowledge, June 17)
5. Qualifying nuance: The A10G Ohio re-spike (+49% in 7 days) is episodic noise and does not invalidate the structural divergence thesis; it is the same burst-and-release behavior observed in April, driven by a single workload event in a thin single-provider pool

Implied Action:

• Position in inference-tier L40S/L4 on-demand (not spot) in EU regions — the rate card stability at $2.00–$2.05/hr Frankfurt is the floor, not the ceiling.
• Monitor A10G Ohio with a 7-day horizon only — the $0.55/hr re-spike will likely crater again within days as the burst job completes; do not confuse it with the EU structural trend.
• Neocloud equity read-through: APLD ($46.59, "strong buy," $73 target, +74% over 3 months, only -8% from 52-week high) has the strongest inference GPU utilization profile without the valuation stretch of ARM/AMD — the pricing data supports continued utilization gains if they have EU exposure.

H3 — B200/H100 Price Ratio Has Found a Functional Equilibrium; H100 Floor Is Firming Confidence: 3 / 5

Thesis: The B200/H100 SPOT ratio in US-Ohio is currently 1.83× ($3.25 vs $1.78/hr), which is analytically coherent given B200's ~1.5–2× inference performance advantage over H100. This is not a distressed ratio — it is a functional market clearing price. The more important finding from the 90-day B200 history is that the current $3.25/hr is not a launch-price correction: B200 actually troughed at $1.86/hr in early February (before H100 had even bottomed), then staged a multi-month rally to $6.35/hr peak on May 20, and has since corrected by 49% to today's level. The $3.25/hr is a mid-cycle re-stabilization, not a collapse. For H100, the evidence of a firming floor is equally compelling: three successive price troughs have each been higher (January start ~$2.19/hr → April trough ~$1.50/hr → current re-test at $1.78/hr holds above April levels), and the 2-provider spread has compressed to 15%, suggesting the Ohio market has cleared. The critical structural risk — Azure or GCP activating B200 catalog listings — has not materialized: B200 remains AWS-only in the depreciation catalog, preserving AWS's unilateral ability to set the price without competitive cross-provider pressure. However, B200 has appeared in Virginia ($2.81/hr), Oregon ($3.58/hr), and Mumbai ($5.83/hr), suggesting AWS is expanding the Blackwell footprint deliberately, not defensively.

Key Evidence:

1. B200 SPOT Ohio 90-day history: $3.56/hr start → $1.86/hr trough (early Feb) → $6.35/hr peak (May 20) → $3.25/hr today; current level is above the February trough, consistent with mid-cycle stabilization (live ticker history)
2. B200/H100 spot ratio: $3.25/$1.78 = 1.83× — within the performance-value range of 1.5–2× (live tickers, June 22)
3. H100 Ohio SPOT: 2 providers, 15% spread, current range $1.66–$1.91/hr — compressed spread indicates market has cleared, not fragmented (live ticker, June 22)
4. B200 now active in 4 AWS regions (Ohio $3.25, Virginia $2.81, Oregon $3.58, Mumbai $5.83) — geographic expansion signals AWS confidence in Blackwell demand, not supply-dump behavior (live tickers, June 22)
5. B200 CATALOG_SURVIVAL: AWS-only; Azure inactive, GCP inactive — no competitive cross-provider pricing pressure on B200 at this time (depreciation data)
6. Key risk: If Azure or GCP activates B200 listings, the current AWS-set floor at $3.25/hr faces genuine competitive pressure that could push it toward $2.00–$2.50/hr, dragging H100 below $1.50/hr in the process

Implied Action:

• H100 neocloud operators are not facing the margin compression the -21% 7-day headline implies. At a 1.83× B200/H100 ratio, H100 remains economically rational for workloads not requiring B200's peak throughput — particularly inference fine-tuning, RAG pipelines, and single-model serving under 70B parameters.
• Watch trigger: Azure B200 catalog activation. This is the single event that would break the H100 floor thesis — an 8-K or product announcement from Azure enabling B200 instances should be treated as a negative catalyst for H100-heavy neocloud operators (APLD, HUT) and a positive catalyst for customers.
• B200 in Mumbai at $5.83/hr (+79% premium to Ohio) is the most actionable geographic arbitrage signal: non-US Blackwell demand is price-inelastic, directly supporting the export-control demand fragmentation thesis.

H4 — Power Constraint Is the Most Underpriced Risk in Neocloud Valuations Confidence: 4 / 5

Thesis: Every neocloud DCF model built in 2024–2025 was predicated on GPU availability as the binding constraint. That assumption is operationally obsolete. Grid interconnection queues for US data centers now average 2,100+ days; ERCOT received 198 GW of large-load applications in Q1 2026 against ~86 GW of current peak load — a 2.3× oversubscription that means the majority of announced Texas data center capacity will never get grid power. The Columbus, Ohio data is the most surgical data point in this brief: 84-month interconnection delay for new Columbus-area facilities means no new Ohio GPU capacity will be grid-connected before 2033. The H100 SPOT Ohio market with only 2 providers at a tight 15% spread is not a coincidence — it is the pricing signature of a protected, grid-moated oligopoly. FERC's unanimous order to expedite AI data center connections (Broadband Breakfast, June 20) sounds positive but contains a structural poison pill: data centers bear the full cost of grid upgrade infrastructure, adding $50–200M per project to capex models that do not currently include this line item. NERC's "Risk Mitigation for Emerging Large Loads" guideline is codifying frequency/voltage trip settings for large AI loads — regulatory compliance overhead that adds months to commissioning timelines. Behind-the-meter gas generation (16–30 months to commission) is the bridge, but it requires pipeline access, firm transport capacity, and basis exposure management — competencies that are conspicuously absent from most neocloud management teams' hiring profiles.

Key Evidence:

1. Columbus, Ohio grid interconnection delay: 84 months — no new GPU capacity can grid-connect in Ohio before late 2032; H100 Ohio 2-provider spread at 15% is the pricing signature of this constraint (Intel feed / regulatory data)
2. ERCOT Batch Zero: 198 GW of large-load applications vs. 86 GW current peak — 2.3× oversubscription; historical 10% build-out rate implies only ~20 GW of the announced pipeline gets built (Ascend Analytics, May 5)
3. FERC unanimous order to expedite AI data center connections: cost of grid upgrades borne fully by data centers, not ratepayers — direct capex headwind (Broadband Breakfast, June 20)
4. NERC "Risk Mitigation for Emerging Large Loads" guideline: codifying frequency/voltage trip settings for large AI loads — regulatory compliance cost and commissioning delay (Intel feed)
5. Neocloud equity betas: APLD β=5.64, HUT β=6.04, IREN β=4.23, BTDR β=2.45 — 2–6× market beta with no explicit power procurement risk premium priced into forward P/E multiples
6. CRWV at $117.95 vs. 52-week high of $187 (-37%): the market has begun pricing something; the question is whether it has priced enough of the power constraint overhang (equities feed, June 22)
7. Qualifying factor: US energy policy environment is actively pro-data-center in 2026; FERC expedite order could meaningfully compress interconnection timelines in favorable regions

Implied Action:

• Relative value trade: overweight neoclouds with existing, powered, grid-connected capacity; underweight neoclouds with announced-but-unpowered 200MW+ expansions. CRWV (Ohio + EU capacity, existing grid connection) is the cleaner long vs. pure-expansion plays.
• Specific short trigger: Any neocloud 10-Q or 8-K that shows capex guidance increase without a corresponding confirmed power agreement — that is the signal that grid delay risk is materializing against an already-spent hardware commitment.
• HUT at $124.44 vs. analyst target of $119 (target is below current price despite +148% 3-month move) — the equity has overrun its fundamental valuation; the power constraint thesis provides a concrete mechanism for mean reversion.
• Watch for: BTM gas or nuclear power agreement announcements from APLD or CRWV — these would be structural moat signals that justify premium multiples.

H5 — US Export Controls Are Fragmenting Inference Demand Globally; EU/APAC Are the Structural Beneficiaries Confidence: 3 / 5

Thesis: The consensus China AI narrative is focused on training restrictions — H100/A100 export bans, Huawei Ascend 910C as the domestic alternative. This misses the inference fragmentation story entirely. The Anthropic export block is the first visible manifestation of a policy vector that is compounding in real time: US models are becoming inaccessible to non-US developers, open-weight alternatives (Zhipu AI GLM-5.2 at 753B parameters, MIT license, released June 17) are filling the void, and the inference compute requirements for running 700B-parameter models dwarfs the requirements for API calls to hosted models. A single GLM-5.2 inference deployment at full precision requires approximately 1.5 TB of GPU memory — achievable only with multi-GPU clusters using L40S (48GB), H100 (80GB), or equivalent. ByteDance's pivot to Iluvatar CoreX domestic GPUs (50,000 "TianGai-100" AI inference chips; Crypto Briefing, June 15) confirms this is not theoretical — Chinese tech giants are actively building parallel inference infrastructure outside the US supply chain. The GDDR7 supply crunch created by Samsung/SK Hynix/Micron diverting fab capacity to HBM for AI training clusters is the structural memory feedback loop: RTX 50 consumer GPU allocations cut 15–20% (Tech Times, June 21), pushing incremental inference compute toward cloud GPU spot markets and supporting L4/L40S spot floors globally. The SK Hynix market cap reaching $1.35T (briefly displacing Samsung as Korea's largest company) is partly an inference memory story, not just a training one.

Key Evidence:

1. Zhipu AI GLM-5.2: 753B parameters, MIT license, 1M-token context — requires multi-GPU inference clusters; released same day as Anthropic export block confirmation (SCMP, June 22)
2. ByteDance Iluvatar pivot: 50,000 TianGai-100 inference GPUs explicitly for domestic AI inference infrastructure — confirms Chinese demand is actively exiting Western GPU supply chains (Crypto Briefing, June 15)
3. GDDR7 supply crunch: 15–20% RTX 50 allocation cuts as Samsung/SK Hynix/Micron divert to HBM — incremental inference compute redirected to cloud GPU markets, supporting L4/L40S spot floors (Tech Times, June 21)
4. SK Hynix $1.35T market cap, ~57% HBM market share: HBM is now the density constraint on inference GPU VRAM per cluster, directly relevant to 700B+ parameter inference scaling (GPU news feed, June 22)
5. B200 SPOT Mumbai at $5.83/hr (+79% premium to US-Ohio): non-US Blackwell demand is price-inelastic, validating the EM inference infrastructure build thesis (live ticker, June 22)
6. DeepSeek Entity List delay: approved but unpublished as of June 21 — policy ambiguity creates a window of continued inference demand from DeepSeek's global API; publication is a binary negative catalyst (Memeburn, June 21)
7. Primary risk: EU AI Act data provenance requirements create real headwinds to Chinese-origin model (GLM-5.2) enterprise adoption in Europe; the EU demand signal may be driven by Anthropic substitution with US alternatives, not open-weight Chinese models

Implied Action:

• Long MU (Micron) as the cleanest expression of both HBM demand (training) and GDDR7 scarcity (inference cloud redirection). MU at +38% over 3 months and -8% from 52-week high is the least crowded of the memory trades.
• SK Hynix ADR (Nasdaq listing reportedly H2 2026) — if priced at premium to Samsung's implied valuation, it is the highest-conviction HBM inference expression in the equity market.
• APAC L40S SPOT — B200 Mumbai premium confirms non-US Blackwell demand; watch for L40S SPOT tickers in Tokyo and Singapore beginning to appreciate as the inference fragmentation trade broadens.
• Avoid: The GLM-5.2 "EU inference boost" narrative as a near-term trade — Chinese-origin model adoption in EU enterprise is structurally constrained by regulation and will take quarters to develop, if it develops at all.

Risk Flags

Immediate (0–30 Days)

R1 — H100 Ohio -21.1% in 7 Days: Supply Flush or Demand Air Pocket? The H100 SPOT price in Ohio has dropped from a June 8 peak of $2.28/hr to $1.78/hr today. This is the second significant leg lower in 90 days — the first was the April trough at $1.50/hr. The 2-provider market with a 15% spread rules out a supply-side auction collapse (that would widen the spread dramatically). The more likely explanation is demand-side softening — enterprises pausing training workloads ahead of summer procurement cycles, or B200 capacity absorption reducing H100 reservation demand. If $1.78/hr cracks through the $1.66/hr min (the intraday floor from live tickers), the April trough at $1.50/hr becomes the next test. Below $1.50/hr, the H100/B200 ratio compression starts to threaten H100 neocloud economics in earnest. Watch: Daily H100 Ohio SPOT; the rate of change matters as much as the level.

R2 — A10G Ohio Single-Pool Volatility Is a Trap The A10G SPOT in Ohio has just re-spiked +49% in 7 days to $0.55/hr, mirroring the identical parabolic pattern that peaked at $0.65/hr on April 25 before crashing -59% back to $0.27/hr by June 11. This is a single-provider spot pool exhibiting burst-workload behavior: one large inference job consumes the pool, driving the spot price up; when the job ends or the operator switches to reserved capacity, the pool dumps. Investors or operators reading the +49% 7-day number as a structural inference demand signal are looking at the wrong ticker. The structural EU inference signal is in L40S Frankfurt and the L4 EU tier. Risk: If you are short the EU L40S thesis and long A10G Ohio as a hedge — the A10G will likely crash again before the L40S moves.

R3 — DeepSeek Entity List Publication Is a Binary Negative Catalyst The US approved DeepSeek's Entity List addition but has not yet published it as of June 21 (Memeburn). When published, it immediately disrupts DeepSeek's global API demand — which currently serves tens of millions of queries and provides baseline GPU demand across inference pools. More significantly, publication creates a policy precedent for broader Chinese AI developer restrictions, potentially affecting Zhipu AI (GLM-5.2) and others. This is a same-week risk. Watch: Federal Register for Entity List publication; any neocloud with Chinese customer exposure is immediately at risk.

Near-Term (30–90 Days)

R4 — Azure B200 Activation Could Break the H100 Floor B200 remains AWS-only in the depreciation catalog — Azure and GCP are both listed as inactive. This single-provider status is the primary reason the current $3.25/hr B200 Ohio floor has structural integrity: there is no competitive pricing pressure. If Azure activates B200 instances (plausible within 90 days given Azure's typical 3–6 month lag behind AWS on Hopper/Blackwell introductions), the B200 price will face cross-provider compression toward $2.00–$2.50/hr, and H100 would need to re-price below $1.50/hr to maintain the ~1.83× ratio. For every H100-heavy neocloud, this represents a ~15–20% revenue-per-GPU-hour compression scenario. Monitor: Azure instance type announcements; any news of ND-series Blackwell additions to Azure's instance catalog.

R5 — Anthropic Export Block Reversal Would Collapse the EU Inference Premium The EU inference pricing thesis (H1 and H3) is substantially dependent on the Anthropic export control block persisting. If the block is reversed — through diplomatic negotiation, a licensing exemption, or a technical workaround like a European data residency arrangement — EU enterprises would immediately de-prioritize sovereign inference infrastructure build-out. OVHcloud's $12.6B sovereign cloud IaaS forecast would be revised downward, and the L40S Frankfurt $2.23/hr equilibrium would face genuine mean-reversion pressure. The base case is that government export control letters are not reversed within 30 days, but this is a policy risk, not a market risk — it can materialize instantaneously. Watch: Any US-EU trade negotiation announcements touching AI model access or technology licensing.

R6 — HBM Supply Concentration Risk (SK Hynix Dependency) With SK Hynix holding ~57% of the HBM market, the GPU supply chain has single-point-of-failure memory risk. A yield issue at SK Hynix's Icheon fab, a geopolitical disruption to South Korean semiconductor exports, or an unexpected Samsung HBM3E yield recovery could each dramatically shift the HBM supply balance within a quarter. Given that HBM supply is the binding physical constraint on H100/B200/GB200 density, any negative SK Hynix supply event flows directly into GPU scarcity and upward price pressure; any positive Samsung supply event does the opposite. Monitor: SK Hynix quarterly earnings (next report expected July) for yield guidance and HBM allocation disclosures.

Structural (6–24 Months)

R7 — Grid Interconnection Backlog Will Strand GPU Capex at Scale The 2,100+ day average grid interconnection queue, ERCOT's 2.3× large-load oversubscription, and Columbus's 84-month delay are not cyclical — they are infrastructure deficits that take 5–10 years to clear at best. The consequence for GPU capital allocation is profound: announced data center capacity that is not currently grid-connected will face 3–7 year delays before it can accept GPU hardware. Neocloud operators who have purchased H100 or B200 inventory for facilities without confirmed power agreements are holding hardware that will depreciate faster than their interconnection timelines permit monetization. By the time their facilities are powered, H100 may be at $0.50/hr SPOT and B200 at $1.50/hr. The NERC regulatory overlay (frequency/voltage compliance requirements for large loads) further extends commissioning timelines. The 6–24 month manifestation: Capex writedowns and missed utilization targets at neoclouds with Texas/Midwest expansion pipelines that are not grid-confirmed.

R8 — Open-Weight Model Size Deflation Could Undercut Inference GPU Demand The GLM-5.2 at 753B parameters is being cited as evidence of massive inference compute demand. But the counter-thesis is equally supported by the evidence: the DeepSeek-R1 paradigm demonstrated that 7B–14B distilled models can achieve near-frontier performance on many enterprise tasks at 1/50th the compute cost. If enterprise inference workloads converge on smaller, more efficient models (MoE architectures, quantized 13B models, distillation pipelines), the per-query GPU memory requirement collapses and the inference GPU demand thesis loses its compute intensity premise. The L40S at 48GB VRAM becomes overkill for a 13B model that runs on an L4 (24GB) or even an A10G (24GB). Watch: Mix of inference model sizes actually being deployed in production — quarterly surveys of enterprise AI spend on model hosting vs. API consumption are the leading indicator.

R9 — GDDR7 Supply Recovery Could Deflate Cloud Inference Margins The current cloud GPU pricing floor for inference is partly supported by the GDDR7 supply crunch diverting RTX 50 GPU allocations away from on-premise inference deployments. If Samsung or Micron resolves GDDR7 supply constraints faster than expected — or if HBM capex is redirected back to GDDR following a training demand plateau — consumer GPU supply normalizes, RTX 50 allocations recover, and on-premise inference becomes competitive again with cloud L4/L40S at the current spot rates. Timeframe: GDDR7 supply normalization is a 12–24 month story, but early signals (Samsung earnings, DRAM spot price on DRAMeXchange) would appear in the next two quarters.

Charts

Four charts rendered above. Summary of what each shows: