The purchase order, if it converts into a formal contract, would rank among the largest in the history of chip manufacturing. Samsung Electronics has won an AI chip foundry deal with Broadcom valued at about $200 billion, the biggest single order Samsung Foundry has ever booked, according to a Reuters report on July 25.
The deal hands Samsung the kind of customer it has chased for years. Broadcom designs custom AI accelerators for some of the largest data center operators in the world, and its silicon — not Nvidia’s off-the-shelf GPUs — sits inside a growing share of the industry’s most ambitious builds. Winning that work puts Samsung directly against Taiwan Semiconductor Manufacturing Co., which has dominated the contract manufacturing business for a decade.
Samsung’s pitch has been capacity and geometry. The company said it has been improving yields on its most advanced nodes — processes at 3 nanometers and below — and is expanding output at its foundry fabs. People familiar with the negotiations said the Broadcom deal covers multiple generations of custom AI chips, with production spread across Samsung’s advanced lines.
The scale of the order reflects the explosive growth of custom silicon. Hyperscalers are designing their own accelerators to cut cost and power consumption, and Broadcom has become the chip architect of record for much of that effort. The company’s XPU business, which designs AI chips for customers such as Google and Meta Platforms, has been the fastest-growing part of its portfolio.
Reuters reported that the agreement directly challenges TSMC’s dominance in AI chip manufacturing. TSMC controls roughly 60 percent of the global foundry market and nearly all of the most advanced AI silicon, a position built on years of flawless execution and a customer list that includes Nvidia, Apple and AMD. Samsung has struggled to dent that share despite massive investment in its foundry division.
The problem has been execution, not ambition. Samsung’s advanced nodes have suffered from yield problems that pushed early customers to defect, and the company lost its most prominent AI chip customer — Nvidia — to TSMC years ago. The Broadcom order suggests that calculus is changing, at least for one customer willing to bet on Samsung’s next-generation process.
Analysts cautioned that a deal of this size carries execution risk. Broadcom’s custom chips are complex, multi-die designs that require flawless manufacturing discipline. If Samsung’s yields on its 2-nanometer-class process fall short, the economics of the order could shift, and Broadcom has alternates — TSMC’s own 2nm line remains an option for the same designs.
The deal also reflects pricing pressure in the foundry business. TSMC has raised prices for its advanced nodes as demand outstrips supply, and large customers have been eager for a credible second source. Samsung has signaled it is willing to price aggressively to win volume, a strategy that could compress margins in a business already known for thin returns on capital.
For Samsung, the order is a vindication of years of investment. The company’s foundry arm has spent tens of billions of dollars building capacity and chasing leading-edge yields, often at a loss, while shareholders questioned the returns. A $200 billion pipeline, even spread over several years, changes the arithmetic.
The broader implications reach beyond Samsung and Broadcom. The deal gives the AI supply chain something it has lacked: redundancy. Every major AI chip currently flows through TSMC’s fabs in Taiwan, a concentration that has worried governments and customers alike. A second viable foundry for advanced AI silicon reshapes the risk profile of the entire industry.
The timing matters too. AI infrastructure spending shows no sign of slowing, and the largest cloud operators have committed to multiyear build-outs. Custom accelerators are expected to take an increasing share of that spending, which means the foundry that wins the custom-chip business will share in the industry’s most reliable growth.
People familiar with the matter said the Samsung-Broadcom agreement was structured in phases, with early production focused on chips for existing Broadcom customers and later phases dedicated to new designs. Neither company has confirmed the financial terms publicly, and analysts said final pricing will depend on yields and market conditions at each stage of production.
The competitive response is already forming. TSMC has moved quickly to lock in its own custom-chip customers, and its advanced packaging capacity — the technology that stitches multiple chips into a single accelerator — remains the industry’s scarcest resource. Samsung has invested in its own advanced packaging, but analysts said it trails TSMC there by years.
Industry executives said the deal could also pressure Nvidia, which has relied on TSMC for its highest-end GPUs. If Broadcom’s custom chips gain traction at Samsung’s fabs, the cost and performance gap between custom silicon and off-the-shelf GPUs could narrow, giving data center operators another reason to design their own hardware.
The order is a test for Samsung in the most literal sense. The company’s advanced nodes have improved, according to people who have audited its facilities, but they have not been proven at the volume and complexity the Broadcom deal demands. Success would make Samsung the second pole of the foundry world. Failure would push Broadcom back toward TSMC and leave Samsung’s foundry ambitions where they have sat for a decade — promising, but unfulfilled.
