Revised AVGO scenarios
What Jalapeño changed: the partnership the stock sold off on (June 3 guide not raised) just produced tangible, marquee execution — a real wafer, a named flagship, a claimed ~50% inference-cost cut, and the fastest high-performance ASIC cycle on record. That lifts the base and bull cases on revenue visibility and cadence. It does not touch the bear mechanics (capex digestion, multiple compression) — and it quietly seeds a longer-term structural risk covered in §02.
| Scenario | Target | vs ~$393 | Illustrative odds | What Jalapeño does to it |
|---|---|---|---|---|
| Bear3–6 months | $300–330 | −18% | ~20% | Unchanged. AI-capex digestion or a soft hyperscaler print still compresses the multiple regardless of one chip launch. |
| Base6–12 months | $420–440 | +9% | ~40% | Strengthened. A shipping flagship with a marquee customer hardens near-term revenue visibility. |
| Bull12–18 months | $540–570 | +41% | ~28% | Strengthened. Faster design cadence → more frequent tape-outs feeding the >$100B FY27 AI target. |
| Stretched bull~24 months | $680–720 | +78% | ~12% | More credible. If models-design-chips compresses the cycle for real, generations stack faster than a linear capex model assumes. |
Anchored at ~$393 (post-earnings), ~4.76B shares. Forward FY26 EPS ~$15 scaling toward ~$20 in FY27. Odds are illustrative weightings, not probabilities of record. Consensus target ≈ $524 (Strong Buy).
The disintermediation timeline
This is the explicit bear thesis, and it is the same flywheel as the bull case read backwards. OpenAI has stated the goal plainly: build the full stack to reduce reliance on external hardware. Today Broadcom owns the hard physical work — RTL implementation, place-and-route, timing closure, signoff — the exact know-how a customer accumulates by doing it once. Each generation, OpenAI can in-source more of it, using its own models to accelerate the design. Broadcom's value-add per program narrows toward what's hardest to replicate: networking silicon and advanced packaging.
Crossover is the whole thesis: by the third turn of the crank, the lower-margin integrator may hold a more defensible relative position than the higher-quality silicon partner — because nobody wants to in-source rack assembly, but everybody eventually wants to own their own physical design.
The single metric that falsifies the AVGO bull case
How much of generation two and three OpenAI (and the other five custom-silicon customers) keeps in-house versus leaves with Broadcom. Watch the ratio over the next few tape-outs — that, not quarterly AI revenue, is the structural ballgame. Networking attach rate is the tell: if Tomahawk content holds even as implementation work leaks away, the moat is intact.
Celestica: the inverted-risk proxy
CLS is the named integrator on Jalapeño — boards, racks, system integration — and it plays the identical role for AMD's "Helios" rack platform and for hyperscaler networking switches. Its leverage to the build-out is just as real as Broadcom's, but its risk profile is almost the mirror image: lower quality business, lower margin, lower IP — and therefore lower disintermediation risk. No hyperscaler dreams of in-sourcing low-margin contract assembly. That is the structural case for owning both, not one.
Why CLS benefits
- Named in the announcement. Celestica industrializes Jalapeño into deployable racks — recurring, program-tied work that scales with units shipped.
- Customer-agnostic layer. Every hyperscaler designing its own XPU still needs someone to bend metal into racks. CLS wins regardless of whose chip wins the socket.
- Already executing. Q1 2026 revenue +53% to $4.05B; adjusted EPS $2.16; FY26 guide raised to $19B revenue / $10.15 EPS; new 800G & 1.6T switches and a co-packaged-optics program ramping 2027.
- Second flywheel. AMD "Helios" rack collaboration means CLS isn't a one-customer bet on the OpenAI program.
Why CLS is the riskier hold
- Razor-thin margins. ~10–11% gross margin is structural to contract manufacturing — there's little cushion if volumes wobble.
- A contract manufacturer at ~50× earnings. The multiple prices CLS like a growth-tech name; any deceleration de-rates it hard. It already ran to a ~$655 ATH and gave much of it back.
- Customer concentration. A large share of revenue sits with a handful of hyperscalers — the same capex cycle that lifts it can cut it.
- Lowest link in the chain. Defensible against in-sourcing, but also the most commoditized and cyclical — it captures volume, not pricing power.
How the two fit together
Broadcom is the higher-quality business with the higher long-tail disintermediation risk. Celestica is the lower-quality business with the more durable relative position. Owning both is a way to hold the AI-infrastructure build-out while hedging the single structural risk that threatens the better company — at the cost of taking on CLS's margin and valuation fragility. They are not redundant; they fail for different reasons.
The postulate, restated
The original line was "the asymmetry skews favorably if you believe AI infrastructure spend compounds." Jalapeño sharpens it into something testable:
The asymmetry skews favorably if AI-infrastructure spend compounds faster than the hyperscalers learn to disintermediate the parts they're paying for.
The compounding now looks real — a shipping flagship, a record design cadence, a demand-expanding cost cut. The open question was never demand; it's durable share of demand. Broadcom holds the high-value share that's most tempting to in-source. Celestica holds the low-value share that isn't. The two together are the cleanest public expression of "own the build-out, hedge the one risk that matters" — with OpenAI itself, the actual catalyst, sitting off the board where retail money can't follow.