Broadcom’s current AI semiconductor revenue sits at roughly $20–24 billion. Wolfe Research projects that figure to hit $200 billion by 2028. The ledger remembers what the headline forgets.
Context: The Prediction and Its Source Wolfe Research, a sell-side firm, published a report claiming Broadcom could see $200B in AI revenue by 2028. The projection is based on the assumption that the custom ASIC (XPU) and networking chip business will capture a dominant share of the exploding AI infrastructure market. Broadcom’s key customers include Google (TPU), Meta, and potentially Microsoft and OpenAI. The report, as relayed by Crypto Briefing, frames this as a “significant reshaping of Broadcom’s business model.”
But numbers divorced from physical constraints are just noise. Pics are noise; the hash is the identity. Let us examine the supply chain.
Core: The Physical Impossibility of $200B To reach $200B in AI revenue by 2028, Broadcom must ship approximately 400–500 million custom AI chips annually (assuming an average selling price of $4,000–5,000). That requires about 50,000–60,000 12-inch wafers per year on TSMC’s 3nm or 2nm process, all with CoWoS advanced packaging. TSMC’s total advanced process capacity in 2025–2026 is roughly 150,000–180,000 wafers per year. NVIDIA already consumes 30–40%, and Apple takes another 20–30%. The remaining capacity cannot support Broadcom’s target.
CoWoS packaging is the bottleneck. TSMC’s monthly CoWoS capacity in 2025 is about 40,000–60,000 units. NVIDIA uses over 60%. To support $200B in Broadcom revenue, CoWoS capacity would need to triple to 150,000–200,000 units per month by 2028—a physically and logistically improbable expansion.
HBM (high-bandwidth memory) supply is another constraint. SK Hynix, Samsung, and Micron control global HBM output. In 2025, total HBM supply is about 50–60 billion GB, with NVIDIA consuming over 70%. Broadcom’s $200B scenario would require an additional 20–30% of global HBM—demanding billions in new fab investment with a 2–3 year lead time.
Customer concentration compounds the problem. Google alone accounts for over 50% of Broadcom’s AI revenue today. To hit $200B, Google would need to purchase roughly $100B in custom chips—equivalent to 30% of its 2024 total revenue. That is not a forecast; it is a fantasy. Silence in the code speaks louder than the pitch.
Contrarian: What the Bulls Got Right The bulls are correct that Broadcom’s custom ASIC route is validated by Google’s TPU deployments. The networking business (Tomahawk, Jericho) benefits from the shift from InfiniBand to Ethernet in AI clusters. These segments are growing at 40–60% CAGR. The $200B number, however, is not a base case—it is a tail scenario that requires every variable to break perfectly: no cyclical downturn in AI capex, no supply chain bottleneck, no competitive response from NVIDIA.
NVIDIA’s upcoming Rubin Ultra architecture could maintain a 1.5–2x performance advantage over ASICs in training workloads. If that happens, ASIC adoption in training stays below 15%, capping Broadcom’s upside. The bull case also ignores that custom ASIC margins are lower than GPU margins, reducing the economic incentive for Broadcom to scale volume at any cost.
Takeaway The realistic range for Broadcom’s 2028 AI revenue is $60–100B—30–50% of Wolfe’s prediction. The $200B figure is a self-serving narrative designed to fuel market enthusiasm. Investors should track the gap between AI capex and AI revenue growth, and monitor TSMC’s capacity allocation. The ledger remembers what the headline forgets. The chain is both map and territory.