A GPU is the easy part now. The hard part is everything that has to exist around it before a single accelerator does useful work, and most of that is in short supply.
Copper Road exists to track that other layer: the packaging, the memory, the power, the cooling, the raw materials, and the equipment that builds the equipment. We score each chokepoint on a single 0 to 100 scale we call the Tightness Gauge, sourced entirely from public data, and we watch it move. Here is the picture on day one.
The squeeze is upstream of the chip
Start with the thing a finished accelerator literally cannot exist without: advanced packaging. TSMC's CoWoS is how logic and high-bandwidth memory get fused into one package, and demand has run ahead of capacity for two years straight. The interposer at the center of it is capped by the lithography reticle, so as packages grow past 3.3 times reticle size, a single wafer yields only a handful of usable units. TSMC's own fix, a shift to panel-level packaging, is not scheduled for volume until 2028 or 2029. That is a long time to wait when the constraint binds today.
Memory tells the same story from a different angle. SK Hynix said its capacity was essentially sold out for 2026 back in October. Makers are pouring wafers into HBM because the margins are several times higher than conventional DRAM, and the result is a server-memory shortage that Goldman Sachs has called the worst undersupply in more than fifteen years, with contract prices climbing more than 60 percent in a single quarter. New fabs do not produce meaningful bits until late 2027 at the earliest. You cannot buy your way out of this one quickly, and everyone building a cluster is bidding for the same supply.
And underneath all of it sits one company most investors never think about. Every leading-edge chip and every advanced memory die is patterned on an extreme ultraviolet lithography machine, and exactly one firm on earth builds them. ASML closed 2025 with a record order book larger than a full year of sales, EUV systems carry lead times beyond twelve months, and Canon and Nikon left the technology more than a decade ago. ASML's annual build rate is, in a real sense, the speed limit on the entire frontier.
Then there is the power
Here is the part the chip-centric forecasts miss entirely. You can have every wafer, every HBM stack, and every package you want, and none of it computes without electricity delivered to the rack.
The grid is the first wall. Interconnection queues in the United States stretch past four years, gigawatts of capacity sit waiting to connect, and the timelines do not bend to quarterly demand. The hardware that moves that power is its own crisis. Large power transformers now carry lead times of four to five years, with demand up sharply against a supply base of only a few serious manufacturers. Gas turbines, the bridge a lot of operators are reaching for, are sold out at the major makers well into the end of the decade. None of these are technology problems. They are heavy-industry problems, and heavy industry does not scale on a software timeline.
This is why we treat power as a first-class part of the compute story rather than a footnote. On our gauge, transformers and gas turbines sit at the very top of the tightness ranking, above the chips.
Why a gauge, and why sourced
Two principles run through everything here.
The first is that constraints are comparable. HBM memory and a 500-ton transformer have nothing in common physically, but both can be the thing that stops a data center from coming online, and both can be scored on the same axes: how badly supply trails demand, how long new supply takes to arrive, how concentrated the producers are, and which way the last 90 days have moved. That is what the Tightness Gauge measures, and it is what lets you rank a whole supply chain on one page.
The second is that every number earns its place. Each figure on the site carries its value, the date it was true as of, and a link to where it came from. If a claim cannot be sourced, it does not appear. The gauge is only as trustworthy as the data feeding it, so we show the working on every page and let you check us.
What to watch
The through-line on day one is that the AI buildout has moved from a chip-supply story to a physical-infrastructure story. The scarce things are increasingly the unglamorous ones: a transformer, a turbine, a packaging slot, a memory wafer, a lithography machine with a two-year queue. These are the points where the whole effort can stall, and they move slowly enough to see coming if you are watching the right layer.
That is the layer Copper Road covers. Fifteen chokepoints are live now, each with a gauge and a sourced trail, and we will add the rest of the map and track every one as it moves.
The buildout moves at the speed of its tightest link. We are here to tell you where that link is.