s silence.
Wolfe Research dropped a number that made the market pause: Broadcom AI revenue could reach $200 billion by 2028. That's more than 1.5 times NVIDIA's entire fiscal 2024 revenue. It's a number that demands a forensic look, not applause.
I've spent years reconstructing ledgers—from ICO wash-trading patterns to LUNA's liquidity divergence. The same principle applies here: the data must speak, not the narrative. And the data on Broadcom's $200B prediction reveals a structural gap between hype and hardware reality.
Context: The Broadcom AI Stack
Broadcom isn't NVIDIA. It doesn't sell a single GPU for every AI workload. Instead, it dominates two niches: custom AI accelerators (ASICs/XPUs) for hyperscalers like Google and Meta, and high-speed networking chips (Tomahawk, Jericho) that connect massive GPU clusters. In fiscal 2025, its AI semiconductor revenue is expected around $20-24B. That's a real business, growing fast, but it's also a concentrated one—Google alone accounts for over 50% of Broadcom's AI chip revenue.
Wolfe Research's prediction implies a jump from $24B to $200B in three years. That's an 8.3x compound annual growth rate, with no historical precedent in the semiconductor industry. NVIDIA's own hypergrowth from $27B to $130B was driven by the once-in-a-decade demand explosion for LLM training. Broadcom's path requires a second, even larger explosion, but with a different set of physics.
Core: The On-Chain Evidence Chain (Physical Constraints)
Let's treat the supply chain as an immutable ledger. The bottlenecks are not opinions; they are capacity constraints written in wafers and power lines.
First, TSMC's CoWoS advanced packaging. In 2025, the total monthly capacity is roughly 4-6K wafers. NVIDIA consumes 60%+ of that. Broadcom's XPU and TPU products also require CoWoS. To support $200B in revenue, Broadcom would need a monthly allocation of 10-15K wafers—a 2.5-3x increase in total CoWoS capacity by 2028. That's not impossible, but the timeline conflicts with NVIDIA's own expansion plans. TSMC's capacity allocation will prioritize the highest-margin customer, and NVIDIA's GPU margins are 40-60% higher than Broadcom's ASIC margins.
Second, HBM memory. Every AI chip needs high-bandwidth memory from SK Hynix, Samsung, or Micron. In 2025, the total HBM supply is about 50-60B GB, with NVIDIA taking 70%+. Broadcom's $200B revenue would require an additional 20-30% of global HBM production—demanding new factories that take 2-3 years to build. The lead time alone makes the 2028 target unrealistic.
Third, power. The chips needed to generate $200B in revenue would consume 100-200 GW of electricity. That's more than the entire global data center power consumption in 2024. Grid infrastructure doesn't scale at chip speed.
Contrarian: Correlation ≠ Causation
The counter-intuitive angle is that Wolfe Research's prediction may be a self-fulfilling signal, not a forecast. If hyperscalers believe the $200B target, they will sign long-term contracts with Broadcom, locking in capacity and validating the demand. But the data shows a different story: the gap between AI infrastructure capex and AI revenue from applications is widening. Cloud providers are spending on AI hardware faster than they are earning from AI services. This is the same pattern I saw in the LUNA model—a divergence between a narrative (stablecoin peg) and the underlying liquidity (reserves). When the divergence exceeds a threshold, the correction is abrupt.
Furthermore, Broadcom's customer concentration is a risk the prediction ignores. If Google alone accounts for 50% of current AI revenue, $200B implies Google would need to buy $100B in Broadcom chips by 2028. That's 30% of Google's total 2024 revenue. No hyperscaler has ever allocated that much of its revenue to a single supplier. The probability is near zero.
Takeaway: The Signal to Watch
I don't write off the AI infrastructure boom. But I do write off predictions that ignore physical constraints. The real signal for investors is not the $200B number, but Broadcom's quarterly AI revenue growth. If it falls below 60% year-over-year in the next two quarters, the Wolfe prediction is already off the table. Watch TSMC's CoWoS allocation announcements. Watch the power grid buildout. And remember: the ledger never lies, but narratives do.