Broadcom's fiscal second quarter strengthened the custom AI chip story. It did not match the valuation case investors had built into the stock.
That gap explains the reaction.
Revenue rose 48% to $22.19 billion. Adjusted earnings were $2.44 a share. AI chip revenue hit $10.8 billion, up 143% from a year ago. The company guided third-quarter revenue to $29.4 billion and AI chip revenue to $16 billion, which would mean growth of more than 200%.
The problem was not the quarter. It was the bar.
Investors wanted CEO Hock Tan to raise the long-term AI chip target above the "more than $100 billion" in 2027 that he laid out in Q1. He did not. He reiterated the number and added more proof behind it.
That sets up the real debate.
Did Broadcom's Q2 strengthen the custom AI chip thesis? Or did it show that hopes have moved ahead of even very strong AI growth?
More Proof, Not a Bigger Number
Broadcom's AI business keeps speeding up.
In Q1, the story was simple. Management said it had a clear path to more than $100 billion in AI chip sales by 2027. The path was paved by custom chips, AI networking, and major customers like Google, Anthropic, OpenAI and Meta.
Q2 did not raise that target. But it added real backing.
The company said it expects $56 billion in AI chip sales this fiscal year. It has six core AI customers. It booked more than $30 billion in AI chip orders this quarter. It shipped $10.8 billion.
That gap between bookings and shipments is the most important data point in the print.
It says demand is not the problem. Customers are ordering far ahead of what Broadcom can ship. The timing is shaped by what else needs to come together: power, data centers, networking, memory and deployment plans.
The market wanted more. Stronger proof was not the same as a bigger number.
Google Visibility Got Much Better
The Google chip deal became more important after Q2.
Tan said Broadcom signed a long-term deal with Google in April. It covers many future versions of TPU chips and AI networking. WSJ reported that Tan called the deal "very substantial" in dollar terms.
That matters because this is not a new test. Broadcom and Google have worked together for about ten years. They have already built eight versions of Google's TPU chip.
That history strengthens the visibility case. Google is not just one more customer. It is the longest proof that hyperscalers will keep building their own AI chips instead of just buying from Nvidia.
Alphabet also raised the bar before the print. Last week, it said it would sell $80 billion in stock to help fund AI spending. Broadcom shares rose on that news.
That likely made the hurdle for Broadcom's guidance higher than it would have been.
The Customer List Got Stronger but Stayed Narrow
The six core AI customers all got more detailed treatment than in Q1.
Anthropic, OpenAI and Meta all got more specific deployment color. Management talked about Anthropic's chip needs, OpenAI's production timing and Meta's own chip plans. The Meta detail was important. In Q1, the company spent time pushing back on doubts about Meta. In Q2, it gave real numbers.
But the customer base is widening, not opening up.
Broadcom is not selling AI chips across the broad enterprise market. It serves a small group of hyperscalers and frontier AI labs. Those are the only customers with the scale and money to make custom chips worth it.
That is a risk. It is also why the economics are big enough to support the 2027 target.
Broadcom Is Not Trying to Be Nvidia
One of the most important comments was a clarification.
Tan made clear Broadcom is not trying to sell full rack-scale systems. It is focused on chips: XPUs, TPUs, networking chips, and the gear that connects them. CNBC called this out.
That narrows the story.
Broadcom is not the next Nvidia in terms of product scope. It is the custom chip and networking option for customers big enough to design their own systems.
That makes Marvell the closer peer in custom chips. Nvidia is still the broader benchmark for the full AI platform.
Networking Is More Important Than People Think
AI networking is no longer a side story.
Management said networking made up almost 40% of AI revenue this quarter. The strength came from Ethernet, switching, optical gear and other ways to link chips together.
Tan said networking could fall back to closer to 30% of AI revenue as chip sales ramp up faster. That is not a bad sign. It just means custom chips are growing fast enough to dilute networking as a share.
The bigger point is that Broadcom is in AI infrastructure two ways: through compute and through the fabric that links it.
Software Pays the Bills. It Does Not Set the Stock Price.
Software is still a cash machine. It is no longer the story.
Software revenue rose 9% to $7.18 billion. That came in below some forecasts. Margins held up well. The company also generated $10.26 billion in free cash flow, or 46% of revenue.
But the mix is shifting fast. Last year, software was 42% of Broadcom's total revenue. Next year, that is expected to drop to about 20% as AI chips grow.
That makes the software miss less damaging to the bull case. It does not make it meaningless. VMware still matters. It funds the cash flow, the margins and the buybacks. It just no longer drives the multiple.
The Bottom Line
Broadcom's Q2 was strong by almost any normal measure.
AI revenue grew fast. Q3 AI guidance was strong. Bookings were far ahead of shipments. Google visibility improved. Anthropic, OpenAI and Meta commentary got better. Networking looks more durable. Software stayed highly profitable.
The stock fell because investors wanted the 2027 target raised, not just repeated.
That puts Broadcom in a more demanding spot. The custom AI chip story is stronger. But the market now expects steady upward revisions, quarter after quarter.
Q2 gave investors more proof.
It did not give them a bigger number.
That is why the quarter was both bullish for the business and a warning for the stock.