A fibre tray is boring and permanent. Once it is pulled and labelled, the topology of a row of racks is fixed until someone with a ladder changes it. free-space optical inter-rack links promises to replace that rigidity with steerable beams, so a cluster can be rewired in software as jobs come and go. The trade is not bandwidth, which both can deliver, but availability: a cable does not care if a floor vibrates, and a beam does.
With installed fibre trays, each link is a physical path whose loss is measured once at installation and then stays within a narrow band for years. With free-space optical inter-rack links, each link is an alignment problem that must be solved continuously. A 400 Gbps optical beam with a narrow divergence has to stay on a receiver aperture across thermal expansion of the rack, fan vibration, doors opening and the slow settlement of a raised floor. Active steering loops can correct most of that, but the correction loop itself has a bandwidth and a failure mode.
Fibre sells a loss budget measured once; a free-space link sells an alignment budget that has to be paid continuously by a steering loop.
Part One: What is actually being compared
These two options are usually presented as alternative implementations of one capability. They are better read as two different answers to the question of where uncertainty is allowed to live. free-space optical inter-rack links absorbs it in one place, installed fibre trays in another, and a comparison that measures only the average outcome misses precisely the reason a team would accept the extra complexity of either one in cluster-scale optical interconnect.
There is an accounting asymmetry worth making explicit. free-space optical inter-rack links tends to put expenditure into capital, where it is spread over years and needs a approval process; installed fibre trays keeps it in operations, where it is easy to approve and difficult to cap. Finance and engineering usually prefer opposite sides of that line, so the decision is often settled by the budgeting calendar rather than by a measurement in cluster-scale optical interconnect.
Part Two: The arithmetic that decides it
The arithmetic shows up in the availability budget. A conventional fibre link might be specified at five nines of availability with a repair measured in hours. If a free-space link loses alignment for a few seconds on each of a handful of vibration events per day, it spends far more time degraded, even if each outage is short enough for a retransmission to hide. At 400 Gbps a one-second interruption is roughly 50 gigabytes of in-flight traffic that has to be buffered or replayed, and the buffer sits on the hosts rather than in the network.
What actually decides the outcome is the slowest link in the chain, not the fastest. free-space optical inter-rack links improves one step, but if the surrounding steps still dominate the total, the improvement barely moves the wall clock. This is the recurring disappointment in cluster-scale optical interconnect: a component-level win that is real, reproducible and almost invisible to the user because it was applied to the wrong term in the sum.
Deployment changes what gets measured. After free-space optical inter-rack links goes in, the metric people optimise tends to shift toward predictability; after installed fibre trays, toward unit price. That shift is why two teams can look at the same production data and reach opposite conclusions about which option won, and why an evaluation should record the metric before the trial rather than after it.
Part Three: Where the simple model breaks
The reconfiguration benefit is real but narrower than it sounds. A software-defined optical fabric helps most when the communication pattern changes between jobs, which is common in research clusters and rare in steady inference serving where the same all-reduce pattern repeats for months. Paying an availability penalty to rewire a topology that never changes is a bad trade. The interesting deployment is therefore a hybrid: fixed fibre for the stable spine, steerable links for a smaller reconfigurable pool.
At 400 Gbps a one-second pointing outage strands roughly 50 gigabytes of in-flight traffic, so short interruptions are not free even when retransmission hides them.
Part Four: The cost that arrives later
There is a maintenance cost that does not appear in a capital quote. A fibre plant is serviced with an optical time-domain reflectometer and a labelling convention; a free-space fabric needs clean apertures, periodic realignment, and a diagnostic that can distinguish a dirty lens from a drifting actuator. Dust is a mundane problem, but it is exactly the kind of problem that turns a clever architecture into a standing operations ticket.
Adoption is unlikely to be a single flip. The first movers are the cases where the deciding variable is most extreme, the middle of the market follows once tooling and reference designs exist, and the long tail stays with installed fibre trays until the cost of switching falls below the cost of maintaining an exception. A reasonable reading of the next several years is coexistence with a slowly moving boundary rather than a replacement event.
Part Five: What this changes outside the engineering team
For a city, the same comparison appears in building-to-building links that avoid trenching a street. A beam across a roof can deliver connectivity in an afternoon where a permit for digging takes a season. That is a genuine social benefit in dense areas, and it is also why the technology keeps resurfacing: the value is not the photons, it is avoiding the civil works. Inside a data hall the civil works are already done, which quietly removes the main advantage.
There is a mundane public-facing consequence as well: reliability and predictability are what people actually notice. A household, a commuter or a small shop does not experience an architecture, only whether a screen updates, a bill arrives on time or a service works after two years. Those are the terms in which the eventual judgement will be made.
A fair evaluation counts the reconfiguration events per month and prices each alignment outage at the buffer and replay cost it causes. If the topology rarely changes, the tray wins before any optical specification is compared.
A market sizing sanity check
A defensible sizing: if reconfigurable optical fabrics reach even a small fraction of new AI cluster builds by 2032, the transceivers, steering assemblies and control software together plausibly sit in the $1B to $4B annual range. The lower bound assumes the technology stays confined to research and high-performance clusters; the upper bound assumes it becomes a standard option in commercial training halls. The dominant uncertainty is how much job-to-job topology variation real workloads exhibit, not the raw link speed.
Treat the numbers as a scale model. They are built to show where the curves cross and which term dominates, and they deliberately round the inputs that vary most between deployments. The ranking is robust to those rounds; the exact threshold is not.
The comparison that matters
The comparison that matters is between the civil and operational rigidity removed and the pointing stability that must be maintained forever. Steerable links win where topology genuinely changes and trenching is expensive; fibre wins where the pattern is stable and dust is plentiful. The beam is a flexibility purchase, not a bandwidth purchase.
A change in regulation, in the supply of a single component, or in the tariff a site faces could invalidate the ranking overnight. The comparison is a snapshot of prices and constraints in cluster-scale optical interconnect, not a property of the technologies, and anyone reusing it in a different year should re-run the threshold rather than reuse the conclusion.
Count topology changes per month before buying reconfigurability: a fabric that rewires a pattern that never changes has paid an availability penalty for nothing.
SOL: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS
Creating this takes real effort — thank you for the support.
Top comments (0)