Liquiditrax
Take the quiz
Liquiditrax Command & Research

AI Capex

Theme
Live note
Note updated

Our current view

State (current)

The theme has moved off the income statement entirely. On 14 August Broadcom fell about 6% on a day it reported nothing, because Bank of America put a number on how its AI buildout is financed: analyst Tom Curcuruto estimated Broadcom's chip-financing vehicle could reach $370B of senior debt by mid-2029 at 20-gigawatt scale, including roughly $150B of new issuance in 2027 alone. None of that is Broadcom's balance-sheet debt; per the latest 10-Q an investor partner assumed the purchase and lease agreements and Broadcom backstops lease payments for five years with maximum exposure up to $29B on the initial transaction. The demand side was never in question, fiscal Q2 revenue was $22.19B (+47.9% y/y) with AI chip revenue +143% and $16B of AI semiconductor revenue guided for the current quarter. The market sold the structure, not the business. The control experiment ran the same day: NVIDIA has a structurally similar roughly $500B financing framework with Goldman Sachs, Blackstone and Apollo, and it did not sell off, because no specific debt figure had been attached and Huang has said NVIDIA would cover as much as 25% of an opportunity with no deals signed at announcement. Same structure, different disclosure, opposite reaction. Alongside it, AMD rose about 4% to 5% on Baird doubling its target to a Street-high $1,250, modelling $147B of AI GPU platform revenue by 2030 on 15% TAM share, so on one screen the market repriced AI leverage down and AI addressable market up. SOXX moved only -0.7% with Intel -2% and AMAT continuing its post-earnings slide -5.6% to $504.67, confirming this was single-name balance-sheet repricing rather than a de-rating of the trade. Read against the rest of the week (SMCI on guidance, CoreWeave on a narrowing loss, Nebius on pricing power, AMAT on cash conversion), the sequence is now explicit: demand is settled, cash conversion is contested, and financing structure is the newest and least-disclosed axis. A Goldman note cited Friday says hyperscalers are using lease commitments and other debt structures to fund expansion, which means the off-balance-sheet layer is a theme-wide feature, not a Broadcom quirk. Tracking that layer now has its own entity, ai-credit.

Timeline

Our calls (journaled)

  • 2026-08-05: AI capex is now a credit story, watch IG/HY spreads + Oracle CDS. Invalidation: spreads tighten back. Outcome: open

  • 2026-08-07: Screen AI names by funding source (P&L vs debt vs equity) and depreciation assumption. Outcome: open

  • 2026-08-08: The buildout is scaling faster than the funding is cheapening (SpaceX $15.8B capex, Switch $50B). Watch capex-to-cash-flow and who has to visit the capital markets. Outcome: open

  • 2026-08-10: Treat this week's optics + WFE prints (LITE, AMAT) and SMCI's backlog as primary evidence on capex durability, not the hyperscaler narrative. A revenue beat that holds guidance confirms the physical-layer demand; a beat that cuts forward guidance is the tell to watch. Outcome: open

  • 2026-08-11: SMCI vs CoreWeave today is the funding split made testable in one day: both roughly double revenue, but one earns and one loses. For an allocator the read-through is not "is AI demand real" (it is) but "who funds it" - underwrite the self-funders on margin durability and the debt/equity-funded names on cost of capital and dilution, not on revenue growth alone. Watch whether CoreWeave's loss narrows or widens per dollar of revenue. Outcome: resolved favorably 08-12 - CoreWeave's adj loss narrowed (-$1.03 vs -$1.20 exp) and SMCI carried on guidance despite a revenue miss; the "who funds it, and is the renter's economics improving" read was the right lens, both rewarded for different reasons.

  • 2026-08-12: The tell that mattered was not the revenue line on either name. SMCI missed revenue and rose on demand guidance; CoreWeave's revenue beat was secondary to the loss narrowing. Next test: whether CoreWeave's path to a smaller loss per dollar holds across quarters (financing risk) and whether SMCI can convert a $65-72B guide into cash without visiting capital markets. Underwrite self-funders on margin, financed names on the slope of the loss, not on revenue growth. Outcome: open

  • 2026-08-14: The funding-source lens now needs a second axis: cash conversion. AMAT is the cleanest illustration yet, because it is nobody's credit story. It is profitable, self-funding, selling into a WFE market its own CEO guides above 30% growth, and it still sold off on a record quarter, because the market has learned to check whether the earnings show up as cash. Five EPS beats against four free-cash-flow misses is a pattern, not a quarter. Read-through for the theme: through this week every layer of the stack proved demand, so demand has stopped being the differentiator and price-for-quality has taken over. The allocator question shifts from "is the buildout real" (settled) to "which layer converts revenue into free cash flow, and what am I paying for that conversion". Watch AMAT's actual cash-flow line versus the EPS line next quarter, and treat any AI name whose beat lives only in adjusted EPS as unproven. Outcome: open (08-15 supportive: AMAT extended its post-earnings decline -5.59%, so the cash-conversion penalty persisted a second session rather than being a one-day reaction)

  • 2026-08-15: The 08-05 call that AI capex is a credit story is now VALIDATED, and by a route we did not specify. We said to watch IG/HY spreads and Oracle CDS, the visible credit market. The repricing arrived instead through an equity, on an analyst estimate of an OFF-balance-sheet vehicle, which means the leverage in this theme is accumulating where spreads and CDS will not show it. Updated framing for the allocator: screen AI names on three axes now, not one. Funding source (P&L, debt, equity), cash conversion (does the EPS beat show up as free cash flow), and financing disclosure (is the buildout financed on the balance sheet you can read, or in a vehicle you cannot). Broadcom and NVIDIA carry structurally similar arrangements and only one has a public number attached, which tells you the risk is disclosure-driven and will arrive as a step function each time an analyst prices it, not as a gradual drift. Practical read: treat any AI name whose growth depends on customer financing it arranges as carrying an unpriced contingent liability until it is quantified. Invalidation: hyperscalers and merchant vendors funding the next leg out of operating cash flow with these vehicles shrinking rather than scaling. Watch whether Broadcom's disclosed $29B backstop grows as the platform scales toward 20GW, and whether NVIDIA's ~$500B framework converts into signed contracts. Outcome: open

Hit-rate

  • 1 hit / 0 miss (1 closed) plus 1 resolved favorably. 08-05 "AI capex is a credit story" VALIDATED 08-15 on thesis, with a process miss on the indicator: we named IG/HY spreads and Oracle CDS, the repricing came through an equity on an off-balance-sheet estimate. Lesson: pick indicators that can see the leverage, not just the ones that are easy to quote. 08-11 funding-split lens resolved favorably 08-12. Open: 08-07 funding-source screen, 08-08 capex-vs-funding, 08-10 physical-layer evidence, 08-12 slope-of-the-loss, 08-14 cash conversion, 08-15 three-axis screen.
LinkedMSFTAMZNGOOGMETAORCLPLTRCLSSPCXCOHRLITESMCICRWVNBISAMATWDAYAVGOAMDNVDAai-creditregime

Recent research on AI Capex

No edition has singled out AI Capex yet. See the latest cross-asset read on Research.

Live figures are delayed and for analytics/education only — not investment advice, not a signal service, no buy or sell recommendation. Journaled calls include their outcomes, wins and misses. Every decision and risk is your own. DYOR.