Delos vs Nvidia, DDR5 Spot Price Spike & Mexico's China Chip Loophole

Jamie Call (0:00)

AI Hardware & Chips Daily News. I'm Jamie Call, thanks for joining me. Today, networking chip startup Delos challenges Nvidia's inference dominance. A startup called Delos Data just raised over $100 million to build networking chips specifically for AI inference, and that's a direct challenge to the infrastructure layer Nvidia currently dominates.

The product is called the Nonstop AI Data Interface chip. Delos claims it delivers 10 times lower latency than competing solutions, and, more importantly, it lets inference cluster scale beyond what a single device can handle. That second point is the strategic one. As AI workloads shift from training to inference at scale, the bottleneck moves from raw compute to how fast data moves between chips and systems. Delos is betting that's where the real value gets captured next.

The key implication is this. Nvidia's inference dominance has rested partly on its closed ecosystem, where the hardware, networking and software stack are tightly integrated. Delos is taking the opposite approach, using open standards, specifically OAM interconnect and OSFP routing formats, and selling its chip in modular formats including PCIe, chiplet and NPO. The open standard bet is designed to reduce Fender locking. Whether hyperscalers actually embrace it over Nvidia's closed stack is still unresolved, and that adoption question is the real test for the company's thesis.

Shift to memory and the picture there is one of genuine supply stress becoming visible in pricing. DDR5 spot prices have spiked to $3100 per 64 gigabytes. That's more than double the fixed contract price of around $1,500. When spot and contract prices diverge that sharply, it signals buyers are scrambling for supply outside normal procurement channels.

The signal here is the inventory number. Industry sources confirm memory maker inventories have fallen below 10 days of supply. That's a thin buffer by any measure, and if it isn't rebuilt quickly, the shortage could persist well into 2027

Samsung's Device Solutions Unit is already projecting operating profit exceeding $370 trillion on the back of these conditions. Record margins at the memory makers reflect structural under supply, not a demand spike. The uncertainty worth watching is CapEx. Whether Samsung, SK Hynix and Micron accelerate spending fast enough to close that gap in time or whether allocations tighten further before new capacity comes online will determine how severe the next 18 months actually get. Here's the thing. SK Hynix is exploring US memory chip production for the first time, with Intel's Ohio facility at the Center of Negotiations, a venture potentially involving Intel and major cloud providers is under discussion. If it materializes, it would mark the first significant international memory co-investment at an American domestic fab. That matters, because memory reshoring is now happening alongside logic chip reshoring, and the geography of global supply is fragmenting faster than most CAPEX plans anticipated. On the policy side, the US is proposing to extend USNCA rules of origin to AI hardware. The practical intent is to block Chinese chip manufacturers from rerouting products through Mexican assembly to avoid US export controls. Enforcement details remain vague. No component percentage thresholds or grace periods had been specified, and Mexico's resistance could complicate implementation. Foxconn and Taiwanese suppliers with Mexican operations are directly in scope. The September 24 Trump-Xi summit is unlikely to produce meaningful AI safety agreements. Chip export controls remain non-negotiable from the US side, and China's cooperation on AI safety continues to function as conditional geopolitical leverage rather than genuine policy alignment. The gap isn't closing. Consider this. The watchpoints heading out of this briefing are straightforward. Does open standard inference networking gain hyperscaler adoption, or does Nvidia's closed stack hold? Does memory CAPEX respond fast enough to prevent a deeper shortage?

And does the Mexico origin rules proposal develop teeth or stool in negotiation?

Those are the signals that will define where this industry moves next. Thanks for listening. This podcast was built using AI technology, a Yes We production.