Broadcom: Custom AI Chips Set to Exceed $100 Billion by 2027 as Hyperscalers Escape Nvidia's "General Compute Tax"
RAISE Summit 2026, Paris — Charlie Kawwas, President of Semiconductor Solutions Group, discusses Broadcom's custom silicon roadmap
Broadcom's custom AI accelerator business is scaling at a pace that even its own leadership frames in superlatives. Charlie Kawwas, who runs the company's semiconductor solutions group, told an audience at RAISE Summit 2026 that the XPU business will exceed $100 billion in revenue within roughly five years of starting from zero, with the company now supplying four of the five largest AI labs in the world. "Zero to 100 billion in five years... remember 100 billion is not good enough. More than 100," Kawwas said, correcting the moderator's framing in real time.
The Economics Driving Hyperscalers Away from Merchant Silicon
The core thesis Kawwas laid out is that Nvidia's general-purpose GPU platform, while technically excellent, imposes what he called a double tax on the largest AI spenders: an efficiency tax from buying hardware built for every workload rather than a specific one, and a margin tax from Nvidia's pricing power. At spending levels of $150 billion to $200 billion annually, which he said now describes each of the top four US hyperscalers, that tax translates into $100 billion to $150 billion per player that could otherwise have gone toward compute output, or in his words, "tokens." That math, more than any philosophical preference for vertical integration, is what is pushing OpenAI, Anthropic, Google, Meta and xAI toward custom silicon. Kawwas was explicit that this is not a rejection of Nvidia's technology. "I'd love to use the Nvidia chip, but I got a general compute tax and then I got a margin tax," he said, paraphrasing customer conversations. The pitch for custom silicon is straightforward: two to three times the compute per dollar, which drives token costs down materially.
A Physical Roadmap: From One Chip to a "Beast"
The most concrete new disclosure came via a literal show-and-tell. Kawwas passed around Broadcom's XPU generations from 2025 through the 2027 design, illustrating the pace of scaling in physical terms. Year over year, memory capacity roughly tripled and compute density roughly tripled again, and over the two-year span from the 2025 part to the 2027 part, memory capacity increased nearly 14-fold. The 2027 chip, which Kawwas nicknamed "the beast," stacks compute dies face-to-face in a double-decker configuration, delivering 16 times the density of the chip Broadcom shipped a year prior, of which the company had already deployed a million units. Four hyperscalers are being built versions of this platform, according to Kawwas. In gigawatt terms, the scaling is similarly steep. One customer is deploying about 1.5 gigawatts of the current-generation chip this year, already contracted for over 5 gigawatts next year, implying a jump to roughly 6.5 gigawatts of cumulative XPU capacity with that single customer within two years.
Why Broadcom, Not a Merchant Vendor, Can Do This
Kawwas attributed Broadcom's ability to match Nvidia's now-annual product cadence, itself accelerated from a two-to-three-year cycle, to a modular, pre-validated chiplet architecture. Memory blocks, Ethernet-based I/O, and compute chiplets are engineered and production-qualified independently, in partnership with foundry and memory partners including Samsung, before a customer engagement even begins. "So when you come and talk to me, it becomes an integration exercise," he said. That approach let Broadcom and OpenAI complete the Jalapeno chip in under nine months, which Kawwas described as the fastest chip development cycle the company has executed. He extended the France-based couture metaphor the moderator opened with, calling Broadcom's approach a design house that removes the general-compute tax: "You come to us, you tell us what kind of tuxedo or dress you want. We will customize it and build it for you." The joke aside, the substantive point is that Broadcom's model depends on co-design not just at the chip level but at the full rack level, including scale-out and now scale-up networking based on open Ethernet standards, an area Kawwas flagged as the "next battle" now that scale-out networking has become open industry-wide, including at Nvidia.
An Emerging Model for Open-Source Inference
Asked whether custom silicon has any relevance to open-source model developers rather than closed frontier labs, Kawwas pointed to Broadcom's multi-generation partnership with SambaNova as a live proof point. SambaNova runs closed, state-of-the-art models on its Broadcom-co-designed XPU for premium use cases, while routing lower-stakes workloads to open-source models such as MiniMax or DeepSeek on the same silicon, at effectively zero marginal token cost. Kawwas said he expects this hybrid open/closed inference pattern to spread across more XPU deployments going forward, a signal that custom silicon economics may benefit open-source model adoption independent of frontier lab spending.
The AGI Framing and What Comes Next
Kawwas tied the competitive stakes to the race for advanced AI capability directly, arguing that the first company to deliver a fully integrated GPU, XPU and CPU platform "potentially can be the first to reach AGI," and that 2028 should bring materially more capable domain-specific agents as this hardware becomes available. Whether that framing proves accurate or is standard industry salesmanship is something investors will need to judge for themselves, but it underscores how Broadcom is positioning its custom silicon roadmap not merely as a cost play but as critical-path infrastructure for the next phase of model capability. On near-term visibility, Kawwas confirmed more custom XPU designs are in development beyond Jalapeno, both from the largest labs and additional startups, alongside gigawatt-scale demand that he described as "insatiable."