Daily Investment Brief — June 25, 2026

Market Pulse

The compute market is bifurcating cleanly along two fault lines: inference versus training, and power-rich versus power-constrained. Mid-tier inference GPUs (A10G, L4) are in active spot reflation — Ohio A10G SPOT is up +120% in 12 days from a June 11 trough to $0.587/hr, completing the second full demand cycle in its 90-day history — while H100 on-demand prices across every major US region remain essentially frozen at $7.73–$11.28/hr with 30-day deltas of 0.0% to -0.4%. This price stasis is not market equilibrium; it is a supply ceiling installed by the power bottleneck. GPU hardware supply improved materially in H1 2026 (B200 ramps, broadening H100 availability) and yet on-demand prices didn't budge — the only mechanism that explains both facts simultaneously is that cloud providers cannot add incremental capacity regardless of silicon availability. Overlaying this domestic stasis, APAC spot markets are running hot across every geography (V100 Tokyo +298% in 30d, Inferentia HK +184%, H100 Jakarta +41%), with a fingerprint consistent with demand displacement from China's export-controlled operators routing through accessible endpoints. Three structural forces — the inference pivot, the power constraint, and the geopolitical compute arbitrage — are now operating simultaneously and reinforcing each other.

Key Movers

Thin-market noise to discount: V100 Tokyo and Tel Aviv A100 oscillations are single-provider AWS markets with structurally exaggerated amplitudes. AMD MI25 Singapore decline is pure obsolescence, not a competitive signal. Amsterdam H100 ON-DEMAND flat at $0.434 is a residual model artifact since April 16 — not a genuine arbitrage.

Investable Insights

H1 — Inference GPU SPOT Re-Inflation Is a Pre-Deployment Lead Indicator

Confidence: 5 / 5

Thesis: The A10G SPOT market in Ohio has now completed two full demand cycles within 90 days, each following an identical structural pattern: a sustained 8–10 week ramp as operators pre-provision inference capacity for forthcoming model deployments, followed by a 7-week unwind as deployed workloads normalize, followed by an accelerating re-ramp as the next deployment wave loads. The current re-ramp — +120% in just 12 days from the June 11 trough of $0.267 to $0.587 on June 23 — is tracking four times faster than the original February–April ramp, consistent with a known deployment calendar creating urgency around pre-provisioning. Critically, the signal is geographically synchronized across multiple regions and providers, ruling out a single-market capacity artifact: Ohio, Oregon ($0.712/hr, +33.7% in 7d), Virginia ($0.506/hr, +26.6% in 7d), and Jakarta ($0.500/hr, +46.9% in 30d) are all moving together. GCP L4 SPOT in Nevada (+234% in 30d) and Iowa (+135% in 30d) provide cross-provider confirmation. The demand catalyst is visible in the news calendar: GPT-5.5 rollout is imminent, Google Gemini 3.5 Pro is in queue for July, and multiple frontier model deployments are loading simultaneously. Inference GPU spot markets are a real-time leading indicator of model deployment pipelines — the pre-provisioning demand precedes go-live by 2–4 weeks, and the current velocity says go-live is imminent.

Key Evidence:

1. A10G SPOT Ohio reached a confirmed peak of $0.651/hr on April 25 after a 10-week sustained ramp from $0.320 on Feb 17 — a +104% move with a smooth, sustained shape (not a spike) confirming genuine workload placement, not hedging. (Live pricing data, June 23 2026)
2. Post-April-25 unwind: prices retreated to $0.267/hr by June 11 over 7 weeks, with spot discount expanding from 61.2% to 68.7% — textbook post-launch normalization pattern. (Live pricing data, June 23 2026)
3. Re-ramp from June 11 to June 23: +120% in 12 days to $0.587/hr — acceleration is 4× faster than original ramp, consistent with a known calendar-driven deadline. (Live pricing data, June 23 2026)
4. The denial condition — A10G Ohio reversing below $0.45/hr by end of June — is currently 23% away from triggering, with momentum firmly to the upside. (Live pricing data, June 23 2026)
5. Vultr publicly disclosed (Data Center Knowledge, June 17) that customer demand shifted from "experimentation to production" over the past 12 months — structural demand validation from a market participant.
6. Qualifying caveat: GCP L4 spot discounts are fixed estimates, so L4 Nevada/Iowa moves reflect demand modeling assumptions rather than real cleared spot prices. AWS A10G data is live market. Weight accordingly.

Implied Action: This is a short-duration tactical long with a 3–4 week horizon. Applied Digital (APLD, consensus Strong Buy, analyst target $73 vs $42 current) reports July 29 — a near-perfect timing alignment between the inference pre-provisioning wave and an earnings catalyst. The A10G Ohio price of $0.587 is currently $0.064 below the April peak; a breach above $0.651 within 7–10 days at current velocity would confirm the thesis is on track. Monitor A10G|SPOT|us-ohio daily. Exit trigger: if price reverses below $0.45/hr before June 30, the deployment pipeline may have shifted timelines. Avoid leveraged exposure on GCP L4 as price confirmation (fixed discount model, not real spot).

H2 — Jalapeño Erodes CoreWeave's H100 Pricing Floor; AVGO Is the Structural Beneficiary

Confidence: 4 / 5

Thesis: CoreWeave's entire $55B equity valuation rests on the assumption that H100/H200 inference pricing holds through 2027–2028 debt maturities — a capital structure containing $35.1B in debt against $2.3B in cash, -$4.7B quarterly FCF, and $7.7B of CapEx in a single quarter (Q1 2026). The H100 SPOT Virginia market has re-accelerated to $2.77/hr through Q2, and CoreWeave's Q1 revenue of $2.08B (+111% YoY) is genuinely explosive. But two concurrent developments are beginning to undercut the foundational assumptions: First, OpenAI and Broadcom unveiled "Jalapeño" on June 24 (Tech Times), a custom ASIC targeting 50% lower cost per inference token versus current Nvidia GPUs, with TSMC-manufactured engineering samples already delivered and 10GW deployment commitments beginning late 2026. Second, and more immediately, the H100 SPOT Virginia market has developed a 57% provider spread — min $2.21/hr (almost certainly Oracle OCI) versus max $3.48/hr (CoreWeave-tier pricing) — that did not exist in January when the market was single-provider at $2.47/hr with 0% spread. This spread is not demand-driven convergence; it is competitive undercutting. OCI's confirmed role as the cheapest provider across multiple EU tickers (Frankfurt H100 min at $0.397/hr vs AWS max at $3.13/hr — a 687% spread) establishes the pattern: Oracle prices GPU compute at aggressive floors to displace enterprise customers from Azure/AWS. The same dynamic is now active in Virginia SPOT. CRWV at $100.88 is already down -13.8% over 5 days and -4.4% in 24 hours — the market is beginning to price this risk. Broadcom (AVGO), conversely, is the direct design/manufacture partner for Jalapeño and trades at only 19.7x forward PE with a PEG of 0.69 — the most attractively valued name in the ASIC transition.

Key Evidence:

1. H100 SPOT Virginia: single provider, $2.47/hr, 0% spread in January → two providers, $2.77/hr median, 57% spread (min $2.21 / max $3.48) by June 23. Spread is expanding, not converging — confirming active competitive undercutting. (Live pricing data, June 23 2026)
2. OpenAI/Broadcom "Jalapeño" ASIC announcement: 50% lower inference token cost vs. Nvidia GPUs, engineering samples delivered, TSMC manufacturing, targeting late-2026 volume production. (Tech Times, June 24 2026)
3. H100 SPOT Virginia 90-day history confirms demand has genuinely absorbed supply through Q2 (sustained ramp from $2.47 → $3.12 peak → $2.77 current), but the two-provider split means the bid side is fragile — any softening in OpenAI/Microsoft utilization commitments triggers repricing. (Live pricing data, June 23 2026)
4. CoreWeave Q1 2026: Revenue $2.08B (+111% YoY), CapEx $7.7B in one quarter, FCF -$4.7B, total debt $35.1B. EV/Revenue ~14x with -$124 forward PE. Zero margin for demand deterioration. (CRWV 10-Q, SEC filing)
5. CRWV equity: -13.8% in 5 days, -4.4% in 24 hours as of June 25. AVGO: +20% over 3 months, PEG 0.69. The equity spread is already emerging as a directional signal ahead of Jalapeño production. (Equities data, June 25 2026)
6. Qualifying caveat: Jalapeño is pre-production. H100 SPOT has not breached the $2.20/hr denial threshold. The thesis is structural and lagged; H100 demand is genuinely strong through Q2 2026.

Implied Action: Short CRWV / Long AVGO as a pairs trade. The immediate catalyst is the CoreWeave Q2 earnings call on August 12 — any guidance revision on H100 cluster utilization rates or customer concentration metrics would reprice rapidly. The spread between CRWV -13.8% (5d) and AVGO +1.4% (5d) is already manifesting directionally. The denial condition is H100 SPOT Virginia holding above $2.60/hr through Q3 with spread compression (not expansion); monitor the spread metric weekly. If H100 Virginia spread exceeds 70%, upgrade confidence to 5/5 and increase short exposure. The risk to the short: if Jalapeño slips past 2027 in production timelines, CRWV's revenue momentum can sustain the valuation considerably longer.

H3 — Power Is the Binding Constraint: Hyperscaler Nuclear Moats Widen for 3+ Years

Confidence: 5 / 5

Thesis: The most underappreciated finding in this dataset is the reason H100 on-demand prices are frozen. GPU supply improved materially in H1 2026 — B200 ramps are underway, H100 availability has broadened — yet US on-demand H100 prices show 0.0% to -0.4% drift across every major region over 30 days. In a normally functioning supply/demand market, improving supply would compress on-demand pricing. The only mechanism that explains persistent price stasis despite hardware supply improvement is a supply ceiling installed by the power bottleneck: cloud providers cannot add capacity regardless of whether they can procure GPUs. The regulatory evidence is unambiguous. FERC issued a unanimous fast-track order (Broadband Breakfast, June 20) but it sets a 60-day clock for RTOs to respond with new study processes — not actual transmission builds, which take 2–7 years. The House committee report released June 24 explicitly states over half of US data center projects are forecasted to be placed on hold. Grid interconnection queues average 2,100+ days (84 months in Columbus, Ohio). ERCOT's new "batch connection process" (June 18 filing) creates a new study queue rather than solving the physical constraint. El Paso is opposing Meta's $551.8M BTM gas plant — even the fastest workaround (behind-the-meter generation) is encountering municipal opposition. Meta's own filing frames BTM gas as a "bridge period of up to five years" while awaiting grid interconnection — five years of premium bridge power costs are exactly the mechanism that structurally elevates cloud compute pricing as the on-prem alternative. The companies that locked in power corridors early — Microsoft, Amazon, Google, all with 10GW+ nuclear/SMR commitments — hold infrastructure moats that are not replicable under current regulatory timelines before 2029 at the earliest.

Key Evidence:

1. H100 ON-DEMAND US regions — Virginia ($8.69), Ohio ($7.73), Illinois ($11.28), Nevada ($10.73), Texas ($10.44) — all show 0.0% to -0.4% 30-day delta despite global GPU supply improvement in H1 2026. Price stasis in a supply-improving environment = physical capacity ceiling. (Live pricing data, June 23 2026)
2. FERC unanimous order directing six regional grid operators to expedite AI data center connections — explicitly framed as US-China AI competition policy. But 60-day RTO response clock addresses study timelines, not physical transmission. (Broadband Breakfast, June 20 2026 / FERC.gov, June 18 2026)
3. House committee report: over half of US data center projects forecasted on hold due to supply chain constraints. (House.gov, June 24 2026)
4. Columbus OH grid interconnection queue: 84 months (7 years). ERCOT creates new batch queue rather than accelerating capacity. (Intel feed regulatory data)
5. CPUC approved one (1) 250MW data center energization in California — "with modifications," months in process. This is the policy ceiling on even the "fastest" approvals. (Intel feed, June 23 2026)
6. El Paso municipal opposition to Meta's $551.8M / 366MW BTM gas plant, citing ratepayer protection — the BTM bypass route is now also facing friction. (El Paso gov filing, June 18 2026)
7. Strengthening signal: California voters polling 70% against new data centers in their communities. Policy headwinds extend beyond regulatory timelines to voter opposition. (Victorville Daily Press, June 17 2026)

Implied Action: Long hyperscalers with secured nuclear/SMR power agreements (MSFT, AMZN, GOOGL) versus short merchant GPU cloud operators dependent on 2026–2027 grid interconnection without disclosed PPAs. The FERC RTO 60-day response deadline falls in mid-August — the first indicator of whether the fast-track order produces measurable queue compression. Watch for any neocloud Q2 earnings commentary disclosing build delays due to power availability — that's the confirming catalyst for the short side. For data center REIT selection, focus on operators in PJM and ERCOT with existing transmission service agreements, not interconnection queue positions. Southeast Asia (India, Philippines) as greenfield power development markets may emerge as the structurally better geography for new capacity precisely because they're starting fresh. The 12–36 month time horizon here is not a limitation — it's a feature; this is a structural position, not a trade.

H4 — China Compute Scarcity Is Inflating APAC Spot Prices; AWS Is the Primary Beneficiary

Confidence: 3 / 5

Thesis: The $82,000 Chinese black market price for A100 servers — a 4–6× premium to list price — is the most direct measurement in this dataset of what compute scarcity under export controls is worth to a buyer. That price signal has a logical extension into legitimate cloud markets: operators in China-adjacent geographies who can access US cloud infrastructure will route inference workloads through available APAC endpoints. The pricing data is consistent with exactly this pattern. V100 SPOT Tokyo is up +298% in 30 days (a multi-week sustained climb from $0.658/hr to $2.621/hr, not a spike). Inferentia HK is up +184% in 30 days to $0.091/hr with a +14.8% 24-hour move still ongoing. H100 SPOT Jakarta surged +41% in 30 days and +27% in 7 days. TPU V6E in Tokyo is up +88% in 30 days. These are not random fluctuations — they form a geographically coherent pattern centered on the Guangdong access corridor (HK, Tokyo, Jakarta) that does not appear in Singapore, where on-demand pricing (-1.1% in 7d, -3.7% in 30d) reflects stable enterprise reserved-capacity buyers, not spot arbitrage. The Inferentia HK signal is particularly clean: there is no conventional production inference rationale for AWS Inferentia to nearly triple in Hong Kong without a demand pull from operators preferring cheap ASIC-accelerated inference through an accessible APAC endpoint. China's "LineShine" supercomputer topping the TOP500 list using only domestic chips (Tech Times, June 24) but ranking 4th on AI-relevant mixed-precision benchmarks confirms the performance gap is real — increasing the incentive for Chinese AI operators to source US cloud capacity through whatever channels remain open.

Key Evidence:

1. V100 SPOT Tokyo: $0.658/hr on March 12 → $2.621/hr on June 23, a +298% sustained multi-week climb (not a single-day spike). Prior 90-day peak was ~$5.54/hr in early January — current price is mid-ramp, with the prior peak as the logical target. (Live pricing data, June 23 2026)
2. Inferentia SPOT Hong Kong: $0.012/hr in January → $0.178/hr peak in March → $0.016/hr trough in June → $0.091/hr on June 23 (+184% in 30d, +108% in 7d, +14.8% in 24h). Single-provider AWS, single-region — genuine spot market dynamics. (Live pricing data, June 23 2026)
3. A100 servers trading at $82,000 on Chinese black markets, 4–6× list price, following smuggling crackdown. (MSN, June 25 2026)
4. China's "LineShine" supercomputer: 2.198 exaflops using domestic LX2/HiSilicon chips, tops TOP500. But ranks 4th on AI-relevant mixed-precision benchmarks — performance gap persists. (Tech Times, June 24 2026)
5. Korea semiconductor exports hit record $62B — broader APAC AI supply chain acceleration. (Tech in Asia, June 22 2026)
6. Singapore H100 ON-DEMAND at $3.16/hr: -1.1% in 30d — conspicuously absent from the APAC spot surge, consistent with the thesis that Singapore is dominated by stable enterprise/reserved buyers rather than spot routing. (Live pricing data, June 23 2026)
7. Downgrade caveat vs. original QCOM angle: QCOM at $197.41 is down -17.1% in one month with analyst consensus target of $186.50 — below current price. Qualcomm's datacenter revenues remain pre-commercial. APAC spot demand is routing through AWS infrastructure today, enriching AMZN/GOOGL, not QCOM. Expression pivoted accordingly.

Implied Action: Long AMZN and GOOGL as the primary expressions of APAC compute demand, not QCOM. Both hyperscalers are the infrastructure through which this demand routes — AWS for the V100/Inferentia spot tickers, GCP for the TPU V6E Tokyo signal. Monitor V100_32GB|SPOT|jp-tokyo for a sustained hold above $3.50/hr — that would confirm the APAC geopolitical demand thesis and validate approaching the prior $4.70/hr peak. The denial condition is US Treasury or BIS issuing guidance explicitly targeting cloud service provider liability for Chinese end-users — any compliance review that causes AWS to restrict APAC GPU availability would collapse these tickers immediately. Set this as a monitoring alert. The QCOM datacenter thesis is real but is a 6–12 month equity story, not a compute pricing signal story today.

Risk Flags

Immediate (0–30 days)

R1 — H100 SPOT Virginia Provider Spread Instability The H100 SPOT Virginia market entered January as a single-provider market at $2.47/hr with 0% spread. It now has two providers, a $2.77/hr median, and a 57% spread between the min ($2.21/hr) and max ($3.48/hr). This is not a market in equilibrium — it is a market where one provider is undercutting and one is holding. If the spread expands beyond 70%, the high-price provider (likely the one CoreWeave routes through) is exposed to demand migration. The denial condition for the CRWV short thesis is H100 SPOT holding above $2.60 with spread compression — currently moving the opposite direction. Watch daily.

R2 — A10G Inference Pulse Exhaustion Risk The current A10G re-ramp (+120% in 12 days) is the highest-velocity move in the inference GPU space. The April peak of $0.651/hr is the technical target. A failure to breach that level followed by a rapid reversal (back below $0.45/hr) would signal that the model deployment pipeline shifted timelines or that AWS is managing spot capacity differently. Given that the April post-peak unwind took 7 weeks, a long position entered here carries meaningful timing risk. The tactical long window is narrow — expect 7–14 days to the test of the April peak.

R3 — APAC SPOT Thin-Market Misread The V100 Tokyo and Inferentia HK tickers are single-provider AWS markets with extremely thin order books. A single AWS capacity management decision (adding one additional instance pool, or a compliance review) can collapse these prices 50–80% in days, as demonstrated by the Tel Aviv A100 pattern (peak $3.04 → floor $0.18, a 94% drawdown, in under 6 weeks). Do not extrapolate APAC spot moves as sustained demand signals beyond a 2-week horizon without corroboration in on-demand pricing.

Near-Term (30–90 days)

R4 — Jalapeño Timeline Slippage The OpenAI/Broadcom ASIC announcement (June 24) is the most consequential single news item in this briefing — but it is pre-production. Engineering samples are delivered; volume production is late 2026. Every month of delay reduces pressure on H100 inference pricing and extends CoreWeave's high-utilization window. The CRWV short thesis is structurally sound but operationally dependent on Jalapeño hitting its timeline. Monitor TSMC APAC order flows and any Broadcom ASG revenue guidance updates — these will be the leading indicators of whether the timeline is slipping.

R5 — FERC Fast-Track Reality Check FERC's unanimous order (June 20) set a 60-day clock for RTO responses. The mid-August deadline will be the first test of whether federal intervention can actually compress interconnection timelines. The most likely outcome is RTOs submitting new study processes without materially faster approval timelines — which would confirm the power bottleneck thesis. A surprise positive outcome (RTOs announcing expedited processes with 12-month timelines) would be a genuine threat to the structural cloud premium thesis and could catalyze a sell-off in hyperscaler premium valuations built on power moat assumptions.

R6 — Oracle OCI Underpricing Sustainability The Frankfurt H100 ON-DEMAND market shows a 687% spread (OCI min: $0.397/hr, AWS/Azure max: $3.13/hr). While the Amsterdam ticker is a residual model artifact, the Frankfurt multi-provider spread is real. If Oracle is pricing H100 at $0.397–0.434/hr as an enterprise displacement strategy, the question is whether this is sustainable. Oracle's June 22 10-K disclosed multiple debt tranches due 2026–2040 — the balance sheet can absorb below-cost GPU pricing as a customer acquisition investment. But if Azure or AWS respond by matching OCI in Frankfurt, the EU H100 on-demand price could compress by 50–70% within a quarter, meaningfully affecting both OCI's economics and the EU neocloud pricing environment.

Structural (6–24 months)

R7 — Custom ASIC Commoditization of Inference Pricing Jalapeño is the headline, but it is part of a broader pattern: AWS Inferentia, GCP TPU V6E, Cerebras wafer-scale silicon, and now OpenAI's custom chip are all targeting the same wedge — transformer inference at lower cost per token than Nvidia GPUs. If even two of these reach volume production simultaneously, the inference GPU spot market (A10G, L4, even H100 inference workloads) faces demand substitution from multiple directions. The current A10G/L4 re-inflation may represent the last cycle of strong inference GPU demand before ASIC displacement begins compressing utilization. Operators pricing multi-year inference contracts on Nvidia silicon today should build ASIC optionality into those contracts.

R8 — US Export Control Tightening Creates Binary APAC Risk The current China compute scarcity premium (APAC spot inflation, $82K black market A100s) exists because Chinese operators can still route through HK/Tokyo/Jakarta endpoints. If the US Treasury or BIS issues guidance expanding cloud provider liability for Chinese end-user compute access — a policy action that is consistent with the current administration's stated China AI competition priorities — AWS would face compliance pressure to restrict APAC GPU availability, potentially collapsing the APAC spot tickers by 50–80% rapidly and eliminating the compute scarcity premium that underpins the H4 thesis. The China LineShine TOP500 win will likely accelerate US policy discussions about compute access. Monitor BIS rulemaking dockets and any congressional hearing outputs on cloud provider export control compliance.

R9 — Neocloud Capital Structure Stress in a Power-Constrained Build Environment Multiple neoclouds have announced GPU cluster expansions predicated on 2026–2027 build timelines. If over half of those projects are genuinely on hold (per the House committee report, June 24), these operators face a scenario where they have signed Nvidia purchasing agreements, are paying carrying costs on hardware, but cannot deploy capacity due to grid interconnection delays. The resulting cash burn without revenue would stress capital structures — particularly for companies without MSFT/AMZN/GOOGL-scale balance sheets to absorb multi-year bridge periods. Watch neocloud Q2 earnings across the board for any disclosure language about "power timing" or "facility readiness" — these are the early signals of a broader capital structure distress wave in the sector.

All prices as of June 23–25, 2026. Spot prices reflect AWS live market data; GCP/Azure spot uses fixed discount estimates and should be weighted accordingly. This brief is produced from systematic cross-source analysis and does not constitute investment advice. Every hypothesis carries explicit denial conditions — monitor trigger levels weekly.