The AI boom is running into a copper problem

Lightmatter says photonics could materially improve AI cluster performance, claiming all-optical connections can link 500 GPUs directly versus four racks with copper and cut model training time by as much as 3x. The startup has raised $850 million from investors including Google, Fidelity, and T. Rowe Price and has joined Nvidia's NVLink Fusion ecosystem, boosting its relevance in AI infrastructure. It is also targeting a reduction in cabling for hyperscale data centers from about 300 miles to 150 miles through its BiDi technology.

Analysis

The near-term winner is not “optics” in the abstract; it is whoever controls the bottleneck between GPU clusters and can translate a technical superiority story into procurement wins at hyperscalers. NVDA gains the most because this makes its ecosystem stickier: if optical interconnect becomes the default upgrade path, the value shifts from raw accelerator silicon toward system-level architecture, and NVDA is best positioned to own the standards layer before rivals can fragment it.

The second-order effect is a capex reallocation inside data centers. If interconnect distance becomes effectively unconstrained, operators can optimize for cooling, serviceability, and power distribution rather than cable reach, which should improve utilization and lower marginal inference/training costs over a 12-24 month horizon. That is negative for legacy copper-heavy suppliers and potentially a slow burn positive for optical component vendors, but the bigger implication is that AI buildouts can scale into more power-constrained sites without waiting for perfect physical layouts.

TROW’s exposure is more indirect but meaningful: the venture/late-stage financing arc benefits if photonics moves from “science project” to infrastructure standard, because this validates deeper private-market deployment into picks-and-shovels AI hardware. The contrarian risk is adoption lag: hyperscalers are notoriously conservative on mission-critical infrastructure, and even a superior technology can take multiple budget cycles to penetrate at scale. If performance claims prove difficult to mass-manufacture or integration raises reliability issues, the market could rerate these names on disappointment within 1-2 quarters.

The consensus may be underestimating how much of the economic value is captured before full optical replacement. Even partial deployment that reduces cabling complexity and rack density can unlock incremental GPU utilization and lower build time, which matters more than the end-state vision. The more overdone part may be the assumption that this is an immediate TAM explosion; in reality, it is likely a staged transition with NVDA as the gating beneficiary and the rest of the ecosystem lagging until standards harden.

Key Decisions for Investors

• Long NVDA vs short SMCI on a 3-6 month horizon: NVDA captures the standard-setting and ecosystem pull-through, while SMCI is more exposed to commoditized rack-level build activity if optical interconnect reduces differentiation; target 1.5-2.0x relative performance if NVDA retains ecosystem control.
• Buy NVDA call spreads 6-9 months out, financed by selling nearer-dated upside: upside is tied to the market re-rating NVDA as an infrastructure platform rather than just an accelerator vendor; risk is capped if photonics adoption takes longer than expected.
• Initiate a basket long in optical-enablement suppliers and pair against copper-heavy interconnect exposure over the next 12 months: express as a relative-value trade because the adoption curve will likely be gradual, but the margin pool should migrate steadily away from legacy cabling.
• Add modest exposure to TROW as a private-markets validation trade, but only on pullbacks: the memo-friendly angle is that successful photonics commercialization supports late-stage venture marks; keep sizing small because the linkage is indirect and sentiment-driven.