Hock Tan does not sell chips so much as he sells inevitability. On Wednesday, the Broadcom CEO stood before the market and projected that his company’s AI-related revenue would eclipse $100 billion by 2027. This is not a mere forecast. It is a declaration of a structural monopoly over the plumbing of the intelligence age. For the first time, Tan pulled back the curtain on a customer list that reads like a ledger of the world’s most powerful compute engines: Google, Meta, ByteDance, Anthropic, and now, finally confirmed, OpenAI.
The math is staggering. Broadcom expects $10.7 billion in AI semiconductor revenue for the second quarter of fiscal 2026 alone, a 140% surge year-over-year. But to understand why Broadcom is winning, one must look past the top-line numbers and into the brutal economics of custom silicon. While Nvidia captures the headlines with general-purpose GPUs, Broadcom has quietly cornered the market for Application-Specific Integrated Circuits (ASICs)—bespoke processors designed to do one thing with terrifying efficiency.
The Custom Silicon Trap
The "why" behind the $100 billion target is simple but profound: the world’s largest tech companies have grown tired of paying the "Nvidia tax." When Google builds a Tensor Processing Unit (TPU) or Meta develops its MTIA chip, they are attempting to claw back the massive margins currently flowing to Santa Clara. However, building a high-end AI chip from scratch is a fool's errand for most. You need IP for high-speed memory interfaces, PCIe lanes, and, most importantly, the networking fabric to connect thousands of these chips together.
Broadcom provides the skeleton; the customers provide the specific "brain" logic. This co-design model creates a level of customer stickiness that is almost impossible to break. Once Google integrates Broadcom’s intellectual property into the TPUv7 (codenamed Ironwood), switching to another partner isn't just a matter of price—it’s a multi-year re-engineering nightmare.
The 10 Gigawatt Reality Check
During the earnings call, Tan introduced a metric that most analysts missed in the scramble for revenue figures: 10 gigawatts. By 2027, Broadcom expects its six strategic customers to have an installed base of compute capacity reaching that threshold. To put that in perspective, 10 gigawatts is roughly the output of ten nuclear power plants, all dedicated to processing prompts and training models.
The industry standard valuation for 1 gigawatt of AI infrastructure is approximately $200 billion. If Broadcom’s clients are indeed deploying 10 gigawatts, the total addressable market is $2 trillion. In that context, Tan’s $100 billion revenue goal is actually a conservative estimate of Broadcom’s take-rate. They aren't just selling chips; they are taxing the very electricity that fuels the AI revolution.
The VMware Tax and the Software Stall
However, the Broadcom story is not without its cracks. While the semiconductor division is screaming ahead, the Infrastructure Software segment—the home of the $61 billion VMware acquisition—is showing signs of digestion issues. Revenue in this segment grew a mere 1% in the first quarter of 2026.
Broadcom’s strategy with VMware has been characteristically ruthless: kill off perpetual licenses, force customers into expensive subscription bundles, and focus only on the top 2,000 global enterprises. It is a playbook designed to maximize cash flow, but it is also driving a quiet exodus. Gartner projects that by the end of 2026, half of all VMware enterprises will be running proofs-of-concept for alternative platforms.
Some customers have reported price hikes of over 300% since the acquisition. This creates a precarious tension within Broadcom's balance sheet. The high-margin software cash is meant to fund the massive R&D required for 2nm compute SoCs and next-generation 102.4 Tbps Tomahawk switches. If the VMware "cash cow" begins to bleed customers faster than Tan can raise prices, the engine of Broadcom’s AI expansion could lose its fuel.
The Networking Moat
If the custom chips (XPUs) are the stars of the show, the networking business is the director. AI clusters are only as fast as the wires between them. Broadcom’s dominance in Ethernet switching is the secret weapon that separates it from competitors like Marvell or even a resurgent Cisco.
The transition to 1.6T Ethernet is the next major battlefield. As AI models grow, the "tail latency"—the time it takes for the slowest piece of data to travel across the network—becomes the primary bottleneck. Broadcom’s Tomahawk and Jericho chipsets are the only products currently capable of handling the massive data throughput required for trillion-parameter models. Tan noted that AI networking now represents nearly 40% of his AI revenue, a clear sign that the physical "pipes" of the internet are being rebuilt from the ground up to accommodate machine intelligence.
The China Contradiction
One overlooked factor in the $100 billion projection is the role of ByteDance and the broader Chinese market. Despite escalating U.S. export controls, Broadcom continues to navigate a razor-thin line. ByteDance remains a core customer for custom silicon, and any further tightening of trade restrictions could immediately jeopardize a significant chunk of the 2027 forecast.
Furthermore, Beijing has begun instructing state-owned enterprises to phase out Western virtualization software—a direct hit to VMware’s footprint in the region. Broadcom is attempting to counter this with a "Sovereign AI" strategy, helping nations build domestic clusters that theoretically bypass some geopolitical friction. Whether this will satisfy regulators in Washington or Beijing remains a high-stakes gamble.
The Concentration Gamble
Broadcom’s AI future currently rests on the shoulders of just six companies. This is an extraordinary concentration of risk. If OpenAI pivots its strategy, or if Google successfully moves more of its design "in-house" (using what the industry calls Customer-Owned Tooling), Broadcom’s $100 billion dream could evaporate.
Tan dismissed these concerns with the weary confidence of a man who has seen every semiconductor cycle since the 1970s. He argues that the complexity of 3.5D packaging and the thermal challenges of 2nm silicon make it impossible for even a trillion-dollar company to go it alone. He is betting that the "physics" of chip design will keep his customers captive.
The sheer scale of the investment is the ultimate barrier to entry. Broadcom spent $1.5 billion on R&D in a single quarter. They are not just building chips; they are building the industrial machinery required to build chips.
The $100 billion figure is a massive number, intended to cow the competition and entice the institutional investor. But it is also a reminder that in the AI era, the winners aren't just the ones with the best algorithms. They are the ones who own the silicon, the software, and the switches. Hock Tan has positioned Broadcom to be the landlord of the data center, and the rent is about to go up.
Would you like me to analyze the specific technical advantages of Broadcom's new 2nm custom SoC compared to the current 3nm generation?
