Hock Tan has extended Broadcom’s custom-chip business into an unusual new role: coordinating technology, customers and private capital for artificial-intelligence infrastructure measured in gigawatts. In June, Broadcom, Apollo and Blackstone established a platform intended to enable more than 20 gigawatts of compute through 2028. An initial $35 billion transaction will finance more than one gigawatt for Anthropic at sites operated by Fluidstack.
The arrangement moves Broadcom closer to the economics of its largest customers without requiring the company to own data centres or fund them from its balance sheet. Its processors and networking become the technical base; asset managers provide capital; specialised operators build and run facilities; frontier laboratories commit workloads. Tan is turning the company’s ability to design workload-specific silicon into the anchor for an infrastructure ecosystem.
Two weeks later, Broadcom and OpenAI unveiled Jalapeño, OpenAI’s first intelligence processor. The accelerator was developed from design to production in nine months, according to the companies, and is intended for gigawatt-scale deployment over several generations. The sequence made Tan’s strategic claim unusually concrete. Broadcom is no longer merely supplying components behind another company’s branded system. It is helping laboratories define, finance and deploy their own computing platforms.
Custom silicon becomes the growth engine
Broadcom’s fiscal second-quarter results show why the model has captured investor attention. Revenue for the period ended 3 May 2026 rose 48 per cent to $22.2 billion. Semiconductor revenue associated with AI reached $10.8 billion, 143 per cent higher than a year earlier, driven by custom accelerators and networking. Tan expects that figure to reach $16 billion in the third quarter, growth of more than 200 per cent.
The group guided to total third-quarter revenue of about $29.4 billion and an adjusted EBITDA margin near 68 per cent. In the latest quarter, adjusted EBITDA was $15.2 billion, equivalent to 69 per cent of revenue. Free cash flow reached $10.3 billion, or 46 per cent of revenue, after only $231 million of capital expenditure. Those numbers illustrate the attraction of Broadcom’s fabless structure: it captures intellectual-property and platform value while manufacturing partners carry much of the physical capital burden.
Tan has spent years concentrating Broadcom on markets where a small number of technically demanding customers will pay for differentiated, mission-critical products. Custom AI processors fit that discipline. A frontier laboratory can optimise an accelerator around its models, memory patterns and inference requirements, potentially reducing power and cost relative to general-purpose hardware. Broadcom supplies design expertise, high-speed interconnect and access to an advanced manufacturing network.
The model challenges Nvidia from a different direction than AMD. Broadcom does not need to persuade millions of developers to adopt a general accelerator platform. It co-designs hardware with a customer that already controls the workload and can commit enormous volume. The resulting processor may have limited value outside that relationship, but the engineering, packaging and networking knowledge can be applied to later engagements.
Jalapeño demonstrates the speed such cooperation can deliver. OpenAI contributed its model requirements and used its own systems to accelerate design work, while Broadcom brought the chip to production. Early testing is said to show better performance per watt than existing alternatives, though independent operating evidence will matter more than launch claims. The real test is stable output across large clusters and a lower cost for useful inference after software, memory, networking and power are included.
Financing solves one bottleneck and creates another
The AI XPV platform addresses the funding requirement behind those clusters. Training and serving frontier models now requires data-centre projects so large that technology companies, laboratories, utilities and private-credit providers must coordinate years ahead. Apollo and Blackstone can supply long-duration capital to facilities backed by customer commitments. Broadcom gains greater certainty that demand for its processors will be matched by buildings and power.
The initial transaction is designed around Anthropic’s previously announced capacity expansion, with more than one gigawatt expected to begin deployment from mid-2026. A framework covering more than 20 gigawatts by 2028 suggests repeatability, including potential use by OpenAI. Yet the scale should be understood as an ambition, not booked Broadcom revenue. Sites must receive permits and power, financing conditions must hold, processors must ship and customers must consume the capacity.
Project finance also moves some risk rather than eliminating it. Credit investors will evaluate the strength and duration of laboratory contracts, residual value of specialised equipment and exposure to technology obsolescence. An accelerator optimised for one model family may be harder to redeploy than a standard server. If model efficiency improves faster than demand grows, a facility conceived during scarcity could open into a different market.
Tan’s position is protected by selling high-value technology rather than owning the asset, but Broadcom still depends on project completion. Delayed power connections or a financing retrenchment can postpone chip revenue. The company’s exceptional forecast is increasingly tied to a handful of laboratories and hyperscalers making capital decisions on a scale few businesses can sustain. Customer intimacy is a moat until one road map changes.
Concentration is built into the strategy
Broadcom’s second-quarter semiconductor solutions revenue was $15.0 billion, representing 68 per cent of the group. AI contributed most of that segment and is rising faster than the rest of the portfolio. The concentration improves growth and margins today. It also makes quarterly performance sensitive to deployment schedules, product acceptance and purchasing by a limited set of buyers.
Tan accepts concentration when it is supported by long design cycles and deep engineering ties. A custom processor can take years to architect and integrate, creating switching costs. The new OpenAI chip reached production unusually quickly, but the companies describe a multi-generation relationship. Meta has also expanded its work with Broadcom to support several gigawatts of its MTIA custom silicon. Each win embeds Broadcom farther into a customer’s infrastructure plan.
The bargaining power runs both ways. Large buyers can demand custom economics, supply priority and road-map influence. They may also develop more design capability internally or use alternative partners. Broadcom needs enough simultaneous programmes to spread research costs without overextending a finite pool of specialised engineers. A failure for one marquee customer would carry financial and reputational consequences far beyond a conventional product miss.
Competition is expanding. Nvidia is integrating processors, networking, storage and software into its DSX factory architecture. AMD is winning multi-gigawatt commitments for its Helios and MI450 platform. Marvell and other designers pursue custom opportunities, while hyperscalers invest in internal silicon teams. Broadcom’s advantage lies in proven interfaces, packaging and delivery, not an exclusive claim on the concept.
Asia manufactures the capital-light model
Broadcom’s light internal capital expenditure is possible because an Asian manufacturing network performs the expensive work. Advanced foundries, packaging providers, memory makers and system assemblers turn its designs into physical infrastructure. Taiwan is central to leading-edge fabrication and packaging; Korea supplies critical memory; Singapore is a major operational and billing hub. Broadcom reported $10.8 billion of fiscal 2025 revenue attributed to Singapore based on customer location, although such geographic data do not necessarily identify final consumption.
The network provides technical depth and flexible scale, but it creates a point of vulnerability precisely when customers want certainty. Advanced packaging and high-bandwidth memory remain constrained. Earthquakes, trade restrictions, shipping disruption or a conflict around Taiwan could break schedules across several customers at once. Project financiers may regard supply assurance as seriously as credit quality when underwriting gigawatt deployments.
Broadcom is broadening relationships across the region. It began shipping a 2-nanometre custom system using advanced three-dimensional packaging for Fujitsu, supporting a next-generation Japanese processor. It is collaborating with Samsung on an integrated 5G and Wi-Fi 8 platform. These projects diversify application exposure and deepen engineering links, but they do not remove dependence on a relatively small number of cutting-edge production facilities.
Trade policy adds complexity. American controls can restrict the processors delivered to China, while customers and suppliers must comply with rules that change faster than chip-development cycles. Broadcom sells into global infrastructure and has significant Asian commercial exposure. Tan must reserve capacity and commit designs years ahead without knowing the precise market-access conditions at shipment.
VMware supplies cash and an enterprise route
Tan’s semiconductor surge sits beside a large infrastructure-software business created by the VMware acquisition. Software revenue was $7.2 billion in the latest quarter, 9 per cent higher than a year earlier. Broadcom has focused the portfolio on private cloud, security and developer platforms, arguing that enterprises will run sensitive inference and agents across their own environments as well as public clouds.
The strategic connection is plausible. VMware Cloud Foundation can provide a governed operating layer for AI workloads, while Tanzu supports the development and monitoring of enterprise agents. Private deployments may use Broadcom networking and, over time, custom or merchant accelerators. The software also supplies recurring cash flow that helps finance semiconductor research and shareholder returns.
Execution requires more than cross-selling. VMware customers judge licensing, support, portability and total cost over long periods. Aggressive portfolio simplification can improve near-term profitability while encouraging some customers to examine alternatives. Tan must show that private-cloud innovation creates enough value to preserve trust, particularly as enterprises want flexibility across Nvidia, AMD, custom silicon and multiple public clouds.
Broadcom’s latest numbers validate Tan’s concentration strategy: record revenue, operating profit and free cash flow, followed by extraordinary guidance. The next proof sits outside the income statement. Jalapeño and the Anthropic build must operate at scale, the financing platform must convert announced gigawatts into powered capacity, and VMware must remain a durable enterprise base. If those pieces align, Broadcom will have become an orchestrator of AI infrastructure without owning the factories. If one fails, the close proximity to a few customers will transmit the shock with equal efficiency.
