Fiber built the AI data center. Is it now holding it back?

Industry forecasts suggest global data center demand could more than triple by 2030, and some U.S. projections have it growing 20% to 25% a year. AI compute workloads are behind most of that growth, and almost all of the workloads run on fiber. Training a large language model means thousands of GPUs exchanging enormous datasets in parallel, at 800 Gbps today and 1.6 Tbps on the way. Copper cannot carry those rates without serious signal loss and heat. Fiber can, which is why it became the backbone of every AI facility.

But I believe this is not the most efficient architecture and it will approach a wall, eventually. Most AI clusters rely on point-to-point cabling, a dedicated run from every GPU server to a switch. A single rack can need hundreds of fiber connections, bundled together and tucked under floor tiles or into overhead trays. The glass itself is inexpensive. Everything around it is not. Every port needs an optical transceiver that can cost hundreds or even thousands of dollars, and installation labor typically runs $1 to $6 per linear foot.

Where point-to-point fiber falls short

Fixed fiber runs into three problems at once. The first is upgrades. Moving from 800G to 1.6T often means pulling out entire cable runs, and in dense bundles that risks damaging neighboring fibers that are still carrying live traffic. The second is reconfigurability. AI workloads are unpredictable. A fat-tree topology may suit today's training job, while tomorrow's real-time inference may call for a mesh. Fixed wiring locks the physical relationship between server and switch, so moving a GPU resource can mean re-running fiber across the room. The third is cooling. Bulky fiber bundles block server exhaust and get in the way of liquid-cooling retrofits, so we end up designing thermal systems around the wires.

Structured cabling and optical patch panels help. Technicians can change network logic by moving a patch cord instead of pulling new glass through the building. But the topology is still made of physical strands, and someone still has to touch them. Is that really the right foundation for workloads that change rapidly?

Why terahertz wireless deserves a serious look

Researchers have proposed an alternative, the terahertz wireless data center. Instead of hopping through several switch layers, data travels over short-range THz links of 1 to 100 meters, beamed from rack to rack in a single hop. Latency drops from microseconds to nanoseconds. The energy picture is encouraging too. Fiber transceivers consume 20 to 40 picojoules per bit, while emerging THz front ends operating under 20 meters are reaching single-digit picojoules per bit. Over the distance of a room, wireless is starting to look greener than glass.

Because beams may be steered electronically, the network layout can change on the fly to fit a specific training job, with no rip and replace. Standards work on these links for AI clusters is underway, and the beam steering algorithms keep improving.

The catch: coherence at terahertz frequencies

So why isn't every hyperscaler building one? The short answer is physics. Conventional systems reach 300 GHz and above by multiplying up a lower-frequency electrical reference, and that multiplication is the main source of phase noise in THz systems. Phased arrays add to the burden. Phase shifters, switches, mixers, and local oscillator distribution networks cannot operate fast enough for nanosecond beam alignment, and they bring limited phase resolution, high insertion loss, and thermal drift along with them. At data center scale the economics break down quickly, since 1,000 racks would need 1,000 high-precision oscillators.

Our approach: generate the carrier from light

The CTO development team at Adeia started with a different question. What if the THz carrier never had to be synthesized electronically at all? At a central coherence unit, we generate a pair of phase-related optical tones whose frequency difference equals the THz carrier we want. Those tones travel over fiber to compact remote radio heads at each rack, where a photodetector converts the optical beat directly into a THz carrier. There is no frequency multiplication and no local oscillator at the rack. The design inherits the natural phase stability of light, and noise common to both tones cancels along the shared fiber path.

The reference travels on what we call a coherence plane, a dedicated WDM overlay wavelength that carries phase instead of data to every node at once, separate from packet transport. A closed-loop calibration mechanism continuously corrects drift from fiber thermal expansion and hardware variation, which keeps a large group of racks in sync over time.

Because every rack derives its carrier from the same optical reference, signals are aligned at the moment they are created rather than reconciled afterward. Cell-free MIMO is not a new idea, but earlier designs depend on heavy digital signal processing to align clocks after the fact, and that is too slow for THz. Aligning at the source makes true distributed MIMO possible, with physically separate racks acting as a single phase-coherent antenna aperture that combines power coherently to overcome THz path loss. And since phase shifting happens in the optical domain at the central hub, we can steer beams in under a nanosecond and bypass electronic phase shifters altogether.

What this means for data center operators

The commercial logic is a cost inversion. In place of one expensive precision oscillator per rack, a single high-stability laser pair feeds the whole fabric through inexpensive passive optical splitters. Spending moves from active electronics to passive fiber. Topology stops being a construction decision and becomes a software decision, executed in nanoseconds, so a cluster could move from a mesh to a torus between training jobs. None of this needs exotic infrastructure either. The reference tones ride a dedicated wavelength on fiber that is already in the building, using WDM techniques operators know and trust.

Less cable also means better airflow and simpler liquid-cooling retrofits, and it leaves facilities better prepared for whatever comes next, quantum systems included. I like to think of it this way: the data center stops being a collection of nodes and becomes a single, phase-coherent electromagnetic entity. As CTO, I see that as more than a network upgrade. It is a new physical layer for the AI era, and it is exactly the kind of architectural rethink Adeia exists to pursue.

Key takeaways

• AI demand could more than triple global data center capacity needs by 2030, and fixed point-to-point fiber is becoming the bottleneck for upgrades, reconfiguration, and cooling.

• Short-range THz wireless links cut inter-rack latency from microseconds to nanoseconds and can beat fiber's energy efficiency below 20 meters.

• The blocker is phase coherence. Frequency multiplication injects phase noise, and per-rack precision oscillators do not scale economically.

• Our architecture distributes a shared optical beat over a dedicated coherence plane, so racks generate phase-aligned THz carriers photonically and operate as one distributed MIMO array.

• Beam steering moves to the optical domain, enabling topology changes in nanoseconds instead of recabling projects.

‍

FAQ

Why is phase coherence so hard at terahertz frequencies?

Conventional THz radios multiply a lower-frequency electrical reference up to 300+ GHz, which amplifies phase noise. The supporting hardware, including phase shifters, mixers, and LO distribution, cannot react at the nanosecond timescales THz beam steering demands, and thermal drift makes it worse.

What is distributed MIMO across server racks?

It is multiple radio heads on physically separate racks transmitting as one phase-coherent antenna array. Coherent power combining lets this virtual aperture overcome THz path loss that would defeat any single small antenna.

Does this eliminate fiber from the data center?

No. Fiber remains the backbone for data ingress and egress, and it takes on a new role carrying the optical coherence reference. THz wireless offloads the short-range, high-churn inter-rack links where fixed cabling hurts most.

How does this reduce data center costs?

It replaces per-rack high-precision oscillators with one central laser pair and passive optical splitters, shifting capital expense from active electronics to passive optics. It also enables reconfiguration without recabling labor or downtime.