Broadcom AI Custom Silicon, September 2026: Record Revenue and the $12.9 Billion Nvidia Signal

Broadcom's AI semiconductor revenue tripled to $16.7 billion in Q3 FY2026 and its co-developed Jalapeño ASIC claimed 1.9x throughput-per-watt over Nvidia's flagship GPU, prompting unconfirmed reports that Nvidia's $12.9 billion Hugging Face acquisition was in part designed to pre-empt a competing Broadcom bid.

The clearest sign of how far Broadcom has moved in the AI silicon contest arrived on September 7, when two specialist outlets reported that Nvidia had paid $12.9 billion to acquire a stake in Hugging Face — reportedly to pre-empt a competing bid from Broadcom. Neither company has confirmed that framing, and the sourcing falls short of a first-hand disclosure, so the claim warrants attribution rather than assertion. But the logic is not implausible. Broadcom has spent two years repositioning itself as a full-stack AI silicon partner, and Hugging Face's open-source model ecosystem and inference tooling would offer any custom ASIC vendor a meaningful software foothold to reduce dependence on Nvidia's proprietary stack. Even if the competitive detail is imprecise, the episode illustrates where Broadcom's threat is now perceived to reside: not just in networking hardware, but in the model-developer relationships that determine who designs the next generation of inference accelerators.

The Q3 FY2026 results, filed with the SEC via 8-K on September 6 and distributed by PR Newswire, gave that repositioning its clearest numerical foundation. Revenue reached $29.6 billion for the quarter, up 86% year over year, with AI semiconductor revenue tripling to $16.7 billion — a figure that approaches Nvidia's core data-center run rate from less than two years prior. The quarter surpassed analyst consensus on both top and bottom lines, extending what Mizuho characterized before the print as a 30-consecutive-quarter growth streak. At Hot Chips 2026 the same week, Broadcom disclosed performance specifications for its Thor Ultra 800GbE Ethernet NIC, asserting its position in the high-bandwidth cluster-fabric market that increasingly determines AI infrastructure unit economics. Analyst projections published on Stocktwits target $230 billion in AI-related revenue by 2028, driven by demand from OpenAI and Anthropic — a figure Broadcom itself has not officially confirmed but has not publicly disputed.

The sharpest competitive statement came from the Jalapeño ASIC, co-developed with OpenAI and benchmarked publicly at Hot Chips 2026. At 700 watts, Jalapeño reportedly achieves 1.9x throughput per kilowatt and 3.6x lower latency relative to Nvidia's 1,400-watt GB300 flagship — a performance-per-watt framing that resonates directly with hyperscalers facing power-constrained data-center expansion. OpenAI confirmed the chip is performing to expectations, according to CNBC. Broadcom's structural role in that project — silicon co-designer and production partner, not model developer — is what makes the relationship strategically durable: it allows frontier labs to reduce GPU dependence without assuming full chip-architecture risk. Jefferies and Mizuho cited this partnership model in their pre-earnings bullish notes; ARK Invest has reinforced its position on Broadcom's custom silicon thesis, while RBC Capital Markets has flagged execution risk as the principal counter-argument. The divergence between those two camps reflects a genuine uncertainty: whether Broadcom can repeat the OpenAI co-development model across enough additional clients to underwrite the long-range revenue projections.

Broadcom's capital structure sits in deliberate contrast to the scale of its ambitions. Slicast's own compilation of SEC XBRL filings records capital expenditure of $623 million in FY2025 — approximately 1% of the $63.89 billion revenue base recorded that fiscal year, the lowest capex intensity among the twelve chip peers in the tracked group, up only 13.7% from FY2024's $548 million. The fabless model, built on TSMC capacity, preserves margins and limits cyclical exposure, but it also means supply ceilings are set by third-party foundry allocation rather than owned fabs. Ad-hoc-news.de reported a $60 billion debt financing program linked to the Google custom silicon alliance; if accurate, leverage implications merit scrutiny given that Broadcom's credit risk metrics are already at record levels, as noted by one credit-focused analyst. The enterprise dimension has also expanded: the VMware-based private AI platform launched in early September targets corporations running inference on owned infrastructure — a segment less exposed to public cloud spending cycles and one that the Google, Amazon, and Microsoft hyperscalers do not serve by design.

The risks several analysts have named are specific rather than theoretical. Gross margin compression from the growing TPU product mix — Google's tensor processors carry lower margins than other custom ASICs in the portfolio — was identified by Trefis as a structural headwind even as aggregate profitability has held. Google's decision to deepen its Marvell partnership for custom chip development introduces a credible second supplier and constrains Broadcom's pricing latitude with its largest single customer. Customer concentration remains the foundational concern: the $230 billion revenue projection through 2028 rests on sustained, accelerating spend from a small number of hyperscalers and frontier labs, each of which has both the engineering capacity and the financial incentive to diversify over time. The early-September sell-off in AVGO shares — attributed to broader semiconductor sector weakness and concerns about custom AI silicon order pacing — demonstrates that the market is not pricing in a linear path to those targets, even after a quarter that beat on every headline metric.

Three signals will determine most of the outcome over the next four quarters. First, whether co-development agreements with additional frontier labs — Anthropic is named in analyst revenue projections — produce publicly announced volume programs; any such disclosure would materially de-risk the long-range revenue case. Second, the trajectory of Google's Marvell engagement: the next round of TPU cluster procurement announcements will indicate whether that relationship is a hedge or a genuine workload shift away from Broadcom. Third, the terms of any large-scale debt financing: if the reported $60 billion program proceeds, bond market pricing and covenant disclosures will reveal how much leverage Broadcom is prepared to carry against a revenue base that, while growing rapidly, remains concentrated among a handful of customers. The Jalapeño benchmarks and the Hugging Face reports, read together, describe an AI silicon landscape in which Nvidia's strategic options are narrowing at the margin — a structural shift that a single record-breaking quarterly print can only partially capture.