Broadcom Is Borrowing $100B for AI Chips. Samsung Foundry Is the Play

The Money Is the Story

Broadcom is raising up to $100 billion in debt. Not because it’s struggling. Because it can’t get chips fast enough to sell.

The company disclosed on August 25 that it’s negotiating a financing package split between roughly $30 billion in subordinated debt and $60 billion to $70 billion in senior secured debt, with Blackstone and Apollo Global Management leading the syndicate. Some of the senior tranches may even carry Broadcom’s own guarantee. The structure — issued through a special-purpose vehicle — is designed to keep the debt off Broadcom’s balance sheet while funding what amounts to a capital-intensive arms race for AI infrastructure.

The logic is stark. Broadcom’s second-quarter AI semiconductor revenue hit $10.8 billion, up 143 percent year over year. Its third-quarter guidance calls for $16 billion, implying 200 percent growth. Management says the full AI chip backlog of $73 billion — part of a total order book of $162 billion — will be consumed within six quarters. That’s over $12 billion per quarter in AI-related revenue alone.

Even that pace may not be fast enough. Broadcom’s CEO has said explicitly that organic cash flow cannot keep up with the speed of customer orders. The company isn’t just selling chips anymore. It’s building out entire data-center infrastructures — custom semiconductors, rack systems, memory, packaging — for clients like Google, Anthropic, and OpenAI.

TSMC Hit a Wall

The financing move lands against a backdrop that matters more than any earnings guide: TSMC’s capacity is no longer a bottomless well.

Broadcom’s own CTO told Reuters in March that TSMC’s production capacity had effectively maxed out. That was not a complaint. It was a factual statement about the architecture of the global semiconductor supply chain. TSMC had been the single point of fulfillment for Broadcom’s most advanced AI silicon, and now it can’t absorb everything.

The math is simple. If Broadcom has $162 billion in orders and a major portion of those chips require leading-edge nodes that only TSMC can reliably produce at volume, then capacity allocation becomes the binding constraint — not design, not demand, not pricing power.

That constraint is exactly where Samsung’s moment opens up.

The $20 Billion Memorandum

In July, Samsung and Broadcom signed a $20 billion memorandum of understanding covering memory, foundry, and advanced packaging. The deal calls for Broadcom to use Samsung’s 2-nanometer and below process for its next-generation wide-bandwidth communication semiconductors, and to source HBM4 and HBM4E memory from Samsung as well.

HBM — high bandwidth memory — is now as critical as the logic chips themselves. Every AI accelerator runs into the same wall: compute outpaces memory. HBM4 and HBM4E are Samsung’s answer to that bottleneck, and they’re also a product Samsung has struggled to commercialize at scale. Securing Broadcom as a anchor customer changes the revenue equation dramatically.

Samsung’s own October earnings call referenced discussions with Broadcom and other major clients about next-generation projects, and projected that 2-nanometer orders would more than double compared with the prior year. That projection now looks conservative if Broadcom’s $100 billion borrowing drive translates directly into foundry commitments at Samsung.

Who Wins, Who Loses

Samsung wins if it can move from memorandum to mass production quickly enough. The 2-nanometer process has beenSamsung’s stated ambition for years, but yield rates and volume delivery have lagged behind TSMC’s promises. Broadcom’s desperation — and willingness to borrow aggressively to secure supply — is Samsung’s leverage.

TSMC loses market share in the near term, not because its technology is inferior, but because its capacity ceiling forces customers like Broadcom to diversify. That diversification, once established, tends to persist. Even after TSMC expands, Broadcom will already have qualified Samsung lines and established HBM supply chains.

Google wins by locking in a custom semiconductor supply through 2031. Anthropic gets committed compute capacity starting in 2027. Both companies are essentially buying insurance against the next round of AI capability leaps that require more silicon than any single foundry can provide.

The market loses nothing directly, but the structural shift is real: the illusion of a single dominant foundry is cracking. The AI boom demanded so much leading-edge capacity so fast that the ecosystem had no choice but to split its bets.

What Happens Next

The critical question is timing. Samsung has publicly admitted that its 2-nanometer ramp is behind schedule. Broadcom’s backlog will not wait forever. If Samsung cannot deliver volume before the next generation of AI accelerators hits the design phase, those orders will drift toward alternative suppliers — or toward TSMC’s next capacity expansion, which is still years away from full delivery.

The $100 billion debt raises another question: what happens if AI demand softens faster than expected? Broadcom’s financing structure spreads risk across investors and special-purpose vehicles, which limits direct exposure to the company itself. But the broader market still bears the risk of a sector-wide demand correction funded by unprecedented corporate leverage.

For now, the signal is clear. Broadcom is borrowing like a company that expects demand to keep accelerating. Samsung is positioned to capture the overflow. Whether it can convert that position into sustained foundry revenue at the leading edge — rather than one-off deals — will determine if this is a temporary supply-chain adjustment or a structural realignment of the semiconductor industry.