Wolfe Research's latest note projects Broadcom's AI revenue hitting $200 billion by 2028. Let's trace the gas trails back to the root cause of this claim. The number is both staggering and suspicious—equivalent to 1.5 times NVIDIA's total FY2024 revenue, and roughly 4x Broadcom's entire 2024 revenue of $51.6B. For a company that generated just $12B in AI semiconductor revenue in FY2024, reaching $200B in four years implies a compound annual growth rate of over 70%. That's a pace no semiconductor company has ever sustained, not even NVIDIA during the ChatGPT boom.
Before we deconstruct the math, we need to understand Broadcom's actual AI business. It sits on two pillars: custom AI accelerators (ASICs, branded as XPU) and high-speed networking silicon (Tomahawk/Jericho Ethernet switches and NICs). Their largest customer is Google, supplying TPU v5/v6 for both training and inference. Meta, Microsoft, and Amazon are also rumored to be in the pipeline, but only Google has publicly confirmed a multi-year partnership. The 2025 consensus for Broadcom AI revenue is around $20-24B—respectable, but a far cry from $200B.
The core of Wolfe's prediction likely rests on three assumptions: first, that Broadcom captures a dominant share of the custom ASIC market beyond the hyperscalers; second, that AI infrastructure investment grows at 40%+ annually through 2028; and third, that supply chain constraints miraculously dissolve. Let's examine each.
Custom ASIC market share: Broadcom currently holds about 55-60% of the custom AI chip design market by revenue, according to Semianalysis. But the total addressable market for custom ASICs in 2028 is estimated at $100-150B, not $200B. Even if Broadcom takes 100% of that market, they'd still fall short. The $200B target implies Broadcom also captures a significant slice of the general-purpose GPU market—which is NVIDIA's fortress. That's not a plausible scenario without a fundamental shift in software ecosystems. CUDA's lock-in remains the single most underestimated barrier in the industry. Based on my audit experience, I've seen how network effects in developer tooling can make or break a hardware platform. NVIDIA's lead is not just hardware; it's a decade of software optimization.
Infrastructure investment growth: The $200B prediction assumes that cloud capex on AI continues to grow at 40%+ CAGR through 2028. But early warning signs are already flashing. In Q1 2025, Microsoft's Azure AI revenue grew at 79% year-over-year, but its capex grew 86%. The gap between AI revenue growth and capex growth is narrowing. Historically, when that gap closes, a capex digestion period follows. The semiconductor industry has seen this before—the 2018 crypto mining bust, the 2022 cloud optimization cycle. The code does not lie, but the auditor must dig. The data shows that hyperscalers are already questioning the ROI of massive AI clusters. If the inflection point arrives in 2027, Broadcom's actual AI revenue could be $50-70B, not $200B.
Supply chain constraints: This is where the prediction becomes physically impossible. Broadcom's AI chips require TSMC's most advanced nodes (3nm/2nm) and CoWoS advanced packaging. In 2025, TSMC's CoWoS capacity is about 4-6k wafers per month, with NVIDIA consuming over 60%. To support $200B in revenue, Broadcom would need at least 10-15k wafers per month—meaning TSMC would have to triple CoWoS capacity by 2028. That's theoretically possible, but it would require massive investment and allocation rebalancing. Moreover, HBM (high-bandwidth memory) is another bottleneck. SK Hynix and Samsung are already struggling to meet NVIDIA's demand. Broadcom's chips would need an additional 20-30% of global HBM supply. The lead time for new HBM fab construction is 2-3 years. The physical constraints are a hard ceiling.
Power consumption: The electricity required to run the chips generating $200B in revenue would be on the order of 100-200 GW, equivalent to the entire current global data center power consumption. Grid infrastructure cannot scale that fast. This is the ultimate ceiling—centers need substations, transformers, and cooling systems that take years to build. In the chaos of a crash, the data remains silent. But right now, the data screams that the power timeline is misaligned with the revenue timeline.
Contrarian angle: The most overlooked risk is customer concentration. Google alone accounts for over 50% of Broadcom's AI revenue. For Broadcom to hit $200B, Google would need to spend $100B on Broadcom chips in 2028—roughly 30% of Google's total 2024 revenue. That's absurd. Even if Broadcom adds OpenAI, Meta, and Amazon as clients, each would need to spend $20-30B annually. Only a handful of entities globally can absorb that scale. The prediction also ignores the possibility that hyperscalers will insource more chip design over time, reducing Broadcom's role to a mere foundry partner. The consensus layer is shifting, one block at a time.
Takeaway: Wolfe Research's $200B is a narrative, not a forecast. It reflects the bull case where every variable aligns perfectly—no capex slowdown, no supply chain snag, no NVIDIA retaliation. The realistic range for Broadcom's 2028 AI revenue is $60-100B, which is still impressive but far from the hyperbolic headline. For crypto investors who follow tech stocks, this matters because the AI infrastructure narrative directly influences the broader risk appetite in the market. If the capex cycle turns, so does the flow of capital into all speculative assets, including crypto. The question to ask is not whether Broadcom can reach $200B, but whether the market will believe it long enough to price in a premium that disappears when reality diverges.