Micron reported a staggering 15-fold profit increase for Q2, with net income hitting $4.2 billion—up from just $280 million year-over-year. The headline number that matters most to investors and AI labs alike: HBM4 high-bandwidth memory revenue crossed the $1 billion threshold for the first time. Gross margin approached 85%, a figure that would make any semiconductor fab salivate. These numbers aren't anomalies; they're the opening act of what's becoming a full-blown toll booth on AI infrastructure.

The Margin Machine

That 85% gross margin isn't luck—it's pricing power. Micron charges roughly five times more per gigabyte for HBM4 compared to standard DDR5 memory, according to industry estimates. Unlike commodity DRAM where margins compress when supply catches demand, HBM4 sits in a different category: it's the critical fuel for NVIDIA's H200 and B200 GPUs, which currently train most of the world's large language models. When you're the gas station on every AI highway, you set the price at the pump.

The Anthropic Play

Micron isn't just selling memory—it's buying into the AI stack itself. On June 22, 2026, the company backed Anthropic's Series H with a multi-year memory supply deal, locking in demand for its future capacity. This vertical integration mirrors what Google is doing with its $11 billion annual commitment to SpaceX compute. By tying AI labs to long-term supply agreements, Micron creates predictable revenue streams while ensuring those labs have guaranteed access to scarce memory during training runs. It's a clever hedge against the inevitable capacity glut when Samsung and SK Hynix ramp their competing HBM4 products.

Competitive Landscape: The Clock Is Ticking

The 85% margins won't last forever—Samsung and SK Hynix are both ramping HBM4 production, and commodity DRAM history shows that three-player markets eventually normalize pricing. But here's the twist: hyperscalers are increasingly designing custom memory architectures to reduce dependence on merchant suppliers. Google's TPU v6 already uses HBM4e, and Microsoft isn't far behind. The top 10 AI labs are separated by just 44 Elo points, meaning more training runs and sustained memory demand—but also means customers have leverage to switch vendors once alternatives exist.

Key Takeaways

  • Micron's $4.2B quarterly profit proves AI memory is the highest-margin business in semiconductors right now
  • The Anthropic deal signals a new era of vertical integration between memory suppliers and AI labs
  • HBM4 pricing at 5x DDR5 per GB won't hold once Samsung/SK Hynix reach full production capacity
  • Hyperscaler custom silicon (Google TPU v6) represents the long-term existential threat to merchant memory vendors

The Bottom Line

Micron just printed numbers that make NVIDIA look conservative—but the window of 85% margins is closing. Watch their late September earnings call for guidance on HBM4 pricing and capacity expansion plans. If they don't announce aggressive capex to defend market share against Samsung and SK Hynix, those fat margins become a target rather than a fortress.