Public Markets Digest
TL;DR
The AI-slowdown question is answered — nobody slowed down — and the debate has moved to the financing side: the big four are on pace for ~$725B of 2026 capex while Morgan Stanley counts $1.75T of external debt needed through 2028. Meanwhile the Fed's second hike is 66% priced for October, the 10-year sits at 5.18%, and the Dow tore to a record anyway. The chip selloff round-tripped, SpaceX absorbed its unlock without a flinch, and Intel is the market's highest-beta bet on AI demand being broader than GPUs.
The bear case migrated from price to credit — and the credit tail is already cracking
The AI complex fully recovered from the Amodei-driven selloff and then rallied through a Fed hike, but the bear argument found a new home in the bond market.
"The most profitable companies in the world, except Apple, are betting everything on this as their free cash flow turns negative and the borrowing pace increases." — Burry's Sept 25 Post
+2 (from [AOL/FinanceBuzz](https://www.aol.ca/articles/michael-burry-sees-loom ing-write-093000000.html))
The divergence is the tell: investment-grade spreads sit under 1% while CCC-and-below high-yield spreads have breached 10%, approaching their March 2025 peak — the fringe of AI financing is being repriced while the majors rally (BigGo Finance). Price action says no bubble yet: NVDA closed Friday at $225.07, just 4.3% off its high, AMD crossed $1 trillion in market value on September 21, and the SOXX correction trigger has moved further away (Semiconductor Bubble Debate). Burry's answer evolved from a short into an accounting audit — net capital investment at 2.07% of GDP, the highest in four decades outside the 2000 peak, with the real exposure off-balance-sheet: Alphabet ~$900B in commitments, Meta ~$700B, Microsoft's uncommenced leases tripled past $300B (Burry's Sept 25 Post
+2). On Microsoft stretching data-center useful lives from 15 to 25 years, his verdict was three words: "Too cute by half."
What to watch: CCC spreads versus IG spreads — if the tail repricing crawls up the credit stack into the majors, the bubble debate converges with earnings season.
The capex slowdown call is officially noise; the fight is now over the return on $1.75T of debt
Two weeks after Amodei's slowdown call, the evidence is unambiguous: spending went up, not down.
"Higher component pricing this year." — Meta CFO Susan Li
The big four are tracking roughly $725B of combined 2026 capex, up ~77% from ~$410B in 2025, and Meta raised guidance citing component costs — memory inflation is pushing capex up, not demand weakness down (Capex Tracker+2). Demand barometers agree: Anthropic's run-rate hit $65B, surpassing OpenAI's $40B, and H100 rentals rose 22% in a month to $3.28/GPU-hour. The new bear case is Goldman's return math: earning a 15% ROIC on Phase-2 capex requires ~$1.42T of cumulative 2028–30 revenue, and if real GPU life is 3 years instead of 5, the mature-phase revenue requirement jumps to ~$6T a year. S&P projects five of six hyperscalers post negative free cash flow in 2027 — Microsoft the exception — while Oracle, the cautionary tale, was downgraded to BBB-, one notch above junk (Capex Tracker
+2).
What to watch: Micron's September 30 earnings — memory pricing is the load-bearing input for both the component-cost inflation story and the Intel/SK Hynix Ohio thesis.
The Fed is being handed its second hike, and Washington might take away its data
The discount-rate headwind is no longer hypothetical: the market is doing the Fed's work for it.
"Markets are pricing in a 66% chance of a rate hike in October, according to CME Group's FedWatch Tool." — Fed Week II
(from FXStreet)
The 10-year sits at 5.18% — above the 5% line and at a one-year high — yet the Dow tore to a fresh record near 51,828.62 on AI megacap strength anyway (Fed Week II). The week's fulcrum is political: Congress faced a midnight Tuesday deadline to fund the government, and failure would delay Friday's non-farm payrolls — taking away the very labor data the Fed is using to justify its path. Oil complicates everything: WTI rebounded 4.9% to $92.45 after Houthi strikes on Riyadh and stalled US–Iran talks, "a direct input into US inflation expectations."
What to watch: Wednesday's core PCE print (consensus 3.4% y/y) — a hot number effectively locks in the October hike two weeks ahead of the meeting.
Intel is the highest-beta expression of "AI demand is broader than GPUs" — and the price tag is the problem
Intel climbed 14.3% last week and is up over 250% in the past year, fifth-best in the S&P 500, at a ~$650B market cap.
"If a major memory company like SK hynix fills part of that space, it would materially help both utilization and visibility on investment returns." — Intel Comeback
Three drivers are stacking, and they're not equally solid. The SK Hynix Ohio talks (stock +12.14% on the Reuters report) were walked back by both sides — "no specific plans or arrangements have been finalized." The agentic-AI demand story is an inference, not an order book: Meta's Muse agent hit 3.4M downloads and grew 55% day-over-day in its first two weeks, and the market read rising agent usage as more Intel data-center chips. The cleanest number is the mark-to-market: SK Hynix's Solidigm — the NAND business Intel sold for ~$9B in 2020 — is weighing an IPO at up to $150B, a 16x return on Intel's discarded business (Intel Comeback+1). Bears note FY2026 consensus revenue is up only ~18% against a +250% stock.
What to watch: Whether the SK Hynix deal gets signed or stays speculation — the foundry-utilization bull case needs a customer, not a rumor.
SpaceX absorbed the unlock; now the catalyst is execution, not share flow
The supply event the market spent two weeks pricing arrived and was a damp squib.
"The previous tranche unlocked on September 9, and the stock fell 3.9% that day." — SpaceX Post-IPO Reality
+1 (from Yahoo/beincrypto)
SPCX closed Friday at $148.68, down just 2.6% over the five days spanning the September 24 unlock of ~328.4M shares, on 0.6x its 20-day average volume — no forced-selling signature (SpaceX Post-IPO Reality+1). The marginal seller so far is exercising insiders, not unlock recipients: Gwynne Shotwell sold ~$52M of stock on September 22. The real catalyst is Starship Flight 14 — the first planned orbital trip (~275 km, ~10 hours) and the vehicle's first revenue-generating mission, carrying 26 Starlink V3 satellites. Morgan Stanley's Adam Jonas holds a Buy with a $300 target (~99% upside) on the orbital-computing option. The valuation still demands the bull case: 85x trailing sales, -$15.95B quarterly free cash flow, and a $1.96T market cap priced for execution.
What to watch: Flight 14's outcome — clean orbital insertion plus satellite deployment is the first commercial proof; failure hands the bears a fresh data point.
Qualcomm is now a momentum position with a binary legal event six days out
QCOM ran to $201.97, +23.2% on the month, above the $194.13 analyst consensus target with RSI at 73 — overbought.
"Qualcomm acquired robotics-software firm PickNik 'to strengthen its position in physical AI and robotics technologies.'" — Qualcomm Runs to $202
(from Yahoo Finance)
The stock has left Amazon's $161.26 warrant strike 25% in the rearview — the customer-alignment narrative from the AWS deal is winning, and Apple renewed its patent license. But the bear math hasn't moved: TTM revenue is still down 4.0% YoY with earnings down 23%. The market is paying for the data-center option, not the current P&L, and the Qualcomm v. Arm Delaware trial begins October 5 — an adverse outcome reprices the entire licensing segment that funds the diversification (Qualcomm Runs to $202).
What to watch: The Arm trial verdict — the asymmetry has flipped from the September 8 setup; the news is in the price, the trial isn't.
What surprised us
- The slowdown-call thread resolved decisively — in the wrong direction for bears. Two weeks after Amodei's call, capex guidance went up, Meta explicitly cited component-price inflation, and H100 rentals rose 22% in a month. The question moved from "will they spend" to "what return on borrowed money" (Capex Tracker
- The unlock-overhang thesis is 0-for-2. Both SpaceX tranches produced sell-the-news dips of under 4% followed by stabilization — mechanical share flow is not setting this stock's path; execution milestones are (SpaceX Post-IPO Reality
- Intel's best comeback evidence is a business it sold. Solidigm at up to $150B is a 16x return on the NAND unit Intel divested for ~$9B — the capital-cycle irony Burry-style critics live for, sitting inside the market's favorite comeback story (Intel Comeback
- The Dow made a record with the 10-year at 5.18%. The "AI cash flows vs. discount rates" tension resolved, for now, entirely in favor of cash flows.