Broadcom’s after-hours drop today is not because the AI story broke—it’s because Wall Street suddenly realized that even a great AI story cannot endlessly borrow from tomorrow. More importantly, this earnings report exposed not a Broadcom problem, but a broader shift in AI semiconductors: the market is moving from “who rises fastest” to “who can actually keep the money.” We only discuss the most valuable logic.
Broadcom reported Q2 revenue of $22.187 billion, up 48% year over year. Adjusted EBITDA reached $15.244 billion, or 69% of revenue. Free cash flow topped $10.2 billion. In plain English: while other companies are still selling a story, Broadcom is already counting cash.
What really impressed the market was AI semiconductor revenue of $10.8 billion, up 143% year over year. The company also guided Q3 AI semiconductor revenue to $16 billion, implying more than 200% growth.
Under normal conditions, a report like this should earn applause, model upgrades, and a chorus of “the AI second leg is here.” Instead, Broadcom fell sharply after hours. So what is Wall Street afraid of?
The answer is simple: Wall Street is not afraid Broadcom won’t make money; it is afraid Broadcom makes too much money for today’s price to make sense. Stocks fear not bad news, but “good news that has already been priced in.”
How crazy has the AI semiconductor rally been over the past few months? Nvidia no longer looks like a company—it looks like the central bank of the U.S. stock universe. Marvell, AMD, Micron, TSMC, Applied Materials, Lam Research, KLA, and many others have all rallied. The Philadelphia Semiconductor Index has kept climbing on Computex and AI hardware enthusiasm. The logic is simple: the closer you are to the AI data center, the more you look like you found a gold mine.
But when everyone knows there is gold, the market won’t sell it to you at gold prices—it will sell it as if you must dig diamonds for the next ten years. Broadcom’s after-hours drop is the market saying: the business is excellent, but I already priced you like half an Nvidia. You proved you are strong, but not strong enough to ignore valuation.
That is the first key point: Broadcom did not lose on earnings; it lost on expectations.
Think of Wall Street as an impossibly demanding date. A normal person says, “Stable job, house, and car—that’s good enough.” Wall Street says: no, I also want surging AI revenue, stable margins, exploding cash flow, locked-in customer demand, and next quarter guidance that beats again. Broadcom delivered most of that, but the market wanted the final “chosen one” storyline.
So why was the AI revenue still not enough? Because Broadcom’s AI business is not Nvidia’s general-purpose GPU universe. It is more about custom AI chips and AI networking. That difference matters a lot.
Nvidia is like selling standard luxury sports cars: anyone with money can buy in, and then join the CUDA ecosystem party. Broadcom is more like building a bespoke high-speed rail line for giant customers. It is not for everyone, and it is not an overnight delivery business. Its strengths are deep customer lock-in, high order visibility, and tight integration with cloud data-center architecture. Its weakness is equally obvious: concentrated customers and long project cycles, so revenue can swing sharply if a big customer changes pace.
That explains Wall Street’s second concern. Broadcom’s AI growth is strong, but it is not a fully open-ended market story. It is more like a super-project customized by a few giant customers. As long as that project keeps moving, the money is huge. But if a major customer changes capex plans or slows an ASIC roadmap, Broadcom feels it directly.
So the market is not confused by 143% growth. It is asking: how long can 143% last? After next quarter’s $16 billion AI revenue, can growth still accelerate? By 2027, will the bigger AI semiconductor target be delivered smoothly, or will the market have already priced the fireworks in advance?
That brings us to the most important conclusion: Broadcom’s drop is not a signal that the AI semiconductor rally is over. It is a signal that the rally is entering phase two.
Phase one was simple: if you touched AI, the market gave you a valuation premium.
Phase two asks three questions: Is your AI revenue real? Is your profit real? And do customers keep paying for the long term, or are they just excited this year and calm next year?
Broadcom has already proven the first two. The third one still needs more proof. After such a big move in semis, investors become very sensitive. Before, it was “buy anything with AI.” Now it is “can AI keep compounding for three more years?” That is the sign of a maturing rally—and the easiest place for retail investors to get the timing wrong.
Some will see the after-hours drop and say AI is a bubble. That is too simple. The real picture is that AI is not losing believers; it is being sorted into tiers.
Tier 1: companies with ecosystem control and pricing power, like Nvidia. Tier 2: companies deeply tied to cloud-custom silicon and networking, like Broadcom and Marvell. Tier 3: toolmakers, equipment makers, materials, advanced packaging names such as Applied Materials, Lam Research, KLA, and TSMC-linked chains. Tier 4: cyclical stocks lifted by AI demand, like Micron, SanDisk, and Western Digital.
My different view is this: for a while, the most interesting opportunity may not be “the next Nvidia,” but who can capture the most stable profit from AI capex spillover. In other words, Broadcom’s report is not the end point—it is a signal that AI is moving from “compute worship” to “infrastructure re-rating.”
Why am I more bullish on memory in the medium to long term? Because AI data centers are not powered by GPUs alone. GPUs are the brain, networks are the blood vessels, and storage and memory are the warehouse and highways. You cannot give a super-chip brain enough power without enough memory and data pipes. The bigger the AI model, the more inference, the more private enterprise data, the more multimodal content—meaning storage is no longer optional, it is essential.
In the past, AI discussions focused on training. Now the real change is inference. Inference is work—it happens every day, and the more users there are, the higher the cost. Every AI answer, every video generation, every database query is not one GPU burning alone; it is the whole compute, network, memory, and storage stack working overtime.
That is why Morgan Stanley warns about “chip inflation.” Memory prices have jumped sharply over the past year, AI data centers are locking up premium capacity first, and traditional phone, PC, and consumer electronics makers are feeling cost pressure. Memory used to be the classic boom-bust pig cycle. This time is different because premium AI demand is changing the supply-demand structure.
Memory used to follow the phone cycle, the PC cycle, and the server cycle. Now there is a new huge variable: the AI data center cycle. And that does not just mean buying more memory—it changes the product structure. HBM, enterprise SSDs, high-capacity HDDs, and data-center storage arrays are all being repriced by AI.
That is why Micron has been so strong lately, and why SanDisk and Western Digital can also move higher. The market has realized that AI is not only making Nvidia richer; it is turning storage companies from “cyclical laborers” into “AI warehouse managers.” In the past, Micron was seen as a weather-dependent stock: prices up meant smiles, prices down meant tears. Now the market is asking whether Micron is still just cyclical, or has entered a much longer repricing cycle.
Of course, storage will not rise forever. That would be a fairy tale, not financial analysis. The sector is still cyclical. If prices rise too fast, capacity expansion will follow, and then supply pressure comes back. The competition among Chinese memory makers, Samsung, SK Hynix, and Micron will all affect future pricing.
But what is different this time is that new capacity cannot appear tomorrow just because someone wants it. Advanced processes, HBM packaging, yield, customer qualification, and long-term supply contracts all take time.
So for me, the medium- to long-term logic for memory is not “prices always go up.” It is “AI has redefined the scarcity of premium memory.” Those are two very different things: the first is a cyclical trade, the second is a structural industry shift.
Back to Broadcom. Why does it matter? Because it shows that cloud providers’ AI capex is entering a more complex stage. First they bought Nvidia GPUs in a direct arms race. Then they found that buying only general-purpose GPUs was too expensive, so they began designing their own custom AI chips. Google has TPU, and other hyperscalers are doing similar things. Broadcom and Marvell make money on this shift.
Is that a threat to Nvidia? Not in the short term. Long term, it is a variable to watch. Nvidia’s ecosystem advantage is still massive, and CUDA is not a moat built overnight. But if cloud providers can use custom ASICs for certain inference tasks to achieve lower cost and better efficiency, Nvidia’s absolute monopoly narrative gets diluted a bit—not broken, but reduced from “the only answer” to “one of the core answers.”
That is both an opportunity and pressure for Broadcom.
Some may ask: if the stock fell, can we just buy the dip? Be calm. An after-hours drop does not automatically mean opportunity or danger. You need to know whether the drop is valuation compression or a change in fundamentals. If the stock had simply run too far and the report triggered a sell-the-news reaction, the pullback may actually give long-term money a better entry point.
For ordinary investors, I do not recommend buying Broadcom emotionally. Jumping in after an after-hours dump can easily make you the exit liquidity. If Broadcom falls but Micron, SanDisk, and Western Digital stay strong, that means the market is not rejecting AI—it is rotating from expensive AI chips to the more directly benefited memory-price trade.
From that perspective, I remain moderately bullish on memory in the medium to long term. Not because Micron has already risen a lot, but because AI is pulling storage from the edge of the cycle into the center of infrastructure.
In the past, memory price increases only mattered to the industry itself. Now they can affect cloud capex, PC prices, smartphone costs, and even the speed of enterprise AI rollout.
That is the point many commentators miss: the next stage of the AI semiconductor rally is not necessarily “find the next Nvidia,” but “find who has pricing power besides Nvidia.” Broadcom has pricing power because it controls custom AI chips and high-speed networking. Micron, SK Hynix, and Samsung have pricing power because premium memory is tight. TSMC has pricing power because advanced process and packaging are unavoidable.
You will find that what is really worth studying is not whether “AI will keep rising,” but which link in the chain is hardest to replace.
Of course, the risks must be stated clearly.
First is valuation risk. Semiconductor stocks have already run a long way. For any high-growth company, if growth is even slightly less explosive, the market can turn fast.
Second is capex risk. AI data centers are hot now, but if cloud providers later see longer payback periods, capex could slow and the whole chain would feel it.
Third is supply risk. The faster memory prices rise, the stronger the future temptation to expand. The end of the cycle will always bring pressure.
Finally, there is geopolitical and export-control risk. Semiconductors are never just an industry—they sit at the intersection of technology, capital, and policy. One gust of wind, and valuations catch a cold.
So today’s conclusion is not “Broadcom fell, buy it now,” and it is not “Broadcom fell, AI is done.” My conclusion is that Broadcom’s report shows the profits in the AI supply chain are still very real, but the market is no longer willing to blindly pay for every AI story.
