Broadcom and AMD both sell AI hardware to the same handful of large buyers, but they sell different things: Broadcom co-designs custom accelerators (which it calls XPUs) for individual customers and supplies the Ethernet switching and routing silicon that links them, while AMD sells its own Instinct GPUs, EPYC server processors and the ROCm software stack. In their latest results, Broadcom reported AI semiconductor revenue of $16.7 billion for the quarter to 2 August 2026, and AMD reported Data Center segment revenue of $8.5 billion for the quarter to 26 September 2026.
This article compares the two companies as suppliers, using their own filings and announcements. It is not investment advice and makes no share price forecasts.
What Broadcom sells for AI
Broadcom does not sell an off-the-shelf AI accelerator. Its main AI business is designing and supplying custom chips with a specific customer, which the customer then deploys in its own data centres. Google’s TPUs are the best-known example. Broadcom’s second AI line is networking: the chips that move data between thousands of accelerators.
Custom accelerators (XPUs)
On its fiscal third-quarter earnings call on 2 September 2026, chief executive Hock Tan referred to six XPU customers and named Google, Anthropic, OpenAI and Meta, according to a published transcript of the call. In that transcript he said XPUs accounted for 73% of AI revenue in the quarter. SDxCentral’s report on the call lists the customer detail Broadcom gave:
- Google: an expanded agreement covering future TPUs and AI networking equipment.
- Anthropic: a 1 GW Ironwood TPU project, 5 GW of TPU version 8i in fiscal 2027, and what Broadcom described as line of sight to a further 10 GW. Broadcom expects Anthropic to become its largest XPU customer in 2027.
- OpenAI: 1.3 GW of OpenAI’s “Jalapeño” chip planned for 2027.
- Meta: named as a customer for its MTIA accelerators.
The OpenAI work builds on a deal the two companies announced on 13 October 2025 to deploy 10 GW of OpenAI-designed accelerators. OpenAI leads the chip and system design; Broadcom co-develops and supplies the accelerators and the Ethernet networking for the racks. Deployment was scheduled to start in the second half of 2026 and finish by the end of 2029.
Ethernet networking: Tomahawk and Jericho
Broadcom’s AI networking chips fall into two families. Tomahawk switches connect accelerators within and across racks; Tomahawk 6 runs at 100 Tb/s, and Broadcom said in September 2026 that it had taped out Tomahawk 7, which SDxCentral reports as a 200 Tb/s part. Tomahawk Ultra targets scale-up links inside a rack over Ethernet.
Jericho is the routing family. Broadcom began shipping Jericho4 on 4 August 2025. It is built on a 3nm process, a single system scales to 36,000 ports at 3.2 Tb/s each, and it is designed to carry lossless RDMA traffic over distances of more than 100 km so that one AI cluster can span several data centres. Broadcom says the chip can link more than one million XPUs. Broadcom also sells optical components, including DSPs and lasers, used in the same networks.
Management said on the September call that it expects AI networking revenue to grow as fast as XPU revenue over the next few years.
What AMD sells for AI
AMD sells merchant hardware: GPUs and CPUs that any customer can buy, increasingly packaged as complete racks.
Instinct GPUs and Helios racks
The Instinct MI450 series is AMD’s current data centre GPU. It ships in Helios, a rack-scale system that pairs MI450 GPUs with 6th Gen EPYC “Venice” processors. In its third-quarter 2026 results on 3 November 2026, AMD said Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure were running models and workloads on the first Helios systems. HPE said Vultr would deploy the first HPE-built Helios systems. The EuroHPC Joint Undertaking chose Instinct MI430X GPUs and 6th Gen EPYC processors for the LUMI-AI supercomputer.
Two large multi-year deals underpin AMD’s GPU business:
- OpenAI: announced on 6 October 2025, covering 6 GW of AMD GPUs over several generations, starting with 1 GW of MI450 in the second half of 2026. AMD issued OpenAI a warrant for up to 160 million AMD shares, vesting in tranches tied to deployment milestones and AMD share price thresholds.
- Meta: announced on 24 February 2026, covering up to 6 GW of Instinct GPUs. The first gigawatt uses a custom GPU based on the MI450 architecture, with shipments expected to start in the second half of 2026, alongside EPYC “Venice” CPUs. Meta also received a performance-based warrant for up to 160 million shares.
The Meta arrangement is notable because it moves AMD towards customised silicon, the area where Broadcom is strongest.
EPYC CPUs and ROCm software
EPYC server processors sit in the same Data Center segment as Instinct GPUs and are used as host CPUs in AI systems as well as in general cloud servers. Oracle’s OCI E7 instances will use 6th Gen EPYC, and Google Cloud and Microsoft made 5th Gen EPYC storage-optimised instances generally available in the quarter.
ROCm is AMD’s open-source software stack for programming Instinct GPUs, and its main answer to Nvidia’s CUDA. AMD released ROCm 10, with a set of AI developer tools called ROCm.AI, during the third quarter. AMD also agreed to acquire World Labs, a model research team, and Taalas, which designs specialised inference silicon; it expects both deals to close by the end of 2026, subject to regulatory approval.
Latest reported results
The two companies use different fiscal calendars and report AI differently. Broadcom discloses a specific “AI semiconductor revenue” figure. AMD does not; its Data Center segment includes both Instinct GPUs and EPYC CPUs, so it is not a pure AI number.
Broadcom’s fiscal Q3 2026 results, published on 2 September 2026 for the quarter ended 2 August 2026, showed total revenue of $29.6 billion, up 86% year on year. AI semiconductor revenue was $16.7 billion, up 221% year on year and 54% on the previous quarter. Semiconductor solutions brought in $20.8 billion and infrastructure software (mainly VMware) $8.8 billion. Hock Tan said: “Demand for our custom AI accelerators and networking continues to be very strong.”
AMD’s Q3 2026 results, published on 3 November 2026 for the quarter ended 26 September 2026 (per its 8-K filing), showed record revenue of $13.4 billion, up 45% year on year. Data Center revenue was $8.5 billion, up 97%. Client and Gaming fell 8% to $3.7 billion, and Embedded rose 29% to $1.1 billion. Lisa Su, AMD’s chair and chief executive, said: “Demand for our Instinct and EPYC products continued to strengthen.”
Company guidance for the next quarter
- Broadcom, fiscal Q4 2026: revenue of about $34.8 billion, with AI semiconductor revenue of about $21.7 billion (up 236% year on year).
- AMD, Q4 2026: revenue of about $17 billion, plus or minus 3%, which AMD says is roughly 66% growth year on year at the midpoint. Su said she expects growth to accelerate “as we scale Helios and ‘Venice’ production”.
Looking further out, Hock Tan said on the September call that Broadcom had “secured the supply to again double AI revenue to approximately $115 billion” in fiscal 2027, according to the call transcript. That is a company target, not a reported result.
Broadcom and AMD side by side
Figures are as of 4 November 2026 and come from each company’s most recent quarterly release.
| Broadcom | AMD | |
|---|---|---|
| Main AI products | Custom XPUs co-designed with customers; Tomahawk switches; Jericho routers; optical components | Instinct MI450 GPUs; Helios racks; EPYC CPUs; ROCm software |
| Business model | Custom silicon built for a small number of large customers | Merchant chips sold to many buyers, plus a custom MI450-based GPU for Meta |
| Named AI customers | Google, Anthropic, OpenAI, Meta | OpenAI, Meta, Anthropic, Microsoft, Oracle Cloud Infrastructure |
| Headline deals | OpenAI: 10 GW of custom accelerators, 2026 to 2029 | OpenAI: 6 GW; Meta: up to 6 GW (both with share warrants) |
| Latest quarter | Fiscal Q3 2026, ended 2 August 2026 | Q3 2026, ended 26 September 2026 |
| Total revenue | $29.6bn (up 86%) | $13.4bn (up 45%) |
| AI-related revenue | $16.7bn AI semiconductor revenue (up 221%) | $8.5bn Data Center segment, GPUs and CPUs combined (up 97%) |
| Next-quarter guidance | About $34.8bn total; about $21.7bn AI | About $17bn total, plus or minus 3% |
| Non-AI business | VMware infrastructure software ($8.8bn); non-AI chips | PC and gaming processors; embedded chips |
Where they compete and where they do not
The companies overlap most at the customer level. OpenAI and Meta buy from both: OpenAI is deploying AMD Instinct GPUs and its own Broadcom-built accelerators, and Meta is taking AMD GPUs while working with Broadcom on MTIA. Anthropic runs work on AMD Helios systems and on Google TPUs made with Broadcom.
Product overlap is narrower. AMD does not sell Ethernet switch silicon on Broadcom’s scale, and Broadcom does not sell a general-purpose GPU. The direct contest is over the accelerator itself: whether a large buyer chooses a merchant GPU such as MI450 or designs its own chip with a partner such as Broadcom. AMD’s custom GPU for Meta shows it is now competing for some of that custom work too.
More on TechAnnouncer: Marvell agrees to buy Lumentum in an all-stock AI optics deal and Ayar Labs and optical I/O chiplets.
Sources
- Broadcom: third quarter fiscal year 2026 financial results (SEC filing, 2 September 2026)
- AMD: third quarter 2026 financial results (3 November 2026)
- AMD: Form 8-K for Q3 2026 (SEC filing)
- OpenAI: OpenAI and Broadcom announce strategic collaboration (13 October 2025)
- AMD: AMD and OpenAI strategic partnership (6 October 2025)
- AMD: AMD and Meta expanded strategic partnership (24 February 2026)
- Broadcom: Broadcom ships Jericho4 (4 August 2025)
- SDxCentral: report on Broadcom’s fiscal Q3 2026 results
- Broadcom Q3 2026 earnings call transcript (Webull)


