Semis in Bear Territory: Nvidia at $200, AMD's July 22 Test & Memory's Paradox

**Jamie Cole** (0:00)
AI Hardware & Chips Daily News A hair word, I'm Jamie Cole. Thanks for joining me. Today correctional rotation. Why semiconductor sector sold off 10%?
The Philadelphia Semiconductor Index is now down 20% from its June peak. That puts it in technical bear market territory, and it happened fast. Nvidia fell another 2.2% on Friday, closing at $202. That's 14% off its 2026 high, and it's testing a key $200 support level that a lot of investors are watching closely. AMD is down 17% from its recent peak. Memory stocks are collapsing. The whole sector gave back roughly 10% in a week. The question worth holding onto here is whether this is a correction inside a bull market, or something that signals a more fundamental shift in how investors are pricing AI hardware demand. Start with what didn't crack. TSMC this week committed an additional $100 billion to its Arizona operations, bringing the total pipeline to $265 billion.
The CFO cited multi-year structural AI demand and confirmed that two-manometer technology is now expected to generate meaningful revenue in the third quarter. That's a significant signal. TSMC is not a company that makes commitments of that scale speculatively. The important distinction is that TSMC's confidence in long-run demand is not the same thing as confidence in near-term valuations. Both can be true at once.
There are also execution risks worth watching. Construction workers shortages, infrastructure constraints, and US fab costs running four to five times higher than equivalent capacity in Thailand. That cost gap will dilute margins, and geopolitical support from Washington isn't guaranteed to hold indefinitely. AMD's 17% drop from its $584 peak down to $486 looks steep. Several analysts raised price targets this week, some as high as $725 even as the stock fell.
The key near-term test is July 22nd. Here's the thing, AMD's advancing AI event runs July 22nd and 23rd. The company is expected to debut Zen 6 Venice EPYC processes built on TSMC 2 nanometer alongside a roadmap update for the Mi455X GPU.
Q2 earnings follow on August 4th. The signal here is straightforward. If the product launches credibly and the earnings guidance holds up, analysts will say the correction was an overreaction.
If either disappoints, the 17% drop starts to look like the beginning of something larger.
The memory picture is the most paradoxical part of this week. Samsung, SK Hynix and Micron are operating in a sold-out environment. HBM capacity is fully omitted. AI infrastructure demand is at record levels. And yet memory stocks are crashing because investors are pricing in a future oversupply before it arrives. SK Hynix this week disclosed something worth paying attention to. Their IMTE architecture stacks HBM, DDR, CXL and SSD into a tiered memory system that improves inference efficiency by 35.7%.
Samsung is racing to mass produce CXL 3.2, which doubles the bandwidth of the prior generation. The implication is that memory innovation may be opening up new capacity for inference workloads without requiring more GPUs. That's a structural shift in how AI systems scale, and it's happening faster than most outside the memory space expected. Two separate pressures are building on Nvidia's competitive position. The first is Chinese AI. Moonshot AI's Kimi K3 model released this week as an open model, and it's being cited as a contributor to the sell-off narrative. The concern is that lower-cost Chinese AI alternatives reduce the dependency on high-end Nvidia GPUs, particularly for inference workloads. China already accounts for only 8% of Nvidia's revenue under current export controls, but the worry is about the trajectory, not the current number. Consider this. The second pressure is internal to Nvidia's biggest customers. Alphabet, Amazon, Microsoft, Meta and OpenAI are all developing custom accelerators. Training still favours Nvidia's architecture, but inference is where the volume goes over time. Custom chips targeting inference could gradually pull that workload away. Jensen Horn's visit to Japan this week produced a partnership with Notra for a Vera-Rubin AI factory, featuring 13,750 Vera CPUs and 27,500 Rubin GPUs. FANUC, Yaskawa and Sony joined on robotics platforms. That's Nvidia expanding its narrative beyond data centers into physical AI and manufacturing automation. Whether that's a genuine growth vector or a story told during a difficult week for the stock is a fair question. The real near-term catalyst for this sector is Hyperscaler earnings. Alphabet, Amazon, Microsoft and Meta Guidance on second half 2026, AI infrastructure spending will either confirm the demand thesis or complicate it. When that visibility arrives, expect the correction vs rotation debate to stay unresolved.
Watch the $200 level on Nvidia, watch AMD on July 22nd, and watch what the memory names do if CXL standardization timelines start to firm up.
Those are the signals that matter. Thanks for listening. This podcast was built using AI technology, a YesOui production.