Research
Long-form work on the things actually moving capital. Every number traces to a primary source — and where the reporting stops and the estimating starts, we draw the line rather than blur it.
SPCX: Everyone Is Telling the Float Story. The Filings Say It Is Wrong.
The SpaceX float-scarcity story is wrong — the 19% pop was ordinary underpricing against a 4x book, every supply mechanic the scarcity thesis depends on has already failed in public, and the real price discovery is not the lock-up cascade but the denominator: a ~95x sales multiple built on revenue that exists only because xAI and X were retrospectively combined into every prior period.
Read the dossier
The AI Financing Complex
5 linked dossiersFive dossiers built on one shared source pool — so every one of them prices the same fact off the same number.
The Multiple Is the Risk: A $965 Billion Price Struck on a 28-Day Denominator Against $225 Billion of Fixed Compute
The brief has it backwards: at $965B on a $47B run-rate (~20.5x), even a physically impossible 100% gross margin still leaves the equity at 20.5x gross profit — so only growth, not margin, can rescue the price; customer concentration is a stale, anonymously-sourced mid-2025 figure ($1.2B of a ~$4B base) measured against a denominator that has since grown roughly 12x; gross margin is not a known risk but an unknown accounting choice (the SpaceX + AWS contracts alone imply a ~47% company-level gross-margin ceiling if compute is COGS, and a 70% margin is reachable only if training compute sits below the line as R&D); and the exposures the brief never names are the ones the primary documents actually show — more than $225B of fixed, multi-year, partly vendor-financed compute obligations against a sworn lifetime revenue of "exceeding $5 billion," a headline metric built by multiplying a 28-day window by 13 under a formula Anthropic has never published, and one product (Claude Code, ~18% of run-rate) in the one category where the largest named customers can build their own models.
OpenAI Isn't Short $1.4 Trillion — It's Long a Free Option. The Duration Mismatch Is on Oracle's, Broadcom's and CoreWeave's Books.
The duration mismatch is real but it is not OpenAI's: the ledger shows OpenAI's compute "obligations" are largely options it is short — only $18.4bn (~1.3% of the $1.4tn headline) is quantified by name in any SEC filing, NVIDIA's $100bn was "never a commitment" and appears nowhere in its 10-K, Broadcom is a term sheet with no dollar figure, one of AMD's six gigawatts is binding, Stargate's $500bn is intent on SoftBank's balance sheet, and management re-anchored the plan from $1.4tn to "roughly $600 billion" and trimmed Stargate sites with no counterparty penalty reported — while OpenAI sits on $73bn of permanent-equity cash with no disclosed balance-sheet debt and is lengthening its revenue duration via Guaranteed Capacity; the fixed, dated, legally enforceable leg sits with the suppliers and their lenders (Oracle's $638bn RPO, 88% of it beyond twelve months, backed by a bankruptcy risk factor about a customer it will not name; Broadcom's own $128,110m of non-cancellable purchase commitments against zero disclosed OpenAI backlog; CoreWeave's $98.8bn RPO with OpenAI named inside its credit-risk factor), which is exactly where the Bank of England and the BIS say the maturity mismatch is being warehoused and mispriced.
SPCX: Everyone Is Telling the Float Story. The Filings Say It Is Wrong.
The SpaceX float-scarcity story is wrong — the 19% pop was ordinary underpricing against a 4x book, every supply mechanic the scarcity thesis depends on has already failed in public, and the real price discovery is not the lock-up cascade but the denominator: a ~95x sales multiple built on revenue that exists only because xAI and X were retrospectively combined into every prior period.
The Depreciation Wall Is Real — But It's Sitting in Construction in Progress, Not in the Footnote
The hyperscaler depreciation shock is already locked in at TODAY'S useful-life assumptions — Alphabet alone holds $108,597M of "assets not yet in service" carrying zero depreciation (up from $78,592M in one quarter), and roughly $281bn sits not-yet-depreciating across Alphabet, Meta, Amazon and Oracle — while the useful-life reversion the bears await was already executed by Amazon in 2025 and cost exactly $0.10 per diluted share, about 1.4% of EPS, in a year its EPS rose sharply.
The Chips Are Not the Collateral, the Tenant Is: GPU Debt Is Term-Matched, Amortising and Domiciled in Delaware — and the Four-Year Number Is a Meta Lease, Not a Chip Life
The filings invert the premise on two of its three legs — every GPU-secured instrument on the record is deliberately SHORT and amortising against its asset (CoreWeave depreciates technology equipment over 6 years, DDTL 4.0 matures March 2032, DDTL 5.0 runs approximately 5.5 years in quarterly instalments, and only $4,338m of $21,615m of principal falls due after 2030), and not one financing vehicle is offshore ("Cayman" appears zero times in CoreWeave's 10-K, Meta's 10-K and Nebius's 424B5) — leaving the one leg that holds decisively as the story: the collateral does no rating work, because the same sponsor, the same GPU collateral and the same arrangers produce A3/A (low) paper at SOFR+2.25% when the customer is investment grade and Ba2/BB+ paper at SOFR+4.50% when it is not, a 225bp gap of pure counterparty credit, while the only genuinely long bond in the chain — Beignet Investor LLC's $27.3bn 6.581% notes due 2049-05-30, against Meta's "initial four-year lease term" and a residual value guarantee with a threshold of approximately $28bn — is a real-estate deal in which the four years is how long Meta has agreed to rent a building, not how long a GPU lasts.
Across the desks
5 reportsWhere the section does not de-rate to zero if AI does — rates, semiconductors, power, healthcare and the dollar going on-chain.
Two Markets, One Fed: Bonds Have Fully Priced Warsh's Hike; Equities Say It Won't Happen
The market has indeed swung from pricing cuts to pricing a hike under Warsh -- 76.06% on the Atlanta Fed's September SOFR window, only a 19.6% chance rates are still at 3.50-3.75% by December 9 -- but this is not an unpriced discount-rate shock: it is fully in the strip, already paid for by a 2-year at 4.26% (51bp above the top of the range), and owned by leveraged funds net short 2,872,406 SOFR contracts; the only people not positioned are equity investors, 83% of whom expect no hike before the November midterms with cash at 3.6%, which makes the dislocation one between two markets rather than between the market and the Fed -- and dangerous in both directions, because every hawkish dot was cast before a June CPI that printed core at -0.02% m/m and 2.57% y/y.
The Cheap Price Is the Volume: Lilly Grew 56% While Cutting Cash Prices, Novo Fell 4% — Same Channel, Opposite Outcomes, and Mix Explains Neither
The cash-pay price cut is not a mix tax that volume fails to offset — it IS the volume engine: Lilly, with self-pay already at 45% of total and 55% of new Zepbound scripts, grew revenue 56% on +65% volume against -13% price, held gross margin at 81.9% (-0.6pp) and RAISED FY2026 guidance by $2bn to $82.0-85.0bn, while Novo — which set its cash price ($349) BELOW its own ~$460-525 implied insured net rather than at parity like Lilly's $449 vs a $450-500 commercial net — fell -4% CER and is losing the units themselves (Ozempic -8% CER, US GLP-1 diabetes -16% CER on "lower volumes"); so the same channel produced a ~35-point divergence in guided growth, which means the operative variable is SHARE, not mix, and the only untested cliff is the Medicare Part D bridge that went live on 1 July 2026 at $245/month with a $50 copay and has touched zero reported quarters.
Nobody Holds the Licence Yet: The Stablecoin Market Stalled at $320bn While Washington Missed Its Own Deadline — and the Wrappers Are Onshoring to the OCC, Not to Cayman
Regulated stablecoin issuance has not started and the float has stalled — as of the as-of date every GENIUS implementing document (OCC Docket OCC-2025-0372; the FinCEN/OFAC joint proposal) is still a proposal, so no 120-day trigger has run, the 18-month backstop of 18 January 2027 controls, and there is not one permitted payment stablecoin issuer to name; two Fed publications a month apart put the market flat at $317bn (6 April 2026) and $320bn (May 2026) with growth "moderated," and Tether calls its own $183bn float "broadly stable" — while the legal wrappers are moving ONSHORE into the OCC's perimeter (Circle's final OCC approval for First National Digital Currency Bank, N.A. on 10 July 2026; Bridge/Stripe's preliminary charter on 12 February 2026; Tether issuing from El Salvador, not George Town), not to Cayman, which GENIUS Sec. 2(23) categorically bars from issuance ("a person formed in the United States"), which has no stablecoin regime or reserve rules to be found "comparable" under Sec. 18, whose VASP register SHRANK to 18 entities in Q4 2025 from 20 in Q1 2025 with exactly 1 registrant for "issuance of virtual assets," and whose only genuinely new digital-asset franchise — tokenised funds — is 12 provisionally registered funds out of 31,145 (0.04%), anchored by a $312,976,695 Circle-linked mutual fund whose token is issued out of Bermuda and sold only to non-US persons.
HBM Taxes Commodity DRAM — It Doesn't Destroy It, and the Wafers Are Already Coming Back
HBM's wafer penalty is real and reproducible — ~2.85x for 2026, derived independently from TrendForce's own 22%-of-wafers / 9%-of-bits shares and landing on Micron's "three-to-one" — but it taxes commodity DRAM rather than destroying it, subtracting roughly 4pp of bit-supply growth in 2026 and ~6.5pp in 2027 from an industry that Micron still guides to grow bits in the "low to mid-20s percentage range," so with quantity up ~20% and ASPs up in the "low-260% range" the price shock is demand outrunning a steepened, cleanroom-constrained supply curve — not a leftward supply shift — and the cannibalisation is already reversing, with Samsung freeing "around 80,000 DRAM wafers monthly" back to DDR5 RDIMM after HBM per-wafer profitability fell below DDR5 64GB RDIMM in 1Q26.
Being the Bottleneck Is Not the Same as Getting Paid for It: The Grid Queue Is the Wrong Metric, the Equipment Shortage Is Real, and the Rent Is Being Collected 60 Points Up the Stack
The AI buildout's durable physical constraint is neither the interconnection queue — a generator queue that shrank in 2025 (750 GW withdrew vs ~600 GW filed), is >70% speculative, and that hyperscalers are not even in, since Microsoft stood up ~2 GW in 2025 while spending $11.1bn on leases — nor the logic die (NVIDIA holds $119.0bn of supply commitments and says it has "secured inventory and capacity beyond the next several quarters"; the CoWoS gap narrows from ~20% to ~10% by end-2026), but the long-lead equipment stack you must buy AFTER you win your slot (four-year transformers, a turbine ordered today first firing in 2031, 100.3 GW of 2025 turbine orders against 60-70 GW of world capacity, Prysmian's HV backlog at ~5.6x segment revenue, Siemens Gas Services printing a 2.55 book-to-bill and taking customer prepayments to hold a slot) — and yet the trade does NOT follow from the constraint, because the bottleneck cannot charge for being the bottleneck (GE Vernova 9.6% adjusted EBITDA margin, Siemens Energy guiding 10-12%, Quanta ~8.7%, against NVIDIA's 74.9% gross margin and TSMC's 66.2%), 56% of GE Vernova's headline 100 GW pipeline is non-binding slot reservations converting at only 6 GW a quarter against demand that Dominion's own book discounts to ~21% (10.4 GW of signed ESAs on a 50+ GW pipeline), and the capacity that ends the scarcity rent is already funded and dated (GE Vernova 50 → 70-80 large frames by late 2026, Mitsubishi doubling through 2028, with Wood Mackenzie calling turbine orders to peak in 2026).
How we build these
The line between what is known and what is estimated is drawn, not blurred.
Every numeral in a dossier resolves to a footnoted primary source, a declared derivation, or a labelled estimate — nothing rests on an aggregator, and a chart changes texture at the exact point a company’s own disclosure stops and someone else’s estimate begins. Where sources conflict, we name both figures and show our adjudication rather than quietly pick one. That boundary is the product.