An anomaly is just a story waiting to be read. On a quiet Tuesday, a sell-side report from Wolfe Research dropped a number that rippled through the semiconductor ecosystem: Broadcom's AI revenue could reach $200 billion by 2028. That is 8x the current consensus estimate of ~$24B for FY2025 and more than double NVIDIA's entire FY2024 revenue of $60.9B. The anomaly is not the figure itself—it is the chain of assumptions required to make it plausible. As an on-chain analyst who has spent years tracing transaction flows, I know that extreme outliers in data often signal a flaw in the model, not a revelation of the truth. Let me trace the past to understand this prediction.
Context: Broadcom's AI Revenue Foundation Broadcom's AI semiconductor business currently rests on two pillars: custom AI accelerators (ASICs/XPUs) and high-speed networking chips (Tomahawk/Jericho Ethernet switches). Its primary customers are hyperscale cloud providers—Google (TPU), Meta, and potentially Microsoft. In FY2024, Broadcom's AI revenue was approximately $12B; FY2025 street estimates range from $20-24B. The Wolfe Research target of $200B implies a compound annual growth rate of roughly 70-90% over three years. For perspective, no semiconductor company in history has scaled revenue at this pace from such a base. NVIDIA's explosive growth from $27B to $130B over 2023-2025 (4.8x) was the largest in the industry, yet Wolfe is projecting 8.3x for Broadcom.
Core: The On-Chain Evidence Chain Every transaction leaves a scar; I map the wound. To validate the $200B prediction, we must examine the physical constraints that act as immutable ledger entries. First, wafer capacity. TSMC's 3nm/5nm capacity in 2025-2026 is roughly 150-180 million 12-inch equivalent wafers per year. NVIDIA consumes 30-40%, Apple 20-30%. To deliver $200B in AI revenue, Broadcom would need approximately 500,000-600,000 high-end ASIC chips per year (assuming a $4-5K ASP). That translates to 50-60 million wafers annually at 3nm, requiring 30-40% of TSMC's total advanced capacity—a share that would compete directly with NVIDIA and Apple. Second, CoWoS packaging. TSMC's CoWoS capacity is the bottleneck for AI accelerators. Current monthly capacity is 40-60k units; NVIDIA takes 60%+. To support $200B revenue, Broadcom would need 100-150k units per month, implying a 2.5-3x expansion of CoWoS by 2028—a heavy but not impossible lift. Third, HBM supply. SK Hynix, Samsung, and Micron produce roughly 50-60 billion GB-equivalent HBM in 2025; NVIDIA consumes 70%+. Broadcom's ASICs would require 20-30% of total HBM supply, necessitating massive new fab investments with 2-3 year lead times. Fourth, power. The compute equivalent of $200B in AI chips would require 100-200 GW of power, far exceeding the current global data center consumption of ~500 TWh/year (AI portion ~100 TWh). Grid infrastructure expansion is measured in decades, not fiscal quarters.
Beyond physical constraints, there is the customer concentration risk. Broadcom's top AI customer (Google) accounts for over 50% of its AI revenue. To reach $200B, Google alone would need to purchase ~$100B in custom chips—roughly 30% of Google's total 2024 revenue of $350B. The remaining $100B would require 5-8 other hyperscalers each contributing $20-30B annually. The universe of entities capable of such spending is fewer than 10 globally. Even if OpenAI's rumored partnership with Broadcom materializes, the revenue contribution by 2028 would likely be $10-20B, not $50B+.
Contrarian: Correlation ≠ Causation I do not predict the future; I trace the past. The contrarian angle is that the $200B prediction might be a self-fulfilling signal if interpreted correctly. Sell-side analysts often produce "blue-sky" scenarios to test the upper bound of a stock's optionality. The real value of the report is not the number itself but the framework it implies: that AI infrastructure spending could sustain 40%+ CAGR through 2028. However, the correlation between capital expenditure and actual AI revenue is breaking. In Q1 2025, cloud providers' AI revenue growth lagged their capex growth by 15-20 percentage points. If this gap persists, the capex cycle will peak in 2026-2027, not 2028. The Wolfe prediction assumes the gap closes, not widens.
Another blind spot: NVIDIA's competitive response. The CUDA ecosystem remains the deepest moat in AI compute. If NVIDIA launches Rubin Ultra in 2026 with 2x performance-per-watt over current ASICs, the incentive for hyperscalers to use custom silicon weakens. Broadcom's ASIC advantage is strongest in inference, but NVIDIA is already deploying TensorRT and Triton to compress that gap. The assumption that ASICs will capture 30%+ of training workloads is heroic.
Takeaway: The Signal to Track The pattern emerges only after the dust settles. The next 12 months will reveal whether the Wolfe prediction is a mirage or a map. Track two metrics: cloud providers' AI revenue growth vs. capex growth (if the gap exceeds 20% for two consecutive quarters, the capex cycle is topping), and TSMC's CoWoS allocation commentary on earnings calls. If NVIDIA locks 80%+ of advanced packaging, Broadcom's ceiling is $60-80B, not $200B. The on-chain data—in this case, the physical supply chain—will tell the true story. Until then, $200B remains a story waiting to be read, not a transaction to be trusted.