Wolfe Research drops a number: $200 billion in AI revenue for Broadcom by 2028. That's not a forecast. It's a statement of faith.
I've seen this pattern before. In 2020, a sell-side note on a DeFi protocol claimed $10B in TVL by 2022. The actual number? $2.5B. The gap between hype and reality is where the smart money gets out.
Let's start with context. Broadcom's current AI revenue sits around $20-24B in fiscal 2025. That's from custom ASICs for Google, Meta, and a few other hyperscalers, plus network chips like Tomahawk and Jericho. The business is real. But the jump to $200B implies a 10x growth in three years. No semiconductor company has ever done that. Not NVIDIA, not Intel. NVIDIA's AI revenue grew from $27B to $130B in two years—about 5x. That was the mother of all demand spikes. Broadcom's target is 8x. The math doesn't add up without a second miracle.
Here's the core analysis. The global AI semiconductor market in 2028 is estimated at $250-300B. For Broadcom to hit $200B, it would need to capture 67-80% of that market. That's absurd. NVIDIA still holds 80-90% of the AI accelerator market. The idea that Broadcom can flip that in three years ignores the CUDA moat, the software ecosystem, and the fact that custom ASICs are only competitive in specific workloads—mostly inference. Training? That's still NVIDIA's playground.
Now look at the physical constraints. Broadcom's custom chips require TSMC's most advanced nodes (3nm/2nm) and CoWoS packaging. TSMC's CoWoS capacity in 2025 is about 4-6k wafers per month. NVIDIA takes 60%+ of that. To support $200B in revenue, Broadcom would need 10-15k wafers per month. That's a 2.5-3x increase in total CoWoS capacity, which TSMC is building, but not fast enough. And HBM? SK Hynix and Samsung are already maxed out serving NVIDIA. Broadcom's HBM demand would require another 20-30% of global supply. That's a multi-year investment cycle.
Then there's power. $200B in AI chips implies an installed base consuming 100-200 GW of electricity. That's more than the entire current global data center footprint. Grid upgrades take a decade, not three years.
The real number is $60-100B. That's the feasible range for Broadcom's AI revenue in 2028, based on customer concentration, capacity constraints, and market share limits. Wolfe's $200B is a bull case scenario—the kind analysts use to get attention, not to guide capital allocation.
Here's the contrarian angle. The market is euphoric because AI capex is exploding. But the gap between AI infrastructure spending and AI revenue is widening. Hyperscalers are spending $50B+ per quarter on AI, but their AI revenue growth is slowing. At some point, CFOs will demand ROI. When that happens, the capex cycle turns. Broadcom's prediction is a bet that the cycle never turns. That's not how cycles work.
Blind spot: customer concentration. Google alone accounts for over 50% of Broadcom's AI revenue. To reach $200B, Google would need to spend $100B on Broadcom chips in 2028. That's 30% of Google's total 2024 revenue. Impossible. The prediction assumes 5-8 hyperscalers each spending $20-30B annually. But there are only a handful of companies with that kind of budget. And each is building its own chips—Google TPU, AWS Trainium, Microsoft Maia. Broadcom is a supplier, not a monopoly.
Signature insight: "The spread was real, but the exit was imaginary." This prediction is a spread trade on hype. The exit will come when the data disagrees.
Let me ground this with my own experience. In 2022, I watched the Terra collapse from my terminal. The on-chain data showed the decoupling days before the price hit zero. I sold my UST position in stages, saving 60% of capital. The lesson: trust the data, not the narrative. The same applies here. Track the on-chain metrics: TSMC's CoWoS allocation, hyperscaler capex guidance, AI revenue growth rates. These will tell you the truth long before Wolfe revises its target.
Another signature: "Alpha decays faster than the code that finds it." The alpha in this prediction is already decaying. By the time retail catches on, the smart money will have already priced in the $60-100B range.
Now, the infrastructure angle. The $200B prediction ignores the physical limits of silicon. Chip supply is not elastic. It takes years to build fabs, secure HBM, and install power. The market is already seeing lead times stretch for CoWoS and HBM. If Broadcom real demand were to materialize, the supply chain would choke. That's not a bullish signal; it's a bottleneck.
Final signature: "I trust the log, not the hype." The log shows the real numbers: $20B today, $60-100B in 2028. The hype says $200B. I know which one I'm trading against.
Takeaway: Wolfe's prediction is a useful sentiment indicator. It tells you that the market is pricing in extreme optimism. But as a trader, you don't trade the headline. You trade the divergence. The divergence between the $200B dream and the $60-100B reality will create opportunity. When the first quarterly miss comes, the spread will collapse. That's your exit.
Watch for these signals: Broadcom's next fiscal 2026 guidance (expected late 2025). If it's below $30B, the $200B case is already dead. Also track TSMC's CoWoS allocation: if NVIDIA secures 80%+ of capacity, Broadcom's upside is capped. And finally, watch hyperscaler AI revenue growth. If it falls below 30% YoY, the capex party is over.
The blind spot is where the money hides. The blind spot here is the assumption that AI infrastructure spending can grow indefinitely without corresponding application revenue. It can't. The math always wins.