Three weeks of chip earnings just settled a question capex planners have been circling all year: the scarce input in the AI buildout is no longer the GPU. On July 7, 2026 Samsung guided to a record quarter. On June 24 Micron posted the highest revenue in its history. And on July 9 SK Hynix priced the largest first-time US share sale by any foreign company. All three run on the same fuel: AI memory.
The signal for anyone budgeting business technology spend is blunt. The choke points in the second half of 2026 are HBM wafer allocation, memory order books already sold out into 2027 and 2028, and grid-connected megawatts. Pricing power has moved with them, away from accelerator buyers and toward the companies that make memory, power, and cooling.
What’s actually new
Samsung’s preliminary Q2 guidance put operating profit near 89.4 trillion won (about $58.4 billion), a roughly 19-fold jump from 4.7 trillion won a year earlier, on revenue of about 171 trillion won (up 129%). It is the third straight record quarter, driven almost entirely by HBM, server DRAM, and NAND. Shares still fell as much as 10% intraday on questions about how long AI demand and capex hold. Full segment results are due around July 30, so treat any Samsung HBM-specific number before then as unreleased. We covered the guidance in more depth when Samsung guided to a record $58 billion quarter.
Micron’s fiscal Q3 is the cleaner tell. Record revenue of $41.46 billion, up 346% year over year, at an 84.9% non-GAAP gross margin. DRAM alone was $31.3 billion, with sequential average selling prices up in the low-60s percent range. Data center SSD revenue topped $5 billion. HBM3E and HBM4 are sold out through calendar 2027, with demand running into 2028, and Micron guided fiscal Q4 to $50 billion (plus or minus $1 billion) at roughly 86% margin. CEO Sanjay Mehrotra told analysts the tightness extends beyond 2027.
SK Hynix supplied the capital-markets exclamation point. It priced a Nasdaq ADR offering at $149 per ADR, raising about $26.5 billion, more than seven times oversubscribed, and the stock jumped roughly 13 to 14% on its July 10 debut. The company’s entire planned 2026 output of HBM, DRAM, and NAND is already pre-sold. SemiAnalysis projects memory will exceed 30% of Nvidia system spend by end-2026 and top 40% in 2027, versus the teens share memory historically took of a server bill of materials.
Micron CEO Sanjay Mehrotra sat down with Jim Cramer on CNBC to talk chips.
— Dan Nystedt (@dnystedt) July 1, 2026
Key comments:
“Micron is firing on all cylinders.”
“Every AI system, regardless of the device it is in, requires more memory at higher performance in order to really bring out the full potential of AI…
What this means for teams buying AI compute
If memory is one binding constraint, power is the other. Nvidia’s Jensen Huang now frames every future data center as power-limited: with fixed gigawatts, throughput per watt equals revenue. Nvidia positions its 2026 Vera Rubin generation at about 10x the inference throughput per megawatt of Blackwell. Next-gen AI data center capex of $50 to $60 billion per gigawatt is a figure attributed to Huang in keynote recaps, not a filed number, so weigh it loosely.
Cooling is the third squeeze. Analyst estimates for the 2026 liquid cooling market diverge widely, from roughly $4.1 billion (MarketsandMarkets) to $6.8 billion (Mordor Intelligence), so there is no single credible number yet. Goldman Sachs estimates, reported secondhand, put liquid-cooled AI server penetration rising from 15% in 2024 to 76% in 2026. The practical read connects to why the buildout keeps scaling, laid out in the AI data center boom in verified numbers: a GPU order confirms nothing until the memory, the megawatts, and the thermal envelope are secured alongside it.
What to watch
Three markers. Samsung’s full breakdown around July 30, for the first HBM-specific revenue split. Whether Mehrotra’s “beyond 2027” language hardens into 2028 supply commitments. And the power side, where projects increasingly arrive with their own generation, a pattern visible when Meta’s first Canadian data center came with its own power plant. The buyers with leverage this cycle are the ones who booked memory and megawatts early.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
