Glass Precede Intelligence
You cannot scale AI without glass. And the glass is running out. The AI Gold Rush of 2024–25 was about GPUs. The infrastructure super-cycle of 2026–2028 is about the physical…
You cannot scale AI without glass. And the glass is running out.
The AI Gold Rush of 2024–25 was about GPUs. The infrastructure super-cycle of 2026–2028 is about the physical plumbing that makes those GPUs useful — and the plumbing is running out.
For every $1 spent on AI silicon, $3–5 is required for the fibre, power, cooling, and switching fabric connecting it. AI clusters do not generate North-South traffic like traditional cloud infrastructure. They generate massive East-West traffic — GPU node to GPU node during distributed training and inference runs. The fibre density requirement is not linear with cluster size. It is multiplicative. Infrastructure topology analysis shows AI Factory fibre density requirements are an order of magnitude greater than equivalent traditional cloud clusters — with estimates placing the differential at 30–40× at hyperscaler scale. Hyperscalers understood this first. Multi-year supply agreements locking in high-purity optical fibre capacity are being signed at the $6 billion level. Unreserved global glass capacity is being absorbed faster than new manufacturing capacity is coming online.
The constraint on AI scaling in 2026 is not compute. It is the glass connecting the compute — specifically, glass pure enough (99.999% silica, sub-0.18 dB/km attenuation) to sustain the 200G-per-lane PAM4 signalling that current 1.6T infrastructure requires. Companies that own their silica preform manufacturing are structurally insulated from this shortage. Companies that buy glass from third parties are not.
Upgrading to 1.6T collapses network tiers — based on vendor specifications and early deployment data: 48% fewer switch ports, 43% better power-per-Gbps, 15–25% faster large model training completion. By 2026–2028, inference is projected to overtake training as the dominant AI workload. Every organisation that has not locked in 1.6T-ready infrastructure before that inflection will be paying premium energy costs per inference token against competitors who secured their physical stack a year earlier.
The software layer that governs, verifies, and optimises this physical stack in real time is not yet owned. That is where the next generation of infrastructure companies gets built.
The picks-and-shovels play in AI is not the GPU. It is the glass, the preform, and the software layer nobody has built yet.