Anthropic's IPO Filing Crowns Broadcom, While Marvell's $115B Lumentum Deal Leaves Investors Nervous

The hook: an AI lab's S-1 just doubled as a Broadcom research report

Public investors rarely see the receipts on what a frontier AI lab spends its money on. Anthropic's S-1 changes that. An S-1 is the disclosure document a company files before going public. The biggest reveal in this one isn't about Claude. It's about Broadcom.

Anthropic has already signed five-year equipment lease agreements with Broadcom worth $161 billion. Those are signed deals, not a forecast. That matters because Broadcom is down about 33% from its early-June high in the recent AI-and-chip selloff. Selling a stock that hard is a bet that demand fades. This filing says a big slice of that demand is already under contract.

The broader point: if phase one of the AI buildout was "buy every GPU you can find," phase two may be custom chips built for one customer's workloads. Anthropic's filing is one of the strongest signs yet that this shift is happening.

Anthropic accounts for roughly 50% of the roughly 32 gigawatts of compute Broadcom said it expects to sell through the start of 2028.

What happened

Two stories landed within days of each other.

**Anthropic and Broadcom.** The S-1 shows the $161 billion lease commitment. Broadcom expects to sell about 32 gigawatts of compute through the start of 2028. Gigawatts measure the power all those chips draw, so think of it as a rough gauge of scale. Anthropic accounts for roughly half of it. The two also struck a new deal last week. And the Wall Street Journal reported that CEO Hock Tan personally courted Bank of America, Citigroup, Morgan Stanley and Blackstone to help finance a $63 billion chip sale between the companies. It's the largest AI debt deal ever.

**Marvell and Lumentum.** Marvell agreed to buy Lumentum, which makes optical components for data centers, in an all-stock deal valued at $115 billion. Lumentum holders get 4.5 Marvell shares for each share they own. The deal isn't expected to close until the first half of 2028. If it does, it would be the biggest merger ever between two publicly listed American tech companies. Day one went badly: Marvell fell 12.7% and Lumentum 18.6%. Both stocks are still well up for the year, with Marvell up 90% and Lumentum up 77%.

Why it matters: custom chips are real, and Broadcom's bet is concentrated

Broadcom builds what it calls XPUs: custom accelerator chips designed for a specific customer's tasks, rather than general-purpose GPUs. It already makes Google's tensor processing units and OpenAI's "Jalapeno" chip, and it works with Meta. Research firm Omdia expects custom chips to pass GPUs by volume as soon as 2028.

Broadcom expects $58 billion in AI chip revenue this year. Tan says that should roughly double next year and double again the year after. That works out to something north of $200 billion. Ambitious targets usually deserve a raised eyebrow, but $161 billion in signed Anthropic commitments makes this one a lot easier to believe.

The catch is concentration. When one customer is about half your expected compute sales, you're partly an Anthropic stock. The good news is that Anthropic looks strong. Noah Smith lays out the bull case in a paid essay. Anthropic now makes an operating profit. It brings in more than three times as much revenue per dollar of compute as it did a year ago. And the average Fortune 500 customer spends less than 0.1% of its expenses on Anthropic, which leaves plenty of room to grow. Smith also lists what could go wrong:

- Demand for AI takes longer to show up than expected. - The government restricts the company. - One player corners the supply of compute. - A price war with OpenAI, Google or Chinese models erodes margins.

Any of those would ripple straight into Broadcom's order book.

Then there's the debt. Banks plan to sell the $63 billion financing to bond investors next year, and the deal could be exposed to another round of interest-rate hikes. Borrowing doesn't make the deal bad. But it does mean the custom-chip boom partly depends on lenders staying willing. That's worth keeping in mind in a year when, per Bloomberg, the AI industry has already raised $300 billion in high-grade debt, meaning bonds from borrowers seen as relatively safe. That's roughly triple last year's total.

Marvell's bet: the plumbing, not the chips

Marvell isn't trying to win Anthropic's chip budget head-on. It already designs custom AI chips for Amazon and Microsoft. In July, Google secured the right to buy up to $12.2 billion of Marvell shares as part of a joint custom-chip deal. Marvell also bought optical startup Celestial AI for $5.5 billion in a deal that closed in February. Lumentum is a much bigger move in the same direction. The bet is that bigger AI chip clusters need better networking, and that light carries more data, faster, than copper wire. That holds no matter whose chips sit at either end.

Nobody publishes why they sold, but Morningstar's take is a good guide to the selloff. It's about price and certainty, not strategy. Morningstar says the deal "makes sense" and sees a bull case in cross-selling and shared costs. It projects Lumentum growing about 46% a year for five years. But it puts the odds of a completed deal below 50%. It also warns that Lumentum's results have been helped by tight supply and high prices, with cheaper Chinese rivals lurking. If that turns out to be a cyclical peak, Marvell "could be overpaying for a weaker asset."

The size matters too. Marvell's whole market value was recently around $150 billion, and it's paying with its own shares. Existing Marvell holders would own a noticeably smaller slice of the combined company. Lumentum holders have the opposite problem. Because they're paid in stock and the deal won't close until 2028, every Marvell dip between now and then shrinks what they actually get.

Timing hasn't helped either. The deal comes right after a jittery October that knocked the Nasdaq 100 down 6.7%, its worst month since March 2025.

The bigger picture and what to watch

**Anthropic's IPO price.** Reports say its target valuation could slip from $2 trillion to $1.5 trillion. A soft debut would be the first real warning sign for suppliers like Broadcom whose order books lean on it.

**The $63 billion bond sale next year.** Strong demand from bond buyers means lenders will keep funding the buildout. Weak demand or higher borrowing costs would show exactly where the strain is.

**OpenAI's expected IPO next year.** A second detailed filing would show whether Broadcom's custom-chip lead holds across the big AI labs or rests mostly on one giant customer.

**Whether Marvell–Lumentum survives to 2028.** Nvidia has a vote of confidence on the table. It invested $2 billion each in Marvell and Lumentum this spring, and Jensen Huang has called Marvell the "next $1 trillion company." But Morningstar's below-50% odds suggest that if you own either stock, you shouldn't treat the deal as done.