Broadcom’s AI Trajectory: De-Risking Silicon Growth and Managing Credit Exposure

Key Highlights

• Broadcom is systematically de-risking customer balance sheets and addressing supply bottlenecks through private credit SPVs (such as the XPV vehicle with Apollo and Blackstone) and vertical integration into advanced packaging substrates, shifting its valuation profile toward counterparty credit exposure tied to customer capital market events.
• The market incorrectly interpreted Broadcom's $230 billion fiscal 2028 figure as strict quarterly guidance rather than a directional trajectory, overlooking management's explicit caveat that customer deployment constraints (site-readiness and gigawatt power availability) will prevent full conversion within the window.
• Gross margin compression in the print stems entirely from higher memory content inside custom XPUs rather than pricing erosion, making it a positive marker of accelerating deployment that is offset by revenue growth outrunning operating expenses.
• Broadcom's custom accelerator business is heavily concentrated across six large-scale customers, with Anthropic expected to surpass Google in XPU volume next fiscal year, granting multi-year design visibility while binding supplier revenue schedules directly to the financial structures and public market timelines of frontier AI labs.
• Broadcom's competitive positioning extends beyond custom silicon due to its unconditioned Ethernet networking attach across both XPU and third-party GPU clusters, alongside strategic vertical integration into Singapore substrate manufacturing to control the binding constraint of advanced packaging capacity.

By the Numbers

• Revenue: $29.591B, up 86% YoY, beat LSEG consensus of approximately $29.36B (Street files for the quarter ranged roughly $29.24B to $29.52B, so the beat is real and small)
• Non-GAAP diluted EPS: $3.32, up 96% YoY, beat LSEG consensus of approximately $3.24; GAAP diluted EPS $2.68
• Non-GAAP operating income: $20.1B, up 92% YoY; non-GAAP operating margin 67.9%, up 240 bps YoY
• Non-GAAP gross margin: 75%, down 210 bps sequentially, above the 74% guide
• AI semiconductor revenue: $16.7B, up 221% YoY and 54% sequentially, 56% of total revenue versus 49% in Q2
• XPU shipments up more than 3.5x YoY, representing 73% of AI revenue; AI networking revenue up more than 2.5x YoY
• Semiconductor solutions segment: $20.839B, up 127% YoY, 70% of total revenue; segment gross margin approximately 67%, down from approximately 70% in Q2; segment operating margin 61%, up 440 bps YoY, on opex at 6% of segment revenue (see note 1)
• Infrastructure software segment: $8.752B, up 29% YoY, versus prior-quarter guidance of approximately $8.9B, up 31% YoY, a modest miss against the company's own guide; ARR growth 15% YoY, decelerating from 17% in Q2; segment gross margin 94%; segment operating margin approximately 84%, up 650 bps YoY
• Non-AI semiconductor revenue: approximately $4.2B, up 5% YoY, flat sequentially, with broadband and server storage up and wireless down
• Free cash flow: $13.665B, 46% of revenue, up 95% YoY; capex $532M
• Balance sheet: cash $23.975B versus $19.628B prior quarter; inventory $4.523B; gross fixed-rate principal debt $59.6B at a 4% weighted average coupon and 7.4 years to maturity (see note 2)
• Capital allocation: $3.103B in dividends at $0.65 per share; $5.6B of long-term debt retired in quarter, plus $1.5B of senior notes at maturity after quarter end (prepared remarks)
• XPV platform: established June 2026 with Apollo and Blackstone, framed as enabling more than 20 gigawatts of compute for OpenAI and Anthropic by end of 2028; first tranche of $35B closed in June against Anthropic's initial one-gigawatt deployment. The ~$29B maximum-exposure figure from the prior quarterly filing was put on the call by an analyst and confirmed by Thuener as still accurate. She declined an aggregate across future tranches.
• Q4 FY26 guide: revenue approximately $34.8B, up 93% YoY, modestly below the approximately $35.03B LSEG consensus; AI semiconductor revenue $21.7B, up 236% YoY; infrastructure software approximately $8.7B, up 25% YoY; gross margin approximately 73%; operating margin approximately 66%; capex $1.4B
• FY26 AI semiconductor revenue: approximately $58B, above prior guidance of approximately $56B
• Out-year framing: FY27 AI semiconductor revenue approximately $115B, raised from the "in excess of $100B" framing given on the Q2 call; FY28 approximately $230B; management states it is on target to exceed $30 in EPS in FY28
• Product disclosures in quarter: Tomahawk 7 taped out at 200 terabit per second, described as an industry first; Tomahawk Ultra scale-up over Ethernet now shipping; TPU v7 (Ironwood) in volume to Google and Anthropic; TPU v8i in production shipment for Google; Jalapeno shipped for OpenAI; MTIA production shipments to Meta expected in Q4
• Next catalysts: Goldman Sachs Communacopia + Technology Conference appearance Tuesday, September 8, 2026; management currently plans to report fiscal Q4 and full-year results after market close Wednesday, December 9, 2026 (both per IR closing remarks)

Note 1: the CFO stated approximately 76% for semiconductor solutions gross margin in prepared remarks, then corrected to approximately 67% during Q&A. We use the corrected number, which reconciles with the stated 61% segment operating margin against opex at 6% of segment revenue.

Note 2: the $59.6B gross fixed-rate principal figure is from prepared remarks and is not the same measure as the sum of short-term and long-term debt carrying values on the balance sheet, which totals approximately $59.5B.

The News

Broadcom posted fiscal third quarter revenue of $29.59 billion, up 86% year over year, with AI semiconductor revenue more than tripling against the year-ago period. The print cleared consensus on both revenue and adjusted earnings, while the fourth quarter revenue guide landed fractionally under the Street. Reported on the third day of VMware Explore in Las Vegas, management framing pointed well past the quarter, laying out a multi-year custom accelerator trajectory and naming Anthropic as the company's largest XPU customer beginning in fiscal 2027. Full results are in theearnings release.

Analyst Take

Two Broadcoms shared a stage this week and only one of them got discussed. The company reported mid-conference at its own flagship VMware event, and the software business that generates roughly a third of revenue was almost entirely absent from the call's question queue. Meanwhile the market read the $230 billion fiscal 2028 figure as guidance. We would read that differently. Management framed it as a trajectory, stated on the call that it does not intend to update the figure on a quarterly basis, and separately conceded that not every gigawatt in the customer roadmap converts inside the fiscal window.

The counter-read is strong: supply is secured, wafer and substrate and memory commitments are in place, and site-readiness analysis is performed jointly with each customer before the outlook is booked. Fair enough, the Company is doing what it can control and being transparent about the rest because the constraint has largely moved off the die. Broadcom is selling jet engines to airlines that have not finished pouring the runways. Roughly 30 gigawatts of stated customer intent converts, on management's own accounting, to deliberately less.

By offloading infrastructure financing to private credit SPVs and vertically integrating into advanced packaging substrates, Broadcom is systematically de-risking its customer balance sheets and supply bottlenecks to capture a massive custom ASIC expansion. However, this strategy introduces a delicate structural trade-off: tying multi-year revenue recognition directly to hyperscaler capital markets events and non-standard lease vehicles inherently shifts the company's valuation profile from pure semiconductor growth to counterparty credit exposure. From our viewpoint, until these off-balance-sheet financing structures prove resilient through public market listings and the joint Singapore fab achieves sustainable yield, Broadcom's $230 billion fiscal 2028 trajectory remains an intricate function of operational execution and financial engineering rather than simple end-market demand.

What the Numbers Mean

Gross margin compression is the most likely misread in this print. The dilution traces to rising memory content inside the accelerators, not to pricing erosion, and management was unusually blunt in redirecting attention to operating margin, where revenue growth continues to outrun operating expense growth. The arithmetic supports that redirection. It also means gross margin degrades further for as long as the XPU ramp accelerates, which is a peculiar thing to call good news and an accurate thing to call it anyway. Investors should note that the segment gross margin figure was corrected mid-call, from a prepared-remarks number to a materially lower one in Q&A, and the corrected figure is the one that reconciles with the stated operating margin.

Customer concentration is the more consequential disclosure. Six accelerator customers, four of which management describes as heading toward very large scale, and a stated expectation that Anthropic overtakes Google in XPU volume next fiscal year. That compresses the sales cycle and gives Broadcom design visibility years ahead of any merchant supplier. It also means the revenue line now inherits the capital structure of a small number of frontier labs. The XPV vehicle established with Apollo and Blackstone closed a first tranche in June against Anthropic's initial deployment, with third parties underwriting the assets and Broadcom limiting itself to what management characterizes as modest residual value guarantees it views as low risk. We would note the adjective is management's, not ours.

The maximum exposure figure attached to that first tranche in the prior quarterly filing was raised on the call and confirmed as still accurate, and it is not a rounding error against the tranche. The CFO declined to give any aggregate figure across future tranches. Management then made the financing calendar legible in an unusual way, noting that Anthropic's path to public markets would change its credit profile and therefore the structure of subsequent tranches. That ties a supplier's revenue schedule to a datable capital markets event.

Almost nobody asked about VMware, and the quarter gives two reasons we think they should have. Infrastructure software compounded well on subscription conversion, but it printed under the company's own prior-quarter guide and ARR growth decelerated. Broadcom used Explore to reposition VMware Cloud Foundation as the control plane for private AI rather than the successor to a virtualization stack, launching Private AI Cloud and an AI Factory construct alongside agent governance and security tooling. The repatriation argument, moving workloads back from public cloud for cost and control, is the quieter revenue story. It is also now the one with a small blemish nobody asked management to explain.

Callout: content per gigawatt, unreconciled

• Derived from disclosed revenue and gigawatt figures by a sell-side analyst on the call: approximately $11B to $12B of Broadcom content per gigawatt
• Stated by management later in the same call: $20B to $30B of content per gigawatt, described as sustaining, since per-chip power rises with performance and fewer accelerators fit inside a gigawatt
• Gap: roughly 2x, not addressed once both figures were on the table
• Why it matters: the two numbers imply materially different gigawatt counts behind the same revenue outlook, and therefore different sensitivity to deployment slippage

Market Analysis

The same-quarter constellation says something the individual prints do not. NVIDIA's July-quarter data center revenue rose 117% year over year with a sequential guide well above it, and management there flagged memory scarcity as the driver of a gross margin trough later in the fiscal year. Broadcom's own dilution traces to the same input. Two suppliers, different architectures, one shared constraint. That is a supply-chain read rather than a competitive one, and it argues the accelerator market is expanding along both merchant and custom paths rather than trading share between them.

Marvell reinforced the custom silicon signal, raising its fiscal 2028 revenue outlook to approximately $18 billion and pointing to a custom ramp weighted to the back half of its fiscal year. That is the second consecutive quarter of out-year raises from that name. Credo's July-quarter revenue more than doubled year over year on connectivity demand spanning copper and optics.

Which frames the networking argument properly. Broadcom taped out its next-generation Ethernet switch during the quarter, and management stated on the call that customers who are not buying Broadcom XPUs are nonetheless deploying its scale-out Ethernet switching, with scale-up Ethernet now shipping into both XPU and GPU clusters. Networking attach, in other words, does not appear contingent on winning the accelerator socket. That is a materially different competitive position than the custom silicon narrative alone suggests.

The gestalt across these prints is that leading-edge accelerator demand is not the segment's open question. Memory allocation is, and packaging is close behind. Micron's fiscal fourth quarter print, expected later this month, is the one we would circle. Our read, and it is a read rather than a disclosed relationship, is that if HBM and high-capacity server DRAM pricing continues to firm into calendar 2027, every accelerator vendor's gross margin guide including Broadcom's functions as a memory pass-through forecast wearing a semiconductor label.

Looking Ahead

We believe that Broadcom presents a compelling value proposition for partners, prospects, and customers by leveraging its multi-generational silicon design expertise and robust supply chain control to establish sharp multi-year visibility across custom AI accelerators (XPUs) and scale-out networking. Its ability to secure scarce advanced packaging and substrate capacity, demonstrated by vertically integrating its Singapore substrate facility, de-risks physical deployment bottlenecks for hyperscalers and frontier AI labs struggling to scale compute clusters. Furthermore, because Broadcom’s high-margin Ethernet switches and connectivity stack attach frcitionlessly to both custom XPUs and third-party GPU clusters, ecosystem partners can standardize on its networking architecture regardless of their chosen accelerator strategy.

Based on what we are observing, the analytical work in this name shifts from silicon roadmaps to deployment physics and supply ownership. The clearest signal in the quarter may have been the least discussed: Broadcom is bringing substrate capacity up in its Singapore facility in fiscal 2027, built jointly with a partner. A company does not vertically integrate into advanced packaging substrates because the merchant supply chain is working. We read that as a considered judgment that packaging capacity, not lithography, is the binding constraint on custom accelerator volume, and as a position that could become a durable advantage, though nothing is proven until the line is up and yielding.

We will be tracking whether Singapore comes up on schedule, whether the XPV structure attracts further third-party tranches on comparable terms, and whether roadmap gigawatts convert at the discounted rate management implied. Two near-term markers: the company appears at Goldman Sachs Communacopia on Tuesday, September 8, where the content-per-gigawatt question deserves a second asking, and management currently plans to report fiscal Q4 on December 9.