Observe the logs. On July 14, Ericsson dropped 10% in a single session. The headline says costs are rising. The real story is that AI's insatiable appetite for HBM memory is systematically cannibalizing the supply chain for everyone else. This isn't a telecom problem. It's a structural signal that the old industrial economy is being sacrificed to feed the new AI machinery.
I don't trade headlines. I trade order flow. And the order flow here is clear: memory prices are exploding, not because of a cyclical uptick, but because Samsung, SK Hynix, and Micron have redirected their most advanced DRAM fabs to produce HBM3e for Nvidia's GPUs. The capacity that used to go into DDR4 and LPDDR5 for network switches and base stations is now locked into 3D-stacked memory for AI training clusters. Ericsson is the first major victim of this resource war, but they won't be the last.
Context: The Victim Profile
Ericsson is not a tech company in the AI sense. They are a telecom hardware integrator. They buy standard memory chips (DDR4, LPDDR4), FPGAs, and ASICs, and they assemble them into 5G base stations and routers. Their end customers are cost-sensitive telecom operators who are already cutting CapEx because 5G isn't generating the ROI they expected. So Ericsson sits in the classic squeeze: upstream, memory suppliers have monopoly pricing power; downstream, their customers hold all the cards.
Smart contracts don't lie. The financial contracts here do. Ericsson's Q2 EBITDA dropped 7%. Their CEO explicitly blamed "input cost increases" and warned that core network margins would be under pressure. Citi analysts pushed that pressure horizon all the way to 2027. That's not a quarterly blip. That's a multi-year structural shift.
Core Analysis: The Memory Bottleneck
Let me trace the on-chain logic of the physical supply chain. AI requires HBM — High Bandwidth Memory — which is essentially a stack of DRAM dies connected through TSVs (Through-Silicon Vias) and micro-bumps. This is advanced packaging. To produce one HBM3e stack, you consume roughly the same wafer capacity as 3-4 standard DDR5 dies. And Nvidia is ordering hundreds of thousands of these stacks per quarter.
I watch the blockchain, not the ticker. But in this case, the ticker tells a story that aligns perfectly with the hardware production cycle. The memory oligopoly — Samsung, SK Hynix, Micron — is running at 100% HBM capacity. To increase HBM output, they must convert existing DRAM fabs. That conversion takes 18-24 months for a new line to reach full production. Until then, the supply of non-HBM memory is shrinking, and the price for standard DDR and NAND is rising.
This is why Citi sees pressure through 2027. The capacity investment cycle for memory is notoriously long. The decision to build a new HBM fab was made in late 2023. The first wafers from that fab won't meaningfully impact the market until late 2025. By then, Nvidia will have doubled its demand again. The math doesn't balance.
Contrarian Angle: The Squeeze Is Intentional, Not Accidental
The common narrative is that Ericsson is just unlucky — caught in a supply-demand imbalance. That's wrong. This is a deliberate resource allocation by the memory oligopoly. They are choosing to allocate their most profitable capacity to the highest bidder: AI. Ericsson, with its low volume and thin margins, is a lower-priority customer. This is not a bug in the market; it's a feature of oligopoly pricing power in an era of scarce advanced manufacturing.
Code is law, but human greed is the bug. The memory suppliers have no incentive to prioritize Ericsson. Their AI clients (Nvidia, AMD, Google) are growing at 100%+ and can absorb any price increase. Ericsson's telecom operator clients are struggling to grow. So the memory suppliers extract maximum rent from the AI side, and Ericsson is left to either eat the cost or pass it on — which is impossible given their customer's CapEx constraints.
This creates a Darwinian filter. The companies that survive the next 3-4 years will be the ones with pricing power or vertical integration. The ones without it — like Ericsson, and likely many other industrial IoT, automotive, and traditional networking firms — will see their margins permanently compressed until the next memory fab wave arrives or demand shifts.
Takeaway
I see a clear trading signal. Watch the memory spot prices versus AI CapEx announcements. If HBM prices stay elevated through 2025, expect more earnings warnings from non-AI hardware companies. Ericsson's 10% drop is not an overreaction. It is a rational repricing of a company whose business model is now structurally squeezed by an AI-driven resource war it cannot win.
I don't hold Ericsson. I hold the memory shorts. But more importantly, I am watching the blockchain for the next victim — the next company that relies on standard memory chips without owning its own supply. The logs will tell us who's next.