Micron’s revenue just quadrupled. Your next laptop is already more expensive. Here’s the mechanism — and what business leaders need to budget for.
On June 25, 2026, Micron Technology reported quarterly revenue of $41.5 billion — roughly four times what it made in the same quarter a year ago. Its stock jumped 15% in a single session. Its market capitalization, in the days that followed, surpassed both Meta and Tesla.
That same week, Apple announced price increases on MacBooks and iPads. Microsoft raised the price of Xbox consoles. Both companies cited the same cause: rising memory chip costs.
These are not separate stories. They are one mechanism. And if you run a business that buys computers, servers, or mobile devices — or plans to over the next two years — you need to understand it.
Every AI GPU — the chips that power ChatGPT, Claude, Gemini, and every large language model in production — requires a special type of memory called High Bandwidth Memory (HBM). Think of it as the short-term workspace that sits directly on top of the GPU die, stacked in layers, moving data at extraordinary speeds. Without HBM, an AI GPU cannot operate at scale.
A single NVIDIA H100 GPU — the workhorse of today’s AI infrastructure — contains 80GB of HBM3. Its successor, the H200, contains 141GB. The Blackwell B200, NVIDIA’s current flagship, uses 192GB of HBM3e. As AI model sizes grow and inference demands increase, that number will only go up.
Here is where the constraint lives. HBM is manufactured by exactly three companies in the world: SK Hynix (South Korea), Micron (United States), and Samsung (South Korea). That is the entire global supply chain for the memory that makes AI possible at scale.
The market shares, as of mid-2026, look roughly like this:
| Supplier | HBM Market Share | Status | Key Customer |
|---|---|---|---|
| SK Hynix | ~50% | Dominant, ramping HBM3e | NVIDIA (primary) |
| Micron | ~25% | Growing fast, high yield | NVIDIA, AMD, custom |
| Samsung | ~25% | Yield issues, behind schedule | Limited, recovering |
Samsung’s yield problems — a manufacturing issue where chips come off the production line defective — have been widely reported since late 2024 and remain unresolved through mid-2026. This has pushed NVIDIA and other hyperscalers to source more heavily from SK Hynix and Micron, concentrating supply further.
The practical effect: the global AI infrastructure buildout is running through two companies. Both are operating near capacity. Neither can scale overnight — a new HBM fabrication facility takes three to four years and tens of billions of dollars to build.
This is the part that surprises most business leaders. HBM is manufactured on the same fabrication lines as standard DRAM — the memory in laptops, smartphones, and enterprise servers. When a memory manufacturer prioritizes HBM production (which earns three to five times the margin of standard DRAM), it pulls fab capacity away from standard DRAM output.
Less standard DRAM supply with steady or growing demand means higher prices. Higher DRAM prices mean higher component costs for Apple, Dell, HP, Microsoft, and every other device maker. They pass those costs on to customers.
This is not hypothetical. In June 2026, Apple raised prices on MacBooks and iPads in India and other markets, explicitly citing memory component costs. Microsoft raised Xbox console prices. These are the first visible consumer-market effects of a supply squeeze that started in the enterprise. They will not be the last.
Micron’s financial results tell the story more clearly than any forecast model:
| Quarter | Revenue | YoY Change | HBM Signal |
|---|---|---|---|
| Q3 FY2024 | $6.8B | Pre-AI ramp | Minimal |
| Q3 FY2025 | $10.3B | +51% YoY | HBM3 ramping |
| Q4 FY2025 | $27.0B | +245% YoY | HBM3e volume |
| Q1 FY2026 | $34.0B | +290% YoY | Full production |
| Q2 FY2026 | $38.0B | +300% YoY | Constrained supply |
| Q3 FY2026 | $41.5B | +303% YoY | Demand exceeds supply |
To put $41.5B in quarterly revenue in context: that is more than Intel earns in an entire year. Micron, a company that most business leaders outside of tech procurement could not have named eighteen months ago, now has a larger market cap than Meta. It passed Tesla in the same week.
The HBM supply oligopoly is not the only concentration dynamic worth watching. The entire AI infrastructure stack is consolidating around a small number of companies in ways that create both risk and opportunity for enterprise buyers.
On the compute side, NVIDIA still controls roughly 80–85% of AI GPU shipments. Qualcomm entered the data center market in June 2026 with its Dragonfly C1000 CPU, already committed to a Meta deployment — but it is a CPU play, not a GPU play, and will take years to be meaningful competition. AMD continues to gain share at the margins. Custom silicon (Google TPU, Amazon Trainium, Meta MTIA) is growing but primarily serves internal workloads at hyperscalers.
What this means: the AI infrastructure supply chain has roughly three single points of failure for enterprise buyers — NVIDIA for compute, SK Hynix and Micron for memory. A disruption to any of these (geopolitical, yield-related, demand spike beyond capacity) would cascade immediately into data center pricing and availability.
| Infrastructure Layer | Effective Suppliers | Concentration Risk |
|---|---|---|
| AI compute (GPU) | NVIDIA (~85%), AMD (~10%), custom (~5%) | Very High |
| HBM (AI GPU memory) | SK Hynix (~50%), Micron (~25%) | Very High |
| Standard DRAM | Samsung, SK Hynix, Micron | High |
| AI cloud (hyperscale) | AWS, Azure, GCP + emerging GPUaaS | Moderate |
| AI cloud (GPUaaS) | CoreWeave, RunPod, Lambda Labs + regional | Growing |
Three developments could meaningfully relieve the HBM supply constraint, and business leaders should track all three:
Samsung’s HBM yield recovery. Samsung has been working to fix its HBM manufacturing yield problems since late 2024. A successful recovery would add approximately 25% more HBM capacity to the global supply chain almost overnight. Every quarter Samsung fails to qualify its HBM at scale is another quarter of tight supply. This is the single most important variable to watch in semiconductor procurement planning.
New fab capacity coming online. Micron is building a new memory fabrication facility in Idaho, with US CHIPS Act funding. SK Hynix is expanding capacity in Korea and building a new packaging facility in Indiana. These investments will add meaningful supply — but not until 2028 at the earliest. Capital expenditure commitments made today translate into supply in three to four years.
Efficiency gains in AI models. If AI models become dramatically more efficient per inference — requiring less memory bandwidth per query — demand growth could moderate even as infrastructure buildout continues. This is the wildcard. The history of computing suggests that efficiency gains get immediately reinvested into running more workloads, not reducing hardware spending. But it remains a genuine mitigating factor.
Business leaders who manage technology procurement need to update their planning assumptions for the next 24 months. The memory cost inflation currently visible in consumer devices (Apple, Microsoft) will work its way through enterprise hardware pricing with a lag of one to two quarters, as existing procurement contracts expire and are renegotiated.
The categories most exposed:
The practical implication is not that you should stop buying hardware. It is that you should accelerate procurement timelines for planned refreshes, lock in multi-year contracts where possible, and build memory cost inflation into 2027 budget models at 15–20% above current prices for DRAM-intensive products.
The daily AI briefing for business leaders — by stack layer and by function. Free.
The Cisco Question: Will NVIDIA Suffer the Same Fate? — We traced 25 years of tech market-cap leadership from Cisco’s peak in 2000 to NVIDIA’s today. The data shows a clear pattern.
The AI Killer App Watch — Which AI solutions are gaining real traction with real enterprises? We track 17 contenders across 5 categories.
The Demand Paradox — More software is being written than ever. IT services stocks are at nine-year lows. The market may be right about the pain but wrong about the permanence.