Copper Can’t Reach the Top of the Rack Anymore

Copper is dead.

Inside the $500 billion AI buildout, a quiet physics failure is forcing the switch to light. That failure is the market POET Technologies serves.

Ten gigawatts. Five hundred billion dollars. One customer, but first consider that 60% of power in AI Data Centers never touches a chip…

That is the shape of the deal Reuters reported on July 26: Nvidia in talks to guarantee roughly $250 billion in financing so OpenAI can lease a data center campus that SoftBank’s SB Energy is building on a former uranium enrichment site in southern Ohio. The guarantee covers the lease and construction debt. The chips inside, worth up to another $350 billion, are a separate conversation. Reuters could not verify the report, and the talks could still collapse.

Whether or not this deal closes, the direction is settled. AI infrastructure spending is set to top $700 billion this year. The scarce input is no longer capital or land. It is power.

The shovels are already in the ground

New Census Bureau data, reported by Data Center Knowledge, shows data centers are now the largest category of U.S. commercial construction, running at a $51 billion seasonally adjusted annual rate. Offices have been passed. Microsoft broke ground in Indiana the same month. This is a building boom, measured in concrete.

But a data center is only as useful as the power it can feed to chips. And here the industry has an uncomfortable accounting problem.

Sixty cents of every energy dollar moves data

Roughly 60 percent of data center energy is spent on data movement, not compute. AI data center power demand is projected to grow 50 percent globally by 2027. In a power-capped campus, interconnect efficiency is not an optimization. It is capacity.

Copper is losing this fight on physics alone. At 1.6 terabit speeds, passive copper cannot span the height of a rack. The industry’s answer is co-packaged optics: put the optical engine next to the switch silicon and cut the electrical path. The result is a 30 to 40 percent reduction in per-port power at 800G and above. Hyperscale buyers have noticed. Watts per terabit now appears on vendor scorecards.

This is why the interconnect layer, long a commodity afterthought, has become strategic. Whoever solves optical power at scale sells into every gigawatt announced this year.

POET’s bet: the laser is the bottleneck inside the bottleneck

POET Technologies (NASDAQ: POET) builds optical engines on its wafer-level Optical Interposer platform, targeting 800G today and 1.6T and 3.2T next. Its sharpest claim, though, is about a single component.

In an April post on the company’s blog, POET argued that the co-packaged optics transition stalls without external light sources built at semiconductor scale rather than as boutique photonics. The market framing it cites: AI Ethernet optics and CPO at $16.5 billion in 2025 and heading to $26 billion in 2026, with the co-packaged component segment growing from $1.3 billion to $2.7 billion by 2028 and external lasers called out as immediate demand. Large-scale deployments accelerate from 2027. Qualification and design wins are being decided now.

Real Talk

No disclosure connects POET to the Ohio campus or any tenant of it, and none is implied here. The connection is arithmetic. Every gigawatt announced makes the watts-per-terabit problem larger, and that problem is the market POET has built its platform to serve.

The proof point is dated. POET reports second quarter results on August 12. The questions that matter: shipment progress under the Lumilens supply agreement, volume traction in 800G and 1.6T engines, and whether customer qualification is converting to revenue. The macro tailwind is now the loudest it has ever been. August 12 shows whether the company is catching it.