Broadcom closed the third quarter of fiscal 2026 with $29.59 billion in revenue, up 85.5 percent year-on-year and up 33.4 percent sequentially. Chief executive Hock Tan used the call with Wall Street to sketch a roadmap that dwarfs the chipmaker's current run-rate.

Tan projected Broadcom's AI-related revenue — custom XPUs, networking ASICs, and the auxiliary silicon that rides with them — would scale from $2.6 billion in fiscal 2022 to $230 billion by fiscal 2028. Two orders of magnitude in six fiscal years. For scale: the United States spent $1 trillion on its military last year.

Tan told analysts not to expect quarterly updates to that figure. The forecast's job, he signaled, was simple. Settle nerves about a GenAI bubble. Big money was rolling in.

Who's buying the silicon?

Six XPU customers, with more on the way. Google, Anthropic, OpenAI, and Meta Platforms are confirmed. ByteDance or Apple probably round out the roster. Amazon Web Services uses Broadcom for packaging and integration on its Trainium 4 silicon.

Across those accounts, total compute capacity runs about 30 gigawatts. The math pencils out at roughly $12 billion per gigawatt for Broadcom-shepherded XPU installations, or $15 billion per gigawatt once CPUs, memory, and racks are added.

Nvidia is the benchmark. Grace-Hopper rackscale systems cost around $18 billion per gigawatt. Vera-Rubin systems land near $40 billion per gigawatt. Tan's pitch to the hyperscalers: custom silicon through Broadcom undercuts Nvidia on price and gives customers something their rivals cannot copy.

Anthropic overtakes Google

Anthropic has emerged as Broadcom's surprise volume leader. The AI lab is deploying 1 gigawatt of "Ironwood" TPU 7 accelerators this year. It will roll out 5 gigawatts of TPU 8i systems in 2027. It wants 10 gigawatts of the follow-on TPU 9 in 2028.

Tan said Anthropic would be Broadcom's largest XPU customer in 2027 and would keep that slot through 2028. That means Anthropic is buying more Google-designed TPU iron than Google itself.

Google still matters. Tan said Broadcom would ship "multi-tens of billions of dollars of TPUs annually over the next several years" to the search giant. Then he cut to the chase on competitive moats: "We have consistently delivered the fastest time to market for TPUs from product definition to production without the need for resubmits. We believe these are very deep moats for any competitor to cross."

Forget Marvell, MediaTek, and Alchip poaching TPU work, Tan implied. It's all about who gets the HBM allocations.

OpenAI, Meta, and the networking wave

OpenAI is in. Its "Jalapeno" XPU is on track for 1.3 gigawatts of deployment in 2027. Broadcom expects another 5 gigawatts in 2028 once the follow-on chip ramps. That trajectory pushes OpenAI past Google into the number-two XPU slot.

The Jalapeno follow-on is approaching tapeout, Tan said. The two companies are already working on a third generation.

Meta has committed to three MTIA generations between now and the end of fiscal 2027. Across those accelerators, Broadcom has line of sight on 3 gigawatts through fiscal 2028.

The networking side carries its own momentum. The 100 Gb/s and 200 Gb/s SerDes flavors of the Tomahawk 6 Ethernet switch ASIC are selling fast. Every hyperscaler using Broadcom-shepherded XPUs runs Tomahawk 6 for scale-out. Tomahawk Ultra — the latency-trimming variant that pulls Ethernet into scale-up coherent memory roles inside rackscale machines — is shipping now. The ramp runs through fiscal 2027.

Behind it sits Tomahawk 7. The 204.8 Tb/s device has not been formally unveiled, but it uses 400 Gb/s SerDes and delivers 3.2 Tb/s per port across eight lanes. It taped out recently. Switches carrying it should surface in late calendar 2027 or early calendar 2028.

What's underneath

Taiwan Semiconductor Manufacturing Co runs the foundry and packaging. Broadcom has locked in HBM memory, wafer capacity, and advanced packaging — including the 3.5D Extreme Dimension System in Package (XDSiP) chiplet stacking — for the forecasted revenue stream. Tan said Broadcom can dial up more if demand runs hotter.

The Q3 numbers themselves told the story. Semiconductor Solutions revenue hit $20.84 billion, up 127.4 percent, with $12.71 billion in operating income. AI-related sales — $12.22 billion for compute, $4.52 billion for networking — totaled $16.74 billion, up 3.22x year-on-year and up 54 percent sequentially. Infrastructure Software, anchored by VMware, contributed $8.75 billion.

Operating income for the quarter hit $15.96 billion, up 2.71x. Net income reached $13.09 billion, up 3.17x. That is a 44.2 percent net margin.

Broadcom ended the quarter with $23.98 billion in cash and $59.42 billion in debt, most of it the leftover VMware bill.

Tan has squeezed more than money out of the businesses he buys. He has squeezed competitors out of the hyperscaler XPU roadmap. If Anthropic, OpenAI, and Meta hit the gigawatt targets Tan laid out, Broadcom will own the bulk of non-Nvidia AI silicon by fiscal 2028 — and the per-gigawatt economics will keep steering the next round of design wins.