O-RAN Didn't Open the Radio Unit.
That's Our Opportunity.

The O-RAN Alliance standardized the fronthaul link between the DU and the RU — the protocols, the planes, the packet formats. But it left the RU's internal baseband processing as a vendor black box. For a company building neural inference inside the radio unit, that's a feature, not a bug.

What O-RAN Actually Specified

When people say "Open RAN," most assume the entire radio access network has been cracked open — modular, interchangeable, fully disaggregated. The reality is more nuanced. The O-RAN Alliance's central contribution is the Open Fronthaul interface: a set of specifications (Control, User, Synchronization, and Management planes) that define how the O-DU talks to the O-RU over an eCPRI/Ethernet link using the 3GPP 7.2x functional split.

This split draws a line through the physical layer. The DU handles high-PHY functions — scrambling, modulation, layer mapping, resource element mapping. The RU handles low-PHY — FFT/iFFT, cyclic prefix insertion and removal, and (for Category B radios) digital beamforming. O-RAN specifies what the RU must do functionally, and how it communicates with the DU. It does not specify how the RU implements its processing internally.

WG7 has published white-box hardware reference architectures for the O-RU, but these are high-level block diagrams — not open silicon you can swap components in and out of. No network equipment provider (NEP) ships an RU where you can replace the DSP pipeline. The RU remains, in practice, a proprietary box.

The ULPI Debate: More Intelligence Moving to the RU

There is one important recent development that proves the industry sees the current split as insufficient. In June 2023, the O-RAN Alliance agreed on a new work item called Next Generation Lower-Layer Split (NG-LLS), combining two proposals aimed at improving uplink performance for massive MIMO.

The core problem: under the original 7.2x Cat-A split, all uplink processing — channel estimation, beamforming weight calculation, equalization — happens in the DU. But cloud-based processors handle these matrix-heavy operations far less efficiently than purpose-built silicon at the radio site. Ericsson demonstrated a 60–80% performance gap between its proprietary Cloud-RAN (which puts the full uplink receiver in the RU) and O-RAN 7.2x.

The agreed-upon compromise included two classes. The Ericsson-led approach (Class A) moves the entire uplink receiver — channel estimation, beamforming, equalization — into the RU. The Qualcomm-led approach (Class B) splits channel estimation across both DU and RU. Both classes were approved, with 7.2x remaining mandatory as a baseline.

Why This Matters for Neural Inference in the RU

At neuraRAN, we're building a neural receiver — a single CNN+GNN forward pass (730K params, INT8) that jointly replaces three legacy DSP blocks: channel estimation, MIMO equalization, and soft demapping. It runs inside the radio unit and outputs log-likelihood ratios (LLRs) directly, collapsing ~25–50 Gbps of fronthaul IQ data down to ~2 Gbps.

The fact that O-RAN didn't standardize the RU's internals creates four strategic dynamics for this approach:

What the Spec Landscape Looks Like

The Risk Isn't Permission — It's Adoption Path

Nothing in the O-RAN specifications prevents you from building a smarter RU. The specifications define the interface, not the implementation. As long as you can interoperate — or define a clean interface of your own — you're free to innovate inside the box.

The real strategic question is the adoption path. In Phase 1, this is straightforward: we're building the FPGA prototype and controlling both the RU and a lightweight DU, so there's no interoperability surface to worry about. The neural receiver needs to prove BER parity with traditional DSP at a fraction of the power and fronthaul cost.

For broader deployment, there are a few options. We could propose a profile or extension within the O-RAN Alliance that defines an LLR-based fronthaul for smart RUs. We could partner with DU vendors (open-source projects like srsRAN or commercial DU stacks) to support our output format. Or we could ship the neuraRAN RU as a vertically integrated solution — smart RU plus thin DU — targeted at private 5G networks where single-vendor simplicity is an advantage, not a limitation.

Private 5G is particularly interesting here. These deployments are greenfield, cost-sensitive, and often fiber-constrained. The value proposition of collapsing fronthaul bandwidth by 10–20× and cutting baseband power by 3–5× is immediate and tangible — and the customer doesn't care about multi-vendor RU–DU interoperability the way a Tier 1 MNO does.

Bottom Line

The O-RAN Alliance opened the link between the DU and RU. It did not open the RU itself. That's precisely the gap we're building into. The industry is already moving more intelligence toward the radio unit — we're just replacing the DSP pipeline with something fundamentally better: a single neural forward pass on purpose-built inference silicon.

The black box is our canvas.