The market spent July 2nd afraid of the wrong thing.
Broadcom fell more than 2% because The Information reported, citing three people with direct knowledge, that Anthropic has reportedly started developing its own AI chips and is talking to Samsung about manufacturing them.
Read the actual details and the fear evaporates. Anthropic's effort is "in its early stages." The company is still deciding how powerful the processor should be and how it might fit into server infrastructure. It's "some distance" from detailed design, testing, and manufacturing.
So the tape sold a chipmaker on a rumor about a customer that might build a competing chip years from now.
And here is the irony sitting in the same coverage: the marquee custom-silicon partnership everyone actually points to — OpenAI's Jalapeno chip — was co-developed with Broadcom.
The customer investors panicked about is theoretical. The competitor Broadcom is actively funding is already on its books.
That inversion is the whole story. Merchant chipmakers like Broadcom should fear their partners far more than their prospective customers, because the partnership itself is the transfer mechanism.
Merchant chipmakers like Broadcom should fear their partners far more than their prospective customers, because the partnership itself is the transfer mechanism.
Broadcom's business isn't selling a finished GPU off a shelf like Nvidia. Its business is being the ASIC co-design partner — the firm that turns a hyperscaler's architecture idea into a chip that can be manufactured at scale.
As one analyst put it, Broadcom brings "the chip design expertise and industry partnerships needed to turn hyperscalers' architecture ideas into chips that can actually be manufactured at scale." That's the pitch. It's also the problem.
Every marquee contract Broadcom signs is, functionally, a technology-transfer agreement. OpenAI gets Jalapeno. Meta reportedly gets its MTIA line — four generations of custom processors deploying through 2027, developed with Broadcom and fabricated by TSMC, per The Tech Panda.
The customer arrives without deep semiconductor expertise. It leaves each design cycle with more of it. That's not a side effect of the relationship; it's the deliverable.
The moat erodes one design cycle at a time, and Broadcom is paid handsomely to accelerate the erosion. It is, in the most literal sense, being compensated to train its own future competitors.
This is why the Anthropic headline was a category error. Anthropic is at the beginning of the exact process that OpenAI and Meta are already deep inside. The threat isn't the press release about a chip that doesn't exist. The threat is the chips that do.
To see where this goes, look back sixty years, because we've run this experiment before.
IBM was the most successful vertically integrated company of the 1960s, making its own hardware components for its mainframes on the theory that owning the specialized parts would improve the end product and the margins. It worked spectacularly: by 1985, IBM reportedly accounted for more than half the market value of the entire computer industry, as Carliss Baldwin documented in "Design Rules."
Then it collapsed. In the 1990s, falling semiconductor costs and the rise of Intel and Microsoft reportedly dug into IBM's moat. By 2000, IBM no longer claimed to be vertically integrated at all.
The lesson isn't that vertical integration fails. It's that vertical integration wins until the merchant market's scale and falling costs make in-house production a margin drag rather than an advantage — and then it reverses, hard.
Today's hyperscalers are early in that arc, not late. Seaport's Jay Goldberg frames the strategic logic plainly: the hyperscalers "recognize there's a grave strategic danger from having a single supplier of AI compute," so "they now have a very strong strategic reason to do their own silicon."
But Goldberg also thinks the swing is already approaching its "far limit." Designing a leading-edge chip is "a massive expense," he says, and "only so many companies can afford that."
That's the self-limiting mechanism. Alphabet, Amazon, Meta, Microsoft, OpenAI — the list of firms that can actually sustain a leading-edge roadmap, a software stack, and a foundry-and-packaging supply chain is short and doesn't get longer easily.
So the customer-turned-competitor narrative is real. It's also slow, capital-intensive, and confined to a handful of players. Anthropic weighing a chip it hasn't designed doesn't change that math.
Which brings us to the number that actually pays Broadcom's bills.
Nvidia runs roughly 75% gross profit, per Cambrian AI's Karl Freund. That margin is the entire prize in this fight. Freund's framing is precise: hyperscalers "get these chips at manufacturing cost," and against Nvidia's 75% markup, "these companies hope to derive significant savings."
Every custom-silicon program — Jalapeno, MTIA, Trainium, TPU, Maia — is an attempt to arbitrage away that 75%. Amazon claims its Trainium3 delivers up to a 60% price-performance benefit over GPUs for inference, per Yahoo Finance.
Here's the part investors should sit with. Broadcom's booming ASIC business exists because that 75% margin is worth attacking. The co-design revenue is downstream of Nvidia's pricing power.
If custom silicon succeeds broadly and the margin dollars compress, the economic incentive that funds Broadcom's co-design contracts compresses with it. Broadcom isn't insulated from the arbitrage — it's the vehicle for it, and the vehicle gets cheaper to build the more successful it becomes.
And then there's concentration. Freund's other observation is the quiet one: "Broadcom only has a few very large customers."
That cuts both ways. A handful of hyperscaler contracts is what makes the business enormous today. It's also what makes it fragile — every one of those customers is contractually acquiring the competence to eventually do the work in-house, and there aren't enough of them to lose one comfortably.
This is the same single-headline overreaction that misread Meta's compute posture on the same tape — one rumor, one 2% move, no engagement with the mechanics underneath. The market keeps pricing announcements when it should be pricing structures.
So the correct read on July 2nd is close to the opposite of how the stock traded. Anthropic's rumored chip is a rounding error. The real questions are structural and slower.
Whether the design competence Broadcom transfers each cycle eventually lets its biggest customers cut it out. Whether the "far limit" Goldberg describes keeps the club of viable in-house designers small enough that Broadcom stays essential.
And, underneath it all, whether Nvidia's 75% gross margin holds — because that margin, the thing every hyperscaler is spending billions to arbitrage away, is precisely the thing paying Broadcom's bills.
The competitor to fear isn't the one issuing a press release. It's the one signing your contract.