How large does custom silicon get as each hyperscaler designs its own chips?

Thesis

Broadcom runs two businesses with opposite shapes. One designs custom AI accelerators and networking for a handful of the largest buyers of compute on earth. The other collects subscription fees from enterprises running VMware and throws off cash with little new investment. In fiscal 2025 semiconductors brought in $36.9 billion and infrastructure software $27.0 billion. In the quarter ended August 2, 2026, the mix had moved to $20.8 billion of semiconductors and $8.75 billion of software.

Management expects custom silicon to get very large. On the September 2026 call it guided AI revenue to about $58 billion for fiscal 2026, $115 billion for fiscal 2027 and $230 billion for fiscal 2028, and said it expects to exceed $30 of earnings per share in fiscal 2028. If those targets hold, software shrinks to a minor share of revenue and Broadcom becomes a custom-silicon vendor with a software cash annuity attached.

Each XPU dollar is a design win that a customer can in principle take to another vendor, and Marvell has just landed a deal with Google that makes the point concrete. The fastest growth now comes from buyers that are not cash-rich hyperscalers: Anthropic and OpenAI need financing, and Broadcom has begun supplying it. Margin is moving too. The company guides consolidated gross margin to about 73% in the fourth quarter against 78% a year earlier (our calculation from fiscal fourth quarter 2025 non-GAAP gross profit of $14.0 billion on revenue of $18.0 billion), because XPU systems and memory dilute it.

At the October 30, 2026 close of $326.08, the shares sit about 34% below the 52-week high of $493.28 reported on October 4 (our calculation), even as quarterly AI revenue has doubled since the first quarter. That gap between results and price is the tension to watch over the next two quarters.

Company History

Broadcom traces its lineage to Hewlett-Packard. The company’s 10-K describes a history of more than 60 years that dates back to AT&T/Bell Labs, Lucent and Hewlett-Packard, and lists the acquisitions that built the modern business: LSI, Broadcom Corporation, Brocade, CA, Symantec’s enterprise security unit and VMware. Neither the 10-K nor the filings we opened give a founding year for the Hewlett-Packard semiconductor operation, so we do not use the 1961 date that circulates in secondary sources. The founded date in the header, August 2005, is the incorporation of Avago Technologies Limited in Singapore (Avago S-1/A, which also says Avago succeeded Agilent’s Semiconductor Products Group, acquired on December 1, 2005). That entity bought Broadcom Corporation for $37 billion and took the Broadcom name, and the registrant today is Broadcom Inc., a Delaware corporation. Hock Tan is President and CEO.

The software strategy came later. Broadcom acquired CA Technologies for $18.9 billion in 2018 and Symantec’s enterprise security business for $10.7 billion in 2019. The VMware deal was announced in May 2022 at about $61 billion plus roughly $8 billion of assumed net debt, and closed on November 22, 2023. The 10-K reports that consideration consisted of about $30.8 billion in cash and 544 million Broadcom shares valued at $53.4 billion. Broadcom later sold VMware’s end-user computing unit to KKR for $3.5 billion in July 2024.

The effect on the income statement was immediate. Annual revenue was $33.2 billion in fiscal 2022 and $35.8 billion in fiscal 2023. After VMware, it reached $51.6 billion in fiscal 2024 and $63.9 billion in fiscal 2025. Infrastructure software went from $7.6 billion in fiscal 2023 to $21.5 billion in fiscal 2024.

The second inflection began in 2025 and is still unfolding. Custom AI accelerators moved from a Google-dominated program to a broader customer set. By the fiscal first quarter of 2026, AI semiconductor revenue was $8.4 billion, up 106% from a year earlier. Two quarters later it was $16.7 billion. The company is headquartered in Palo Alto and had about 33,000 employees as of November 2, 2025, with roughly 57% in research and development roles.

Product

Broadcom sells into two segments: Semiconductor Solutions and Infrastructure Software. The semiconductor segment includes AI products (custom XPUs and AI networking) and a non-AI portfolio of broadband, wireless, storage and industrial chips. Management guided non-AI semiconductor revenue to about $4.3 billion in the fourth quarter, a small and slow-moving base next to AI.

Custom AI accelerators (XPUs)

An XPU is an accelerator designed by or for one customer. Broadcom supplies the design engineering, the high-speed interface blocks, packaging and manufacturing coordination, then ships finished parts or systems. In the third quarter, XPUs made up 73% of AI revenue, and XPU shipments rose more than 3.5 times year over year. Broadcom has six XPU customers.

Google is the anchor. Management described an expanded multi-year agreement to supply future TPU generations and AI networking, which it sized at multiple tens of billions of dollars of TPUs a year (per Benzinga’s call coverage). Management said on the call that Ironwood (TPU v7) shipped in high volume to Anthropic and Google and that production shipments of the TPU v8i for Google have started (call transcript).

Anthropic is the fastest-growing account. Management expects 1 gigawatt of Ironwood TPUs in 2026, about 5 gigawatts of TPU v8i in 2027 and sees line of sight to a further 10 gigawatts in 2028. It expects Anthropic to be its largest XPU customer in 2027. OpenAI is the second named lab: its first-generation chip is on track for 1.3 gigawatts in 2027 and, with a successor chip, line of sight to more than 5 gigawatts in 2028. Meta runs three generations of its MTIA chip through 2027, with line of sight to 3 gigawatts deployed through 2028.

Management put Broadcom’s content at $20 billion to $30 billion per gigawatt and expects that figure to stay stable as chip power rises. That is the conversion rate between the gigawatt plans above and the dollar targets.

AI networking

Networking revenue grew more than 2.5 times year over year in the third quarter. In the first quarter it had grown 60% and represented a third of AI revenue. The product set is Ethernet switch silicon (the Tomahawk family), optical digital signal processors and lasers, and PCIe and connectivity chips.

Tomahawk 6 is a 100 Tb/s Ethernet switch that management said is deployed at nearly all hyperscalers building XPUs with Broadcom (call transcript). Tomahawk 7, a 200 Tb/s Ethernet switch, has taped out. Management said indium phosphide laser capacity will more than triple year over year. Networking matters strategically because it is the layer where Broadcom competes with NVIDIA’s proprietary NVLink and InfiniBand using open Ethernet, and it keeps Broadcom in the cluster even if an XPU socket moves to another designer.

VMware and infrastructure software

Infrastructure software includes VMware Cloud Foundation (VCF), mainframe software from CA, and Symantec security. According to StorageNewsletter commentary, Broadcom cut VMware’s catalog from 168 product SKUs to four bundles and moved it to subscription-only licensing. VCF 9.1 became generally available in May 2026 adding memory tiering and vSAN deduplication. Management says software gross margin is about 94% and operating margin about 84%.

Hock Tan has argued that AI raises VMware demand. He said he is confident that generative and agentic AI “will create the need for more VMware, not less.” Customers complain about pricing, particularly in the mid-market, and the strategy focuses on roughly the top 2,000 enterprise accounts.

Market

Customer

The customer base is concentrated. Broadcom’s 10-K says distributors accounted for 48% of net revenue in fiscal 2025 and that its top five end customers generated about 40%. With AI revenue now 56% of the total, so the fiscal 2025 share may understate current concentration, but the filing gives no update. The 10-K does not name the customers, and the filing’s excerpt we read gives no single-customer percentage.

The buyers fall into three groups. Hyperscalers that design chips for internal use (Google, Meta) fund their own purchases from operating cash flow. AI labs (Anthropic, OpenAI) need outside capital to pay for compute. Enterprises buy VMware as a private-cloud platform, and many of those are locked in by switching costs.

Capital intensity at the first group is the engine of the whole system. In May 2026, a Reuters host quoted by 247Wall St said big-hyperscaler spending for 2026 had topped $700 billion of 2026 spending, including Meta’s raised range of $125 billion to $145 billion.

The labs are the less certain group. Anthropic’s IPO filing, reported on October 1, 2026, shows Broadcom has agreed to lend Anthropic up to $42 billion, which could fund about a third of Anthropic’s $125.2 billion five-year TPU compute commitment. We return to this under risks.

Market Size

Three data points frame the market. Broadcom itself says its AI revenue will reach about $115 billion in fiscal 2027. NVIDIA, the incumbent, reported data center revenue of $89.0 billion in a single quarter ended July 26, 2026, which is more than five times Broadcom’s quarterly AI revenue. And Marvell, which competes for custom silicon, cited at a June 2025 event a 2028 data center market of $94 billion in which custom compute is $55.4 billion, of which custom XPUs are $40.8 billion.

These numbers do not agree. Marvell’s 2028 custom XPU estimate is a fraction of Broadcom’s own 2028 AI target of $230 billion, and the two definitions differ (Broadcom’s figure includes networking and covers rack-scale systems). The wide gap tells us the market is being defined by a few contracts rather than by a settled third-party estimate. Our working assumption is to treat Broadcom’s gigawatt-based plans as the best available guide to its own revenue and to treat third-party total-market numbers as a cross-check on share, not as a ceiling.

Competition

Competitive Landscape

Competition operates on three levels. In merchant AI compute, NVIDIA sets the benchmark: its Q2 fiscal 2027 gross margin was 75.0% on revenue of $96.2 billion. In custom silicon, Broadcom competes with a small group of design partners for each hyperscaler socket. In networking, Broadcom’s Ethernet silicon competes with NVIDIA’s proprietary fabrics and with other merchant switch vendors.

The most important structural fact is that the buyer controls the architecture. A hyperscaler that designs its own chip can dual-source the physical design work, bring more of it in-house, or move a generation to a cheaper variant. Broadcom’s defense is the accumulated stock of high-speed interface technology, packaging know-how and delivery capacity, plus supply it has pre-arranged. Management said it has secured the supply to meet its outlook.

In software, VMware competes with other virtualization and private-cloud platforms and with public cloud providers. The pricing strategy increases margin per customer but also invites migration, particularly among mid-market customers.

Competitors

NVIDIA

NVIDIA is the standard against which every custom chip is measured. It guided its next quarter to revenue of $108.0 billion and a gross margin of 74.0%. Broadcom’s fourth-quarter AI guide of $21.7 billion is a fifth of that scale. The relationship is partly a substitute (an XPU replaces a GPU) and partly complementary (many clusters mix both). NVIDIA’s networking business competes with Broadcom’s Ethernet.

Marvell

Marvell is the closest direct rival. Its July 2026 agreement with Google, disclosed in an 8-K filed August 19, 2026, gives Google a warrant for up to 58.97 million Marvell shares at an exercise price of $206.58, vesting in 240 tranches as Google buys custom products. Futurum calculates that the full vesting ceiling equals $120 billion of purchases through fiscal 2033. Marvell targets more than $10 billion of custom revenue in fiscal 2029 and a 20% share of the custom market by 2028. Neither company’s filing allocates TPU volume between them, so the effect on Broadcom’s Google business is unknown. Futurum reads the Marvell deal as covering parts that attach to the TPU rather than the core compute die, and notes Broadcom’s TPU supply agreement runs through 2031.

MediaTek and in-house teams

MediaTek supplies lower-cost TPU variants for Google, according to the same Futurum analysis. The longer-term risk is that large customers staff their own silicon teams for the work Broadcom does today. We have not found a source that sizes this risk, and we do not estimate it.

Business Model

Broadcom earns revenue three ways. First, it sells chips and systems. For XPUs the pricing is negotiated per program with each customer, and the revenue arrives as shipments ramp. Second, it sells merchant networking and connectivity silicon at standard list prices to a broader group of customers. Third, it sells software on multi-year subscriptions.

The margin structure shows the difference. Software runs at about 94% gross margin. Consolidated gross margin in the third quarter was 75.0% non-GAAP (our calculation from the reported non-GAAP gross profit of $22.2 billion), and management guides the fourth quarter to about 73%. The reason is mix: XPU systems carry lower gross margin and more memory content. Management argues that non-GAAP operating margin, guided to about 66% of revenue, is the right measure because operating costs do not scale with revenue.

Capital allocation is plain. Broadcom pays a quarterly dividend of $0.65 per share, which totaled $3.1 billion in the third quarter. It repurchased about $7.85 billion of stock in the first quarter, authorized a new $10 billion program, and bought back none in the third quarter. It repaid $5.6 billion of debt in the third quarter.

The newest element is financing. In June 2026, Broadcom set up a financing platform with Apollo and Blackstone, and the first $35 billion tranche closed for Anthropic’s 1 gigawatt deployment. Management described any residual value guarantees as modest. The direct loan to Anthropic of up to $42 billion is a larger step: it turns Broadcom into a lender to its own customer.

Financials

Quarterly results show how quickly the mix has shifted. Total revenue was $19.3 billion in the first quarter of fiscal 2026 (ended February 1, 2026), $22.2 billion in the second quarter and $29.6 billion in the third, up 86% from $16.0 billion a year earlier. Trailing twelve-month revenue is $89.1 billion (our arithmetic: $18.0 billion in the fourth quarter of fiscal 2025 plus the three fiscal 2026 quarters above).

The third quarter produced non-GAAP operating income of $20.1 billion, up 92%, and GAAP net income of $13.1 billion, up 216%. GAAP diluted EPS was $2.68 and non-GAAP was $3.32. Operating cash flow was $14.2 billion and capital spending was only $532 million, giving free cash flow of $13.7 billion, or 46% of revenue. Capital intensity is low because Broadcom designs chips and TSMC and others manufacture them. Management guided fourth-quarter capex to about $1.4 billion.

AI revenue is the driver. The sequence of reported AI semiconductor revenue is $8.4 billion, $10.8 billion and $16.7 billion across the first three quarters, and the company guides to $21.7 billion for the fourth. The four quarters sum to $57.6 billion (our arithmetic), consistent with the raised full-year guide of about $58 billion. The third quarter exceeded the $16.0 billion guided a quarter earlier.

Fiscal 2025 provides the full-year base: net revenue of $63.9 billion, GAAP operating income of $25.5 billion, GAAP net income of $23.1 billion and free cash flow of $26.9 billion. Tax incentives lowered the fiscal 2025 tax provision by about $2.7 billion, worth $0.56 of diluted EPS, which shows how much the tax rate matters to reported earnings. The company guides the fourth-quarter non-GAAP tax rate to about 16%.

The software segment grew more slowly and then re-accelerated. It rose 1% in the first quarter, 9% in the second and 29% in the third. On the call, management put software annual recurring revenue growth at 15%. We found no reconciliation of the 29% revenue growth with the 15% recurring-revenue growth in the sources we opened, and we treat the third-quarter figure as possibly including timing effects until the fourth quarter confirms it. Management guides software to about $8.7 billion next quarter, which would be flat to down sequentially.

The balance sheet carries debt but is well covered. At August 2, 2026, cash was $24.0 billion against total debt of about $59.4 billion (short-term debt of $2.3 billion plus long-term debt of $57.2 billion), for net debt of roughly $35 billion (our calculation: $59.4 billion less $24.0 billion). Aggregate indebtedness was $67.1 billion at the end of fiscal 2025. Annualized free cash flow at the third-quarter run rate covers the net debt in under a year.

Fourth-quarter guidance is revenue of about $34.8 billion, up 93%, and a diluted share count of about 4.94 billion before any buybacks. Management said the market should not expect quarterly updates to the long-term AI figures.

Valuation

Market value is the October 30, 2026 closing price times shares outstanding from the latest 10-Q. Broadcom closed at $326.08 that day, and the 10-Q for the quarter ended August 2, 2026 reports 4,773,629,865 shares outstanding as of August 28, 2026. The product is a market capitalization of $1,556.6 billion, or about $1.56 trillion (our calculation). Data providers differ on price and share count: Trefis, for example, listed $335.13 and $1,599.8 billion for the same date, which we do not use. The 10-K cover reported 4.74 billion shares on November 28, 2025.

On trailing numbers, the market capitalization is 17.5 times trailing twelve-month revenue of $89.1 billion (our calculation). Trailing GAAP diluted earnings per share are $7.83, the sum of $1.74, $1.50, $1.91 and $2.68 across the last four quarters, which puts the shares at 41.6 times GAAP earnings (our calculation). We do not show forward multiples because we have no primary source for consensus estimates.

Two comparisons help. Against its own history, the trailing price-to-earnings ratio was about 45 times in early October, so the multiple is lower now, reflecting a lower share price and higher trailing earnings. The share price is about 34% below the 52-week high of $493.28 reported on October 4 (our calculation from $326.08 and $493.28). In the 10-K, the aggregate market value held by non-affiliates on May 2, 2025 was $939.2 billion, a measure that excludes affiliates’ shares, against a total market capitalization of about $1.56 trillion at the October 30, 2026 close.

Against management’s own target, fiscal 2028 EPS above $30 implies about 11 times that figure at the October 30 close (our arithmetic: $326.08 divided by $30 is 10.9). A multiple that low on the target implies the market is applying a sizable discount to it, or doubting that it will be reached on time. This memo makes no fair-value claim. It records what the price implies: a low multiple on fiscal 2028 EPS depends on belief in $230 billion of AI revenue, and the price today does not show full belief.

Key Opportunities

Anthropic and OpenAI scale into the largest XPU programs

The gigawatt schedule is the opportunity. Anthropic alone is guided to 1 gigawatt in 2026, about 5 gigawatts in 2027 and 10 more in 2028. At management’s content estimate of $20 billion to $30 billion per gigawatt, 5 gigawatts would be $100 billion to $150 billion of revenue on a single account (our arithmetic). The upper end exceeds the whole fiscal 2027 AI target, and management said on the call that not all announced gigawatts will necessarily reach production within the period, so the dollar targets are the better gauge. OpenAI’s line of sight to more than 5 gigawatts in 2028 would make it the second-largest customer (SDxCentral). Both labs depend on outside capital.

Ethernet wins the scale-up network

Broadcom is making Ethernet a credible option for the links inside a rack, a role that NVIDIA’s NVLink has held. Tomahawk 7 has taped out, and networking revenue is growing faster than 2.5 times a year. If scale-up moves to Ethernet at the volume the roadmap implies, Broadcom sells networking silicon into every cluster regardless of whose XPU sits at the end of the link. This opportunity reduces the importance of any one XPU socket.

VMware as private-cloud AI infrastructure

VCF could benefit if enterprises move AI inference to private infrastructure to control cost and data. Management says ARR grew 15% and that the software margin is about 84% at the operating line. VMware revenue grew 13% in the first quarter of fiscal 2026, and total contract value exceeded $9.2 billion. A durable growth rate in the mid-teens on a $35 billion annualized base (our arithmetic: four times the third-quarter $8.75 billion) would add several billion dollars of near-pure profit each year. The caution is that the segment’s reported growth has swung between 1% and 29% in three quarters.

Key Risks

Customer concentration and in-sourcing

Five end customers generated about 40% of fiscal 2025 revenue, and AI is now a larger share of sales, though the filings we read give no updated concentration figure. The largest customer in 2027 is expected to be Anthropic, a company that has not yet completed its IPO. Each XPU program is a multi-year design win, but the customer retains the right to bring a second vendor, as Google’s agreement with Marvell shows. A single lost or reduced socket on a program the size of the Google TPU would be visible in the reported results. In the sources we read, management gave no gigawatt figure for Google, which limits our ability to size that exposure.

Circular financing and balance-sheet exposure

Broadcom has become a supplier, lessor, lender and possible shareholder to the same customer. Anthropic’s IPO filing warns of “potential conflicts of interest.” Management declined to give a maximum for future guarantees. An analyst estimated the first Apollo and Blackstone tranche’s maximum exposure at about $29 billion, and the direct loan could reach $42 billion. Both are large relative to cash of $24.0 billion. If Anthropic’s funding or IPO falters, the revenue forecast and the loan book weaken at the same time. Management has called the guarantee risk low, and we have no evidence to the contrary, but the exposure is not yet fully disclosed.

Gross margin dilution

Gross margin is guided to 73% in the fourth quarter from 78% a year earlier. If the mix shifts further toward rack-scale systems with high memory and third-party content, gross margin falls, and the non-GAAP operating margin of 66% depends on spending growing much more slowly than revenue. Memory and substrate costs are outside Broadcom’s control, and management itself cited memory content as a driver.

Supply constraints and execution

Management said it has secured the supply to meet its outlook. The fiscal 2027 and 2028 targets still depend on third parties delivering chips, memory and substrates on schedule, and a delay in any of them moves revenue out even if demand stands. A Singapore substrate facility is expected to begin deploying in fiscal 2027.

Software pricing and AI disruption

VMware’s growth depends on price increases and bundling decisions that customers resist. The sources we read describe pushback, especially from mid-market customers, and the segment is now 30% of revenue and will fall as a share. A durable loss of mid-market accounts would reduce the cash annuity that supports the dividend and buybacks.

What Would Prove Us Wrong

Our reading is that Broadcom’s custom-silicon business will keep scaling at the pace management has laid out, with some risk on margin and financing. These signposts would contradict it.

1. AI semiconductor revenue below $21.7 billion in the fiscal fourth quarter (ending November 1, 2026, reported in December 2026). That would mean the quarter fell short of guidance after a $16.7 billion third quarter and would weaken the case for the fiscal 2027 target of $115 billion.
2. Consolidated gross margin below 70% in either of the first two quarters of fiscal 2027 (quarters ending January and April 2027). The company guided the fourth quarter to about 73%, so a drop of more than three points would indicate that mix dilution is larger than management expects.
3. Disclosed residual value guarantees, loan draws and other Anthropic-related support that together exceed the $42 billion loan facility, in filings through the fiscal 2026 10-K (due December 2026) and the fiscal 2027 10-Qs. That would show the financing is growing faster than the revenue it supports.
4. Infrastructure software revenue growth below 10% year over year in the fiscal first quarter of 2027 (quarter ending late January 2027, reported in March 2027). That would suggest the third-quarter 29% was a timing effect and that the segment’s growth is closer to the 1% and 9% rates of the first two quarters.

Summary

Broadcom reported record fiscal third-quarter 2026 results: revenue of $29.6 billion, AI semiconductor revenue of $16.7 billion and free cash flow of $13.7 billion. It guided the fourth quarter to $34.8 billion of revenue and set targets of about $115 billion of AI revenue in fiscal 2027 and $230 billion in fiscal 2028.

Three things stand between those targets and delivery. Google, Meta, Anthropic and OpenAI have to keep Broadcom as primary design partner while Marvell and others compete for sockets. Anthropic and OpenAI have to raise the capital to take delivery, with Broadcom now financing part of the purchase. And gross margin has to hold near current levels as XPU systems take a larger share of revenue.

VMware is now a minority of the business, but its high-margin cash supports the dividend and the balance sheet while the AI ramp is financed. The October 30 close of $326.08 is roughly 41.6 times trailing GAAP earnings (our calculation) and well below the 52-week high. The December 2026 fourth-quarter report is the next test.