Broadcom

Broadcom Inc. (NASDAQ: AVGO) is a US semiconductor and infrastructure-software company headquartered in Palo Alto, California, and led by chief executive Hock Tan. The current company was formed in 2016 when Avago Technologies acquired the original Broadcom Corporation (founded 1991) and adopted the Broadcom name; it later acquired the enterprise-software firm VMware in 2023. As of 2025, roughly 58% of revenue came from semiconductor products and 42% from infrastructure software (Source: en.wikipedia.org).

Role in AI compute

Broadcom's relevance to AI policy comes from two semiconductor businesses that sit underneath the frontier-model build-out. First, it is the leading external design partner for hyperscaler custom AI accelerators — generically termed XPUs — co-developing application-specific silicon that lets cloud operators reduce dependence on Nvidia GPUs. It has a long-standing role co-designing Google's TPU line and works with Meta and other large customers. Second, its Ethernet switching and networking silicon (the Tomahawk and Jericho families) is widely used to build the high-bandwidth fabric that connects accelerators inside AI data centers.

Broadcom has guided to a substantial expansion of its AI revenue, with Hock Tan publicly framing a fiscal-2027 AI revenue target near $100 billion supported by a large committed customer backlog, and reporting custom-silicon and AI-networking growth through fiscal 2026 (Source: hashrateindex.com). These figures are company guidance and analyst estimates rather than audited results and are tracked here at medium confidence.

Fiscal third-quarter 2026 results and AI guidance

Broadcom reported fiscal third-quarter revenue of $29.59 billion on September 2, 2026, against an LSEG consensus of $29.36 billion and up 86% from $15.95 billion a year earlier, with semiconductor revenue more than tripling to $16.7 billion. Its fourth-quarter forecast of $34.8 billion fell below the $35.03 billion consensus (Source: cnbc.com).

On the earnings call Hock Tan said Broadcom expects Anthropic to deploy 5 gigawatts of Google TPU 8i chips in 2027, with line of sight to a further 10 gigawatts, and a 1.3 gigawatt deployment of the OpenAI-designed Jalapeño chip in 2027, with line of sight above 5 gigawatts across Jalapeño and its second generation. He said Broadcom is targeting AI revenue of $115 billion in fiscal 2027 and $230 billion in fiscal 2028 — above the roughly $100 billion fiscal-2027 figure recorded above from earlier guidance. Finance chief Amie Thuener said Broadcom may provide residual value guarantees, which are contingent liabilities, to "bridge the gap between their current cash flow and the significant upfront investments required" by customers (Source: cnbc.com). The deployment figures are Broadcom's expectations for its customers rather than announcements by those customers. See AI Bubble Debate, Circular Financing in AI.

OpenAI custom-silicon partnership

Broadcom is the design partner for OpenAI's custom-chip program. Reporting through late 2025 and 2026 described a multi-gigawatt custom-accelerator commitment between the two companies, and on June 24, 2026 OpenAI and Broadcom unveiled Jalapeño, described as OpenAI's first custom inference accelerator, co-developed from design to tape-out in nine months and built specifically for large-language-model inference rather than adapted from general-purpose AI silicon. Broadcom said the partners would deploy the chip at gigawatt scale with Microsoft and other partners beginning in 2026, with engineering samples already running workloads including GPT-5.3-Codex-Spark; OpenAI reported early performance-per-watt substantially better than the prevailing state of the art (Source: openai.com).

OpenAI presented first benchmark results at the Hot Chips conference on August 25, 2026 and published them the same day. Measured on SemiAnalysis's InferenceX benchmark across GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems, which the appendix identifies as Nvidia GB200 (GPT-OSS) and GB300 (DeepSeek and Kimi) parts. Every ratio is normalized by rated chip power — 700 W for Jalapeño against 1,200 W and 1,400 W for the Nvidia parts — rather than stated per chip (Jalapeño's first results show industry-leading speed and efficiency in AI inference). Richard Ho, OpenAI's head of hardware, put deployment at the end of 2026 "in very small volumes," with more significant deployment in 2027 — a more conservative schedule than the gigawatt-scale ramp beginning in 2026 described at the June announcement (Source: techcrunch.com). OpenAI's own post puts the start of deployment inside its compute infrastructure by the end of 2026, and describes Jalapeño as the first of a multigenerational platform with a second generation in development and a third taking shape (Jalapeño's first results show industry-leading speed and efficiency in AI inference).

Apple partnership

Broadcom disclosed in a July 6, 2026 securities filing that it and Apple extended their technology collaboration through 2031, with new multi-year agreements to develop and supply custom chips — including ASICs supporting Apple's AI servers planned for deployment as early as 2027, reported as Apple's first chips designed for AI servers; Broadcom shares rose 4% in pre-market trading on the disclosure (Source: bloomberg.com). Apple announced the deal on July 8, 2026 as worth more than $30 billion, covering custom silicon and wireless components designed and manufactured in the U.S.; Broadcom will invest $1.5 billion to expand its Fort Collins, Colorado plant to produce at least 15 billion chips, and Apple framed the arrangement as part of its $600 billion four-year U.S. investment commitment (Source: channelnewsasia.com; ca.finance.yahoo.com).

Samsung memory and foundry agreement

On July 24, 2026, Broadcom signed a memorandum of understanding with Samsung Electronics valued at more than $200 billion, covering HBM4 and HBM4E high-bandwidth-memory supply and 2-nanometer foundry manufacturing for Broadcom's next-generation AI chips through 2030. Subsequent reporting described the scope as spanning memory, foundry and advanced packaging on Samsung's 2-nanometer and below processes, and as including joint development of HBM4E and HBM5. The agreement was signed at an AI summit in San Francisco (Source: thestar.com.my; cnbc.com).

AI debt financing

Broadcom was in talks with a group of lenders as of August 20, 2026 to raise more than $60 billion in debt for an AI chip financing deal that reporting said would benefit Anthropic and other companies, according to people with knowledge of the matter. The financing was described as still being worked out and as potentially including a junior debt tranche of roughly $30 billion. Only the article lede was retrievable, so the structure, the counterparties beyond Anthropic, and the intended use of proceeds are not established (Source: bloomberg.com). See AI Infrastructure Capex.

Policy relevance

Broadcom is a central node in the shift toward custom, non-Nvidia AI silicon, which bears on the economics of inference, on AI-infrastructure capital expenditure, and on competition questions about concentration in the compute stack. Its networking and accelerator businesses are part of the AI compute supply chain that US export-control and industrial policy increasingly treat as strategically significant.

China

China's State-owned Assets Supervision and Administration Commission (SASAC), which oversees central state-owned enterprises, has been surveying the number of Broadcom switches installed in state-backed data centres as part of a push for domestic reliance, the Financial Times reported on September 22, 2026, citing people familiar with the matter. According to one of them, the survey found that the penetration rate of Broadcom switches among state-owned companies could be as high as 90 per cent, which could lead SASAC to issue informal guidance to reduce their use; Broadcom, Nvidia and Huawei sell high-end data-centre switches in China. The Financial Times article body was not available; these details come from its search excerpt and a Benzinga summary of it (Source: ft.com; tradingview.com). See US-China AI Competition: Different Races, Different Metrics.

Relationships

• related: Nvidia & TSMC — AI Compute Infrastructure, OpenAI, Google DeepMind, Meta AI, Microsoft, Inference Economics and Token Pricing, AI Infrastructure Capex, Semiconductor Supply Chain, AI Antitrust.