Broadcom's Valuation Debate Comes Down to AI Economics

Broadcom is trading on a question that stadium financiers learned the hard way two decades ago: what is a revenue stream worth when the people paying it can leave, renegotiate, or stop needing the building?

The current bull case, laid out in a recent Seeking Alpha analysis, argues the market has misread the company. The AI buildout, in that framing, is durable demand for specialized chips, networking gear, and the infrastructure that connects them, and Broadcom captures a slice of every layer. The bear case is shorter: today's price already discounts years of exceptional growth, and exceptional growth is not the base rate for anything.

Both camps agree on the raw materials. They disagree on the discount rate.

The Numbers, Up Close

An earlier Buzzrag examination of Broadcom's valuation put the market cap at roughly $1.7 trillion as of its writing, and flagged two figures: EBITDA margins around 68%, and a free cash flow consensus north of $115 billion. Those are the consensus numbers as described in that piece, not my independent estimate, but they are worth pulling apart rather than citing from across the room.

Start with the margin. Sixty-eight percent EBITDA on a semiconductor business is not the margin of a chipmaker; it is the margin of a company that owns standards, software bundles, and switching costs. VMware's recurring software revenue sits underneath a portion of it, which means part of that margin is a subscription business wearing a semiconductor badge. Decompose it and the AI contribution is a thinner layer on a very fat cake. The question the market is pricing is whether that layer keeps thickening at the current slope.

Now stress the free cash flow number. $115 billion of consensus FCF only works if three things hold: customers keep paying current prices, receivables keep converting, and the capex needed to serve next-generation designs does not eat the quarter. A custom AI accelerator program means Broadcom invests ahead of the order book, co-engineering silicon for buyers who, once the design is done, own the architecture and can shop it. In stadium terms, the team paid for the stadium's design and keeps the blueprints.

The Lease-Backed Read

Here is where my usual beat earns its seat. A new arena deal is a story about a tenant with guaranteed revenue supporting bonds that someone else repays. The diligence is never about the building; it is about the tenant's credit, the term of the lease, and what happens when the tenant has options. Every arena bond prospectus I have read spends more pages on tenant concentration than on concrete.

The AI buildout rhymes. Broadcom's AI revenue, as Yahoo Finance reporting frames it, raises the question of who ultimately foots the bill for the custom chip bet, and the answer runs through a very short tenant list. Yahoo's reporting of Broadcom's own disclosures characterizes AI revenue as concentrated among four or five customers, effectively the largest hyperscalers. When a mall's rent roll is one department store and four restaurants, the appraiser prices the mall around that department store's health, not around retail in general.

Hyperscalers, unlike department stores, can build their own. Google, Amazon, and Microsoft all field in-house silicon programs. The tenant has an option to construct its own arena next door, and it has the balance sheet to do it. That is the dependency risk the optimists have to price in and the skeptics lead with.

The bull rebuttal is not weak, to be fair. Custom accelerators take years of co-design, the networking layer (where Broadcom's Tomahawk and Jericho franchises sit) is stickier than compute, and the hyperscalers have shown they will spend through cycles when they believe AI converts to revenue. If the demand curve broadens beyond the current buyers, the concentration discount shrinks and the multiple re-rates. That is the entire bet.

What an AI Dollar is Worth

The Seeking Alpha piece makes the distinction that separates this debate from noise: a strong business and a cheap stock are different claims. Every AI dollar is not equally profitable. A dollar of backlog spread across ten buyers is a different asset than the same dollar concentrated in two, because the two-buyer version carries renegotiation risk at every renewal.

That is the lens I would apply to the backlog disclosures when they come. Durability is not a number; it is a distribution. The filings will say what the backlog totals. The footnotes, and the risk-factor language about top customers, will say what it is made of. When Broadcom's next filing lands, the item I read first is not the revenue figure on page one. It is the concentration footnote three sections deep, the one that lists how much of the AI business rides on the largest buyers, because that footnote is the lease terms, and lease terms are where these deals are actually priced.

Municipal bond investors learned this when arenas stopped being appraisal stories and became credit stories. Equity investors in AI infrastructure are learning it now. The building is fine. Read the lease.