Nvidia is no longer just designing chips. It is now managing upstream materials—a signal of how deep the fracture in the AI supply chain has become.
On June 8, Nvidia went directly to Co-Tech Development, a Taiwanese copper foil manufacturer, to discuss long-term capacity planning for HVLP4 copper foil. It bypassed its traditional intermediaries—the copper-clad laminate (CCL) makers—altogether. The move is quiet but seismic: the world's most powerful chip company is now behaving like a mining company, securing raw inputs more than a year ahead of need.
What HVLP4 Is, and Why Nothing Else Will Do
HVLP stands for Hyper Very Low Profile. The "4" denotes a generation of copper foil with surface roughness as fine as 0.6 to 1 micron—thinner than a human hair by orders of magnitude.
At those tolerances, electrons traveling along the foil surface lose dramatically less energy. In AI servers running at 800G or 1.6T networking speeds, that efficiency isn't a luxury—it's a structural requirement. Traditional copper foils introduce signal loss that makes high-frequency AI interconnects unreliable. There is no workaround. You either use HVLP4, or you don't build the server.
Every major AI accelerator platform depends on it—Nvidia's GB200 and GB300, AMD's lineup, and Huawei's Ascend chips alike. Compounding the pressure further up the chain: an AI server consumes roughly eight times more CCL material than a conventional server.
The Numbers: A Gap That Widens Into 2027
The HVLP4 supply shortfall for 2026 is estimated at approximately 1,500 metric tons—over 40% of total demand. By 2027, that gap is projected to widen to 2,500 metric tons. Monthly demand is already forecast to exceed 560 tons, more than the market's dominant supplier, Japan's Mitsui Kinzoku, can produce alone.
The technical barrier is the moat and the trap at once. Achieving the ultra-low roughness, uniformity, and adhesion that HVLP4 demands—at industrial yield rates—takes 12 to 18 months of qualification per customer. Nobody can simply flip a switch on new capacity. Mitsui is expanding, and so are Taiwan's Co-Tech and Korea's Lotte Energy Materials. None of it moves fast enough to close the gap before 2028.
Downstream, CCL makers are already rationing output to PCB and IC substrate manufacturers. Lead times on high-end laminates have stretched to six months. Prices are moving too—Mitsui has already pushed through roughly a 12% increase on copper foil, with further hikes likely as demand accelerates through the second half of 2026.
The Investment Landscape: Discrimination, Not Discovery
The easy money is gone. The trade has shifted from discovery to discrimination, and that distinction matters.
Mitsui Kinzoku remains the institutional-grade expression of this thesis. At roughly 16.8x earnings—against peers priced like pure-play AI semiconductor names—the market still treats it as a diversified Japanese materials conglomerate rather than a gatekeeper to AI infrastructure. The stock has already surged roughly 855% over the past year, so most of the multiple re-rating has happened. What's left is a question of earnings durability—and if the shortage holds through 2027, that durability looks credible.
Co-Tech has already been discovered. At roughly 109x trailing earnings after a nearly ninefold run, the stock is priced for flawless execution: clean yields, customer acceptance, and pricing power that Nvidia doesn't push back on. Nvidia's direct approach validates Co-Tech's relevance—but it also guarantees that margin discipline, not blank-check pricing, will govern the relationship going forward. That's the risk the current price doesn't reflect.
Lotte Energy Materials is the asymmetric option. It has already supplied HVLP4-grade foil to Doosan Electronics for AI accelerators and has announced a 1,800 ton-per-year mass-production system. But its income statement still shows negative profitability, and the stock trades on optionality, not earnings. This belongs in a catalyst sleeve, not a quality-compounder portfolio.
The insight the market keeps missing: Nvidia's intervention is not unambiguously good news for suppliers. When the platform owner goes straight to the foil maker, it brings demand certainty—but also procurement discipline, in equal measure. Prepayments, consignment arrangements, forced dual-sourcing, and long-term volume agreements will shape, and cap, supplier economics from here. The shortage premium is real. Its duration is commercially and politically bounded by the very customers funding it.
One more note on positioning: raw copper exposure—COMEX near $6.50/lb, close to record highs—is the wrong instrument for this trade. HVLP4 scarcity is a qualified-process-capacity thesis, not a cathode-copper thesis. Buying copper to express it is like buying sand to express EUV lithography scarcity.
What Comes Next
The bottleneck persists through 2027. Prices keep climbing, though not without limit. Capacity announcements will underdeliver on timing—specialty materials don't scale the way semiconductor fabs do. And the trade will increasingly split between incumbents with proven yield and challengers whose ramps remain unvalidated.
At this point, the contrarian position isn't contrarian anymore: the shortage is real. The edge now belongs to investors who can tell qualified scarcity apart from press-release capacity—and who understand that Nvidia's move upstream is as much a constraint on supplier upside as it is confirmation that the bottleneck is real.
not investment advice