Broadcom leans fully into AI systems as backlog swells to $162 billion

AI engine powers a record year

Broadcom closed fiscal 2025 with the kind of numbers that underline just how central it has become to the AI infrastructure build‑out.

For the year, revenue rose 24% to $63.9 billion, with organic growth accelerating rather than fading. AI was the main culprit: AI-related revenue grew 65% to $20 billion, helping drive semiconductor sales up 22% to $36.9 billion. Infrastructure software – now dominated by VMware – grew even faster, up 26% to $27 billion, propelled by adoption of VMware Cloud Foundation.

The fourth quarter showed that momentum is still building. Revenue hit a record $18 billion, up 28% year on year and above guidance, while adjusted EBITDA climbed 34% to $12.12 billion, or 68% of revenue, a point above the company’s target margin.

CEO Hock Tan was unusually explicit about the scale and granularity of Broadcom’s AI franchise. AI semiconductor revenue in Q4 reached $6.5 billion, up 74% year on year, and has grown more than 10‑fold over the eleven quarters since the company began disclosing the metric. Semiconductor revenue overall was $11.1 billion, up 35% year on year; infrastructure software contributed $6.9 billion, up 19%.

Free cash flow in fiscal 2025 rose 39% to $26.9 billion – 42% of revenue – allowing Broadcom to return $17.5 billion to shareholders via $11.1 billion in dividends and $6.4 billion in buybacks and eliminations. The board has now approved a 10% increase in the quarterly dividend to $0.65 per share (or $2.60 annualized) for fiscal 2026, marking the fifteenth consecutive annual increase since 2011.

From components to AI systems – and a backlog that redefines visibility

The more intriguing story, however, sits in the order book rather than the income statement.

Broadcom now holds $162 billion of consolidated backlog. Of that, $73 billion is tied to AI data centers – a figure Tan said encompasses XPUs (custom accelerators), AI switches, DSPs, lasers, PCIe switches and other AI‑centric components – and is “anticipated” to be shipped over the next 18 months. By his own description, this is a moving snapshot, not a peak; lead times run six to twelve months, and new AI bookings are “accelerating”.

That $73 billion AI backlog, nearly half of the total company backlog, has at least three notable features:

• Concentration in compute: Roughly $20 billion of the AI backlog comes from networking and optical (switches, 1.6T DSPs, lasers, PCIe switches); the remainder – the majority – is in XPUs and custom compute.
• System‑level pivot: For at least one large hyperscale customer (widely understood to be Anthropic, though Broadcom stayed away from names), Broadcom is no longer just selling chips. It is now selling complete “systems” or racks – including XPUs and a basket of its own networking and optical silicon – and assuming responsibility for certifying the full system. Tan framed this as a natural extension given Broadcom’s breadth of content in an AI rack.
• Customer count and diversity: The TPU business at Google remains one anchor, with Google now exposing TPU capacity to external AI tenants such as Apple, Cohere and others. Separately, Broadcom has disclosed four other XPU customers. This quarter, it added a fifth custom-accelerator customer via a $1 billion order for delivery in late 2026. The OpenAI relationship, referenced obliquely on the call, is being built around a multi‑year 10‑gigawatt infrastructure plan from 2027–2029; its XPU program is described as “at a very advanced stage,” but not yet a 2026 revenue driver.

On the networking side, demand appears even more frantic. Order backlog for AI switches alone now exceeds $10 billion, anchored by Broadcom’s 102Tb/s Tomahawk 6, which Tan called “one of the fastest‑growing products in terms of deployment” the company has ever seen. The 1.6Tb/s DSP portfolio, and the lasers and optical components that pair with it, are seeing what he described as “record orders” as customers build out fabric capacity ahead of GPU/XPU deployment.

In aggregate, Tan characterised the last three months of AI bookings as something Broadcom has “never seen” before in its history.

Margin architecture: lower gross, higher operating leverage

The natural question for investors is what all this do‑it‑all rack ambition means for profitability.

CFO Kirsten Spears laid out the mechanics: AI revenue – and especially systems revenue – carries structurally lower gross margins because Broadcom passes through significant third‑party content. That trend will intensify in the back half of fiscal 2026 as rack shipments ramp, with more non‑Broadcom components included in the bill of materials.

In Q4, consolidated gross margin was a robust 77.9%, flattered by software mix and favourable product mix within semiconductors. Semiconductor segment gross margin, at 68%, already reflects the AI headwind. Software, by contrast, posted a towering 93% gross margin, up two points year on year as VMware integration was completed.

Yet, Broadcom continues to push operating margin up by squeezing more revenue through a relatively fixed operating cost base. Q4 operating margin reached 66.2%, up 70 basis points sequentially despite a 50bp dip in gross margin, as operating expenses grew far slower than revenue. At the segment level, semiconductor operating margin expanded 250bps year on year to 59%, even as R&D spend in that segment rose 16%. Infrastructure software operating margin surged to 78% from 72% a year ago as VMware synergies flowed through.

Looking forward to fiscal 2026, management signalled a familiar pattern: gross margins should drift down as AI and system mix rise, but adjusted EBITDA margin is expected to hold around 67% in Q1 and remain structurally high as operating leverage more than offsets gross margin dilution. In Tan’s words, “gross margin dollars will go up… operating margin dollars will go up, but the margin itself… will come down a bit.”

The one notable P&L headwind is tax. Spears guided to a step up in the non‑GAAP tax rate from 14% to about 16.5% in Q1 and for the full year, driven by the global minimum tax and a shift in geographic income mix. That will clip net income growth relative to operating profit.

Segment dynamics: AI strength versus muted enterprise

Under the AI halo, Broadcom’s non‑AI franchises look pedestrian – though not broken.

In Q4, non‑AI semiconductor revenue was $4.6 billion, up just 2% year on year but up 16% sequentially, buoyed by normal wireless seasonality and a “solid recovery” in broadband. Wireless was flat year on year, while other end markets – notably enterprise‑oriented franchises – remained soft amid constrained enterprise spending. For Q1 2026, Broadcom expects non‑AI semiconductor revenue of about $4.1 billion, flat year on year and down sequentially, again mostly due to wireless seasonality.

The picture in infrastructure software is healthier. Q4 revenue of $6.9 billion grew 19% year on year, comfortably above the company’s prior $6.7 billion outlook. Bookings were particularly strong: total contract value booked in the quarter reached $10.4 billion, up from $8.2 billion a year ago, lifting infrastructure software backlog to $73 billion from $49 billion. VMware Cloud Foundation remains the primary growth engine, as customers commit to large, multi‑year platform deals rather than piecemeal license purchases.

For Q1 2026, Broadcom expects infrastructure software revenue of approximately $6.8 billion, up 2% year on year, reflecting seasonally lower renewals. Over the full fiscal year, Tan still anticipates low double‑digit percentage growth in infrastructure software, fundamentally tied to VMware expansion.

Balance sheet and capital returns: muscular but disciplined

The combination of high margins and rapid growth continues to bulk up Broadcom’s balance sheet.

The company exited Q4 with $16.2 billion in cash, up $5.5 billion sequentially, against $67.1 billion of gross fixed‑rate debt carrying a modest 4% weighted average coupon and 7.2‑year average maturity. Capital expenditures remain light relative to revenue – just $237 million in Q4 – even as Broadcom ramps its own advanced packaging facility in Singapore to secure supply for multi‑chip AI packages.

Working capital remains tightly managed. Days sales outstanding edged up to 36 days from 29 a year earlier, still conservative for a business leaning so heavily on hyperscale customers. Inventory rose 4% sequentially to $2.3 billion, but days of inventory fell to 58 from 66 as AI shipments accelerated and Broadcom deliberately rationed inventory across the ecosystem.

On capital returns, the company’s formula remains constant: pay a steadily rising dividend and opportunistically repurchase stock within the confines of leverage targets. With the dividend moving to $0.65 a quarter, the implied annual payout of $2.60 per share is up 10% year on year. The board also extended the share repurchase authorization, with $7.5 billion remaining, through the end of calendar 2026.

For now, Broadcom’s story is one of extraordinary AI‑driven scale layered onto an already high‑margin software and infrastructure franchise. The risk – and opportunity – lies in how far the company can push system‑level AI ambitions without eroding the margin architecture that has made it an investor favourite.