To The Moon

Most investors still talk about Broadcom as if it were simply another beneficiary of the AI boom, a sort of well-positioned tenant riding Nvidia’s coattails. I think that framing is not just lazy, it is structurally wrong. Broadcom is no longer competing for floorspace inside the AI data centre. It increasingly owns the land, writes the zoning laws, and collects rent whether the buildings are fashionable or not. That distinction matters enormously when valuations start to feel uncomfortable.

Broadcom isn’t renting space — it’s writing the zoning laws

What makes $Broadcom(AVGO)$ interesting in 2026 is not raw compute performance but architectural control. This is an infrastructure story masquerading as a chip story, and the market is only just beginning to price that in.

The Ethernet Crossover: The 'Public Highway' Strategy

Most AI coverage treats networking like plumbing. In my view, networking is the strategic gatekeeper of AI scale. Nvidia’s InfiniBand was the dominant architecture for high-performance AI networking, and it was effective precisely because it was proprietary. Proprietary networks create walled gardens, and walled gardens are profitable.

But walled gardens eventually invite rebellion. The Ultra Ethernet Consortium is the revolt, and Broadcom is not a passive member; it is the ringleader. This is where the 'public highway' metaphor becomes useful. Broadcom is turning Nvidia’s private toll road into a standardised freeway, and it is selling the asphalt, the signage, and the maintenance.

Those 102-terabit switches aren’t just fast; they reduce latency and cost per bit at hyperscale, allowing clusters to grow from hundreds to thousands of GPUs without the network becoming the bottleneck. Broadcom’s switches now power a meaningful share of new hyperscale AI networking deployments, and the shift to Ethernet is already visible in how quickly $Alphabet(GOOGL)$ and $Meta Platforms, Inc.(META)$ are standardising on the open stack.

The practical outcome is that hyperscalers can now build large AI clusters using open Ethernet at scale, with Broadcom’s 102-terabit switches supporting the architecture. Once that happens, Nvidia’s networking moat becomes less relevant, and Broadcom becomes the default supplier for the networking layer of AI infrastructure.

The telling part is that $NVIDIA(NVDA)$ still headlines as the AI hero in public discourse, while Broadcom quietly becomes the infrastructure equivalent of the power company: invisible, boring, and absolutely essential.

The Hock Tan Doctrine: Turning Software into a Utility

The most important thing Broadcom has done since the VMware acquisition is not 'integration.' It is discipline. It is ruthless customer pruning. It is turning software into a utility, which is the only way to kill the semiconductor cycle.

Broadcom did not simply acquire VMware and continue the old model. It went through the installed base with surgeon’s precision and cut away low-margin, high-maintenance customers. The result is a dramatically improved margin profile. VMware’s operating margins have been pushed towards 77–78 per cent, levels that would make most SaaS CEOs laugh in disbelief, and private equity managers nod in approval.

To put that into perspective, top-tier SaaS companies typically operate with 25–35 per cent operating margins. $Broadcom(AVGO)$ has taken a business that was once a typical software company and converted it into a high-margin infrastructure utility. The growth story here is not simply revenue expansion; it is margin transformation and cash flow predictability.

This matters because AI hardware is cyclical, but utility software is not. Combined, this gives Broadcom something rare in semiconductors: a business model with a shock absorber built in. That combination makes Broadcom’s financial model unusually resilient, and it explains why the market is willing to assign such a high valuation multiple despite the obvious risks.

The Custom XPU Moat: Broadcom as the Ghostwriter

The third pillar of the thesis is the one people understand least: Broadcom’s role in designing custom AI chips.

Everyone talks about Google’s TPUs or Meta’s MTIA chips like these companies are building them in a vacuum. In reality, Broadcom is often the invisible partner, co-designing the silicon, power delivery, memory architecture, and networking integration. That is not a vendor relationship; it is an architectural partnership.

This is where the double-dip becomes real. For every dollar a hyperscaler spends on a custom AI chip, Broadcom captures a slice of the silicon revenue and the networking revenue. And because these chips are integrated into the data centre architecture, the switching cost is enormous. Replacing Broadcom is not like swapping out a GPU supplier; it is like redesigning a building’s foundation while still living in it.

The evidence for this is the backlog. Broadcom’s total backlog sits around $162 billion, with $73 billion (45%) tied to AI-related commitments. These are not short-term orders; they are multi-year, non-cancellable commitments. The scale is further highlighted by recent hyperscaler commitments that add billions in contracted revenue, reinforcing that this is not speculative demand—it is structural.

The Competitive Landscape: Why Nobody Can Copy the Model

This is where the thesis becomes difficult to dispute. Broadcom’s advantage is not that it is better at one thing. It is that it is embedded across multiple layers of the AI stack.

Nvidia is the king of compute, but it is not the king of networking, and it cannot own the open standards movement without alienating its customers. AMD can compete on chips, but it cannot replicate Broadcom’s networking dominance or hyperscaler integration. Intel is powerful in servers and CPUs, but it is not the default networking vendor for AI clusters. Marvell and Cisco can sell networking gear, but they lack Broadcom’s silicon depth and hyperscaler design influence.

Even if a competitor could match Broadcom in one layer, replicating the combination of custom XPU co-design, hyperscaler networking dominance, and high-margin software utility is an entirely different challenge. It is not a matter of 'if.' It is a matter of 'how many years of architecture do you need to rebuild.'

In short, Broadcom is not just competing in a market; it is designing the market’s structure.

The Bear Case: The Nightmare Scenario and Why I’m Still Comfortable

The most honest risk is customer concentration. Broadcom’s AI strength is heavily skewed toward a small number of hyperscalers. If one of them pauses AI capex by 30 per cent or decides to internalise silicon and networking completely, Broadcom’s double-dip could collapse into a single point of failure.

To make this concrete, let’s imagine a 30 per cent reduction in AI capex across the top hyperscalers. Broadcom’s AI-related backlog would not vanish, but it would slow sharply. The impact on revenue would be immediate because these are multi-year contracts tied to specific programs. If one major customer shifts strategy, Broadcom would feel it in the next 12–18 months.

What is the probability of this scenario? I would assign it a moderate probability, perhaps 20–25 per cent. I base this on hyperscaler commentary around slowing AI ROI timelines and rising margin pressure from shareholders and regulators. A pause is plausible, but not imminent, because these companies are still competing on AI leadership.

So why am I still comfortable? Because Broadcom is not a single-product story. Even in a slowdown, the VMware utility generates predictable cash. Even if one hyperscaler reduces spending, Broadcom still has a diversified infrastructure business and multiple hyperscaler partners. The worst-case scenario is painful, but not catastrophic.

AI’s most valuable real estate isn’t compute — it’s infrastructure

Verdict: Structural Monopoly, Utility Cash Engine — Worth Paying Up For

Broadcom at ~30x forward earnings is expensive by traditional semiconductor standards, but cheap for what it actually is. I’m not buying a cyclical chipmaker — I’m buying an architectural landlord with a stable utility cash engine and multi-layer hyperscaler lock-in. The concentration risk is real, but it’s also the source of the moat. I’m buying architectural sovereignty, not diversification. The only question is whether 2026 macro headwinds create a better entry point — but at current multiples, I’m comfortable building a position now.

In a world where AI infrastructure is becoming the most valuable real estate in technology, Broadcom is not just participating. It is owning the ground beneath it. And that, to me, is a very different kind of investment story.

@TigerStars @Daily_Discussion @Tiger_comments @Tiger_SG @Tiger_Earnings @TigerClub @TigerWire