Bloomberg Live: Broadcom's Hock Tan Says M&A Is Dead to Him — and OpenAI Chips Are On Track for Production by Year-End
Bloomberg Tech 2026, San Francisco — June 5, 2026
Hock Tan, Broadcom's President and CEO, sat down with Bloomberg's Tom Giles at Bloomberg Tech 2026 in San Francisco for one of the more candid CEO conversations of the year. Two disclosures stood out: Broadcom has all but suspended its historical M&A playbook in favor of riding the AI wave organically, and the OpenAI custom silicon partnership — which some press reports had suggested was hitting snags — is on track to enter production before year-end.
The M&A Machine Goes Quiet
For an acquirer that built its empire deal by deal, Tan's remarks on M&A were striking in their directness. "In the last two years, between '24 and '26, I would be doubling my revenues. I would create over $50 billion per year annualized in revenue. I'm looking around — what can I buy that even comes close to that?" The math, he argued, simply doesn't work. Any acquisition brings regulatory friction, integration distraction, and a year or more of management bandwidth consumed — all while organic demand for AI compute remains, in his words, "almost insatiable." When pressed on whether any area, such as photonics or optics, might tempt him to make an exception, Tan deflected: "I ran a business model for 20 years. I try very hard to avoid bright, shiny objects."
OpenAI Silicon: Denying the Snags, Confirming the Timeline
Asked specifically about a report suggesting the OpenAI chip partnership required Microsoft's sign-off on purchase commitments before moving forward, Tan was unequivocal. "No, not at all." He confirmed the custom AI accelerator silicon is "working very well in their labs, in the data centres," and that Broadcom is "on track to go into production late this year." The broader context matters here: Tan revealed Broadcom now has exactly six hyperscaler customers on the custom silicon journey, each at a different stage. Google is the furthest along, having started earliest, while OpenAI — engaged for over two years — is among the newer entrants. "It's pretty remarkable how much we have achieved," he said, framing each relationship as a generational progression, with first-generation chips giving way to increasingly optimized second and third iterations over time.
Google's In-House Push Is Competitive, Not Existential
Tan addressed the question of Google pursuing customer-owned tooling — effectively designing more of its own chip stack in-house, potentially with smaller partners like Marvell — without any sign of alarm. He described those smaller partners as providing "angled guidance" to Google in its effort, and framed the dynamic straightforwardly: "We compete against our own customer in a different part of it." His confidence rests on engineering depth. Broadcom operates 17 semiconductor product divisions, each aiming to be number one in its specific domain, and Tan's view is that the real forcing function keeping Google investing in ever-better custom silicon is Nvidia. "As long as Nvidia keeps coming generation after generation with superb technology, Google has to create their equivalent technology to match that. And that's where we come in."
The Anthropic Bet Was a Leap of Faith — One That Paid Off
Tan was notably open about the uncertainty that surrounded Broadcom and Google's joint decision, roughly a year ago, to supply TPU-based compute capacity to Anthropic. "We were making a leap of faith that Anthropic and generative AI, and Anthropic through its business model of addressing enterprises with coding tools, would make a difference." He acknowledged the bet has paid off, pointing to Anthropic's surging valuation and its confidential IPO filing as validation. But he was also candid about how rapidly the picture is shifting: "I know I'm probably a fraction of what I know today, and what I know today I'll probably be a fraction of what I'll know in six months time."
AI Productivity at Broadcom: No Token Throttling Yet
On internal AI adoption, Tan offered a concrete and somewhat unexpected illustration of the productivity calculus driving enterprise AI demand. Broadcom uses tools including Anthropic's Claude — he referenced Opus 4 specifically — for engineering design and code assistance. His framing of the return on investment was blunt: "You can get one very senior engineer to produce an application design in one week, what it would otherwise take ten engineers, each paid $300,000 a year, three months to produce the same thing." Broadcom has not imposed token usage limits on its engineers, though Tan acknowledged that throttling decisions ultimately come down to return on investment by application. His broader read on enterprise AI adoption is that cost-sensitivity around tokens, while real and discussed among CEOs, reflects a market that is still in "early innings." The efficiency of Chinese large language models such as DeepSeek was raised, but Tan said he does not see it materially changing his own usage patterns or those of his engineers at present.
Networking Overflow Goes to Cisco — Gladly
On the networking side of the business, where AI cluster infrastructure demand has made Broadcom's switching products increasingly critical, Tan was relaxed about Cisco's ambitions to compete more aggressively. Broadcom, he noted, is selective about which customers it prioritizes — those with durable, multi-generational demand for its products. Overflow, by definition, goes elsewhere. "I'm happy for you to go to Cisco," he said, with the confidence of someone whose order book does not need filling.