Record Quarters, Red Screens — Broadcom's AI Revenue Doubled, the Stock Fell 13%, and the Sector Held While the Crowded Name Bled

The Lead

Issue 10 promised this issue would carry the Broadcom verdict. Here it is, and it is the most instructive print of the year — not because Broadcom missed, but because it didn’t, and the stock fell anyway.

Broadcom reported Wednesday after the close. Revenue was $22.19 billion, up 48% year-on-year. The number that Forecast [2026-05-31-001] hung on — AI semiconductor revenue — came in at $10.8 billion, up 143% year-on-year, above the company’s own guide, driven by exactly what we said it would be: custom AI accelerators (the hyperscaler ASICs) and AI networking. Hock Tan reiterated “in excess of $100 billion” of AI revenue in fiscal 2027. Non-GAAP EPS of $2.44 beat the $2.40 consensus. Free cash flow was $10.26 billion, up 60%. By any standard this was a record quarter.

The stock fell roughly 13%.

Three things did it, and they are the whole story of the week. First, the Q3 AI guide — $16.0 billion, more than 200% year-on-year growth — landed below the roughly $17.2 billion the Street had quietly penciled in. When the bar is “more than triple,” merely doubling-and-a-half is a miss. Second, Tan acknowledged on the call that surging AI revenue is diluting gross margins — the AI silicon mix carries lower margin than the software business, so the faster AI grows, the more it drags the blended number. Third, and most pointed, he conceded that Google would “draw on multiple chip suppliers” — an admission that the custom-silicon moat has a side door, and that Marvell is already through it.

So Forecast [2026-05-31-001] confirms: AI/custom-silicon revenue grew 143% — not just past the >50% bar but more than double it — and it was the fastest of the fortnight’s ASIC reporters (Marvell’s data centre line grew 27%; Broadcom’s AI line grew 143%). The call on the number was right. The call anyone made on the stock — if they bought the consensus expecting a great print to pay — was wrong by 13% in a session. A forecast being right is not the same as a trade being right. Hold that distinction; the rest of the issue is built on it.

Threads

The Broadcom verdict — +143% AI revenue, −13% stock, and why both are true

The financials, from the Q2 FY26 release (quarter ended 3 May), pending the 10-Q:

The bull case is intact and, if anything, strengthened: AI revenue is compounding faster than the company itself forecast, the $100bn-2027 target was reaffirmed, and Tan said Broadcom is “very comfortable” it has secured supply through 2027 and will ship 10 gigawatts of compute next year. None of that broke.

What broke was the expectation. The stock entered the print up 38% year-to-date, having re-rated through Issue 08’s NVDA quarter, Issue 09’s Marvell beat and Dell blow-out, and Issue 10’s optical melt-up. By Wednesday, “great” was the floor, not the ceiling. The Q3 AI guide missing the whisper number, plus the twin admissions on margin and on Google multi-sourcing, was enough — in Barron’s framing, a record quarter that just wasn’t good enough.

This is the precise configuration Issue 09’s overcrowding thread warned about: “when the fundamentals are this good and the stock reactions are this muted [or negative], you are being paid less and less to hold the consensus.” Broadcom was the most-consensus AI long going into the week — seven-of-eleven-fund-NVDA’s natural partner, the cleanest custom-silicon proxy, the name everyone owned. Maximum conviction has no marginal buyer left to convert, and any blemish produces an outsized reaction. It did.

The sector held while the crowded name bled — SOXX +6%, AVGO −13%, MRVL +54%

Here is the tell that matters most for how we actually trade this, and it is the cleanest real-world vindication of the discipline we’ve spent the past fortnight building.

Across the week (close 28 May → close 4 June), as Broadcom fell:

The semiconductor sector rose almost 6% on the week. The single most crowded name in it fell 13% on the day. The gap was bridged by Marvell, which ran 54% — partly the Huang “next trillion-dollar company” bump from Issue 10, and partly a direct transfer: the very admission that sank Broadcom (Google multi-sourcing) is Marvell’s thesis printing, because Marvell is the second supplier. The hyperscaler-ASIC pie didn’t shrink; a slice moved across the table.

The lesson writes itself, and it is the one the signal-lab work has been screaming: owning the sector, sized to survive, beat owning the crowded name. If you held AVGO into the print on conviction, you were down 13% in a session. If you held the sector, you were up 6% on the week — the index absorbed Broadcom’s air pocket because the demand is real and it simply re-routed to the supplier gaining share. This is the entire argument for trading the theme over the signal-by-signal call, and it played out in four trading days.

Beat-and-fall is now the regime — Ciena +40% and −9%, Palo Alto’s FCF held and −5.6%

Broadcom was not alone; it was the loudest instance of a pattern that repeated all week.

Ciena (the optical-transport read-through Issue 10 flagged) reported Thursday pre-market: revenue $1.57bn, +40% YoY, adjusted EPS $1.64 versus $1.40 expected — up 290% — adjusted gross margin expanding to 44.9%, and full-year guidance raised to $6.3bn. A near-flawless print. The stock fell ~9%. The reason was identical to Broadcom’s: the shares had run ~30% into the quarter, the guide-raise wasn’t large enough to clear the bar the rally had set, and AVGO’s plunge dragged the whole optical complex down with it. Beat, raise, fall.

Palo Alto Networks answered Issue 10’s open question — does the free-cash-flow profitability filter that took Zscaler down 19% also bite the cybersecurity leader? The answer is no, on the fundamentals: PANW held an adjusted FCF margin of 38.5% trailing-twelve-month (up 430 basis points year-on-year, including the CyberArk and Chronosphere drag), generated $910m of free cash in the quarter (+57%), grew revenue 31% to $3.0bn, and reaffirmed the path to a 40% FCF margin by FY28. The ZS de-rating was idiosyncratic; the leader’s cash engine is intact. And yet the stock still fell ~5.6% on the day — a GAAP operating loss (acquisition stock-comp and integration), plus the broad tech retreat. Notably, PANW finished the week up (≈257 → 279); the fundamental verdict and the one-day tape diverged.

Three flawless-to-strong prints, three negative one-day reactions. Issue 10’s Credo counter-tape — 157% growth, stock down 10% — was not an outlier. It was the leading edge of the regime: the AI complex is now priced such that beating consensus is the entry fee, not the reward. The fundamentals are validating the thesis at every layer; the trade of buying the consensus into the print is where the money is being lost.

Follow the debt — the AI buildout went on credit, and that’s the backdrop to the selloff

There’s a question this Digest has under-asked across eleven issues of “AI demand is real and the capex is enormous”: how is the buildout being paid for? Increasingly, the answer is debt — and that changes the nature of the risk.

Alphabet is the cleanest example. To fund roughly $185 billion of 2026 capex (nearly double 2025), Google’s parent — a company famous for its cash pile — has turned to the bond market at a pace that would have been unthinkable a year ago: a record ~$32 billion bond sale in February (upsized from an expected $20 billion as demand poured in, including a rare 100-year sterling tranche), on top of $25 billion last November. Its long-term debt quadrupled in 2025, to roughly $46.5 billion — and on 11 May it disclosed plans to sell yen-denominated bonds for the first time. The borrowing isn’t slowing.

It isn’t just Google. Oracle opened 2026 with an ~$18–25 billion offering; Meta has lined up one of its largest ever (up to ~$30 billion); Amazon is in the market too. UBS estimates hyperscaler debt issuance will reach $230–240 billion this year — against a 2020–24 average near $28 billion. CNBC’s framing is the one to hold: the bond binge “shatters the unspoken contract” that AI capex would be funded out of cash flow.

That shift is the quiet engine under this week’s tape. While the buildout was equity- and cash-funded, the AI trade was speculative equity risk — the downside was a lower multiple. Once it’s debt-funded, it becomes a credit question: the bonds need steady cash flow to service, and that cash flow is exactly what gets squeezed if the AI-revenue ramp stalls or margins dilute — precisely the worry Hock Tan handed the market on the Broadcom call. The same institutions reassessing whether $700 billion of annual AI spend is sustainable are now also its lenders. When a record-quarter stock falls 13%, part of what’s being repriced is not the demand but the financing of the demand.

Methodology note: this Digest has tracked who owns the AI names (13F) and what they earn (10-Ks) — but not how they finance the buildout. That’s a gap we’re flagging and closing; the debt channel is now the single largest source of AI capital, and it carries a credit risk the equity layer doesn’t.

A ‘healthy pause’ and a rotation — tech retreats, financials and defence bid on US–Iran

The macro backdrop flipped from Issue 09’s. That issue’s record highs leaned partly on US–Iran ceasefire hopes; this week the wires carried “Stocks Fall on Escalation of US–Iran Hostilities.” As tech took what one outlet called a “healthy pause,” money rotated to defensives and beneficiaries: in our own anomaly layer on 4 June, JPMorgan, Goldman Sachs and Citi each jumped ~3–5%, and RTX (defence) rose ~3.9% — all well outside their normal range. When the most crowded trade exhales and geopolitics turns, the bid goes looking for somewhere less owned. That is rotation, not capitulation — but it is worth watching whether the AI complex’s “pause” is the dip that gets bought or the first lower high.

Since Issue 10 (2 June)

Watchlist Updates

The Week Ahead

Forecast [2026-06-05-001]: at least one of CrowdStrike (~9 Jun) or Oracle (~10 Jun) beats consensus revenue and still closes lower the next session. Confidence 0.55, horizon 9 days. If both rise on beats, the expectations reset is over and the complex has found its floor; if they fall on beats, “great isn’t enough” is the regime and you keep being paid to wait, not to chase.

Methodology and Disclosures

Filings Intel Digest is a bona fide financial publication. Nothing here is personalised investment advice. The editor may hold positions in companies discussed. All claims are sourced to publicly filed documents or identified press sources. Calibration is partial — treat any forward statement as a hypothesis, not a forecast.

Frozen as of Thursday 4 June 2026 (US close), covering 30 May – 4 June. No new primary filings were ingested in the window and the extraction pipeline returned 0 structured signals; this issue is sourced from earnings releases (AVGO/CIEN/PANW), press, and the DB-verified price/anomaly layer (the week’s moves and the 4 June financial-sector anomalies are from price_bars/price_anomalies). The AVGO, CIEN and PANW financials are from each company’s results release and press coverage (WSJ, Barron’s, Reuters, Yahoo, Seeking Alpha) and are directional pending each 10-Q. Positioning is unchanged from the Issue 10 convergence layer.