AVGO Follow-Up: The Buyside Wanted $130B — Why the Flush Is the Trade

1. The print beat the Street and missed the room

The numbers beat every published line — revenue $29.59B vs $29.36B LSEG, adjusted EPS $3.32 vs $3.24, semis $16.7B vs $15.2B StreetAccount, op margin 67.9% vs ~66.7% — yet the stock dropped to $350.99 (-4.6%) at 5:14pm ET before recovering to ~$369.90 by 5:39pm as Hock Tan laid out the multi-year targets. That whipsaw is the whole story: the published bar was cleared and the unpublished bar was not.

TMT Breakout's EOD wrap, the cleanest buyside read I found, put numbers on the whisper: Q4 AI $21.7B vs $21.3B Street but short of the >$22B buyside bogey; Q4 op-margin guide 66.0% vs ~66.5–67.0% wanted; and investors came into the call wanting roughly $130B or more for FY27 AI, not the $115B Tan gave. The quarter is fine in absolute terms but underwhelming stacked against Nvidia's print and elevated AI expectations — their words, and I agree with the framing even though I disagree with selling it.

The consequence is mechanical: in maxed-expectation AI names, the marginal quarter sets the price, not the multi-year vision. June did -12.6% on a record; PANW beat-and-raised for -9.3% and MDB for -13.5% the same evening; SNOW soared ~20% on a clean 5-point raise because software breadth was new news. AVGO's $115B was a genuine raise from June's ">$100B" floor, but it landed below the whisper — so the tape sold first and listened second.

2. What the call actually gave you: $350B with supply secured

I read the full Q3 FY26 transcript (prepared remarks + Q&A, via the SingjuPost transcription of the Sept 2 call). Three disclosures matter more than the Q4 $250M delta, and none were in the press release.

First, Tan committed to ~$115B FY27 and ~$230B FY28 AI semis — "secured the supply" for both, demand already exceeding the 2027 outlook — and framed it as ~$350B shipped over two years "with a pretty high degree of confidence," explicitly distinguishing shipping chips from all ~30GW being deployed in-window. Bernstein's Stacy Rasgon did the GW math live on the call (~$11–12B per GW vs a competitor's ~$40B claim); Tan's answer — custom XPUs at less than half GPU cost — is the margin story in one line.

Second, the customer stack is now named and ranked: Anthropic 5GW TPU 8i in 2027 + incremental 10GW in 2028, on track to be the largest XPU customer both years; OpenAI Jalapeno 1.3GW in 2027, >5GW with gen-2 in 2028, gen-2 near tape-out, gen-3 in development, claimed at half GPU cost and ~comparable to Vera Rubin on OpenAI workloads; Google multi-tens-of-billions per year with TPU 8i now shipping ahead of MediaTek's v8t; Meta MTIA at scale in Q4, three generations through end-2027, 3GW line-of-sight. XPUs were 73% of AI revenue with shipments up 3.5x YoY; Tomahawk 7 (200T) just taped out.

Third, Tan was blunt about the constraint that governs all of it: land, power and shell dictate timing, not wafer starts — with substrates (Singapore fab starts FY27) and HBM as the known sub-bottlenecks. In my opinion this is why the $115B is conservative rather than promotional: it is a judged number against site readiness, and Tan said customers want more.

3. The financing is the thesis — and the risk

CFO Amie Thuener's exact wording deserves quoting precisely because it is the load-bearing sentence for the $30+ FY28 EPS target: Broadcom is "empowering two of our most strategic customers, the leading AI labs, to bridge the gap between their current cash flow and the significant upfront investments required," and "may provide modest residual value guarantees, which are contingent liabilities we view as low risk." When Melius asked for maximum off-balance-sheet exposure (~$29B cited from last quarter), she declined: "nothing new to announce today."

The primary-source scaffolding is public and I verified it: the June 9 Broadcom IR release establishing the AI XPV platform with Apollo and Blackstone for >20GW through 2028, launching with a $35B Anthropic >1GW tranche on Fluidstack sites from mid-2026; and Apollo's matching release calling it the largest private financing ever executed, with GS/WFC/Citi and MS/JPM advising. August press reports of $60B+ further debt talks for Anthropic fit the same vehicle. Tan added the credit hinge on the call: Anthropic's expected IPO ("open secret") changes its investment-grade standing; OpenAI's path is less clear — so each tranche gets tailored, case by case.

My opinion: this is the AI-debt loop personified — vendor silicon plus private-credit capital plus lab ARR, with Broadcom's balance sheet ($24.0B cash, $13.7B quarterly FCF, 46% margin) as the backstop. It works while labs monetize per-GW; it breaks if ARR per GW disappoints or guarantees get called. That is why the 10-Q's contingent-liability and receivables lines are the next catalyst, not the next headline. Any RVG/backstop figure >$5B uncollateralized flips my stance from buy-the-flush to sidelines.

4. Smart money already sold — that is bullish for the flush

Two pre-print exits clear the overhang rather than warn of it, in my view. Third Point's Aug 14 13F showed its ~50,000-share AVGO position (plus ~190,000 NVDA shares) eliminated by June 30, rotating into Warner Bros Discovery and Alphabet; Druckenmiller likewise exited AVGO pre-print into Alphabet. These are June-quarter decisions — before the $115B/$230B raise existed — and AVGO sits ~25% below its $495 June 3 high with YTD +6% vs S&P +12%. The weak hands sold the uncertainty; the flush leg buys the resolution.

Insider flow leans the same way: June–July Form 4s show co-founder Henry Samueli and officers selling into $377–401 (routine strength-trimming, no open-market buys in the 30-day pre-print window I pulled). No signal of distress, no signal of conviction — neutral, and worth stating so nobody over-reads it.

Net: no contradiction with the open book. The flush branch ($330–348) buys a confirmed raise at a discount after the sellers sold; the breakout branch (>$385) buys consensus capitulation toward $30 EPS. Nothing in between. The hawkish-rate cluster (IWM short, OBDC long, puts) remains the explicit hedge if the 10Y clears 4.85% twice — disclosed conflict, deliberate two-sided structure around the Sept 16 FOMC.

5. The trade, the scenarios, and what kills it

Execution is unchanged from the armed branches — limit orders, not chase:

• Bull 30%: sell-side re-models FY28 to $30+ within days; Q4 AI run-rate confirms >$22B pace; close >$385 triggers toward $455.
• Base 50%: rate tape (10Y ~4.78%) + software lumpiness drags a retest of $335–345; flush fills, recovers toward $430 into year-end.
• Bear 20%: hawkish NFP Friday (Sept 4, consensus ~+50–55k; ≥+120k locks the hike) or 10Y ≥4.85% twice compresses AI multiples 10–12%; weekly close <$322 opens $295–305.

What would prove me wrong: (1) weekly close <$322; (2) 10-Q RVGs/backstops >$5B uncollateralized or receivables accelerating without cash conversion; (3) Anthropic/OpenAI defer 2027 GW deployments; (4) Q4 op margin <63.5% vs ~66% guide. Any single one cuts the flush leg immediately.