On plan, struggling

U.S.-Iran peace deal signed; Strait of Hormuz reopening

Risk assets rallied and crude fell as the Hormuz reopening eased the oil supply premium. Equity-beta names led the book; the energy hedge lagged on oil moving down.

Nvidia prices 25 billion bond, AI capex financing signal for semis

Nvidia priced a 25 billion investment-grade bond on June 15, its first debt since 2021, more than three times oversubscribed and upsized from 20 billion, including a 30-year tranche maturing in 2056. Hyperscalers are funding AI infrastructure with debt rather than equity, a constructive longer-horizon signal for the semiconductor complex even with a weak near-term tape. Reaches the book through TSM.

AI-related stocks now about 39 percent of the S&P 500

AI-related names are roughly 39 percent of the S&P 500, an extreme concentration that amplifies single-factor market beta for any tech-overweight book. For NFAI this concentrates risk in the semiconductor exposure (TSM): the position carries more of the index's swings than its weight implies. Reinforces the systematic-versus-specific split shown on the risk page.

Korea's leveraged-ETF blowup hits semis; TSM caught in the feedback loop

South Korea's FSS admitted its May approval of single-stock leveraged ETFs on Samsung and SK Hynix was hasty after a 9.99 percent KOSPI collapse on June 23, with both names down more than 12 percent and circuit breakers tripped; the KOSPI has now triggered five market-wide halts in 2026, nearly half of all such halts since 2000. Contagion hit the US the same session: Nasdaq down 2.21 percent, the Philadelphia Semiconductor index down about 8 percent. Goldman holds a 12,000 KOSPI target, framing it as a valuation correction, not an AI-demand reversal. Meanwhile retail has rotated about 12 billion out of gold and bitcoin ETFs and roughly 20 billion into semis since April, with SOXX up about 81 percent and SMH about 60 percent. TSM has double exposure: it fabs for Samsung and SK Hynix's HBM competitors and trades as a KOSPI-adjacent sentiment proxy, so any LEAPS entered before Q4 carries elevated Korean feedback risk.

Fed pivots hawkish: BofA now sees three 2026 hikes to 4.25 to 4.50 percent

Bank of America reversed its hold call and now expects 25bp hikes in September, October, and December 2026, targeting a 4.25 to 4.50 percent funds rate, after Chair Warsh's hawkish June 17 FOMC and a May CPI of 4.2 percent, the hottest since April 2023. CME FedWatch prices September near 73 percent, October near 81 percent, and December near 88 percent; Deutsche sees two hikes, JPMorgan a hold, Goldman pushed cuts to 2027. This is the single biggest structural shift for the book: higher real yields compress high-multiple and long-duration valuations most. Growth and spec names like RDDT and SOUN carry the largest compression risk, PYPL is rate-sensitive, and the TSM January 2027 LEAPS thesis must be re-stress-tested against a 4.25 to 4.50 percent rate by Q4.

PJM grid emergency: record peak demand and a 50x spot-power spike in data-center country

DOE issued a Section 202(c) emergency order June 30 and PJM hit a record peak near 166,241 MW on July 2, breaking the 2006 record of 165,563 MW, with Maximum Generation and Load Management alerts running through the July 4 weekend. Backup-generator curtailment orders are active for data centers, and Northern Virginia spot power printed above 2,000 per MWh against roughly 40 normal, a 50x spike. The viral 'EEA2 grid failure' framing is imprecise: this is an emergency order plus max-generation alerts, not a full emergency-level collapse. Read: data-center power cost and reliability is now a live, quantifiable constraint on the AI buildout. The book holds no grid or power names, so this feeds as a risk factor on the AI capex nexus, reaching the book through TSM, not as a standalone equity signal.

Korean leveraged-ETF exposure to Samsung and SK Hynix is escalating: a real signal with a volatile number

Retail leveraged-product exposure concentrated in Samsung and SK Hynix is escalating fast, though every snapshot dates quickly: verified figures within weeks of each other show 10B dollars (Kobeissi), 16.8B for the Hong Kong-listed 2x SK Hynix product alone (now Hong Kong's largest ETF, up from 0.24B HKD in about eight months), and 45B market-wide across Korean leveraged products, up roughly 800 percent since early 2026. The trend is the durable fact; no single dollar figure is. Korea's FSS governor Lee Chan-jin publicly criticized the products on June 22 as enriching brokerages at retail's expense, and the breakage is documented, not theoretical: single-stock leveraged ETFs have printed 40 to 50 percent moves AGAINST their underlying on liquidity and tracking failures. The two names carrying this leverage are the same HBM/memory demand proxy behind the record Korea export print logged July 4, and it compounds the June 23 KOSPI feedback loop already on file: an unwind would be disorderly and independent of fundamentals. Adds a standing volatility premium to the semis complex; reaches the book through TSM.

LLM token-spend index rolls over: the AI trade's core pricing signal is cracking

Bloomberg's Silicon Data LLM Token Expenditure Index (dollars per million tokens across the market, the cleanest proxy for marginal willingness-to-pay under the 700B-plus AI capex boom) is down about 20 percent from its May peak after roughly doubling since December. Two readings coexist and both are held: bull, token prices have fallen over 90 percent since 2023 while total spend doubled, so cheaper tokens are expanding the market (digestion); bear (Andreas Steno Larsen), a durable pullback means AI vendors are losing pricing power with cost-sensitive customers, which breaks the transmission chain that funds the hardware: token economics feed AI software demand, which feeds hyperscaler capex, which feeds GPU demand, then HBM, then DRAM and NAND, then the supplier complex, reaching this book through TSM. Concrete stress signals: Amazon killed an internal AI-usage leaderboard after employees gamed it, and GitHub Copilot's June 1 move to token billing hit some users with roughly 19x cost increases. Treat as a PRIMARY LEADING INDICATOR for the AI-infra and semis factor: it fronts the memory-sector volatility cluster logged July 4 and pairs with the Korea export print logged the same day as its physical-demand counterpart. The two signals measure different points in the value chain and are never collapsed into one.

Korea's June exports break 100B with semis up 199 percent: physical AI demand is still white-hot

June exports printed 102.25B dollars, the first 100B month for any economy in the region besides Germany, China, and the US, up 70.9 percent year over year unadjusted (59.5 percent working-day-adjusted), the fastest export growth since 1978. Semiconductors alone were 44.82B, up 199.5 percent and above 40B for the first time; HBM specifically rose 171 percent to 12.68B. Verified nearly to the decimal against Bloomberg, Yonhap, and UPI. This is the PHYSICAL DEMAND CONFIRMATION LAYER: point-of-sale evidence that AI-driven HBM and DRAM deployment is accelerating now. It deliberately sits in tension with the token-spend index logged the same day, which measures whether the economics FUNDING that demand hold up. The fork to watch: does the token pullback stay a pause, or does it surface in Korea's July and August prints as lagging confirmation. Leading indicator (token spend) versus lagging confirmation (exports): the engine tracks both, separately.

Warsh on record: no comfort above 2 percent; markets price a 70 percent September hike

Chair Warsh, verbatim and confirmed across Reuters, CNBC, PBS, and AP: "If there were people... who thought that this central bank was going to be comfortable with an inflation objective above 2%, well, I guess they'd be disappointed." Core PCE is 3.4 percent (May), headline 4.1 percent. With funds near 3.6 percent, markets now price roughly 70 percent odds of a hike to about 3.9 percent at the September 15 to 16 meeting; the next FOMC is July 28. This upgrades the hawkish posture logged July 1 from interpretation to a market-priced probability: rate-cut base cases in any discount-rate assumption are pulled, and compression risk stays concentrated in high-multiple, long-duration names.

Memory correction: SNDK off about 26 percent from its June high as supply fears meet an unwind

SNDK is down about 26 percent from its June all-time high as of July 2, MU similar, and the DRAM ETF fell over 10 percent in two sessions. Confirmed drivers: Samsung and SK Hynix supply additions threatening 2027 pricing, plus a Morningstar analyst (Lorraine Tan) publicly warning of a 30 percent AI-stock correction, very likely the origin of the viral, unconfirmed 'Burry shorts Micron' screenshots, which remain NOT ingested. BofA simultaneously raised its SNDK target to 2,500 from 2,100: sell-side and tape are diverging, which reads as a leverage and positioning unwind (traders cite SK Hynix's Nasdaq listing next Friday as the technical catalyst) on top of a real multi-year supply debate, not a settled fundamental repricing. Treat NAND/DRAM as its own volatility cluster, distinct from the broad AI-capex thesis; with the PJM power constraint logged July 2, two independent inputs now converge on that thesis. Reaches the book through TSM.

Data-center lease commitments hit 850B, up 204 percent; Oracle is the leveraged tell

Bloomberg counts 850B dollars in confirmed future data-center lease commitments, up 204 percent year over year and 31 percent quarter over quarter. Oracle carries the largest book, roughly 250 to 261B against 34B of current cloud revenue, about 7.4x lease-to-revenue coverage, funded with 58B in new debt while the company cut roughly 20,000 jobs in a single email on March 31 despite record fiscal-2026 revenue. The mechanism: these are fixed multi-year obligations underwritten by an assumption of sustained AI demand growth, so apply the same DSCR lens the valuation engine uses on any levered bet. This is the FINANCING AND COMMITMENT LAYER of the AI capex chain the engine already tracks at two other points: the token-spend index logged July 4 (willingness to pay, now softening) and Korea's export print logged July 4 (physical demand, still accelerating). If token economics keep softening while lease obligations stay fixed, coverage stress lands on the most concentrated balance sheet first; Oracle is that name, versus Meta and Microsoft whose commitments are small relative to their balance sheets. The Nvidia 25B bond logged June 15 was the early marker of the same debt-funded buildout. Persistence: structural through the AI capex cycle, with the commitment count updating quarterly. Invalidation: cloud-revenue growth closing the coverage gap, or commitment growth decelerating alongside a token-index recovery. No Oracle position in the book; reaches it through TSM, whose demand these commitments underwrite.

NVDA authorizes an 80B buyback on a 3B beat and falls anyway: the market has stopped paying for AI beats

What changed: NVIDIA's FQ1 2027 (May 2026) printed 81.6B dollars of revenue (+85% year over year), 75.2B in data center (+92%), guided 91B, authorized an 80B buyback, and raised the dividend 25-fold - and the stock fell ~1.5% after-hours and ~4.3% over the following week. Why it matters: price/news asymmetry to the downside on the single largest index constituent is a late-cycle exhaustion signal - a beat that guides up, authorizes a record buyback, and still falls means the marginal buyer for AI beats is gone. Transmission: this is the equity leg of a single correlated bet the feed argues is mis-modeled as four separate factors - AI capex is now simultaneously a marginal driver of GDP growth, a marginal driver of investment-grade issuance, 20 to 30% of private-credit assets (the credit event logged the same day), and no longer rewarded in equity prices. If the correlation matrix treats 'AI equity,' 'IG credit,' 'private credit,' and 'GDP' as independent factors, it structurally understates portfolio VaR. It pairs with the token-spend rollover (July 4, willingness-to-pay softening) and sits in tension with Korea's still-hot HBM exports (July 4, physical demand) - the fork the engine already watches. Persistence: a positioning and late-cycle signal, not a fundamental one; NVDA's numbers were strong. Invalidation: AI names beating and RALLYING again (the marginal buyer returns), or the token index re-accelerating. The Burry 'buybacks vanishing' and NVDA-accounting claims circulating alongside this are NOT ingested - a non-filing Substack source (no 13F since Q3 2025) contradicted by filed data (buybacks +44% year over year) and by NVDA's own public rebuttal; the beat-and-fell tape is the citable signal, not his thesis. Reaches the book through TSM.

Alphabet prints its first negative quarterly FCF since the 2004 IPO on $44.9B of capex, raises the 2026 guide to $195-205B - and falls 7% anyway

What changed: Alphabet's Q2 2026 - revenue +24% to $119.8B, Cloud +82% to $24.8B, cloud backlog above $514B - came with $44.9B of quarterly capex, pushing free cash flow negative (−$5.9B) for the first time since the 2004 IPO. The full-year capex guide was raised from $180-190B to $195-205B and 2027 was guided to 'increase significantly.' The stock fell ~7% in the July 23 session regardless. Why: the AI-capex cycle crossed a line the market prices differently - from investment funded by operating cash to cash consumption - and a record-fundamentals beat repriced DOWN because the marginal revenue dollar now carries the highest capex absorption in the company's public life. Transmission: this is the cash-flow leg of the single correlated AI bet the book already tracks - it joins the beat-and-fell tell (July 13, NVDA), the data-center lease-commitment / Oracle leverage item (July 5), and the token-index rollover (July 4); hyperscaler capex staying up is TSM-demand-bullish while the de-rating and the debt-financed funding path are the bearish leg - hence mixed. With MSFT, META, and AMZN reporting within two weeks, the same test repeats across the complex. Persistence: through the hyperscaler earnings cycle and the 2027 capex guides. Invalidation: a megacap printing heavy capex WITHOUT the de-rating, or capex guides coming down. Reaches the book through TSM, whose order book the capex is.

Tesla goes FCF-negative the same day: capex +142%, operating margin 1.4%, and a $30B debt facility - the debt-financed leg of the AI buildout widens

What changed: Tesla's Q2 2026 - operating cash flow $4.7B against capex up 142% to $5.79B, free cash flow −$1.09B (from +$1.44B in Q1); GAAP operating income down 57% to $398M with operating margin compressed to 1.4% from 4.1%; record ~480K deliveries; and a secured debt facility allowing borrowing up to $30B against a $43.5B cash pile. The stock fell ~14% in the July 23 session. Why: the same mechanism as Alphabet on the same day - AI and robotics capex crossing from funded-by-operations to balance-sheet-supported - but on a far thinner margin base, so the equity repriced harder. Two of the index's largest constituents printing negative FCF in the same 24 hours for the same reason is a regime marker for how the market now prices AI capex, not two idiosyncratic misses. Transmission: joins the NVDA $25B bond (June 15) and the Oracle lease-leverage item (July 5) as the debt-financed leg of the AI buildout; a 1.4% operating margin leaves no buffer if pricing or deliveries soften, which is what makes the capex debt-financed rather than optional. Persistence: through the capex cycle; the facility is standing. Invalidation: FCF back positive with capex sustained (operations catching up), or the capex plan cut. Reaches the book through TSM.

Semis lose ~$1.5T with record ETF inflows on the way down: a drawdown without capitulation, and memory at 6.8x forward earnings

What changed: SOXX fell ~13.2% over four weeks (~15% off the late-June peak), roughly $1.5T of semiconductor market value destroyed since June 25 (Micron ~$350B of it; SanDisk, Intel, Applied Materials and Lam Research each over $100B). Micron, Samsung, SK Hynix and the memory complex are all more than 20% off recent highs - memory is in a bear market. Yet SOXX is still up ~84.6% YTD versus ~17.1% for QQQ, and Micron trades near 6.8x forward earnings against a ~16.8x long-run average (SOXX PEG ~1.26x, lowest since 2016, vs QQQ ~1.56x) while the fundamentals did not break: MU printed FQ3 revenue $41.46B (+346% y/y) with HBM sold out for CY2026. Why it matters: the trigger was reports that SK Hynix will slow HBM capacity expansion and shift toward commodity DRAM, plus AI-capex sustainability doubt - a demand-expectation shock, not a demand print. The bearish tell sits INSIDE the cheap valuation: SOXX, SMH and DRAM funds drew RECORD ETF inflows in July while prices fell the most since 2008, and Micron insider selling is at its highest since 2010. Retail bought every dip; there has been no capitulation, and drawdowns that end usually end with one. Transmission: this is the semis leg of the Korea cascade logged the same day (Samsung/SK Hynix are both the margin-debt collateral and the HBM supply), and it feeds hyperscaler-capex sustainability doubt (the July 22 Alphabet/Tesla negative-FCF prints). Persistence: through the memory-pricing cycle. Invalidation: capitulation (outflows on down days) marking a washout, or SK Hynix reversing the capacity guidance. Reaches the book through TSM.

Korea's leverage bubble unwinds: record 38.6T-won margin loans force-liquidate into a bear market - the live top of the Seoul-to-memory-to-hyperscaler chain

What changed: the KOSPI margin-loan balance, a record 38.63T won on June 24, is unwinding violently - the index is roughly 30% below its June 19 all-time high near 9,385, in a technical bear market, with more than 1.2 million leveraged accounts hit by margin calls and about 360,000 forcibly liquidated as the forced-liquidation rate jumped above 10% from a ~2.1% norm (about 2.3T won of forced sales over two and a half months). Why it matters: this is a LIQUIDATION, not a fundamentals repricing - Samsung Electronics posted a record quarter and the index still fell, because four names (Samsung and SK Hynix among them) carried an outsized share of the ~27.4T won in KOSPI margin debt, so when they dropped brokers issued calls simultaneously across the market. The book already half-tracked this (the July 4 Korea-exports and leveraged-ETF-concentration notes); it is now the live event, not a watchlist item. Transmission: Korea is the top of a chain that runs Seoul to memory (Samsung/SK Hynix) to SOXX to hyperscaler AI-capex demand - the same AI-capex-funding stress the July 22 Alphabet/Tesla negative-FCF events and the July 13 credit dispersion track, now with a forced-seller at the front of it; Nikkei weakness rhymes through the same semiconductor and carry channels. Persistence: acute while forced deleveraging runs (watch the margin-loan balance and the daily forced-liquidation figure); a leveraged-ETF deposit-rule tightening was reported to compound it. Invalidation: margin balances stabilizing and the forced-liquidation rate falling back toward its ~2% norm. Reaches the book through TSM (the memory/semiconductor demand chain); note the honesty limit - this moves the engine as KNOWLEDGE, not as a mechanical VaR input, since NFAI's correlations come from price returns, not headlines.

Held: 32 @ 404.25

Thesis breaks if: The AI capex cycle rolls over.