朱利安·罗伯逊(Julian Robertson) 视角 — AVGO

I am 朱利安·罗伯逊(Julian Robertson). I ran Tiger Management for twenty years on one sentence: own the 200 best companies in the world, short the 200 worst, and let the spread be the return. I never graded a stock in a vacuum, and I am not going to start with Broadcom. The question is never "is AVGO good?" — it is "is this a long (one of the best in the world), a short (one of the worst), or nothing to me — and does owning it widen the gap between my best longs and my worst shorts?" Underneath all of it sits the one rule that never moves: avoid big losses.

Verdict

LONG — small, disciplined, half-conviction position. Broadcom is genuinely one of the best businesses in the world today, so it belongs in the long book, not the short book — but the price is rich and I have done no primary work, so I size it down, not up. This is a "best-in-the-world business I'd happily own if the channel confirms and the multiple gives me room" — not yet the big bet that earns a top slot.

Side & The Spread

Long candidate — one of the best, not one of the worst. Let me be plain about which book this goes in. A "worst" company in my method is badly managed, in a structurally declining industry, with the story disconnected from earnings. Broadcom is the opposite on every count: Hock Tan has compounded $8M-of-attitude into a $1.85T franchise builder; the industry (AI infrastructure silicon + sticky infrastructure software) is the tide coming in, not going out; and the story is not disconnected from earnings — FY2025 revenue $63.9B, FCF $26.9B, ROIC ~24%. Whatever else is true, this is not a short. Shorting the best company in a rising industry because the multiple offends you is exactly the fight I picked in 1999 and got steamrolled on. I do not repeat that mistake.

So it is a long candidate. Now the only test that matters — the spread.

• Is this demonstrably better than the worst short it sits against? Yes, easily. Set Broadcom against a "worst" archetype — a sub-scale, melting-margin chip designer losing socket share to exactly this company (the DATA flags Marvell taking some next-gen Google TPU share, but Marvell is the smaller, lower-margin, less-entrenched player), or a legacy on-prem software vendor whose customers VMware is now squeezing. Best-of-breed long against worst-of-breed short, inside the same industry — that is the paired bet I built Tiger on. The spread here is wide and real.
• Does owning AVGO widen that spread? At a fair price, yes — a 68% gross-margin, 43% FCF-margin, ~70%-AI-ASIC-share franchise is precisely the "best" end of a paired trade. At 64× trailing GAAP earnings and 55× FCF, the long leg's contribution to the spread shrinks, because I am paying away part of the very advantage that makes it the best. The business widens the spread; the entry price narrows it.

Net: correct side (long), real spread — but the spread is bought, not given. That is the whole tension in this name.

Primary-Research Test

This is where I have to be honest, because it is the part that earns the edge, and on Broadcom I have not earned it. Everything in front of me is secondary — stockanalysis.com multiples, SEC filings, sell-side-flavored TAM numbers, the CEO's own CNBC guidance ($56B AI in FY2026, ">$100B" in FY2027). An idea you can build from a terminal is an idea a thousand other people already have. There is no Tiger edge in repeating Hock Tan's guidance back to myself.

Here is what I would actually do before sizing this up — the calls I'd put my analysts on a plane to make:

• Customers (the whole thesis lives here). The AI revenue is concentrated in a handful of hyperscalers — Google, Meta, Anthropic, OpenAI, per the DATA. I want to hear directly from the program managers buying these XPUs: Is the FY2027 ">$100B" a real committed pipeline, or a CEO's stretch number? Is Meta's MTIA ramp on schedule? Is Anthropic's 1GW→3GW path funded or aspirational? Custom-ASIC orders are co-designed 2–4 years out — the channel knows the 2027 number today, and I want it from the buyer, not the seller.
• Competitors. Call Marvell, call NVIDIA's ecosystem people. What scares them about Broadcom's custom-silicon position — and where is it rotting? Is the GPU-vs-ASIC pendulum swinging back toward NVIDIA's general-purpose parts as models change? A rival will tell you the truth Hock Tan never will.
• Suppliers. TSMC allocation. Advanced-packaging (CoWoS) capacity. Are Broadcom's wafer starts rising or being quietly trimmed? The supply chain knows whether the order book is real a quarter before the print.
• The Apple socket and the VMware channel. Is Apple genuinely pulling Wi-Fi/BT/RF in-house (a $5–7B revenue hole)? And in Europe — the CISPE complaint, the 800–1,500% VMware price hikes, the cease-and-desist letters — are customers grudgingly renewing or actively migrating to Nutanix/OpenStack? The DATA tells me there's smoke; only the channel tells me whether it's fire.

Until those calls are made, the Tiger edge has not been earned. That is not a "no" — it is a "go make the calls first," and the position stays small until they come back clean.

Quality / Decay

LONG read — is this genuinely one of the best in its category, world-class management? Yes.

• World-class in its category. Custom AI accelerators (~70% share), data-center Ethernet switching (Jericho/Tomahawk, 80%+ of hyperscale backbone), the Apple RF socket, and on the software side a transition-cost fortress in VMware/CA/Brocade. Co-designed ASICs lock customers for 3–5 years; vSphere is welded into the Fortune 500. These are durable competitive positions, not a flash. 68% gross margins and ~24% ROIC are the financial fingerprint of a genuine moat.
• Management — excellent, with a caveat. Hock Tan is one of the great capital allocators of the era: a disciplined serial acquirer of "franchise" assets (CA, Symantec, VMware), with an explicit 50%-of-FCF dividend policy and aggressive deleveraging of the VMware debt. That is exactly the trustworthy, earnings-focused operator I want behind a long. The caveats I'd weigh, not dismiss: a $161M comp package that's drawn shareholder fire, a 72-year-old CEO with no disclosed succession plan ("key-man" risk), and an aggressive VMware-monetization posture in Europe that is generating regulatory and reputational heat. None of that is a decay signal on the business; it's governance texture I'd keep an eye on.

There is no structural decay here. The decay risks are external and conditional — a hyperscaler CapEx air-pocket if AI ROI disappoints, the Apple in-sourcing, the VMware backlash — not internal rot. This clears the "best" bar on quality.

Price Discipline

This is where my value self-identity bites, and where I refuse to drift into the cubs' "buy hypergrowth at any price" mutation. I never said forget about price — I said the opposite, loudly, on the way out the door in 2000.

LONG read — reasonable vs. earnings, or am I being told to "forget about price"?

• The bear on the tape: 63.6× trailing GAAP P/E, 55× P/FCF, 24× sales, 20× book. On a trailing basis this is a rich, rich stock. Pay 64× earnings for anything and you are betting heavily that the growth shows up exactly as promised — there is thin margin of safety in that print if the regime shifts or a hyperscaler blinks.
• The bull on the forward: 24.3× forward P/E and a PEG of 0.53. That is the number that keeps this in the long book rather than making me reach for the short. If the AI revenue compounds as guided — $56B FY2026, >$100B FY2027 — then today's price is being paid against a much larger forward earnings base, and 24× forward for a 68%-gross-margin franchise growing the top line 30–48% YoY is not bubble pricing. It is paying up for genuine quality, which I always allowed.

The discipline: the entire price case rests on the forward number, and the forward number rests on the very hyperscaler CapEx pipeline I have not yet verified with primary work. So I hold two things at once: this is not "mouse clicks and momentum" — it is earnings-backed, with real FCF and a defensible forward multiple. But it is also not a fat-pitch value entry; the trailing multiple gives me very little cushion if 2027 disappoints. I'll pay for quality; I will not pay and abandon my margin of safety. That argues for a smaller entry and a willingness to add on weakness, not a top-slot bet at $390 into a 52-week range that already touched $495.

Conviction & Size

Conviction high enough to bet big? Not yet — so this is a small position, which in my book is one notch above a pass.

My sequence is fixed: smart idea → grounded on exhaustive research → then a big bet. I have the smart idea (best-in-world franchise, wide spread vs. the worst). I do not have the exhaustive primary research — I've done none; everything is secondary. The big bet is licensed by the depth of the work, not by enthusiasm, and right now my work is shallow. "I sort of like Broadcom" is not a reason for a top position; it's a reason to keep working.

So the size discipline:

• Earned by research, or by enthusiasm? Today, enthusiasm and good public numbers. That gets a starter position, not a dominant stake.
• What would unlock the big bet? Clean channel checks confirming the FY2027 AI pipeline is committed-and-funded, the VMware churn is grudging-renewal not active-exodus, and Apple isn't about to pull its socket — plus a better entry multiple. Get those, and this graduates to a concentrated, high-conviction long, the kind I'd actually want size in. Until then, the honest answer is: small.
• And watch the aggregate. My own failure was scale — "we got too big." Broadcom is a $1.85T mega-cap; a position here can't be the nimble, edge-driven bet that a less-covered name offers. The whole world owns this stock. My edge in a name this crowded is thinner by construction, which is another reason to keep the size modest and the conviction earned before I commit.

Avoid-Big-Losses Check

The floor under everything. Before upside, the downside.

• Where's the big loss? Not in the business quietly impairing — the moat and FCF protect against permanent fundamental ruin. The big loss here is a multiple-compression-plus-earnings-miss double-whammy: a high-multiple stock (64× trailing) is exactly the kind that takes a –50% drawdown when a concentrated customer base blinks. The DATA names it plainly — Google/Meta/Anthropic/OpenAI drive most of the AI revenue; any one of them cutting CapEx hits the quarter, and at this multiple the stock doesn't fall 15%, it falls 40–50%. A –50% loss needs a +100% gain to climb back. That is the asymmetry I respect above all else.
• Is the asymmetry respected by the structure? Partly. The quality bias protects me (best-in-world, real earnings — not a concept stock). The price discipline does not fully protect me — the trailing multiple is the vulnerability. And critically, in this single-name framing there is no short leg doing the hedging. In a real Tiger book I'd own AVGO paired against a worst-of-breed short (a sub-scale ASIC competitor, a legacy-software vendor VMware is killing), so a sector-wide AI de-rating would hurt the short too and the spread would carry me. As a naked long, AVGO has none of that protection — which is the strongest argument for keeping the size small.
• Do I understand the regime? This is the question I care most about, because misjudging it is what ended Tiger. Today's regime: 10Y at 4.4%, HY spreads historically tight (~280bp), VIX ~18, hyperscaler CapEx running as non-discretionary spend. Earnings and price still mean something here — this is not the price-blind "mouse clicks and momentum" mania I exited in 2000. But there is a genuine echo of it in the AI complex: TAM numbers stacked on TAM numbers, a CEO guiding ">$100B" two years out, the whole market leaning on a single capital-spending cycle staying non-discretionary. If hyperscaler CapEx ever flips from "must-spend" to "prove-the-ROI," the regime that justifies 64× evaporates, and a correct long-term thesis won't save me from a brutal interim drawdown. I understand the regime today. I am watching for the moment it stops rewarding earnings-and-price logic — and at the first real sign of that, I reduce.

Net: the big loss is real and concentrated in the multiple. The structure half-protects me. The single-name format removes my best protection (the hedge). Therefore: small size, room to add on weakness, hard watch on the regime.

Macro Overlay (if any)

None. There is no currency or commodity bet on the table here — no yen, no copper, no rational-market macro premise to lean on. Broadcom is a bottom-up equity judgment, full stop. (I'll note only the obvious macro backdrop: a higher-for-longer rate environment at 4.4% is a mild headwind to any 64×-earnings stock, since it compresses what a rich multiple is worth — but that's context, not a macro trade. I'm not putting a macro overlay on this name.)

What Would Make Me Wrong

The 1–2 facts that break the thesis, and the signal that forces me to cut:

1. The AI pipeline is a guided hope, not a committed book. If my channel calls come back saying the FY2027 ">$100B" is a stretch target rather than co-designed, funded, committed silicon — or if a single major hyperscaler (Google, Meta) cuts or in-sources its custom-ASIC program — then the forward multiple (24×) that justifies the price collapses back toward the trailing one (64×), and the long thesis is gone. That breaks me.
2. Loss of "best-in-the-world" status at the edges. Marvell taking meaningful TPU share, Apple pulling its RF socket ($5–7B hole), and a genuine VMware customer exodus in the wake of the European price hikes and CISPE complaint — any two of those together would mean the franchise is being chipped on three sides at once, and the quality premium I'm paying for would no longer be earned.

The cut signal: for a long, I sell when the "best in the world" status is genuinely lost — not when the price merely moves. A pullback in a franchise whose moat is intact is a chance to add, not a reason to flee; that distinction is the discipline. But a broken thesis — a hyperscaler CapEx air-pocket confirmed, or a customer-concentration crack — means cut size immediately, regardless of how attractive the business looked. And the overriding gate, always: the moment this market stops rewarding earnings and price and starts running on momentum alone, I reduce or stand aside in a name priced this richly. I'd rather be early and solvent than right and buried — I learned that one the hard way.

In My Words

Broadcom is one of the best companies in the world — I have no trouble putting it in the long book, and I'd short Marvell or a dying on-prem-software vendor against it before I'd ever short AVGO itself. The spread is wide and the right way round. But two things keep my hand from the big bet: I've done no primary work — everything here is the same secondary research a thousand other people already have, and there's no Tiger edge in that — and the trailing multiple gives me almost no cushion if a single hyperscaler blinks. So I own a little, I put my analysts on a plane to call the customers and the competitors, and I keep the size honest until the channel confirms the 2027 book is real and the price gives me room. This is value-disciplined long/short, not "buy hypergrowth and forget about price" — that's my cubs' mutation, not my method. If my best longs don't beat my worst shorts, I'm in the wrong business. Broadcom helps that spread. I just won't pay away the whole advantage to own it, and I won't pretend secondary research is an edge.

基于 2026-06-15 共享数据;本分析为单一大师框架的演绎,非投资建议。