Most financial coverage of Broadcom leads with VMware. The $69 billion acquisition, the price hikes, the customer complaints. It's a reasonable story but it's half the company, and not the half that's growing fastest.
Broadcom in 2026 is two separate businesses running under one ticker. One designs custom AI chips and networking silicon for the largest data center operators in the world. The other runs enterprise infrastructure software for companies that built their IT stack on VMware. Both generate cash at a scale that doesn't fit the typical "diversified chip maker" framing most analysts still use.
This article covers the business model: how each side makes money, why the margin structure looks the way it does, and what makes both sides structurally difficult to replace.
Two Businesses, One Ticker
Broadcom has two reportable segments.
Semiconductor Solutions covers networking ASICs, custom AI accelerators, storage controllers, broadband chips, and wireless connectivity silicon. Customers are primarily hyperscalers, enterprise data centers, and consumer electronics manufacturers. This segment generates roughly 60–65% of total revenue depending on the quarter.
Infrastructure Software is primarily VMware the enterprise virtualization platform Broadcom acquired in November 2023 plus earlier acquisitions CA Technologies and Symantec's enterprise security division. Customers are enterprise IT departments that run their compute, storage, and networking virtualization on VMware's stack. This segment generates 35–40% of revenue.
The combination produces adjusted EBITDA margins around 68%. That's unusual for a company with significant hardware revenue, and the reason it's possible goes to the structure of each segment which is worth understanding separately before looking at how they combine.
The Semiconductor Side: Networking First, Then AI
Broadcom's semiconductor business started with networking. It's worth understanding that foundation before getting to the AI piece, because most coverage skips it.
Networking ASICs are the chips that move data inside hyperscaler data centers. Broadcom's Tomahawk and Trident product lines dominate Ethernet-based data center fabric the physical switching infrastructure that connects servers inside a Google, Amazon, or Microsoft data center. Jericho handles routing at scale.
When a hyperscaler expands its data center footprint, it needs more switching capacity in direct proportion to that expansion. Broadcom doesn't sell into a discretionary budget line it sells into infrastructure that must scale with compute. That's a different demand profile than most chip companies.
To understand how semiconductor companies make money and why networking ASICs have the margin profile they do, the short answer is: Broadcom designs the chip, TSMC manufactures it, and Broadcom captures the design margin without owning a fab. Capital expenditure runs below 1.3% of revenue. That's the fabless model working as intended.
The Semiconductor Side: Custom AI Silicon
The second business inside Broadcom's semiconductor segment is newer and, right now, growing faster than anything else on the income statement.
When a hyperscaler wants to train AI models at scale without paying NVIDIA's prices indefinitely, it builds a custom AI accelerator a chip optimized for its own specific workloads rather than a general-purpose GPU sold at market price to anyone who wants one.
"A chip optimized for a specific training workload delivers better cost per computation than a general-purpose GPU at scale. The tradeoff is that it takes years and hundreds of millions of dollars to build."
That's where Broadcom comes in. It provides the chip design engineering resources. Google's TPU and Meta's MTIA are the most publicly known examples both chips Broadcom has had design involvement with. TSMC manufactures them. The hyperscaler gets a chip tuned to its workload. Broadcom gets a multi-year design engagement and the revenue that comes with it.
These chips are called XPUs custom AI accelerators built for one customer's architecture rather than sold on the open market.
Why XPU Relationships Don't Break
This is the part of Broadcom's semiconductor business that I think gets under-analyzed.
The stickiness isn't contractual. It's architectural. Once a hyperscaler deploys a training cluster built around a chip Broadcom helped design we're talking about clusters with hundreds of thousands or millions of chips switching to a different chip design partner requires rebuilding the entire architecture. Not updating it. Rebuilding it. That process takes 3 to 5 years minimum and costs more than most companies want to spend when their existing cluster is already operational.
Nobody at Google or Meta is mid-deployment on a billion-dollar cluster and deciding to start over to save money on design fees.
In December 2024, Hock Tan disclosed that three hyperscaler customers are developing next-generation XPUs with Broadcom, with the serviceable addressable market from those three customers alone reaching $60 to $90 billion by FY2027 assuming each deploys a one-million-chip cluster by that date.
Whether those deployments track to schedule is a monitoring question, not a business model question. The model itself design engagement, architectural lock-in, manufacturing through TSMC doesn't change quarter to quarter.
The Software Side: What Broadcom Actually Did to VMware
Broadcom acquired VMware in November 2023 for $69 billion. VMware was already profitable when it was public. The acquisition wasn't about adding revenue it was about converting the revenue model.
Here's what Broadcom changed:
Product consolidation. VMware had dozens of products. Broadcom eliminated most of them. The portfolio now centers on VMware Cloud Foundation a single subscription bundle covering compute, storage, and networking virtualization. Less complexity to support, fewer products to maintain, higher margins on what remains.
Revenue model conversion. VMware sold perpetual licenses. Broadcom converted customers to multi-year subscriptions. Perpetual license revenue lands in one period; subscription revenue spreads over the contract term. That conversion creates a short-term revenue recognition lag that makes the software segment look slower than it may actually be.
Price increases. Broadcom raised prices on the retained product portfolio. Some enterprise customers objected publicly. Some started evaluating alternatives. The customers who stayed are paying more per seat than they were before the acquisition.
The logic is classic Hock Tan: customers who are embedded in VMware's hypervisor cannot easily leave. Moving off VMware requires rearchitecting years of production workloads compute, storage, networking, all of it. The switching cost is real, and Broadcom priced accordingly.
Whether that bet holds through the FY2026 renewal cycle is the open question for this segment. The business model assumption is that switching costs are high enough to retain the customers that matter. The numbers will confirm or deny that over the next few quarters.
Hock Tan's Acquisition Playbook
Hock Tan has run Broadcom since 2006. Roughly fifteen major acquisitions. The playbook has not changed once.
Buy a company with a large installed base and high switching costs. Cut R&D and headcount after close. Raise prices on the core products. Convert perpetual license customers to subscriptions. Use the resulting cash flow to service acquisition debt, pay dividends, and fund the next deal.
That's it. Every acquisition Brocade, CA Technologies, Symantec's enterprise division, VMware follows that exact sequence.
The model has two hard requirements. The acquired company needs customers who genuinely cannot leave without significant pain, and the market needs to be large enough that even a stripped-down version generates substantial cash flow. VMware satisfies both. Enterprise infrastructure is not a market where customers switch vendors because a competitor runs a better promotion.
R&D on a non-GAAP basis runs at roughly 8% of revenue. For context, most semiconductor companies with active product roadmaps spend 15–25%. Broadcom extracts value from acquired technology it doesn't build from scratch. That's a deliberate choice, not a limitation.
The constraint on this model is supply. There are not many companies left at VMware's scale with the same combination of large installed base and high switching costs. The acquisitions get harder to find, not easier.
The Margin Structure: Why 42% FCF Is Possible in Hardware
The number that stops most analysts is the free cash flow margin. Hardware companies don't generate 42% FCF margins. Software companies do. Broadcom does both, and the combination is what makes the margin structure possible.
To understand why foundries like TSMC run lower margins than chip designers, the answer is capital intensity. TSMC spent over $40 billion in capital expenditure in FY2025. Broadcom spent less than 1.3% of revenue. The fabless model transfers manufacturing capital requirements entirely to the foundry.
Add a software segment running at margins closer to 70–80% on a subscription base with minimal incremental cost to serve, and the blended margin moves well above what the semiconductor side alone could produce.
The FY2025 full-year numbers show what this structure produces at scale:
| Metric | FY2025 |
|---|---|
| Revenue | $63.9B |
| Gross Margin | 67.8% |
| Free Cash Flow | $26.9B |
| FCF Margin | 42.1% |
$26.9 billion in free cash flow on $63.9 billion in revenue. That's the two-segment structure working as designed semiconductor design margins blended with software subscription margins, against a capital expenditure base that stays below 1.3% of revenue because Broadcom doesn't own the manufacturing.
For which financial metrics matter most when analyzing a semiconductor stock, FCF margin at this revenue base is near the top of the list. Revenue can be managed. Cash flow is harder to fake.
Balance Sheet: The Debt Is Manageable, Not Comfortable
Broadcom carries significant debt. The VMware acquisition was financed primarily with borrowing, and the balance sheet reflects that.
| Item | Amount (as of Q1 FY2026) |
|---|---|
| Cash and equivalents | $14.2B |
| Total debt | ~$66.1B |
| Net debt | ~$51.9B |
| Goodwill | $97.8B |
Net debt of roughly $51.9 billion sounds alarming until you put it against the FCF run rate. At $26–30 billion in annual free cash flow, net debt sits at roughly 1.7x FCF. That ratio has been declining since the VMware close.
The $97.8 billion goodwill balance is the premium Broadcom paid above book value across its acquisition history primarily VMware. It's a real number to watch. Goodwill impairment flows directly through the income statement if the acquired businesses underperform what Broadcom paid for them. It's not a footnote.
Competitive Position
On networking ASICs: Marvell Technology is the direct competitor. Both companies design data center networking chips and compete for the same hyperscaler design wins. The market largely splits between them, with Broadcom holding the dominant position in Ethernet switching fabric.
On custom AI silicon: The competitive frame here is different from what most coverage implies.
NVIDIA's business model is built around selling general-purpose GPUs at market prices to any buyer. A hyperscaler buying NVIDIA H100s is not choosing between NVIDIA and Broadcom. A hyperscaler building a million-chip training cluster with a custom chip at the center is making that choice and once the design is committed, NVIDIA is largely out of that conversation.
The comparison that matters: NVIDIA captures the customers who need GPU performance now and don't want to wait 3–5 years for custom silicon. Broadcom captures the customers who are large enough to justify custom architecture and willing to invest in the design cycle. Both markets are growing. They're not the same market.
On enterprise software: VMware's main competition post-acquisition comes from customers evaluating alternatives like Nutanix, Red Hat, and public cloud-native virtualization. The switching cost argument holds until it doesn't which is why FY2026 retention data matters more than the model itself.
It's also worth understanding Broadcom's manufacturing dependency here. The entire semiconductor side depends on TSMC's manufacturing capability at advanced nodes. Broadcom is fabless by design, but that means it cannot produce chips if TSMC can't or won't. That's a structural dependency, not just a supply chain footnote. TSMC's competitive moat is relevant to Broadcom's business model in a way that doesn't get discussed enough.
The framing that misses in most Broadcom coverage isn't about the numbers it's about what the company actually is.
Broadcom built a machine that extracts cash from two types of customers who find it genuinely difficult to leave: enterprise IT departments locked into VMware infrastructure, and hyperscalers locked into chip architectures Broadcom helped design. Both sides have high switching costs. Both sides generate recurring revenue. Neither side is going anywhere quickly.
What I find interesting is that the two businesses don't obviously belong together. A chip design shop and an enterprise software business are not natural partners. But from a capital allocation standpoint, they share the same logic: find customers who can't easily leave, extract as much margin as the relationship allows, use the cash to service debt and buy the next thing.
The business model risk that doesn't get enough attention isn't VMware churn or even Apple wireless chip internalization. It's the acquisition pipeline. Hock Tan's model needs large, sticky targets. Those are finite. At some point Broadcom runs out of companies large enough to move the needle at its current revenue base, and the model has to transition from acquirer to organic grower. That's a different company with a different margin structure. It hasn't happened yet but it's the long-term question the current numbers don't answer.
For now, the two-segment structure is intact, the FCF engine is running, and the XPU design engagements are the highest-growth part of the business. Those three things are what the model rests on.
Current Performance
This article covers the business model how Broadcom is built and why the structure is defensible.
For how that model is translating into recent quarterly results, guidance, and earnings beat/miss analysis, see the Broadcom earnings analysis →
For how the current macro environment interest rates, hyperscaler capex cycles, export controls is affecting Broadcom's near-term outlook, see the macro conditions piece →
FAQs
How does Broadcom make money?
Broadcom generates revenue through two segments. Semiconductor Solutions designs and sells networking ASICs, custom AI accelerators, storage controllers, and wireless chips primarily to hyperscalers and enterprise data centers. Infrastructure Software sells VMware virtualization and cloud management subscriptions to enterprise customers. In FY2025, the two segments combined to generate $63.9 billion in revenue at a 42% free cash flow margin.
What does Broadcom actually do?
Broadcom designs chips and runs enterprise software. On the chip side, it makes the networking silicon that moves data inside hyperscaler data centers, and designs custom AI accelerators for companies like Google and Meta that want chips tuned to their specific workloads. On the software side, it sells VMware virtualization subscriptions to enterprise IT departments. It doesn't manufacture anything TSMC handles production.
What is an XPU and why does Broadcom design them?
An XPU is a custom AI accelerator built for one company's specific workload, as opposed to a general-purpose GPU sold at market price to any buyer. Google's TPU and Meta's MTIA are examples Broadcom has had design involvement with. Hyperscalers build custom chips because a workload-specific design produces better cost per computation than a general-purpose chip at the scale they operate. Once a hyperscaler is mid-deployment on a cluster built around a Broadcom-designed chip, switching design partners requires a full architecture rebuild a 3 to 5 year process minimum.
Why does Broadcom have such high profit margins?
Two reasons working together. First, Broadcom is fabless it designs chips but outsources manufacturing to TSMC, so capital expenditure stays below 1.3% of revenue. Second, the VMware software segment runs on subscription economics with minimal incremental cost to serve. Blended, the two segments produce adjusted EBITDA margins around 68% and FCF margins around 42%. Neither side alone would get there.
How does the VMware acquisition fit Broadcom's strategy?
VMware is a textbook application of Hock Tan's acquisition playbook: buy a company whose customers are too embedded to leave easily, cut product complexity, raise prices on what remains, convert perpetual licenses to multi-year subscriptions. VMware's customers run their entire compute and networking virtualization on VMware's hypervisor. Moving to a competitor requires rearchitecting years of production infrastructure. Broadcom bought that switching cost, then priced against it.
How does Broadcom compare to NVIDIA?
They're not competing for the same decision. NVIDIA sells general-purpose GPUs to any buyer at market prices. Broadcom designs custom accelerators for specific hyperscalers and sells networking ASICs broadly. A hyperscaler buying NVIDIA GPUs is not choosing between NVIDIA and Broadcom. A hyperscaler building a custom training cluster with a chip Broadcom helped design is making that choice and once that cluster is deployed, NVIDIA is largely out of the conversation for that workload.
What are the main risks in Broadcom's business model?
Three risks matter structurally. First, Apple customer concentration: Apple is a significant semiconductor revenue contributor and is developing more of its own wireless chips internally. Second, XPU customer concentration: the AI semiconductor growth runs through a small number of hyperscaler design commitments if one pauses or changes architecture, the revenue impact is direct. Third, VMware retention: Broadcom raised prices significantly post-acquisition, and whether retained customers generate enough incremental revenue to offset those who left will be visible in FY2026 renewal data. There's also a longer-term model risk: Hock Tan's acquisition playbook requires large, sticky targets, and those targets are finite.
Is Broadcom a good stock to buy?
Finovian doesn't give buy or sell recommendations or price targets. What the model shows: a business generating 42% FCF margins with two segments that benefit from genuine switching costs, a chip design business with architectural lock-in at hyperscale, and a capital allocation model that has returned significant cash to shareholders while reducing acquisition debt. The open questions VMware retention and XPU deployment pace are monitoring questions, not model questions. Whether that profile fits your investment criteria is your decision. For current performance data, see the earnings analysis →.