Daily Investment Brief — June 15, 2026

GPU & Compute Market Intelligence Brief

June 15, 2026 — Final Edition

Market Pulse

The GPU compute market is fracturing along three simultaneous fault lines: geography, generation, and bottleneck layer. The dominant narrative of "H100 oversupply" is demonstrably false at regional resolution — Montreal SPOT H100 has surged +107% in 30 days to $4.98/hr, compressing the SPOT discount against on-demand from 69% to just 34%, while Dublin simultaneously sits at $1.28/hr with a 6,573% intra-market spread reflecting near-zero inventory dumps. These are not different markets in different cycles — they are the same product, the same pricing tier, diverging because regional demand is outrunning regional supply unevenly. Layered on top of this geographic bifurcation: the B200's arrival has exerted only 3.38%/yr depreciation pressure on H100 pricing versus a historical average of 17.5%/yr, confirming Hopper and Blackwell are being treated as non-substitutable by the market. And beneath both: every major capacity signal — a 14-year Northern Virginia interconnection queue, 14 states with pending data center legislation, Korea's GPU buildout hitting grid limits — confirms that power, not silicon, is the medium-term binding constraint on the entire supply stack.

Key Movers

Noise filter note: Legacy/fringe hardware (MI25, M60, K80 DEV) dominating the 24h mover list are single-provider illiquidity artifacts with zero spread. UNMAPPED vCPU/RAM tickers showing -99.999% 7d deltas are taxonomy pipeline shifts, not price collapses. Neither category carries any signal value.

Investable Insights

H1 — H100 SPOT Is a Regional Market, Not a Global One: Montreal/Stockholm vs. Dublin/Frankfurt Confidence: 5 / 5

Thesis: The consensus view treating H100 SPOT pricing as a single global market is analytically wrong and is generating mispriced risk for operators and investors on both sides. The data reveals a cleanly bifurcated geography: a scarcity cluster centered on North America and Scandinavia, and a glut cluster across Western Europe. Montreal's SPOT H100 at $4.98/hr has risen +107% in 30 days, compressing its discount to on-demand ($7.67/hr) from 69% to 34% — and its spread between providers has exploded from 9.4% to 197%, meaning one provider is now quoting $7.45/hr (near on-demand) while another quotes $2.51/hr. That intra-market spread expansion of nearly 2,000 percentage points in 30 days is not noise — it is two providers experiencing the same demand shock but carrying different inventory positions. Stockholm tells the same story more cleanly: $3.54/hr, +63% in 30 days, spread expanding from 45% to 76% with 2 active providers, and still grinding upward (+11% over the last 7 days alone). A100_40GB in Montreal has independently confirmed the dynamic with a +132% SPOT surge, ruling out H100-specific idiosyncracy. Meanwhile Dublin sits at $1.28/hr with a 6,573% spread — one provider quoting $0.038/hr, another at $2.53/hr — which is a textbook inventory dump, not a competitive market. Frankfurt is at $1.79/hr with zero spread and a single provider, down -27%. These are simultaneously the same GPU tier and entirely different markets.

The 1-year reserved rate in Montreal at $7.01/hr now sits below the SPOT price trajectory — reserved is cheaper than where SPOT is heading. This is the clearest possible signal: operators currently relying on SPOT for bursting compute in constrained regions should be reserving capacity now before the discount fully compresses to zero.

Key Evidence:

1. H100|SPOT|CA_CENTRAL_MONTREAL at $4.98/hr, +107% over 30 days, provider spread expanded from 9.4% → 197% (live ticker data, June 15, 2026)
2. H100|SPOT|SE_STOCKHOLM at $3.54/hr, +63% over 30 days, spread 45% → 76%, +11% in last 7 days (live ticker data, June 15, 2026)
3. H100|SPOT|IE_WEST_DUBLIN at $1.28/hr, -49% over 30 days, 6,573% intra-market spread indicating near-zero inventory dump by one provider (live ticker data, June 15, 2026)
4. H100|ON_DEMAND|CA_CENTRAL_MONTREAL stable at $7.67/hr (-0.24% over 30d) — SPOT has risen from 31% to 65% of on-demand in the same market in 30 days (live ticker data, June 15, 2026)
5. A100_40GB|SPOT|CA_CENTRAL_MONTREAL +132% over 30 days, corroborating that scarcity is regional, not H100-specific (live ticker data, June 15, 2026)
6. No NatEx validation available for any H100 SPOT tickers — residual decomposition cannot be independently validated; prices should be treated as directional rather than precision estimates

Implied Action: Operators in Montreal or Stockholm with SPOT-dependent workloads should convert to 1-year reserved pricing immediately — at $7.01/hr in Montreal vs. a SPOT trajectory that has risen $2.40 in 30 days, the reserved rate could be below SPOT within 3–4 weeks at the current pace. For investors: any neocloud with concentrated North American SPOT revenue exposure is sitting on geographic pricing upside that may not be fully reflected in consensus models; conversely, any operator disclosing European SPOT heavy exposure faces ongoing ASP compression. Monitor H100|SPOT|SE_STOCKHOLM — a move above $4.50/hr in the next two weeks would confirm it is entering the same scarcity regime as Montreal, and would broaden the geographic scarcity thesis beyond a single market.

H2 — Blackwell Has Not Repriced Hopper: H100 Depreciation Running at One-Fifth the Historical Rate Confidence: 5 / 5

Thesis: The H200→B200 generation transition is producing the lowest annual depreciation rate ever recorded in our dataset at just 3.38%/yr — compared to a cross-sectional market average of 17.52%/yr (16 observations, p=0.041) and prior generation transitions that ranged from 23.25% (H100→H200) to 74% (K80→P100). The implied economic useful life of H200 under the B200 transition rate is 29.6 years — an obviously absurd number that simply encodes a market reality: B200's market presence is too thin and too fragmented to exert substitution pressure on Hopper-era hardware. The mechanism is well-documented. The Colossus 1 incident — where SpaceX's mixed Hopper/Blackwell architecture created inter-generational latency incompatibilities severe enough that it had to monetize the entire facility to Anthropic for $1.25B/month — is real-world evidence that B200 and H100 are not interchangeable in distributed training clusters. They require clean generational cohorts, which means B200's existence does not reduce H100 demand; it creates a parallel market. The B200 catalog availability data makes the mechanism concrete: B200 is currently listed as active at AWS only — not at Azure, not at GCP. GB200 is available only at Azure. This fragmented Blackwell availability means the vast majority of cloud providers have no B200 to substitute, removing any competitive pricing pressure on H100 from the supply side.

On-demand H100 pricing across all 20 global regions confirms the flatness with near-clinical precision: London -0.095% over 30 days, Tokyo -0.141%, Montreal -0.244%, Stockholm -0.121%, Seoul 0.0%. The H100 launched in March 2022 is 4.25 years into a model-implied 5.7-year useful life for training GPUs, running on a depreciation curve that is barely moving.

Key Evidence:

1. H200→B200 generation gap annual decay: 3.38%/yr vs. market average 17.52%/yr — running at 19% of historical pace (depreciation model, June 15, 2026)
2. Prior transitions for reference: A100→H100 at 53.65%/yr; H100→H200 at 23.25%/yr; K80→P100 at 74%/yr (depreciation model, June 15, 2026)
3. B200 catalog status: is_active: true at AWS only; GB200 at Azure only — Blackwell fragmented across providers, not yet a systemic H100 substitute (component catalog data, June 15, 2026)
4. H100 on-demand 30-day deltas across 20 regions: range of 0.0% to -0.244%; no region showing meaningful price deterioration (live ticker data, June 15, 2026)
5. SpaceX leased Colossus 1 to Anthropic for $1.25B/month through May 2029 — inter-generational Hopper/Blackwell latency incompatibility cited as forcing event (Stocktwits, June 14, 2026)
6. CoreWeave (CRWV) 10-Q filed June 2026 confirms H100-dominant infrastructure model with 590MW contracted from Core Scientific (SEC filing, June 2026)

Implied Action: CoreWeave (CRWV) at $100.55 — down 46% from its 52-week high of $187 and 28% below its analyst consensus target of $140.18 — is mispriced relative to the actual H100 depreciation signal. The market appears to be pricing in H100 commodity decay that the depreciation data says is not happening at anything close to historical rates. With earnings August 12 and H100 ASPs holding flat globally, a positive utilization print could be a meaningful catalyst from deeply oversold levels. The position thesis is simple: H100 is stickier than consensus models assume, CRWV's revenue base is H100, the stock has already discounted a depreciation scenario that isn't in the data. The key risk to monitor: any AMD earnings commentary (August 4) on MI450 yield rates — if MI450 ships at scale into the open cloud market, the non-substitution thesis begins to erode. Watch for MI300X or MI450 tickers appearing in the cross-provider pricing dataset as the clearest leading indicator; their current absence is itself evidence AMD's cloud ecosystem penetration remains thin.

H3 — AMD at $511 Is Pricing Perfect Execution Into a Warrant-Milestone Trap Confidence: 4 / 5

Thesis: AMD at $511.57 sits 5.7% above analyst consensus of $483.94 following a +183% run from $193 in early March — one of the most aggressive 90-day moves in large-cap semiconductor history. The Meta deal at its core is structurally reflexive: up to 160 million share warrants are tied to MI450 shipment milestones, not revenue milestones, which means TSMC CoWoS packaging yield and HBM3e allocation from SK Hynix and Micron are now directly embedded in AMD's equity risk profile. AMD doesn't get to miss a delivery milestone and explain it away as revenue recognition timing — the warrants vest on delivery, making physical execution the equity event. The bull case is real: $5.77B in Instinct GPU data center revenue (up 57% YoY), Meta's first 1-gigawatt cluster already deployed on AMD hardware, $2.57B in free cash flow in Q1 2026 (+253% YoY), and a PEG of 1.175 that is defensible for this growth rate. But the stock has already run the bull case. At 171x trailing P/E, AMD at $511 requires near-flawless execution on a chip that has never shipped at hyperscale volume in a CoWoS-constrained packaging environment where TSMC capacity is simultaneously being bid for by Nvidia Blackwell, Intel, and multiple custom silicon programs. Critically, the cloud pricing market provides an independent reality check that the equity market appears to be ignoring: despite $5.77B in data center GPU revenue, AMD's MI300X has zero visible market share in cross-provider cloud pricing — every dollar appears to be going into private hyperscaler clusters, not the open cloud. The CUDA moat is not just software; it is the entire ecosystem of tooling, model libraries, and operator familiarity that AMD has not yet displaced.

Key Evidence:

1. AMD at $511.57, up +183% in 90 days from $193; 52-week high $546 on June 3, -6.4% pullback to current (equities data, June 15, 2026)
2. AMD trailing P/E: 171x; PEG: 1.175; analyst consensus: $483.94 — stock is 5.7% above consensus after the run (equities data, June 15, 2026)
3. Meta deal: multi-year 6GW Instinct GPU rollout, AMD's largest AI chip deal ever, up to 160M share warrants tied to shipment milestones (Yahoo Finance, June 14, 2026)
4. AMD Q1 2026 Instinct GPU data center revenue: $5.77B, +57% YoY; FCF $2.57B, +253% YoY (AMD earnings data via AOL, May 17, 2026)
5. MI300X component listed in ticker dataset but zero active cross-provider pricing tickers — no open cloud market presence despite $5.77B in stated revenue (pricing database, June 15, 2026)
6. Next AMD earnings: August 4, 2026 — MI450 yield and shipment guidance will be the binary event
7. Broadcom (AVGO) at $382, PEG 0.70, analyst target $522 (+37% upside) — benefits from both AMD success (networking) and AMD failure (custom silicon acceleration) (equities data, June 15, 2026)

Implied Action: AMD above $520 resistance is an asymmetric short with August 4 earnings as the natural catalyst — any acknowledgment of CoWoS yield or HBM3e allocation constraints in the MI450 context would reset the stock to pre-deal levels. The more risk-controlled expression is long AVGO / short AMD: Broadcom at 0.70x PEG vs. AMD at 1.175x PEG while both are exposed to the same hyperscaler AI capex cycle, but Broadcom benefits from AI networking scale-up regardless of whose GPUs win the training cluster contest. The straight AMD long remains viable but is only a new entry below $480 (below consensus), where margin of safety exists. At current levels the risk/reward is unfavorable unless MI450 initial shipment confirmations appear in August.

H4 — Power PPA Scarcity Is Priced as Utility-Grade, Worth Semiconductor-Grade Confidence: 4 / 5

Thesis: The Helix Digital Infrastructure launch (KIA, Nvidia, KKR, Vistra — $10B+ committed, led by ex-AWS CEO Adam Selipsky) is an explicit statement by the most sophisticated AI infrastructure capital allocators in the market that signed power purchase agreements are now structurally scarcer than GPU inventory. Vistra's designation as preferred power supplier across 18 states in Nvidia's DSX AI factory framework is not a vendor relationship — it is an infrastructure co-design partnership where Vistra's capacity pipeline across PJM, MISO, and ERCOT territories becomes the physical substrate of every Helix AI factory. Yet VST trades at $148 — down 32% from its 52-week high of $219.82 and -6.9% over the last 3 months — at 13.5x forward P/E and a PEG of just 0.44. The market is pricing Vistra as a traditional power utility while the strategic reality is that it holds one of the most valuable scarce resources in AI infrastructure: existing interconnection rights and operating capacity at scale, in a market where new interconnection takes 14 years in Northern Virginia. The regulatory data leaves no ambiguity about the structural nature of the constraint: FERC's 2025 State of Markets report explicitly addresses AI/data center large-load dynamics; NERC reliability guidelines on "Risk Mitigation for Emerging Large Loads" have progressed to formal public documentation, meaning regulators now treat data center load profiles as grid stability threats rather than planning edge cases; CAISO's interconnection queue continues to back up with no relief in sight. Florida's SB 484 (signed by Governor DeSantis) — prohibiting utilities from passing data center costs to residential customers — is a template that creates adversarial utility-operator dynamics in new markets, making PPAs outside of existing agreements harder to structure. The Anthropic/Colossus 1 deal at ~$4.17/MW-hr bundled compute+facility provides a market data point on what locked capacity commands when demand is inelastic: it is not priced like a commodity.

Key Evidence:

1. Helix Digital Infrastructure launched with $10B+ committed capital; Vistra designated preferred power supplier across 18 states within Nvidia's DSX AI factory framework (IndexBox, June 14, 2026)
2. VST at $148, analyst consensus $225.29 — 52% upside to consensus at 13.5x forward P/E, 0.44x PEG, beta 1.4; down 32% from 52-week high $219.82 (equities data, June 15, 2026)
3. Northern Virginia ISO interconnection queue: 14-year wait time for new large-load interconnection (FERC/intel feed data, June 15, 2026)
4. 14 states with pending legislation to ban or pause new data centers; Florida SB 484 signed, prohibiting utilities passing data center costs to residential ratepayers (CNBC, May 9, 2026; intel feed)
5. NERC reliability guideline "Risk Mitigation for Emerging Large Loads" filed at formal public documentation stage (intel feed, June 2026)
6. Anthropic leasing Colossus 1 at $1.25B/month through May 2029 (~$45B total), implying ~$4.17/MW-hr for bundled locked capacity — frontier lab demand effectively inelastic (Stocktwits, June 14, 2026)
7. CoreWeave 10-Q: contracted 590MW from Core Scientific — confirms scale of power commitment required for neocloud operations (SEC filing, June 2026)
8. VST next earnings: August 6, 2026 — first earnings with Helix partnership as a disclosed factor; Helix announcement has not yet flowed into sell-side models

Implied Action: VST at $148 with $225 analyst target and a freshly disclosed Helix partnership that the market has not yet repriced is the cleanest asymmetric risk/reward in this entire brief. The long thesis does not require GPU pricing to cooperate — it only requires data center construction to continue, and every signal in this analysis (Anthropic's lease, Korea's 260,000-GPU buildout, CoreWeave's 590MW contract, Helix's $10B commitment) points in the same direction. Position sizing should account for VST's 1.4 beta — it will move with market risk appetite — but the structural catalyst (Helix) is independent of the macro cycle. The August 6 earnings call is the first opportunity for management to quantify the Helix PPA backlog, which could be the catalyst that begins repricing VST from utility multiples toward infrastructure multiples. Paired trade: long VST / short any neocloud disclosing it is in the interconnection queue rather than having executed PPAs — their capacity additions are 14 years away, not 14 months, and current valuations do not reflect that distinction.

Risk Flags

Immediate (0–30 Days)

R1 — Montreal H100 SPOT Inversion: Reserved Pricing Now Below SPOT Trajectory The 1-year reserved H100 rate in Montreal at $7.01/hr is already below where SPOT is heading if the current 30-day pace continues (SPOT at $4.98/hr, up $2.40 in 30 days, two providers now quoting $7.45/hr at the high end). Operators running significant SPOT workloads in Montreal are facing a structural cost inversion within 4–6 weeks if scarcity persists. The immediate risk: workload schedulers optimized for spot economics will misallocate if spot parity is reached, causing unexpected cost spikes in production pipelines. Watch the min price in H100|SPOT|CA_CENTRAL_MONTREAL — when it crosses $5.50/hr, the inversion is locked in.

R2 — Dublin H100 SPOT Inventory Dump May Be a False Floor Dublin's $0.038/hr H100 SPOT quote (the low end of the 6,573% spread) appears to be a near-zero residual from an instance-type reclassification or a batch capacity flush. If that provider removes the listing or reclassifies the instance, Dublin's floor price resets sharply higher, and any workload that has been scheduling to that $0.038/hr endpoint will face a sudden cliff. Operators relying on Dublin SPOT as a cost backstop should validate which provider is quoting the floor rate and assess its inventory durability.

R3 — SPAC/Blank-Check Overhang in AI Compute The SEC filing landscape shows an active wave of blank-check companies filing investor presentations in AI compute: SharonAI Holdings (SHAZ), Spark I Acquisition Corp (SPKL), Axiom Intelligence Acquisition Corp (AXIN), Axe Compute (AGPU). Capital formation in frothy SPAC mode around GPU infrastructure historically precedes a correction in private-market valuations when operational data fails to match prospectus assumptions. No immediate trading signal, but the SPAC overhang creates a secondary market risk: if any of these vehicles fail to close, the noise around "AI compute demand" narratives gets louder and less reliable.

Near-Term (30–90 Days)

R4 — AMD MI450 August 4 Earnings: Binary Yield Event AMD's August 4 earnings call is a singular event risk. The Meta warrant structure ties milestone payments to MI450 shipment volume, making TSMC CoWoS packaging yield a direct equity catalyst. If AMD discloses any yield, HBM3e allocation, or delivery schedule constraint on MI450, the warrant structure amplifies the downside — Meta's financial stake in AMD's execution means any miss creates uncertainty about the meta-deal's financial architecture, not just the hardware roadmap. Position management around AMD should account for August 4 as a binary date, not a routine quarterly.

R5 — CoreWeave August 12 Earnings: Customer Concentration Risk CoreWeave's 10-Q discloses meaningful customer concentration in its H100 infrastructure book. If utilization or ASP data disappoints, the stock's recovery from $100.55 (already -46% from its 52-week high of $187) stalls and the CRWV long thesis in H2 degrades. However, given the H100 depreciation anomaly (running at one-fifth historical rates), the ASP risk is lower than the market appears to be pricing in. The concentration risk is the real threat — a single large customer reducing commitments would be an earnings event regardless of market pricing dynamics.

R6 — Stockholm H100 SPOT Breakout: Second Scarcity Market Crystallizing Stockholm SPOT H100 at $3.54/hr (+63% in 30d, +11% in 7d, spread still expanding from 45% to 76%) is on an identical trajectory to Montreal 6–8 weeks ago. If Stockholm crosses $4.50/hr in the next 30 days, the scarcity regime has broadened beyond a single regional market, which would be a material signal for both EU-based operators (convert to reserved) and investors in neoclouds with European footprint.

Structural (6–24 Months)

R7 — Power Grid as the True Supply Ceiling: No Capital Solution Exists The 14-year Northern Virginia interconnection queue is not a bottleneck that additional capital expenditure can fix. It is a regulatory and physical grid infrastructure problem that operates on decade-scale timelines. With 14 states considering data center legislation, Florida already having signed cost-allocation reform, and NERC filing reliability guidelines for "Emerging Large Loads," the regulatory apparatus is formally acknowledging that data center power demand is a grid stability threat. The structural implication: the GPU supply stack's effective ceiling over the next 6–24 months is not chip production capacity (Nvidia/TSMC are scaling Blackwell) — it is power availability. Any neocloud or hyperscaler growth plan that relies on greenfield data center additions without executed PPAs and secured interconnection should be treated as speculative. The Korean grid stress story (260,000+ GPUs needing 20kW each vs. 3–5kW for traditional servers, 70%+ of capacity concentrated in Seoul) is a preview of what grid-constrained scaling looks like in practice.

R8 — Chinese GPU Market Self-Closure Is Permanent ByteDance's procurement of 50,000 Iluvatar CoreX inference chips — its third domestic Chinese GPU supplier after Huawei and Cambricon — combined with Chinese makers capturing 41% of China's AI accelerator server market (with Nvidia's share described as "effectively zero" by Jensen Huang) confirms that China is no longer a demand relief valve for Western cloud GPU pricing. Critically, at ~$1,775 per Iluvatar Zhikai chip versus $30,000+ for H100, Chinese domestic chips are viable for inference but not pre-training — meaning this market bifurcation is workload-stratified. Western GPU pricing for training-tier hardware will not receive Chinese demand pressure to normalize excess supply, but inference-tier pricing faces the permanent loss of a large addressable market that was already gone before this cycle. This is a medium-term structural headwind for any company whose GPU revenue thesis includes a China re-entry scenario.

R9 — Blackwell Cohort Adoption Creates Non-Substitution Cliff The H200→B200 non-substitution dynamic that is currently protecting H100 pricing will become a systemic risk in 12–24 months when B200 availability broadens beyond AWS-only and operators begin building clean-generation Blackwell clusters. The same inter-generational latency incompatibility that is currently supporting H100 pricing will — once Blackwell is broadly available — make H100 clusters economically stranded assets. Operators who lock into H100 at current sticky pricing without factoring in a 12–24 month depreciation cliff are mismodeling their TCO. The signal to watch: when B200 becomes active at both Azure and GCP (currently inactive at both), the non-substitution protection dissolves and H100 pricing should reprice sharply — the historical precedent (A100→H100 at 53.65%/yr) is the relevant template for what comes after the current anomaly resolves.

This brief was generated on June 15, 2026, using live cross-provider pricing data, SEC filings, regulatory intelligence, and equity feeds. All price references are point-in-time and subject to intraday movement. This analysis is for informational purposes and does not constitute investment advice.