Share with your CTO
AI infrastructure demand isn’t softening, it’s stratifying. Across a recent industry discussion covered by Moomoo’s market community, executives from Nebius, Cerebras, and Lumentum all said demand exceeds supply, with Lumentum claiming its optical networking products are sold out five years forward. Former Intel CEO Pat Gelsinger named power availability, not customer appetite, as the binding constraint. The headline concern, that Meta and xAI renting out spare compute signals overbuilding, was dismissed by multiple executives as an outlier dynamic rather than a market-wide signal.
What this means for your business
Where you sit on the infrastructure stack determines how this reads. If your organization is still in early AI deployment, the supply squeeze is a real planning problem, long lead times on GPUs and data center capacity mean your 2026 compute needs have to be in procurement conversations now. If you’re already running workloads at scale, the more pressing signal is the industry-wide shift from “tokenmaxxing,” using AI as much as possible, toward return-weighted deployment, which changes how you justify the budget you already own.
The tokenmaxxing-to-valuemaxxing shift is the more consequential trend buried under the infrastructure headlines. Enterprises that spent 2023 and 2024 giving employees broad AI access are now being asked to show what that access actually produced. Open-source models from Alibaba and DeepSeek are making it easier to run cheaper, task-specific models rather than routing every workload through expensive frontier APIs from OpenAI or Anthropic. The practical effect is a tiered model strategy: frontier models for high-stakes, high-complexity tasks and lighter models for everything else. CTOs who haven’t built that routing logic into their architecture are overpaying today.
The executives quoted here, speaking at an industry event and reported through a platform whose audience skews toward retail investors, have an obvious interest in talking demand up, but the five-year optical networking sellout at Lumentum is a hard operational fact, not a sentiment reading, and it’s the kind of data point that survives the promotional frame. The falsification condition worth tracking is capex guidance from Microsoft, Google, and Amazon in Q3 earnings. If hyperscaler infrastructure spending reverses there, the “demand is unlimited” consensus breaks. Until that happens, infrastructure constraints are real and the build cycle has years left.
Concept deep-dive: Optical networking in AI infrastructure
Optical networking moves data between servers inside a data center using light pulses through fiber rather than electrical signals over copper, the way a fiber optic cable brings internet into a building rather than a phone line. In AI training clusters, where thousands of GPUs must exchange enormous amounts of data continuously, optical interconnects handle the traffic that would overwhelm traditional wiring. Lumentum’s sellout position signals that even the connective tissue of AI infrastructure, not just the chips, is constrained.
Based on reporting from $NVIDIA (NVDA.US)$ AI Infrastructure Boom Shows No Signs of …, originally published 2026-07-12 01:29:00.
