Nvidia and Broadcom: The Two AI Infrastructure Trades That Hold Up

The five largest cloud operators are on track to pour somewhere near $660 billion to $690 billion into AI infrastructure this year, almost double the 2025 level. That flood is why the "AI infrastructure" label gets slapped on everything from GPU vendors to cable tray makers. But the money does not all flow through one pipe. It supports two structurally different businesses, and judging them with the same yardstick is how retail investors end up paying for hype.

There is a quick way to tell them apart. Ask not who shows the best spec sheet, but who owns the cost per token — the price of actually serving a model at scale. That single question separates the two companies worth owning from the name everyone is talking about.

The trade-off at the center of the buildout

At data-center scale, the binding constraint on an AI workload is rarely peak teraflops. It is how many tokens you can push for a dollar of capital and a watt of power, and for a high-volume serving workload that cost is dominated by memory bandwidth and utilization efficiency. That is the economics that explains the market's shape.

Nvidia sells the general-purpose answer: one GPU platform, wrapped in a software ecosystem, that runs almost any model. A hyperscaler too big to settle for that can instead pay for a purpose-built chip, designed around its own specific workload, and get more tokens per dollar — but only because it is big enough to amortize the design cost. That is the custom-silicon carve-out, and it is the single largest structural threat to Nvidia's dominance, larger than AMD. Both approaches can be profitable. They are just different margins and different risks.

The platform that keeps its margin

Nvidia still takes roughly four of every five dollars in the AI accelerator market. Its real moat is not the GPU die; it is the software and interconnect stack around it. Independent analysis puts Nvidia's software-stack efficiency near 93% against about 80–85% for rivals' competing software, and that gap is worth more than any peak-spec headline.

That moat shows up in the financials. Nvidia is converting about 47% of revenue into free cash flow, on top of roughly 74% gross margins and revenue growing at an annual clip above 80% — and it still trades near 27 times trailing earnings. A software moat plus that conversion efficiency at that multiple is unusual for a company growing this fast.

The current proof-of-execution is the Vera Rubin platform, which Nvidia says will be its fastest ramp ever — guided to roughly $20 billion of data center revenue in a single quarter, about a fifth of the segment, with rack-scale systems priced between $5 million and maybe $7.8 million apiece. A CFO telling you a product is ramping faster than anything in company history is a claim worth testing, not taking on faith. But the purchase orders have already gone out to every major buyer, and the first units are shipping. So far the story holds.

The custom backbone that tripled

Broadcom is the other pole. It does not compete with Nvidia on the general-purpose GPU; it builds the custom chips that hyperscalers design for themselves, plus the networking fabric that ties them together. Its AI semiconductor revenue more than tripled to $16.7 billion in the quarter ended in July, and its custom accelerators now represent about 40% of total company revenue. The customer list — Google, Meta, OpenAI, Anthropic — is the who's-who of AI buildout.

This is the TCO argument in action: these companies are big enough to buy their way out of Nvidia's margin, trading a one-time design cost for control of per-token economics at scale. Broadcom is their arms dealer, and it is nearly as profitable as Nvidia sells — around 43% free-cash-flow margin — though it carries a higher multiple, near 45 times earnings.

The catch is the flip side of concentration. When power sits in a handful of buyers, the narrative moves on their news. Broadcom fell about 5% when Google struck a custom-chip deal with rival Marvell, attaching a warrant that lets Google buy up to roughly $12 billion of Marvell stock through fiscal 2033 — a sign even a loyal customer wants a second supplier. Then its own quarterly guidance, missing analyst expectations for this quarter by a few hundred million dollars, knocked another 5% off the stock in extended trading. Growth this fast does not come with a guarantee it stays exclusive.

The hottest name fails the same test

That leaves the name retail commentary keeps pushing: AMD, up more than 130% this year. On a spec sheet it looks like a genuine rival — its MI350X matches Nvidia's B200 on FP8 compute and carries more memory, which is exactly the recipe for cheap inference.

But the deployed reality lags the paper. Independent testing finds AMD chips clock-throttling hard under sustained load, and its ROCm software stack runs at 80–85% of Nvidia's utilization efficiency, producing a real-world training-performance gap of 10–25% versus the spec sheet. The memory advantage still gives AMD a genuinely competitive cost per token on inference — it has a real niche. The question is what the market is charging for it: roughly 129 times trailing earnings, against a 23% free-cash-flow margin and an 8% return on capital. That multiple is paying for a transformation that the margins and execution have not yet delivered.

Here is the thing about the per-token test. Nvidia passes it because it owns the software that makes a general-purpose chip efficient, and it gives that up at 27 times earnings. Broadcom passes it because it is how hyperscalers seize control of their own serving costs, and it tripled its AI revenue to prove the demand is real. AMD passes it only on inference, in a niche, at a price that assumes near-total success — which is why, of the three, it is the hardest to call compelling.

The buildout is real; the question is what you pay to stand on it. Own the moat at a discount multiple, or own the custom-silicon backbone and accept that its billion-dollar users are also its billion-dollar weather. Both are defensible. Paying the richest multiple for the thinnest margin in the group is the one that needs the most faith.