Broadcom's 30% Drop, AMD's 7% Surge & Huawei's 18-Month Chip Pull-Forward

Jamie Call (0:00)

AI Hardware & Chips, Daily News.

I'm Jamie Call. Thanks for joining me.

Today, why Broadcom stock plunges 30% on 236% AI revenue growth? Broadcom's AI revenue just grew 236% year over year, and its stock is down 30% from its June peak. That gap is the most important signal in semiconductors right now. The numbers from Broadcom's third quarter are genuinely striking. AI semiconductor revenue came in at $16.7 billion, up over 220% from the same quarter a year ago. Custom silicon shipments tripled. Management guided fourth-quarter AI revenue to $21.7 billion and projected $115 billion for fiscal year 2027 The committed order backlog across six hyperscaler customers runs to $350 billion through 2028 And yet the stock sits at $339, 30% below where it was trading in June.

Here's what matters. Markets aren't pricing Broadcom on what is shipped. They're pricing it on what they believe will actually be deployed. There's a meaningful difference between a committed order and a chip installed in a lag data center. Land constraints, power availability, and data center readiness are all real friction points that Broadcom's own management flagged. Investors are waiting for December 9

When the next earnings report will either validate the deployment timeline or raise harder questions, the margin picture adds to the hesitation. Gross margins fell from 78% a year ago to 73% guidance for the current quarter. That's a consequence of the custom silicon mix shifting upward. XPU production is lower margin than software and infrastructure products. The more Broadcom ships, the more the margin compresses, at least for now. Add $60 billion in fixed-rate debt, and the cushion, if any hyperscaler slows its ramp, is thin.

The concentration risk is real too. Six customers. One of them, Anthropic, has drawn public attention for a slower-than-expected ramp. Coctan said publicly he's not worried about AI slowdown chatter, pointing to sustained inference demand rather than frontier model training as the real driver. That's a deliberate refraining. Whether it holds is what December 9th will test. Away from Broadcom, AMD moves sharply in the other direction. The stock jumps 7% on September 17th, leading a three-day rebound across semiconductor names. The SOXX index gained 3% vs. 2% for the NASDAQ 100 AMD is up 156% year-to-date compared to Broadcom's roughly flat performance. Here's the thing, the important distinction is what's driving that. AMD's gains are partly about data center share and partly about being the highest beta name in the sector when momentum turns positive. That cuts both ways.

A strong session pulls AMD up fast. A weak one reverses it just as quickly. The rebound is worth noting, but conviction requires an earnings catalyst, not just set to rotation. On the geopolitical side, Huawei moved its Ascend 960DT AI chip launch forward from the third quarter of 2027 to the first quarter. That's an 18-month acceleration. The Atlas 960 SuperPoD was announced at Huawei Connect on September 17

The signal here is confidence. Beijing is signalling that its domestic chip roadmap can compete on its own timeline, independent of where US export controls land. There's a caveat worth tracking. The in-out SuperPoD ships with 4096 chips, not the 15,000-plus that earlier roadnaps described. That discrepancy raises legitimate questions about commercial deployment readiness versus announced milestones. Intel and SK Hynix are in exploratory talks about producing advanced high-bandwidth memory at Intel's Ohio facility. The deal structure, whether a lease or a joint venture, remains unresolved. The estimated value runs between $13 billion and $19 billion.

Consider this. The strategic logic is clear.

HBM is increasingly the binding constraint in AI infrastructure, and producing it inside allied nation fabs reduces exposure to concentrated supply chains. The complication is that South Korea's government may resist allowing advanced memory manufacturing to relocate to the United States. Seoul's position on this will determine whether the talks advance or stall.

One quieter result worth flagging, Bull won the EuroHPC contract for the Lumi AI supercomputer in Finland, beating out HPE. The system uses 1,240 nodes with 4,960 AMD Mi430X GPS and is expected to deliver 10 times the AI performance of the existing Lumi-G partition by the second half of Lumi. 2027 It's another data point on AMD's position in government-backed AI infrastructure and another contract where Nvidia wasn't the selection. The real test across all of this converges on December 9th. Broadcom's Next Earnings Report is effectively a proof of concept moment for its entire deployment narrative. If hyperscaler customers are pulling chips into live infrastructure on the timeline's management described, the stock's 30% discount to June looks like a setup. If deployment slips, the margin compression and debt load become harder conversations. Huawei's 2027 timeline, the Intel SK Hynix deal structure, and whether AMD's rebound holds without an earnings catalyst are the three other threads worth watching in T end of the year. Thanks for listening. This podcast was built using AI technology, a yes we production.