A sell-side note lands in my inbox. Wolfe Research projects Broadcom’s AI revenue could hit $200 billion by 2028. That’s eight times the consensus for 2025. I stop reading. I start auditing.
Numbers like that don’t come from financial models. They come from narratives. The same narratives that drove ICO whitepapers promising a million TPS. The same narratives that made NFT floor prices ‘guaranteed’ to moon. Now, a sell-side analyst is promising $200 billion in AI revenue for a semiconductor company. The number is so absurd it demands attention.
I’ve seen this movie before. It was 2017. I was in Bangkok, running a Telegram group called ChainLogic. I manually audited 15 ICO whitepapers. Found red flags in 8. The market didn’t care. The narrative was stronger than the code. Today, the narrative is stronger than the physics.
Context: The Broadcom Story
Broadcom (AVGO) is not a blockchain company. It’s a semiconductor giant. Custom AI accelerators (ASICs) and Ethernet switches. Its current AI revenue is around $20-24 billion for fiscal 2025. The Wolfe Research prediction demands that this number grows to $200 billion by 2028. That’s a compound annual growth rate of 70-90%. No semiconductor company in history has done that.
Code doesn’t lie, but narratives do. The narrative here is that AI infrastructure investment will continue at a 40%+ annual clip, that Broadcom will lock in every hyperscaler as a customer, and that physical supply chains will bend to the will of demand. It’s a beautiful story. It’s also a textbook example of bullish euphoria masking technical flaws.
Core: The Technical Audit
Let’s break down the numbers. $200 billion in AI revenue by 2028. For context, NVIDIA’s total revenue in fiscal 2025 was about $130 billion. Broadcom would need to exceed NVIDIA’s current scale by 50%. The global AI semiconductor market in 2028 is estimated at $250-300 billion. Broadcom would need to capture 67-80% of that. That’s monopoly-level share.
But the real constraints are physical. I’ve spent years auditing chip supply chains for crypto mining rigs. Physical limits are the hardest to fudge.
Wafer Capacity
TSMC’s 3nm and 5nm capacity is about 150-180 million wafers per year (12-inch equivalent). NVIDIA consumes 30-40%. Apple takes 20-30%. Broadcom’s $200 billion revenue would require roughly 50-60 million wafers per year for its own chips. That’s before factoring in network chips. TSMC cannot allocate that much to one customer without starving others. And TSMC prioritizes high-margin customers like NVIDIA. Broadcom’s margins on custom ASICs are lower. The allocation game is stacked against them.
CoWoS Packaging
Advanced packaging is the bottleneck. TSMC’s CoWoS capacity in 2025 is about 4-6 million wafers per month. NVIDIA consumes 60%+. Broadcom’s TPU and ASIC products also need CoWoS. To reach $200 billion, Broadcom would need 10-15 million wafers per month. That’s 2.5-3 times current capacity. TSMC is expanding, but not that fast. Not in three years.
HBM Memory
AI chips need high-bandwidth memory. SK Hynix, Samsung, and Micron produce it. NVIDIA consumes 70%+ of global HBM supply. Broadcom’s $200 billion revenue would require 20-30% of total HBM output. That’s additional capacity that takes 2-3 years to build. The supply chain is already stretched.
Power Constraints
$200 billion in AI chips means deploying 100-200 GW of computing power. That’s more than the total electricity consumption of all data centers today. Grid infrastructure doesn’t expand in three years. It takes a decade. The power wall is the ultimate ceiling.
Customer Concentration
Broadcom’s AI revenue is heavily dependent on Google. Alphabet is its largest AI customer, contributing over 50% of Broadcom’s AI chip sales. To reach $200 billion, Google would need to buy $100 billion in chips from Broadcom by 2028. That’s 30% of Google’s 2024 total revenue. Impossible. Broadcom would need 5-8 other hyperscalers each spending $20-30 billion annually. The list of companies with that kind of budget is short: Amazon, Microsoft, Meta, maybe OpenAI. And each is developing its own chips.
Competitive Response
NVIDIA is not sitting still. Its Rubin architecture will be out by 2026. If NVIDIA maintains a 1.5-2x performance advantage over ASICs, the custom chip market shrinks. Broadcom’s differentiation evaporates. NVIDIA has a history of aggressive pricing when threatened. It can crush ASIC economics.
The Crypto Parallel
I’ve been through this before. During DeFi Summer in 2020, I partnered with SushiSwap to audit their fork mechanism. I organized workshops in Bangkok, teaching 200 developers how to use Uniswap and Aave. I tested liquidity mining strategies personally. Lost 15% to impermanent loss. That failure taught me something: when everyone is certain about exponential growth, the risks are hiding in plain sight.
The Wolfe Research prediction is the same. It’s a narrative built on assumptions that no one is stress-testing. The AI infrastructure capex cycle is running ahead of application-layer revenue. Cloud providers are spending billions on GPUs, but their AI revenue growth is lagging. The gap is widening. When that gap becomes too large, the capex cycle peaks. That’s when predictions like $200 billion become $50 billion.
Contrarian: The Pragmatism Test
Here’s the counter-intuitive angle. The $200 billion prediction is not just optimistic. It’s a signal that the market is overpricing AI infrastructure. The real risk is not that Broadcom fails to deliver $200 billion. It’s that the entire sector corrects when the hype cycle breaks.
Look at crypto. In 2017, the ICO market promised decentralized everything. Projects raised billions. Most failed. The survivors? They built on fundamentals, not narratives. The same will happen in AI. The companies that win are the ones that prove economic returns, not just demand for chips.
But there’s another blind spot. The Wolfe Research report assumes that AI infrastructure investment will continue at 40%+ CAGR. That assumption ignores the possibility of a demand shock. What if AI models stop improving? What if regulation slows adoption? What if the cost of inference drops so fast that the hardware market shrinks? Each of these scenarios would kill the $200 billion prediction.
The contrarian take: the real bottleneck is not supply. It’s demand. The AI application layer is not generating enough revenue to justify the capex. The same was true for crypto infrastructure in 2018. Projects built before users arrived. They failed. The ones that waited for product-market fit survived.
Takeaway: Vision Forward
Trust is the new currency. But trust in sell-side predictions is as fragile as trust in a DeFi protocol without an audit. The $200 billion number is a signal, not a target. It tells us that the market is pricing in an AI revolution that hasn’t yet proven its economic returns.
As a crypto native, I’ve learned to value proof over promise. The alpha is hidden in the noise of the hype cycle. The real opportunity? Build infrastructure that is resilient to the inevitable correction. That means focusing on physical constraints, customer concentration, and competitive moats. Not on analyst projections.
Are you betting on the narrative, or on the code?
I’ll be watching the next Broadcom earnings call. If AI revenue growth slows below 60% year-over-year, the $200 billion dream starts to fade. If NVIDIA’s Rubin architecture exceeds expectations, the ASIC narrative crumbles. If the AI capex-to-revenue gap widens for three consecutive quarters, the cycle is peaking.
Until then, the $200 billion signal is just noise. But noise can be profitable if you know how to filter it. The alpha is in the noise. And the code never lies.