Broadcom AI Chips: Google TPUs, OpenAI Jalapeño, Benchmarks

What Broadcom sells

Broadcom is a semiconductor and infrastructure software company led by president and chief executive Hock Tan. Its chip business spans Wi-Fi, Bluetooth and networking silicon; its software business runs mainframes and virtualized data centers, expanded substantially by its $69 billion acquisition of VMware, completed in November 2023. In fiscal 2025 the two segments split roughly 58 percent semiconductors and 42 percent infrastructure software, and the company crossed a $2 trillion market capitalization in April 2026, the sixth company in history to reach that valuation. [1][3]

The custom chip business

Within semiconductors, Broadcom's fastest-growing line is custom AI accelerators, which it calls XPUs: chips designed collaboratively with a specific customer for that customer's own workloads, rather than sold as a general-purpose product like an NVIDIA GPU. Broadcom has co-developed every generation of Google's Tensor Processing Unit since the line began, and it has worked with Meta on the Training and Inference Accelerator chips Meta first introduced in 2023. It also supplies networking silicon built for AI data centers, including the Tomahawk switch family, whose fifth generation reaches 51.2 terabits per second. [3]

The scale of that business shows up in Broadcom's own numbers. AI semiconductor revenue reached $6.5 billion in the fourth quarter of fiscal 2025, up 74 percent year over year, and the company guided to roughly double that, $8.2 billion, for the first quarter of fiscal 2026. Chief executive Hock Tan attributed the growth to "custom AI accelerators and Ethernet AI switches." [4]

Why it matters to AI

Atlas interpretation: Broadcom sits inside the argument, made loudest by the big cloud providers and now by frontier labs, that NVIDIA's GPU dominance is a cost problem worth designing around. A custom chip built for one company's specific models can be cheaper to run at scale than a general-purpose GPU, if that company has the volume and the engineering budget to design and use one, and Broadcom is the firm most of them have gone to for that design and manufacturing partnership. That makes Broadcom's order backlog, rather than any single chip, the clearer signal of how many large AI buyers are betting on custom silicon over buying more NVIDIA hardware. [3][4]

Broadcom on the timeline

Broadcom's most direct appearance on the timeline is as OpenAI's development partner on Jalapeño, OpenAI's first custom chip, first announced in October 2025 and developed, in OpenAI's own words, "in close collaboration with Broadcom." Neither company's public materials break down which parts of the design, verification or manufacturing process Broadcom specifically contributed, beyond that general collaboration. Jalapeño serves inference only and is not for sale outside OpenAI's own infrastructure. [2][5]

OpenAI presented Jalapeño's first benchmark results at the Hot Chips conference on August 25, 2026, reporting 1.5 to 1.9 times more throughput per watt and 1.7 to 3.6 times lower latency than comparison NVIDIA GB200 and GB300 systems on three open-weight models, using a public SemiAnalysis benchmark called InferenceX. The comparison normalizes by each chip's rated power draw rather than measured power, and OpenAI ran both sides of the test itself, which the event's own page treats as a real but bounded result rather than an independently verified one. [2]

Atlas interpretation: For Broadcom, the more durable fact than any single benchmark is the pattern it fits: OpenAI joining Google and Meta as a hyperscale customer designing a custom accelerator with Broadcom's help. Jalapeño will not be sold, so it does not compete with Broadcom's other chip customers directly, but it extends the same argument those customers are making, that on sufficiently large and specific workloads a co-designed chip can beat a general-purpose one on cost or efficiency, which is the argument underneath Broadcom's entire XPU backlog. [3][2]