Master Thesis

Claim: Memory and adjacent input costs are inflating at a rate and persistence the market treats as cyclical and which is structural, and the shortage is now measurable in sovereign national accounts rather than only in corporate cost lines.

What changed since V9. The origin is corrected. V9 ran the cascade from Iran — war disrupts transport, transport inflates inputs. That is wrong and NVIDIA's CFO said so. Colette Kress: "Memory scarcity today is being driven in large part by the AI buildout itself." The demand is the cause. This is the single most important correction in V10.

Evidence.

• Kress guided NVIDIA gross margin down to 71–72%, bottoming Q4 FY27, from 75.0%, partially on memory prices — 300–400bp of compression at the most profitable company in the buildout.
• Bank of Korea raised 2026 GDP to 3.3% from 2.6% on an unprecedented semiconductor boom while hiking to 3.00% (6-1). Its issue note measures GDI +13.2% on terms of trade.
• US July goods deficit widened 17.2% to $118.8bn against a $100.5bn forecast, on capital goods imports +11.3% in the month — biggest since 1993, computers and semiconductors specifically. Korea's windfall is the American import bill.
• Marvell prepaying ~$1bn to suppliers in FY27, operating cash flow down sequentially because of it. Murphy: pervasive industry-wide constraints. ~$22bn of customer prepayments locked across the complex.
• Murphy on the buyer response: hyperscalers are modifying plans to use more CXL memory expansion because of scarcity — at multiple hyperscalers, extremely high volumes, additional design wins in two quarters. Transmission as architecture, not as a cost line.
• DRAM contract +90–95% in a quarter. 2027 sold out. Relief 2028.
• Capacity cannot be pulled forward. Samsung targeting ~50% HBM capacity increase by end-2026; SK hynix raising Yongin from ₩128tn to ₩600tn (~$410bn). The constraint is back-end — TSV throughput, hybrid bonding, TC bonder availability, KGSD yield. Money cannot buy the date.
• HBM4 ~$550/36GB 12-hi stack (~$15.28/GB), revised down from ~$600. No Micron or SK hynix primary source exists — all estimates. Basis warning: factory-gate per-stack and revenue-implied per-accelerator run ~2x apart and must never be mixed. HBM is now 30–40% of accelerator manufacturing cost, up from under 20% two generations ago.

Strongest counterargument. This is a classic memory cycle with an AI label. Every prior cycle produced sold-out claims, prepayments and capacity announcements at the top. The 2028 relief date is precisely when supply arrives, and the equity market — see the next page — is already trading it that way.

Claim: Unresolved. The commodity and the equities have decoupled, and the decoupling is itself the best falsifier the file has for whether chipflation persists.

What changed since V9. V9 flipped memory from structural long to fade on July 23. That flip is neither confirmed nor falsified — it is chopping, and V10 explicitly refuses to score it either way.

Evidence. MU 924.37 against a 1255 high, −26%. SNDK 1481.43 against 2354.39, −37% — after beating by six dollars, $39.25 against $33.275 estimated. SKHY 161.24 against 194.80, −17%. Weekly MACD histograms rolled negative on MU and SNDK. MU weekly RSI 59 off the 80s, stochastic 41.8 and falling. Two and a half months of maximum bullish news flow and not one has made a new high.

Strongest counterargument. This is digestion after a parabola, not distribution. MU ran 114 → 1255 in roughly fifteen months and still holds far above its 20-week channel. A fade looks like lower highs and lower lows with momentum expanding downward. This is a range.

The read that matters. The people closest to memory pricing will not pay up for it. Either the equity market is front-running the second derivative — memory names historically top when the rate of change peaks, not when price peaks, and +90–95% in a quarter with 2028 relief dated is a defensible peak print — or it is a mispricing that resolves violently.

Claim: Not "Treasury has capped the long end." Precisely: Treasury is attempting to suppress duration transmission, and the observable question is whether term premium breaks through. This conditions every other vein and remains conditional.

What changed since V9. Vein 10 was "rates and the long end." Renamed to the actual mechanism.

Evidence — the intent is on the record with four instruments. Aug 4: Bessent floated buybacks and cash management as objectives. Aug 13: reversion to quarterly refunding. Aug 25: reported preparation to temporarily limit long-term issuance. Aug 27: an official states the objective outright — focused on bringing long bond yields down. And the ESF operation (see Vein 9) adds a fourth instrument.

Who is winning — Treasury, as of today. On the day the Chair reopened a hiking cycle, TNX rose 1.16% and TYX rose 0.39%. The belly sold off three times harder than the long end. 30s−10s compressed 52bp → 48.4bp. 30s−5s fell 9.56% to 71.9bp, near the bottom of a 52-week range of 5.34–14.63, down from ~146bp a year ago. Everything repriced to the hawkish news except the piece Treasury is buying.

Auction evidence supports maturity-specificity, not general aversion. 20Y tailed, dealers 12.5%. 30Y TIPS stopped through, dealers 2.1%, indirects 84.4%. 2Y stopped through, dealers 10.9%. 7Y on Aug 27 stopped on the screws, cover 2.500 vs 2.490, dealers 12.26%, directs 26.96%, indirects 60.78%. The belly clears cleanly.

Strongest counterargument — and it comes from a Fed president. Goolsbee, Aug 28: "I don't think Fed and Treasury are at cross purposes." That directly contradicts this vein's framing, on the record, one day after it was written. Logged at full weight. Second counter, from Citi's Willer: asset-swap spreads are well behaved, and that is where fiscal stress should show first. He is nonetheless trading the cap — removed the Treasury underweight after Bessent, bought gold, short dollars. TYX at 5.211 sits 12bp below the 5.30% cap.

Claim: The yen carry trade is at extreme position, priced against a defence Japan can no longer credibly conduct alone, and the resolution mechanics moved from BOJ-only to a joint sovereign operation without a scheduled unwind.

What changed since V9. Concrete. On Aug 22, coordinated intervention. Japan spent ~$96.5bn defending 160; USDJPY sits at 160.14 today. The defence did not hold — and it did not fail cleanly either. A Treasury statement subsequently described the operation as an ESF swap: US dollars for JGBs, structured explicitly to avoid displacing Japan as the largest single sovereign holder of Treasuries. First US intervention to buy yen since 1998.

That makes the yen defence a duration operation. Yen instability forces Japanese Treasury liquidation, which lifts the US long end — the exact thing Treasury is spending buyback capacity to prevent. Vein 9 and Vein 10 are one operation, not adjacent threads. Constraint nobody has published: how much ESF capacity remains. It was an asset swap, not new money.

Configuration is carry-on, not unwind. AUDJPY 114.58 against USDJPY 160.14 implies AUDUSD ~0.7155 — the funding currency is weak against everything, not the dollar specifically. That is position being built. The unwind signature is violent yen strength and today is its opposite. Which is worse, not better: more position, less credible backstop.

TIC. Japan's US Treasury holdings $1,116.7bn in June, −$26.4bn m/m, ~$86bn below the November 2025 peak of $1,202.7bn. But June predates the intervention. The July print in mid-September is the actual test — a decline is near mechanically guaranteed, so the question is valuation versus genuine liquidation.

Layer 1 crossover. Japan imports ~95% of its energy from the Middle East. Yen weakness compounds the import bill directly into the energy exposure.

Strongest counterargument. The carry trade has been "spring loaded" in this file since V6 and has not sprung. A crowded trade can stay crowded for years, and the BOJ hiking into it in September is the orderly resolution, not the violent one.

Claim: Narrowed from V9's multi-theater crude call with Brent >$100. The claim that survives is refining-capacity-specific and product-specific.

What changed since V9. V9's own falsifier was Brent sustaining below $85 with all theaters active; Brent settled $89.31 and the broad crude claim is drifting toward it. The distillate claim got stronger while the crude claim got weaker.

The war as a standing condition — not a tracked vein. Iran is unresolved and trending toward a frozen conflict rather than a resolution — closer in shape to Russia/Ukraine than to something that clears. That is worse for a supply thesis than resolution, not better: the premium never clears and never resolves. It therefore stops being a catalyst to check and becomes a permanent cost floor written into this page. The Aug 2026 Hormuz filter stands: routine kinetics are FILE only. Currently open and unconfirmed: Rezaei says mediators formally requested Iran's conditions and Tehran is preparing a list, with a corridor agreed with Oman partly in Omani and partly in Iranian waters. The Oman framework was logged FALSIFIED on Aug 19 when Iran rejected ceasefire proposals and fired at the UAE; this is the first constructive movement since, and it is Al Manar via interpreter.

Evidence (Aug 28 settles). WTI $83.40, Brent $89.31 — both down. Diesel $4.3567/gal, up from $4.2787. Gasoline $3.4899. Diesel crack against WTI ≈ $99.58, a new high above the ~$98 print of Aug 13. Against Brent ≈ $93.67. Gasoline crack ≈ $57.27. Product spread ≈ $36.40/bbl between two products refined from the same barrel. Brent–WTI near $6–8, roughly double normal, says US crude is landlocked behind the bottleneck. Distillate stocks at the lowest late-August level since 1996. Russian refining capacity down more than half. Crude fell and products rose on the same session — the cleanest possible demonstration that this is a refining story.

New theater, and it lands on the narrow Claim: Trump confirms a naval blockade assisted from space; WaPo reports the US would take stakes in 17 Venezuelan oil fields. Venezuelan crude is heavy sour — distillate feedstock. A blockade tightens exactly the barrel that makes diesel. AMBER: blockade confirmed, oil-field stakes prospective.

The asymmetry. A corridor opening moves crude but does not build a refinery. If Hormuz reopens and Brent drops, the distillate crack widens — the input cheapens while the bottleneck is untouched. Consensus trades short crude on de-escalation headlines; this vein says long the crack into them.

Strongest counterargument — and the board already flags it. Price status: EXTENDED. A $99 crack is not an entry. The mechanism is the best-confirmed on the board and the vehicle is the worst-priced. Refining margin also mean-reverts violently once any capacity returns, and the file's own prior crack call was directionally wrong.

Open, cutting against a stated hypothesis. July industrial supplies exports fell 11.2%, which weakens the Aug 20 ledger hypothesis that US diesel exports explain the thirty-year-low inventory. Industrial supplies is broader than distillate — this narrows rather than kills it.

Claim: The credit market has begun pricing AI offtake quality explicitly, and the price is observable, large, and almost entirely unremarked. This is the board's only confirmed-and-unpriced vein.

What changed since V9. V9's framework said Layer 4 cracks at the weakest balance sheet first. That is retracted — it arrived at investment grade. What replaced it is better: the market is not cracking, it is differentiating, and the differential is quotable.

Evidence — the spread. IREN is running two GPU debt stacks simultaneously against the same collateral:

Same company, same year, same collateral, 300 basis points apart. The only variable is who signed the offtake. Banks take the Microsoft paper; private credit takes the rest at 9% against assets whose useful life is the most contested assumption in the buildout.

Supporting. IREN FY26 revenue $707.0m with AI Cloud revenue of only $128.8m against $1bn claimed operating ARR and $4bn contracted. The net loss of $702.6m is NOT the finding — $638.8m is non-cash impairment from scrapping mining hardware in the ASIC-to-GPU transition. Do not repeat the loss-equals-revenue framing.

Evidence — the SpaceX node, verified and now circular. V9 flagged this as inference requiring filing verification. Done, and it grew.

Alphabet holds ~551m SpaceX Class A shares marked at $94.176bn as of June 30 — over 95% of its 13F portfolio. $80.0bn is under short-term selling restriction and $14.1bn is locked until Q3 2027. Alphabet cannot sell into a decline; the mark runs through earnings while the position is immobile. V9's composition claim is retracted — Alphabet's disclosure reads SpaceX and a private company, reported to be Anthropic, whose valuation moved $350bn → $965bn in the same quarter. The split is undisclosed and Anthropic may be larger.

Scale. Q2 net income $112.2bn on revenue of $119.8bn. The $99.0bn equity gain added $77.1bn to net income and $6.26 to diluted EPS. Alphabet earned more marking two positions than operating a search and cloud business — and raised capex guidance to $195–205bn from $180–190bn in the same release.

The circle, with Alphabet at both ends. SpaceX (now listed, ~$140.87) reported Q2 revenue $7.8bn (+92%), AI revenue $2.6bn (+247%), capex $18.4bn in the quarter — up from $10bn in Q1, $15.8bn of it AI infrastructure — with capex held near those levels for two more quarters, against a $541m quarterly loss following $4.3bn in Q1. It discloses $14.1bn of contracted cloud agreements including Google and Anthropic. It committed to build exclusively on NVIDIA Vera Rubin, with Musk saying SpaceX will receive a very significant percentage of NVIDIA's GPUs next year, targeting 20GW by end-2027 against 1.4GW at Q2 end. Munster estimates SpaceX at ~5% of NVIDIA's Q2 revenue (~$4.8bn), up from 3%.

So: Alphabet's balance-sheet gain depends on SpaceX's valuation. SpaceX's revenue partly depends on Google as a customer. SpaceX's capex funds NVIDIA's forecast. NVIDIA's guide — which produced a $442bn single-day gain — is ~5% underwritten by a company burning $18.4bn a quarter against $7.8bn of revenue. SpaceX is treated as hyperscaler-tier by narrative; its cash-flow profile is a neocloud's.

Carried and still live. The buyback bid defended with leverage (Alphabet issuing into negative FCF; LVS repurchasing $787m and raising authorisation to $6.0bn into a miss). Off balance sheet: $904bn of leases not started, $1.52tn of purchase commitments.

Strongest counterargument. The 300bp spread is rational credit pricing working correctly, not stress — it is what a functioning market should do. Private credit at 9% against 90% LTV with contracted offtake may simply be well-compensated. And SpaceX's exclusivity is a commercial commitment from a company with $100bn of annualised revenue in sight, not a subprime borrower.

Claim: The binding constraint on the buildout is permission and interconnection, not capital or fuel. This is the highest-lead-time vein on the board — moratoria are announced months before they bind.

What changed since V9. Promoted to first class with mandatory tagging. One tagged ledger row against two published cards and a dated catalyst was a tracking failure, not a small vein.

Evidence. Pennsylvania became the third state to restrict data-centre developers, naming IREN among those hit. PJM backstop auction Sept 30 – Oct 21, with FERC having threatened to impose reforms if PJM did not act by September. October natural gas settled $2.888/MMBtu — the constraint is demonstrably not fuel cost. IREN's pipeline exceeds 5GW with grid-connected power secured above 4.5GW, targeting 0.3GW of IT capacity in 2026 and 0.8GW in 2027, power ramping from 2028. SpaceX targeting 20GW by end-2027 against 1.4GW today.

Crossover from Layer 3. DeepSeek now prices inference by time of day, off-peak rates, on the logic that the cheapest token is the one that can wait. Inference is being priced like electricity because it is electricity. If compute adopts power-market structure, power scarcity transmits into token prices on a daily cycle and the single list price stops being the unit of the market.

Strongest counterargument. Behind-the-meter and on-site generation route around interconnection entirely, and the operators with capital are already doing it. Permission constrains the grid-connected tier, not the buildout as a whole.

Claim: Demoted from HIGH. The output-price-collapse leg is falsified. What survives is narrower: infrastructure eats the cost and the control layer captures the margin — but only where the margin is GAAP-real.

What changed since V9 — the file's largest error, owned at full weight. V9 cited the Token Price Index at $1.58/M "and falling." That was a Q1 2026 reading quoted in a document written July 23, when the index stood near $2.60. It is $2.34 today, up 98.8% year over year. Wrong in level and wrong in direction, and the series was public, free and weekly the entire time. The output blade is not falling.

Evidence. TPI $2.34/M blended, +98.8% YoY, +4.8% MoM. Input $1.92, output $9.37, 22 members across 11 providers. Crossed $2 for the first time in late May when Google tripled Fast-tier pricing; peaked near $2.75 mid-July.

Mechanism caveat that must travel with the finding: the TPI is composition-driven and rises when more capable, more expensive models enter — Fable 5 entering in June at $22 blended lifted the index 12.3% in a single week. The honest reading is not that a fixed unit of intelligence rose 99%; it is that the cost of accessing the frontier rose 99% because the frontier keeps repricing upward. Fatal to the output-collapse claim either way.

The sharper finding is at the floor. The dearest-to-cheapest multiplier halved this week, 121x → 62x — but it closed because the floor rose, not because the ceiling fell. DeepSeek V4 Flash is now $0.352 after a price increase; the index's own July headline was "the cheap one raised the price." If the price leader has pricing power, inference supply is tight — which makes this series the missing cost-pass-through measurement for the commodity leg.

The counter, live this week. Z.AI's Ox Alpha — confirmed to Bloomberg Aug 26 as a GLM-series model, open weights the same night — processed ~26 trillion tokens on OpenRouter in four days (~11.6tn in the first 72 hours), roughly 2.6x the prior single-model launch record, taking the #1 slot and more than doubling DeepSeek's usage. A frontier-class competitor gave away product at record volume in the same week the paid floor rose. Both can be true — a time-boxed promotional preview is not a price — but this is the live test of whether the floor holds.

Corrections for the record: the 100tn tokens/day figure is an aggregate service-capacity claim from the OpenCode gateway, not measured traffic and not a per-user quota; a different provider claimed one quadrillion and the two have never been reconciled. Do not carry it as throughput. Do not carry the benchmarks either — one viral 10-task DeepSWE run scored 80% against Fable 5 at 65% and GPT-5.6 at 52%, while a private benchmark rated it near two-generation-old models. Irreproducible and conflicting.

And the tape does not support general margin compression. NVIDIA guided ~70% revenue growth for FY28, added $442bn of market cap in a day (second-largest in market history, behind Microsoft's $450bn last month), and stands at $5.5tn. US corporate profits hit a record $4.8tn in Q2, +22.8% YoY — the largest annual increase since Q4 2021 and, excluding the 2020 and 2008 recoveries, the largest in 21 years. But NVIDIA simultaneously guided gross margin down 300–400bp on memory. Volume and price are both winning; the input cost still bites. The demotion rests on the output-price falsification, not on an absence of cost pressure.

What survives, with evidence both ways. Marvell supports it: GAAP net income $308.0m against non-GAAP $865.9m — a $558m gap on $2.74bn of revenue — before the Google warrant (58.97m shares at $206.58, ~6.7% of the company) fully lands. And the custom-silicon margin cost is permanent: Q2 non-GAAP gross margin 58.9%, Q3 guided 57.5–58.5% on custom mix, Q4 held there, FY28 guided to the same range through a doubling of the custom business. ~90bp is the price of the ramp and it does not recover; all operating leverage now comes from opex growing at half the rate of revenue. Salesforce remains the cleanest single datapoint: Agentforce ARR above $1.5bn growing 240% against a $46bn base — ~3% — with organic growth of 6–7% ex-Informatica and FX.

Strongest counterargument. Record profits, accelerating guides and a rising frontier price are simply not what a monetization gap looks like. The gap may be a transition cost the file has mistaken for a structural condition.

Claim: The consumer is cracking at the bottom of the income distribution and holding above it. K-shaped, not broad — and the trigger is energy at the pump, not employment.

What changed since V9. V9 had this CONFIRMED but undifferentiated. The differentiation is now measurable in the same week's prints.

Evidence — the crack. Walmart missed comps for the first time in five years with traffic growth halved, and named retail gasoline at $4.03 as the mechanism. Local read: Costco — the cheapest gas available — has gone 4.99 → 5.20 → 5.40. Geography caveat: California runs well above the national figure; this is a directional escalation signal, not a national print.

Evidence — the hold. Ulta beat on revenue ($3.04bn, +8.9%) and EPS ($6.55, +13.3%), held operating margin flat at 12.5%, grew operating income 10.1% to $379.6m, kept inventories flat, repurchased $791m in H1, and raised full-year guidance. Discretionary beauty did not see what Walmart saw. Logged as a disconfirm on any broad consumer-crack read and as evidence for the K-shape. Gap: the comp number, which the release did not lead with — comps versus Space NK versus new stores is what would settle how much is organic.

Strongest counterargument. Beauty is famously defensive and Ulta is a share-gainer, so one beat proves nothing about the tier above Walmart. And gasoline is a transient input — a crude retreat resolves the Walmart mechanism without any structural consumer damage. Note the internal tension with Vein 5: this vein wants crude lower, that vein does not.

Claim: Positioning is transmission. It determines when and how a fundamental vein reaches price. It is never evidence that a fundamental vein is correct, and "the market did not move because gamma" must never become a universal defence.

The question, carried from V9 and still the file's central open one. Not "will the buyers crack" — they are, publicly, on the cash flow statement, at investment-grade names. Does the damage stay dispersed, or does the momentum coil release into a correlation spike that takes both legs of the expression down at once?

Evidence. V9's measurement: one-month realised vol ratio of Momentum over SPX at the 99th percentile on nearly every lookback, higher only during Covid, against correlation near multi-year lows. Everything since has tightened it.

• The dispersion falsifier fired where the price test cannot see. V9 recorded it as "did not fire." Axioma now has a statistical AI factor explaining more Russell 1000 variance than the market factor since mid-June, while COR3M sits near range lows. Options price record diversification; the risk model says it is illusory. Reframe, not a non-event.
• Jackson Hole was absorbed without a price response. Warsh hit every item on the published bearish scorecard. Rates repriced, FX repriced — the yen went through 160. Equities moved 0.1%.
• Dealer positioning is the reason and it is measurable. Gamma very long at this node and getting longer to the downside. VVIX closed at its lowest level year to date and third lowest since August 2024, now 86.91. VIX 14.5 with September a coin flip. Friday's straddle went out at 53bps (~$42) and realised a 60-point range with a 20-point net move — range delivered, direction did not.
• Flow is positioned for the pin to hold. 90k VIX Sep 16 puts bought at 0.55. ~2m vega of SPX hedge rolled from Dec 25d puts into Mar 25d puts. SPX Dec/Mar 7300 1.5x1 put spread traded 1k at 22.1 against 7742.

Rule 27 territory, sharper. Right on the map, exposed on the vehicle. Both legs of the spread are crowded. The index is not mispricing the news out of ignorance — it is absorbing it mechanically. That absorption is a supply of gamma, and supplies run out.

• tokenpriceindex.com — weekly. Track the headline blended number AND the floor-to-ceiling multiplier separately; they tell different stories. Current: $2.34, 62x, floor $0.352, ceiling $22.00.
• TYX:CGI−FVX:CGI and TYX:CGI−TNX:CGI — daily closes only. The two indices update asynchronously and the spread prints artifacts intraday. Current 71.9bp and 48.4bp. Want them wider.
• SOFR − IORB — NOT FIRED. Aug 28: IORB 3.65, EFFR ~3.627, SOFR ~3.638. SOFR crossed above IORB three times since late July (Jul 30–Aug 3, Aug 17–19, Aug 24–25) but by ~1–1.5bp. The test is 3–5bp for three consecutive sessions. Magnitude fails. Baseline has shifted though: SOFR ran 12bp below IORB in early July. Tripwire three may be closer than tripwire one — EFFR ~2bp under IORB, from ~2.7bp in early July, and the test is EFFR closing its 1–2bp gap. That is the one that preceded September 2019. Lapse caveat: during any appropriations lapse, SOFR above IORB is a TGA artifact. Check the balance first.
• Spot-to-contract memory spread — in the Constitution, not being logged. Start.
• ICE BofA CCC OAS — the transmission test for whether the hiking path reaches the marginal borrower.
• COR3M paired with the Axioma AI factor share.
• EIA distillate stocks and export volumes against the five-year range; the WTI/Brent backwardation crossover month (currently into April–May 2027).
• State data-centre moratoria count; turbine and on-prem versus IPP divergence.
• Retail gasoline, national and local.