Nvidia's $81.6B Record, China DUV Ships & HBM4 Pricing Signals

**Jamie Call** (0:00)
AI Hardware & Chips Daily News I'm Jamie Call, thanks for joining me.
Today Nvidia's Margin Fortress under pressure, $81.6 billion beats StockSyncs.
Nvidia just posted the largest quarterly revenue in semiconductor history, and the stock fell. That's the signal worth paying attention to today. $81.6 billion in Q1 FY 2027 revenue, up 85% year-over-year. Data Center alone hits $75.2 billion, which is 92% of total revenue. Guidance for Q2 came in at $91 billion.
By any conventional measure, this is an extraordinary business, printing extraordinary numbers.
And yet the stock is trading below $205, stuck in a range between $194 and $205, having broken below a key uptrend. The key implication is this, the market isn't pricing Nvidia on what it earned, it's pricing it on what it might lose. China export restrictions have effectively zeroed out all data center compute revenue from that region in Q2 guidance.
The actual size of that hole isn't fully quantified. The customer defection risk isn't either. What investors are doing is discounting against an unknown, and unknowns with geopolitical origins tend to be punished harder than fundamentals want. There's also the valuation pressure sitting underneath all of this. 75% gross margins are elite, but at Nvidia's scale expectations are baked in deep. A beat that doesn't dramatically exceed guidance doesn't move the multiple. It just confirms it. And confirmation it stretched valuations isn't a catalyst. One figure that deserves more attention than it's getting. Networking revenue inside data center surged 199% year-over-year. From roughly 6.2 billion to 14.8 billion. That's InfiniBand and Ethernet bundled into Nvidia's integrated stack. The signal there is that customers aren't just buying GPUs. They're buying the interconnect, the software, the full solution. That's what a switching cost looks like in practice. Away from Nvidia there's a development in lithography that shifts the competitive narrative. A Chinese manufacturer unnamed has begun shipping domestic immersion DUV machines. Five units this year, with 20 planned for 2027 These are now concrete deliveries to facilities including SMIC and CXMT, not prototypes or roadmap slides. The important distinction is where the technology actually sits. ASML ships around 130 DUV units annually. Built on two decades and roughly $10 billion of accumulated R&D, the Chinese machines are unproven in yield and field reliability.
Yield data from imported DUV tools at SMIC already lags TSMC materially. Establishing a reliable domestic alternative takes years, not quarters. Here's the thing. That said, the geopolitical narrative has shifted. The question was whether China could build these tools at all. That question now has a partial answer. The remaining test is whether they can scale and perform.
TSMC raised its 2026 capex commitment to $60-64 billion.
Up from a prior plan of $59 billion. The stock has retreated 17% from its record high of $480.
And is now testing support around $390.
Investors are asking a straightforward question. At 70-80% allocation toward advanced processes, does the return timeline hold if China advances in mid-range nodes, and competitors like Sansun and Intel continue ramping? There's no clean answer yet. The ROI visibility on that level of CAPEX intensity, over a 2-3 year horizon, is limited. That uncertainty is what the stock is pricing. CSK Hynix, confirmed HBM4 shipments, begin in the second half of 2026, with yield levels at HBM3e parity. More consequential is what the company said about pricing.
Negotiations for 2027 HBM supply are underway, and the signal from those discussions is sharp price increases. Tied to the broader DRAM market.
If commodity DRAM holds or rises, HBM premiums expand. If it falls, that premium could compress. Customer pushback is a real risk here. Consider this. Separately, CSK Hynix is developing a NAND-based KV cache solution that mounts high-capacity flash storage near GPUs. The reasoning is straightforward. Inference workloads are growing faster than HBM capacity can keep up. NAND becomes the overflow layer. That's not a Mitch product decision. It's a signal that the memory hierarchy for AI inference is being actively restructured. One more data point worth filing. A 4GB variant of the AMD Radeon RX 9005 has leaked, the first modern 4GB consumer GPU in four years. That's not a product choice driven by strategy. That's memory supply pressure cascading into the entry-level segment. The watchpoints from here are clear. Nvidia's actual China revenue exposure needs to be quantified in Q2 results. TSMC's CapEx return timeline needs either confirmation or revision as advanced no-demand data comes in. And the Chinese DUV yield story needs time to develop. Until there's real performance data, the threat is real, but unproven. Thanks for listening. This podcast was built using AI technology. A Yes We production.