The ledger never lies, only the narrative obscures.
A Wolfe Research note has surfaced claiming Broadcom (AVGO) could see $200 billion in AI revenue by 2028. That number is not a forecast—it is a fantasy built on a stack of unverified assumptions. After spending years auditing on-chain data during the 2017 ICO boom and modeling yield sustainability in DeFi Summer, I've learned that extreme projections usually collapse under the weight of their own physics.
Let me state this clearly: Broadcom's current AI revenue is roughly $20-24 billion for fiscal 2025. To reach $200 billion in three years requires a compound annual growth rate of 70-90%. No semiconductor company has ever scaled that fast. NVIDIA's meteoric rise from $27 billion to $130 billion in two years was driven by the largest demand explosion in tech history—and even that was only a 4.8x multiple. Broadcom would need an 8.3x multiple on a larger base.
Context: The Two-Pronged Engine
Broadcom's AI business is built on two pillars: custom AI accelerators (ASICs/XPUs) for hyperscalers like Google and Meta, and high-speed networking chips (Tomahawk, Jericho) that connect GPU clusters. The custom ASIC route has been validated—Google's TPU series, designed with Broadcom, powers a significant portion of its AI workloads. But this is a niche play. The total addressable market for custom AI chips in 2028 is estimated at $250-300 billion. Wolfe's prediction implies Broadcom capturing 67-80% of that market. For context, Broadcom's current share of the custom ASIC design services market is roughly 55-60%. That is not a monopoly—it's a competitive arena where Marvell, Alchip, and in-house teams at AWS and Microsoft are all contending.
Core: The Physical Constraints
Correlation is a suggestion; causality is a truth.
I built a Python script during the 2021 NFT whale tracking saga to map 500,000 transactions and expose wash trading. That experience taught me to trust supply-chain data over hype. Let's apply that rigor here.
First, wafer capacity. TSMC's 3nm/5nm capacity in 2025-2026 is about 150-180 thousand wafers per year. NVIDIA consumes 30-40% of that, Apple another 20-30%. To ship the roughly 500,000 custom AI chips needed to support $200 billion in revenue (assuming an average selling price of $4-5K per chip), Broadcom would require 50-60 thousand wafers annually—on top of its existing networking products. That leaves almost no room for other customers. TSMC allocates capacity based on profit margins, and NVIDIA's GPUs generate higher revenue per wafer than Broadcom's ASICs. Broadcom will not get priority.
Second, CoWoS packaging. TSMC's advanced packaging capacity is the most severe bottleneck. Current monthly output is around 40-60 thousand wafers, with NVIDIA taking over 60%. To hit $200 billion, Broadcom would need 100-150 thousand wafers per month—a 2.5-3x expansion of the entire CoWoS ecosystem. This is physically impossible within three years, even with aggressive investment.
Third, HBM memory. SK hynix, Samsung, and Micron control the supply of high-bandwidth memory. In 2025, total HBM output is about 50-60 billion GB-equivalents, with NVIDIA consuming over 70%. Broadcom's ASICs would require an additional 20-30% of global HBM supply—demanding billions in new fabrication plants that take 2-3 years to come online.
Fourth, power. The electricity required to run the chips implied by $200 billion in AI revenue would be on the order of 100-200 GW. That is more than the entire current global data center power consumption. Grid infrastructure cannot scale that fast.
Contrarian: The Narrative Trap
Whales don't buy the narrative; they follow the flow.
Wolfe's prediction is likely a "bull case" scenario designed to stimulate institutional interest, not a baseline expectation. The note conveniently omits the most critical risk: the AI infrastructure capex cycle may peak by 2027-2028. Cloud providers are already spending more on AI hardware than they earn from AI services. The gap between AI revenue and capex is widening. If application-layer revenue fails to catch up, capital expenditure will be slashed, and Broadcom's revenue will fall far short of $200 billion—probably landing in the $60-100 billion range.
Furthermore, the prediction assumes that Broadcom's top customers—Google, Meta, Microsoft—will continue to rely on custom ASICs rather than develop their own in-house chips. But Google already designs its own TPU architecture; Broadcom is just a design services partner. Microsoft's Maia chips are increasingly self-owned. The moment a hyperscaler decides to bring more design in-house, Broadcom's revenue takes a step-function decline.
Takeaway: The Signal in the Noise
Trust the hash, not the headline.
The $200 billion figure is a useful indicator of market sentiment, not a realistic target. The real signal to watch is Broadcom's own guidance for fiscal 2026, expected in late 2025. If management guides for AI revenue below $30 billion, then even Wolfe's own analysts are selling a story, not a forecast. The physical constraints of silicon, packaging, and power will not bend to a research note. The ledger never lies.
