The number is out. $200 billion. By 2028. For Broadcom's AI revenue. And the market is drinking it like oxygen.
Let me stop you right there. I've been tracking this space since 2017—when I modeled Filecoin's storage capacity against market hype in four hours flat. I've seen sell-side analysts throw out numbers that make headlines, not sense. This one? It's a masterpiece of wishful thinking wrapped in a CAGR calculation.
Context: The Current Reality
Broadcom's AI revenue today sits around $20-24 billion for fiscal 2025. That's not small—it's a beast. But it's driven by custom ASICs for Google (TPU v5/v6) and Meta, plus network chips (Tomahawk, Jericho) that power the ethernet backbones of hyperscale GPU clusters. The company is a legitimate player in the AI infrastructure game. But the gap between $24B and $200B is not a linear step—it's a chasm.
Wolfe Research's prediction, as reported by Crypto Briefing, implies a compound annual growth rate of roughly 70-90% for three years. In semiconductor history, no company has achieved that from a $20B+ base. NVIDIA's own meteoric rise—from $27B in FY2023 to $130B in FY2025—was about 4.8x. Broadcom would need 8.3x. That's not just a different order of magnitude. It's a different universe.
Core: The Physics of $200B
Let's break down what $200B in AI revenue actually requires.
First, the chip count. At an average selling price of $4,000-5,000 per custom AI accelerator (ASIC), Broadcom would need to ship 400-500 million chips annually. That's 400-500 million units. For perspective, NVIDIA shipped about 3.5 million H100s in 2024. Even if you assume each Broadcom chip is cheaper (say $2,000), you're still looking at 100 million units. That's 30x NVIDIA's current volume.
The wafer capacity alone kills this. TSMC's 3nm/5nm capacity is about 1.5-1.8 million 12-inch wafers per year across all customers. Broadcom's $200B target would require roughly 500,000-600,000 wafers annually—just for AI chips. That's 30-40% of TSMC's total advanced node output. And they'd have to compete with NVIDIA, Apple, AMD, and every other high-value customer. TSMC allocates capacity based on margin. NVIDIA's GPUs generate higher ASP per wafer than any ASIC. Broadcom would be fighting for scraps.
Then there's CoWoS packaging. The advanced packaging that stacks memory and logic is the bottleneck of the AI era. TSMC's CoWoS capacity in 2025 is about 40,000-60,000 wafers per month. NVIDIA consumes over 60%. Broadcom's current TPU and ASIC products also need it. To support $200B in revenue, they'd need 100,000-150,000 wafers per month—meaning TSMC would have to triple its CoWoS capacity in three years. That's physically possible, but only if Broadcom signs long-term agreements and pays a premium. Even then, the timeline is tight.
HBM memory is another brick wall. Every AI chip needs high-bandwidth memory, and the supply is controlled by SK Hynix, Samsung, and Micron. Total HBM capacity in 2025 is about 50-60 billion gigabytes. NVIDIA takes 70%+. Broadcom's $200B scenario would require 20-30% of global HBM supply—an additional 10-15 billion GB. That means new fabrication lines, which take 2-3 years to build. The lead time alone makes the 2028 target unrealistic.
And power. Oh, the power. The chips needed for $200B in revenue would consume roughly 100-200 GW of electricity. That's the equivalent of 100-200 nuclear reactors. Global data center power consumption today is about 500 TWh per year, with AI using about 100 TWh. To support Broadcom's chips, you'd need to double the entire world's AI power budget in three years. Grid infrastructure doesn't scale that fast. Not for any price.
The Contrarian Angle: What Everyone Misses
The sell-side is selling a story. And the story is that AI capex will keep growing at 40%+ forever. But the data tells a different tale. Look at the hyperscalers: Microsoft, Google, Amazon. Their AI revenue growth is lagging behind their AI capex growth by a widening margin. In Q1 2025, Microsoft's Azure AI revenue grew 80% YoY, but its capex grew 120%. The gap is real. At some point, investors will ask: "Where's the return?"
When that question comes, capex cycles turn. And when they turn, Broadcom's $200B dream evaporates. The real risk is not that Broadcom fails to execute—it's that the entire AI infrastructure buildout hits a demand wall in 2027-2028. The Wolfe Research prediction assumes the demand side is infinite. It's not.
I saw this play out in 2022 during the Terra crash. The narrative was all about algorithmic stability and decentralized money. But the fundamentals—the liquidity reserves, the collateral quality—were ignored. When the music stopped, the losses were catastrophic. The same pattern is visible here: a headline number that ignores the physical and economic constraints.
Another blind spot: customer concentration. Google accounts for over 50% of Broadcom's AI revenue. To reach $200B, Google alone would need to spend $100B on Broadcom chips in 2028. That's 30% of Google's entire 2024 revenue. No hyperscaler has ever allocated that much of their budget to a single supplier. Not even close.
The Real Play: Positioning for the Chop
This is a sideways market for AI stocks. The hype is baked in. The multiples are stretched. Broadcom trades at 35-40x forward earnings, implying the market already prices in 150-200% annual AI revenue growth. The Wolfe prediction is not a catalyst—it's a validation of existing sentiment.
For traders, the chop is for positioning. The contrarian move is to watch for the gap between AI capex and AI revenue. If that gap narrows, the bull case strengthens. If it widens, the whole thesis cracks.
The Takeaway
The $200B prediction is a sell-side fantasy, but it's a useful fantasy. It tells us that the market is pricing in extreme optimism. The real question is not whether Broadcom can hit $200B—it's whether the AI infrastructure cycle can sustain the current trajectory. The answer will come in 2026-2027, when the first capex pullbacks hit.
Liquidity flows where fear turns into opportunity. Right now, the fear is missing out on the next NVIDIA. But the real opportunity is in understanding the constraints. Speed is the only hedge in a real-time world. The chart whispers, but the volume screams. And right now, the volume is saying: this is too fast, too high, too soon.
We didn't see the cliff until we were falling. Don't let it be you.