2026-10-28 — EPFR: $12.4 billion into semiconductor funds in the week to October 21

• EPFR reported that $12.4 billion flowed into semiconductor funds in the week ending October 21, while flows into artificial intelligence and data center funds had not yet been dented — the same note says concerns over how AI spending is being financed, and whether it will pay off, had investors fretting, and that equity funds took in $55.5 billion overall that week. (EPFR) Claims: kat-claim-d3fcfca3f977

2026-09-23 — Chipmakers and AI hyperscalers fall together as yields hit a 19-year high

• Chip stocks and the biggest AI cloud spenders fell together on Wednesday 09-23 as Treasury yields hit their highest since 2007: Nvidia closed down 1.4% at $225.51 and Alphabet fell 3.8% to $337.83. The pairing is the opposite of what this thread’s rotation call predicts (money leaving chipmakers for hyperscalers): the larger drop was in a hyperscaler, and the move tracked the rate shock. (Barron’s, Trading Economics, biggo)

2026-09-22 — A single hyperscaler consumer product launch drives the sharpest chip rally of the run — CPUs, not GPUs

• Chip stocks surged Monday 09-21 on a hyperscaler’s own consumer-AI-product traction rather than any hyperscaler capex disclosure: AMD rose 9.95% to close at $615.52 (crossing $1 trillion in market value, the fourth US chipmaker to do so), Arm rose 17.2% and Intel rose 12.1%, after Meta’s new consumer AI agent “Muse” hit #1 on Apple’s US App Store and reached 448,000 daily active users just ten days after launch — a faster ramp than ChatGPT took to reach comparable download levels. Traders read the uptake as evidence that inference (running a deployed model for live user requests) needs more CPU capacity alongside GPUs, not just training-side GPU demand — a different mechanism than every prior entry on this thread, which has tracked capex guidance, earnings beats, and financing structures. Meta itself rose more than 11% the same session, and the Nasdaq closed at its first record high since June. This is Meta (a hyperscaler already inside this thread’s own “monetization-proven” bucket) pulling the chip complex up with it via product success, rather than the hyperscaler-vs-chipmaker reallocation the thread was set up to track. (Bloomberg, Semafor)

2026-09-21 — JPMorgan’s Dimon puts hyperscaler AI spending on course for $1 trillion in 2027

• JPMorgan chief executive Jamie Dimon said spending across the hyperscaler ecosystem has more than doubled from about $300 billion last year to around $700 billion this year and could reach $1 trillion next year. Speaking to CNBC-TV18 at JPMorgan’s India investor conference, he called the surge “like 1% increase to GDP each year” and warned it may also add to inflation. It is a direct counterweight to this thread’s premise that AI infrastructure capex is peaking, from the head of a lender to the sector. (CNBC, CNBC-TV18)

2026-09-08 — Qualcomm rises ~10% on an Amazon custom-silicon deal, but the “$4 billion” headline figure is a warrant, not a contract

• Qualcomm rose about 10% on a multi-generational custom-silicon agreement with Amazon for AI data-centre infrastructure. ⚠️ The “$4 billion” figure in most headlines is the value of a warrant Qualcomm issued Amazon for up to 25 million of Qualcomm’s own shares — not cash consideration, and not a disclosed contract size. Most coverage collapses that distinction; it matters for this thread because it means the print is a rotation signal (Amazon adding a second custom-silicon path alongside its in-house Trainium chips) rather than a disclosed revenue number for Qualcomm. (CNBC)
• Qualcomm’s roughly 10% premarket pop faded to +2.91% by 15:07 ET, at $173.65 against Friday’s $168.74 close. The morning entry already caught that the “$4bn” headline was a warrant rather than a contract — Amazon takes warrants on 25 million shares at $161.26 — and the afternoon tape is the market reaching the same conclusion within a session. (Yahoo Finance)

2026-09-04 — Chip-vs-tech split held through Friday’s close, did not reverse

• Friday’s chip-vs-tech decoupling survived the full session and shows up in the closing tape. The PHLX Semiconductor Index (SOX) closed up 3.38% at 11,735.26 and the iShares Semiconductor ETF (SOXX) held its gains into the bell, while the S&P 500 (-0.38% to 7,718.60), Nasdaq Composite (-0.29% to 26,506.99) and Dow (-0.51% to 53,414.25) all finished the day lower on the hot August jobs print that pushed CME September-hike odds toward 58-65%. Rate-sensitivity logic said high-multiple chip names should have been hit hardest by a hawkish repricing; at the close they were the only part of the tape that was up. ⚠️ Individual Intel, AMD and Nvidia closing prices are not stated here because no closing-price page could be opened; the morning’s intraday reads (Intel +4%, AMD +3%, Nvidia +2%) are the last figures on record. (TheStreet, Investrade market review)

2026-09-02 — Broadcom’s Q3 AI revenue nearly triples and beats, but the stock falls on a guidance miss under 1%

• Broadcom’s fiscal Q3 2026 AI semiconductor revenue hit $16.7bn, up 221% year-on-year and 54% sequentially, on total revenue of $29.6bn (+86% YoY) and adjusted EPS of $3.32 (+96% YoY) — a clean beat on every headline number. Q4 guidance called for $34.8bn total revenue and $21.7bn AI revenue (+236% YoY), which came in just under 1% below the $35.03bn consensus — enough to drop the stock roughly 3% in Thursday trading despite Wednesday’s beat, because “at this point the market expects exceptional” and the stock trades above a 60x P/E with little room for deceleration. This is exactly the dispersion this thread tracks: monetization-proven AI demand still getting punished the moment forward guidance falls short of an already-elevated bar, the same dynamic that hit Meta in July. A secondary risk flagged in the same coverage: Google’s recent custom-chip deal with Marvell signals large customers diversifying away from single-vendor dependence on Broadcom. (Broadcom Inc. investor release, Motley Fool, CNBC)

2026-08-31 — Nvidia recovers premarket while the broader tape falls on Iran and further Fed-hike repricing, a mild discriminator against Friday’s selloff being mainly about Nvidia’s own financing pullback

• Monday’s open put Nvidia on the opposite side of the broader market’s move from Friday. NVDA opened around $219 (up roughly 0.3-0.7% from Friday’s $217.55 close), while the S&P 500 (-0.45% to 7,677), Dow (-0.62% to 53,229) and Nasdaq (-0.43% to 26,290) all opened lower on a combination of the fresh Larak Island/Jordan/UAE exchange (see Red Sea Shock) and continued Fed-hike repricing. That is a live discriminator between this thread’s two competing explanations for Friday’s -4.74% Nvidia session: if the financing-program pullback were still the dominant driver, NVDA would be expected to keep lagging; instead it’s the outperformer against a macro-driven broad-market dip this morning. One premarket session is not conclusive, and Micron/Intel/Broadcom are described in same-day coverage as “choppy” rather than clean movers either way.

• The Fed side of the split kept drifting in the same direction already on the record, without a new catalyst. CME FedWatch’s September hike probability rose again from Friday’s 57.5% to roughly 60-62% this morning (CNBC’s 08-31 write-up on Warsh’s analyst-reaction roundup cites 60.4%; other 08-31 coverage frames it as “62%, up from ~40% a week prior” without naming the exact source tool) — continued momentum from Warsh’s keynote rather than a new weekend data point or Fed speaker. Kalshi’s own September-hike contract sits close to flat at 47% (vs. 48% Friday, up from ~30% pre-Warsh) — the two markets’ roughly 13-15 point gap persists, unresolved, not closing. (CNBC, Yahoo Finance live blog)

• The premarket split this morning’s run flagged held up through the first four hours of trading. As of ~15:04 ET, the PHLX Semiconductor Index (SOX) was essentially flat on the day (+0.06%, 11,476.7) and NVDA was modestly positive (+0.62% to $218.89) while the S&P 500 (-0.46%), Nasdaq (-0.42%) and Dow (-0.63%) all remained pinned near their lower open. A single premarket print could have been noise; four hours of the chip complex not re-joining the broader-market decline is a stronger version of the same discriminator against “Friday’s -4.74% Nvidia session was mainly about Nvidia’s own vendor-financing pullback” — if that were still the dominant story, NVDA/SOX would be expected to underperform, not hold flat-to-positive against a macro-driven (Iran/oil/Fed) broad-market dip. (Google Finance, Investing.com real-time quotes, ~15:03-15:05 ET 2026-08-31)

2026-08-29 — Friday’s close confirms the dispersion, and a second explanation for it lands the same day

Nvidia closed Friday 08-28 down 4.57% at $217.55, matching rather than reversing the -4.74% intraday level this thread logged that afternoon, while the S&P 500 (-0.25% to 7,711.76), Nasdaq (-0.52% to 26,402.42) and Dow (-0.02% to 53,559.99) all gave back earlier gains to close lower. This is the closing-level confirmation the prior entry explicitly flagged as missing — it was logged intraday, market still open, decimals not asserted. (Yahoo Finance live markets blog)

A second, competing explanation arrived the same day: Fed Chair Kevin Warsh’s Jackson Hole keynote read as hawkish on inflation, pushing the market’s implied odds of a September rate hike from 35% to 57% and lifting the 10-year yield 4bp to 4.72%. This complicates rather than confirms the prior entry’s “likeliest proximate cause” — that entry attributed the selloff to Nvidia’s reported financing-programme pullback; a same-day macro shock (rates, not one company’s antitrust exposure) is now an equally plausible driver, and the two aren’t disentangled by anything found here. (Yahoo Finance live markets blog)

2026-08-28 — Six points of dispersion between Nvidia and Microsoft in one session

The Philadelphia Semiconductor Index fell 3.45% while Microsoft rose 1.84% — Nvidia -4.74%, CoreWeave -3.86%, AMD -1.58%, Broadcom -1.23%, Oracle roughly flat — on a day yields rose across the curve. This is the dispersion the thread exists to measure, and the likeliest proximate cause is not the Fed but Nvidia’s reported pullback from its revenue-share financing programme. ⚠️ One session is not a rotation and this is logged as a single day’s tape, read directly from live quote feeds at 15:04 ET with the cash session still open — intraday levels, not closing prints.

• CME Group’s own FedWatch tool put September rate-hike odds at 57.5% on Friday, up from 35.4% the day before — a figure CNBC cites directly from the tool rather than through a paywalled or single-outlet intermediary. This resolves the ⚠️ hedge already sitting on this thread’s 08-29 entry, which had the same directional move (35%→57%) sourced only to an unreachable FXStreet article. A second, independent prediction market gives a materially different number for the same question: Kalshi traders priced the odds of a hike at roughly 30% before Warsh’s keynote (nearly 70% for a pause) and about 48% after — a smaller absolute move and a lower post-speech level than CME’s futures-implied figure. Both are real, both are dated to the same Friday, and they do not reconcile; reported as two measures of the same event rather than one adopted number. This bears on the thread’s own framing (a hawkish repricing was the “second, competing explanation” for Friday’s chip selloff, alongside Nvidia’s financing pullback) — the size of the rate move now has better sourcing but not full agreement across markets. (CNBC, citing CME FedWatch)

2026-08-27 — The rotation shows up as concentration instead: one stock up, the average stock down

• Nvidia’s beat lifted the indices while the market underneath fell. Nvidia closed +8.74% at $227.98, pulling the Nasdaq up ~1.5% and the S&P ~0.7% — but tech was the only advancing S&P sector, a slim majority of US stocks fell, and only 6 of the Dow’s 30 components rose while the index closed up ~0.2%. This sharpens rather than confirms the thread’s thesis: what showed up was not a sector reallocation from hyperscalers into chipmakers but a single-name earnings event carrying the index. On the other side of the same tape, Alphabet’s slide from its May peak reached roughly $700bn, about 15% off its 05-13 all-time-high close of $402.38, closing 08-27 at $340.65. The vendor selling the picks repriced up and one of the largest buyers repriced down, on the same day and the same set of AI facts — which is a cleaner statement of the rotation than any sector chart. (Bloomberg on Alphabet, Nvidia investor relations)

2026-08-27 — The beat carries into the cash session and lifts the complex

• Nvidia’s beat pulled the chip complex and the indices up in Thursday’s regular session — Nvidia roughly +7-8% intraday, Marvell +5.8% and Micron +4.5% on the premarket read-across, with the S&P 500 adding ~0.4-0.7% and the Nasdaq ~0.7%. ⚠️ Index-level figures rest on a single secondary aggregator — CNBC’s own 08-27 live page returned a 403 — so the decimals are not asserted; the direction is solid. Worth noting what set the level: the index going into a Fed chair’s first Jackson Hole keynote is being priced off one company’s guidance rather than any rate expectation. (CNBC)

2026-08-22 — The rotation thesis meets a 15% input-cost rise

• Nvidia is raising AI server prices more than 15% on early-2027 shipments, citing memory costs. This thread tracks the analyst call to rotate out of chipmakers into hyperscalers on a thesis that infrastructure capex is peaking. A cost-push increase cuts against that cleanly: it raises what hyperscalers must spend to hold plan while raising what the silicon chain collects, and the beneficiary is neither side of the rotation but the memory makers behind it. Nvidia’s Q2 earnings on 08-26 is the first test. (Bloomberg, Fortune)

2026-08-19 — Google pays a custom-chip partner in optionality instead of cash, and a memory maker answers the rotation debate with a buyback

• Google took a warrant on up to $12.2B of Marvell stock (~7%), vesting against future custom-chip revenue rather than paid upfront — up to ~59M shares at $206.58, ~1.4M vesting year one, the remainder across 240 tranches through fiscal 2033, one tranche per $500M of custom-chip revenue Marvell books from Google. Marvell jumped 11%+ premarket; Broadcom, Google’s incumbent custom-chip partner, fell 2-3%. A new financing-structure genre on this thread’s ladder: a demand guarantee priced in optionality rather than debt or equity. (Bloomberg)
• SK Hynix approved a $28.6B buyback and raised its shareholder-return pledge to >50% of 2025-2027 free cash flow (~$170B), answering investor pressure on memory makers to return more of AI-boom profits rather than plow all of it into capacity. Shares rose sharply premarket. (Bloomberg)

2026-08-17 — Micron through $1,000; an independent silicon effort ends

• Micron crossed $1,000 a share intraday, up about 5% and roughly 240% year-to-date. JPMorgan’s Jay Kwon and KeyBanc’s John Vinh both attributed the move to memory demand broadening from GPU to CPU and outstripping supply through 2027 — a demand-mix thesis rather than momentum, which retires this map’s morning caveat that no Monday-specific catalyst was identifiable. (Yahoo Finance)
• ⚠️ One circulating explanation is excluded as unverified: that the memory move was tied to a lifted Huawei export restriction. The primary write-ups on the $1,000 crossing make no mention of export controls and no source for the claim survived checking. Recorded as rejected, not omitted.
• Groq raised $350M at $3.5B, half its $6.9B September valuation, having abandoned its own LPU silicon to run Nvidia systems. Nvidia hired founder Jonathan Ross and the senior silicon team in a $20B licensing deal in December; the remaining company is now an Nvidia customer — 13 data centres, 6M+ developers, scaling from 54 MW to over 200 MW in 2027, with Nvidia a planned participant in the round. The rotation this thread tracks usually shows up as share shift between incumbents; here it shows up as an independent alternative ceasing to exist. (Bloomberg)
• Nvidia itself was flat on its largest-ever financial commitment: +0.21% at $225.64. ⚠️ A circulating ~4.5%-drop claim is misdated late-July reporting on the earlier $250B figure; not carried.

2026-08-08 — The scare that triggered the rotation debate starts unwinding

• South Korea’s SK Hynix-led selloff began easing as roughly $1.4B in leveraged trades unwound, a day after a brief pre-market “flash crash” (a 30%-limit-down print on an alternative exchange, lasting under an hour) had rattled the KOSPI. The rout traced to Wall Street spillover — SanDisk’s strong earnings but soft guidance sparked broader memory-sector jitters — layered onto South Korea’s own leveraged-ETF exposure to AI/semiconductor names. The easing doesn’t resolve this thread’s core question (is the rotation real or noise), but it does mark the specific volatility episode this week’s “peaking” debate was reacting to as a market-structure event (leverage unwind), not a fresh fundamentals signal. (CNBC)

2026-08-05 — A named fund manager pushes back on “peaking”: this is a reset, not a rotation

• Harvest ETFs co-CIO James Learmonth called last week’s chip-stock selloff (over $1 trillion in combined market cap lost, Nvidia, SK Hynix, Samsung and Micron the biggest losers) a healthy reset within an early-to-mid-cycle buildout, not evidence AI capex is peaking: “the expectation that spending is going to continue to increase is still very much there… the AI infrastructure build out is likely still relatively early, mid to early cycle” — while conceding “the likely growth of this sector will not be as steep as it was before.” Semiconductor-tracking ETFs remain up ~80% YTD and over 100% on a trailing 12-month basis despite the pullback. This is a fresh, named, on-the-record voice landing on the “capex not peaking” side of the split this thread already tracks, not a new data point from fund-flow data itself. (Wealth Professional)

2026-08-03 — Megacap chip names bounce, but SOXX breadth lags its own leaders

• The AI-megacap chips rallied while the chip index lagged — on Monday’s risk-on tape NVDA rose +2.93%, QCOM +2.68% and AMD +1.78%, but SOXX gained only +0.55% (ARM −0.26%), so July’s rotation-away stabilized at the megacap-leader level without breadth across the broader complex recovering. (Yahoo/Google Finance, stockanalysis.com, 08-03)

2026-08-01 — The chip complex fell 22% in July while Nvidia went sideways

• SOXX fell 22.1% in July, its worst month since December 2002, against +67.7% year-to-date. (CNBC)
• Nvidia was essentially flat for the month — roughly +0.33%, from a 06-30 close of $200.09 to a 07-31 close of $200.75, and +2.93% on 07-31 alone — a divergence from its own index worth reconciling rather than assuming NVDA fell with the group. ⚠ Single-source (stockanalysis.com); flag as needing a second check if load-bearing. (stockanalysis.com)
• Broadcom was roughly flat over a trailing 30-day window (−0.04%) despite a −3.1% single session on 07-24.
• ⚠ Clean July monthly percentages for AMD, TSMC, Micron, Arm and Qualcomm could not be obtained from a reliable source in this pass — no such figure should be treated as confirmed if it appears elsewhere.

2026-07-30 — Samsung answers half the watch question; the four prints show a bifurcated verdict, not a clean turn

• Samsung posted a blowout Q2 (chip-division profit ~250x YoY per multiple outlets, AI-driven memory/HBM shortage pricing) and its stock rose ~6-8% — the opposite reaction to Arm/Qualcomm’s post-earnings selloff two nights earlier, despite all three being “chipmakers.” Reinforces this thread’s monetization-proven-vs-unproven reframe rather than breaking it: Samsung’s profit is realized, Arm’s was bookings. ([Reuters])
• Amazon answered the other half: AWS grew 37% YoY, the fastest in 18 quarters — a clean beat against the ~31-33% consensus, completing this thread’s own watch question. Full detail on AWS Capex. The monetization-proven-vs-unproven line now holds across all five prints checked this week (Microsoft, Samsung, Amazon rewarded on visible revenue conversion; Meta, Arm, Qualcomm sold on bookings/legacy exposure without it).
• Restating the four-print verdict with today’s fuller detail: Microsoft (+8-9%, Azure $100B annualized, cloud growth fastest since 2022) was rewarded for capex with visible revenue conversion; Meta (-10%, FCF collapsed 91% to $784M despite +28% revenue, raised the LOW end of its capex guide) was punished for capex framed as a bet without near-term return; Qualcomm (EPS missed by $0.02 despite a revenue beat, guided to a faster Apple-modem decline) and Arm (beat-and-raise, fell anyway on a smartphone-royalty cut) both sold off on legacy/smartphone-exposed revenue. The dividing line is monetization-proven vs. monetization-unproven, confirmed again by Samsung’s opposite reaction.
• Macro crosscurrent stacked on the same tape: Fed Chair Kevin Warsh’s credibility was publicly questioned after Tuesday’s 9-3 split vote; the Iran conflict widened further overnight with oil back above $90 — both pulling on the same market as the earnings verdicts.

2026-07-29 — Opened: analysts call the rotation explicitly

• Split from CHIPS Equity Pivot — that thread is CHIPS Act grants-to-equity policy; this is the separate chipmaker-to-hyperscaler reallocation trade.
• Analysts are framing the selloff as a rotation from chipmakers to hyperscalers, on fear that AI infrastructure capex is peaking faster than expected — Morgan Stanley, UBS and Forrester; BofA dissents and calls it mid-innings. (Yahoo Finance)
• This is the first time the thesis has been stated as an explicit allocation call rather than inferred from price.
• The four prints answered, and split the thesis in two. “Capex is peaking” is undercut — both hyperscalers that disclosed guidance moved it up: Meta’s floor rose to $130-145B (its second raise this year), Microsoft guided FY27 to “grow year-over-year” plus $50B+/quarter. Neither showed a peak.
• “Rotation to hyperscalers” is real but was the wrong axis. Arm and Qualcomm both fell on genuinely strong AI-specific numbers — Arm’s AGI-CPU bookings doubled to $2B+, Qualcomm’s FY2029 data-center target nearly doubled to $40B — capital left them regardless of the fundamentals. But Meta, a hyperscaler, was hit as hard as the chip names, on an EPS miss and near-zero free cash flow. The line the market actually drew tonight was monetization-proven vs. monetization-unproven spend, not hyperscaler vs. chipmaker: Microsoft bought tolerance for its capex with 43% Azure growth; Meta did not.

← Backstory