WuBlockchain’s WhiteLine Daily highlighted four themes across AI hardware and semiconductor markets: OpenAI’s first disclosed benchmark results for its in-house inference chip Jalapeño, a sharp rise in South Korean DRAM export pricing, Morgan Stanley’s tally of long-term AI infrastructure commitments, and accelerating data center revenue at Semtech.
OpenAI releases first measured results for Jalapeño
OpenAI disclosed the first benchmark results for Jalapeño, its first in-house inference chip. The company used SemiAnalysis’ public InferenceX benchmark to test GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.
According to OpenAI, Jalapeño delivered peak throughput per watt that was 1.5x to 1.9x higher than the comparison systems across the three tests, while end-to-end latency was 1.7x to 3.6x lower. In the GPT-OSS test, throughput per watt was about 1.9x higher than Nvidia’s GB200. In the DeepSeek R1 test, it was about 1.7x higher than GB300.
SemiAnalysis said its team verified benchmark execution on site at OpenAI’s lab, although the performance figures were still provided by OpenAI.
WhiteLine Daily said the release marks the first time Jalapeño has produced quantifiable performance data. In its view, OpenAI’s path to lower inference costs has now extended from models and software to self-developed ASICs. The note said the more direct impact on Nvidia would be in inference share, while for Broadcom the results add support for demand in custom AI chips.
South Korean DRAM export prices extend their rise
ChosunBiz, citing data from the Korea Trade Statistics Promotion Institute, or TRASS, reported that South Korea’s DRAM export unit price, excluding modules, reached $92,183 per kilogram during Aug. 1-20, up 401% from a year earlier and about 12.5x above the low seen in January 2023.
During the same period, DRAM export value rose 504.8% year over year to $9.807 billion.
Goldman Sachs expects the global DRAM supply gap to widen from 5.0% this year to 5.9% next year, with shortages potentially lasting through 2028. TrendForce expects HBM to account for about 30% of DRAM wafer input at Samsung, SK hynix, and Micron by the end of next year, while representing only about 13% of bit supply.
WhiteLine Daily said HBM continues to take up DRAM capacity and that supply constraints in conventional DRAM are already showing up in pricing. The report also noted that export unit prices are affected by changes in product capacity and mix, but added that if contract prices keep rising, profit sensitivity for MU, Samsung, and SK hynix could still move higher.
Morgan Stanley puts AI infrastructure commitments above $3.1 trillion
Bloomberg, citing a Morgan Stanley report, said Alphabet, Amazon, Meta, Microsoft, and Oracle, together with Nvidia and Broadcom, have accumulated more than $3.1 trillion in AI infrastructure-related commitments through purchase obligations, leases, guarantees, and other financial support.
Within that total, the five cloud companies alone have more than $2.7 trillion in undiscounted commitments, roughly equal to three years of their current operating cash flow. Alphabet had the highest figure among the five at about $890 billion.
Morgan Stanley expects capital expenditure at the five hyperscalers to reach about $1.2 trillion in 2027.
WhiteLine Daily said AI compute expansion is no longer visible only in current-period CapEx, because procurement, leasing, and guarantee commitments are already locking in spending years in advance. The report added that when cash flow growth fails to keep up with capital deployment, financing costs and credit spreads become more important valuation variables.
Semtech posts record data center revenue
Semtech reported fiscal 2027 second-quarter results early on Aug. 26 Beijing time. Revenue came in at $341.9 million, up 33% year over year.
Data center revenue reached a record $100 million, up 91% from a year earlier and 39% from the prior quarter, driven mainly by demand for 800G and the initial volume ramp of 1.6T CopperEdge and FiberEdge products.
Semtech said data center revenue in the third quarter is expected to rise about 45% quarter over quarter and about 160% year over year. Management said 1.6T products already account for more than 50% of third-quarter data center revenue.
WhiteLine Daily said that suggests the optical interconnect upgrade cycle is moving out of product roadmaps and into reported revenue.
Main thread of the day
The report’s central takeaway was that HBM continues to consume DRAM capacity, keeping memory prices on an upward path, while 1.6T optical interconnect is beginning to show up in revenue and OpenAI is using in-house ASIC development to improve inference efficiency. At the same time, more than $3.1 trillion in long-term procurement, leasing, and financial commitments shows that the compute buildout is testing both hardware efficiency and financing capacity.