Nvidia results put HBM and export controls at the center of the AI chip cycle

Moki

2 min read·2026/08/27

Semiconductors Add to Google preferred sources

Moki

0:00

--:--

Nvidia’s record quarter confirmed that AI infrastructure demand remains far above supply, but the constraint has shifted from accelerator demand alone to HBM, server memory, advanced packaging, power and compliance. Memory suppliers rallied, OpenAI pushed a custom inference-chip path, and a U.S. probe into alleged China diversion kept export controls in focus.

Image / AP News

Nvidia’s record quarter resets the AI chip demand bar

Nvidia reported fiscal second-quarter revenue of $96.2 billion, up 106% year over year, with Data Center revenue of $89.0 billion, up 117%. Its third-quarter outlook calls for about $108.0 billion in revenue, plus or minus 2%, while assuming no Data Center compute revenue from China. The numbers confirm that cloud providers, AI labs, enterprises and sovereign customers are still expanding AI infrastructure, but they also raise the execution bar for every supplier tied to Vera Rubin, Blackwell Ultra, networking and advanced packaging.

Read the full story

Image / Seoul Economic Daily

HBM and server memory become the visible choke point

Nvidia’s CFO commentary showed supply and capacity commitments jumping to $279 billion from $119 billion in the prior quarter, primarily for memory procurement. Seoul Economic Daily reported that management sees memory tightness lasting at least through fiscal 2028, and Korean memory shares reacted immediately: Samsung Electronics and SK hynix rose sharply, while Micron gained after hours. The industry implication is that scarce HBM is now dictating how quickly AI systems can ship, while standard server DRAM and enterprise SSD markets are also being squeezed as Samsung, SK hynix and Micron allocate capacity toward higher-margin AI memory.

Read the full story

Image / Tom

OpenAI’s Jalapeño shows the custom-ASIC route, but not an escape from supply constraints

OpenAI’s Broadcom-developed Jalapeño inference ASIC became a second major AI-chip signal this week. Tom’s Hardware reported OpenAI-claimed InferenceX results showing 1.5x to 1.9x higher throughput per kilowatt and 1.7x to 3.6x lower latency than Nvidia GB200 and GB300 systems on selected inference workloads. The caveats matter: Vera Rubin was not in the comparison, Jalapeño targets inference rather than training, and the chip still uses six HBM4 stacks, 216 GiB of memory, TSMC-class leading-edge manufacturing and advanced packaging. Custom silicon may reduce dependence on Nvidia for some inference loads, but it also creates another large buyer for the same scarce HBM4 and packaging capacity.

Read the full story

Image / The Straits Times

Apex probe keeps the China export-control risk attached to AI hardware

The Straits Times, citing Bloomberg, reported that U.S. authorities are investigating Singapore-based Apex Logistics over alleged transport of Nvidia-powered Super Micro AI systems to China through routes involving Taiwan, the United States, Southeast Asia, Hong Kong and mainland China. The Bureau of Industry and Security is reportedly examining 47 Apex-handled shipments from 2024; Apex said it is cooperating and the matter remains an investigation, not a final finding. For the semiconductor chain, the message is that export controls are moving deeper into logistics and server-distribution compliance, while Nvidia’s own outlook still assumes no China data-center compute revenue.

Read the full story