Carrier neutral connections and cross connects for AI data centers

In short

A carrier neutral AI data center lets every tenant choose from many different telecom carriers, cloud providers, and network operators on equal terms. The physical link that makes that possible is a cross connect. A cross connect is a dedicated point to point cable, usually fiber optic, that runs from a tenant’s equipment to a carrier’s gear inside a secure meet me room. The traffic never touches the public internet, so the connection is private and fast. Federal rules ban exclusive contracts and revenue sharing deals that would lock a building to one carrier. 47 C.F.R. § 64.2500. AI training clusters push these interconnects to the extreme. They need 800 Gbps to 1.6 Tbps of bandwidth with under a microsecond of latency between GPUs. A single dropped packet can ruin hours of expensive compute time. APNIC blog on the catastrophic cost of a single dropped packet. The legal and business terms around cross connects, especially anti competitive lock in and weak tenant protections, are now a major deal point for any company building AI infrastructure.

What is a cross connect and why does it matter for AI workloads?

A cross connect is a physical cable that joins two termination points inside a single AI data center. The most common type is a fiber optic or copper cable running from a tenant’s server cabinet to a telecommunication carrier’s equipment in a central meet me room. Once installed, the cable is a simple cable run, and these monthly charges are almost pure profit for the AI data center. Metanet Hosting cross connect blog. The data stays on that dedicated line and never enters the public internet. That gives the tenant a private link with consistent low latency and no shared bandwidth. Volico cross connect definition.

For an AI training run, a cross connect does far more than connect a server to the cloud. In a modern AI cluster, most of the data movement is east west traffic, flowing between GPUs inside the same building rather than north south traffic to the outside world. East west traffic can make up 70 to 90 percent of all data center traffic. Mulcas east west traffic analysis. The network links that carry that traffic must sustain extremely high bandwidth, often 800 Gbps to 1.6 Tbps, with latency under a microsecond. Volico AI networking blog.

AI workloads create what engineers call elephant flows. These are a few very large, long lasting data streams that move in synchronized bursts during operations like All Reduce. If a single packet drops, the entire group may need to resend massive blocks of data or restart the training job from its last checkpoint. Intelligent Visibility Ethernet vs. InfiniBand guide. Because of that, the quality and speed of every link, including the cross connect to a carrier or cloud on ramp, directly controls how many useful GPU hours a cluster delivers. One industry analysis reports that 20 to 50 percent of total AI training time is spent simply moving data between GPUs across the network fabric. A throughput gain of just 8 to 10 percent can save millions of dollars. Intelligent Visibility Ethernet vs. InfiniBand guide.

What makes a data center carrier neutral?

An AI data center is carrier neutral when it operates independently of any single telecom provider. The facility does not own or resell a carrier’s network service. It does not force a tenant to use a particular carrier. It gives every carrier the same physical and operational access on equal terms, and it publishes transparent cross connect pricing. Flexential carrier neutral blog. The heart of a carrier neutral facility is the meet me room. That is a secure, heavily monitored space where multiple carriers, internet service providers, and cloud on ramps place their equipment and terminate their networks. A tenant uses a cross connect to link its own cabinet to any carrier it selects inside that room. Digital Realty Bersama MMR explanation.

Carrier neutrality is not just a marketing claim. Federal policy strongly disfavors exclusive arrangements in multi tenant buildings. A common carrier cannot sign a contract that stops a commercial multiunit premises owner from letting any other common carrier access and serve the tenants. 47 C.F.R. § 64.2500(a), (b). The FCC’s 2022 Multiple Tenant Environments Order went further by banning graduated revenue sharing agreements and requiring clear disclosures when a carrier holds an exclusive marketing right. 87 Fed. Reg. 17181.

The largest carrier neutral AI data center hubs give tenants access to dozens or hundreds of networks. The Equinix Ashburn campus in Virginia, the densest interconnection hub in the United States, contains over 38,000 interconnections across 15 buildings with more than 32,000 cabinets and 111 MW of power. Dgtl Infra carrier neutral overview, Equinix blog. Flexential, a mid sized provider, lists more than 300 on net carriers across its U.S. facilities. Flexential carrier neutral blog. That choice lets a tenant switch carriers or add new ones without relocating.

What federal rules protect carrier access and cross connect openness?

Several layers of federal law work together to discourage an AI data center from becoming a captive facility for one carrier.

The rule that applies most directly to commercial multi tenant properties is 47 C.F.R. § 64.2500. It does three things.

First, it forbids any common carrier from entering a contract that stops a building owner from allowing other carriers to serve tenants. Second, it bans graduated revenue sharing and exclusive compensation for access deals, where a carrier pays the building owner for the right to be the only provider inside. Any new deal of that kind has been prohibited since April 27, 2022, and any existing one became unenforceable after September 26, 2022. 47 C.F.R. § 64.2500(c)-(d). Third, a common carrier that has an exclusive marketing arrangement must state on all written marketing materials directed at tenants or prospective tenants of the affected multiunit premises that the arrangement does not mean it is the only provider available. 47 C.F.R. § 64.2500(e).

The rule defines a commercial multiunit premises as any contiguous area under common ownership or control with two or more distinct units used primarily for non residential purposes. A commercial multiunit premises therefore means any multiunit property where non-residential use predominates, the area is contiguous, and there are at least two units under common ownership or control. 47 C.F.R. § 64.2501. That definition seems to cover a multi tenant colocation AI data center. No published FCC enforcement action has applied it specifically to an AI data center cross connect dispute, however.

Other federal rules also push toward open interconnection, but with limits. The Telecommunications Act requires incumbent local exchange carriers (ILECs) to interconnect with other carriers and to offer physical collocation at their premises on just, reasonable, and nondiscriminatory terms, and virtual collocation where physical collocation is impractical. 47 U.S.C. § 251(c)(2), 47 U.S.C. § 251(c)(6). Expanded interconnection rules at 47 C.F.R. § 64.1401 require certain local exchange carriers to provide physical or virtual collocation at their central offices. 47 C.F.R. § 64.1401. These duties, however, apply only to the listed carriers. They do not reach a private, non carrier AI data center operator. They are a background principle, not a direct regulation of every colocation building.

Beyond the FCC rules, the Sherman Act outlaws contracts in restraint of trade and monopolization. 15 U.S.C. §§ 1-2. If a dominant AI data center operator used its control to force carriers not to serve rival buildings, that conduct could be challenged as an unreasonable restraint of trade. Industry practitioners call that practice anti tethering, and it is discussed below.

How are AI workloads reshaping cross connect demands?

AI training clusters do not look like ordinary server rooms. A large language model training run uses thousands of GPUs exchanging data in tight, coordinated bursts. The network must be lossless and fast enough to keep every GPU fed, or the expensive compute cycles sit idle.

The data is overwhelmingly east west. In a typical enterprise data center, most traffic flows north south, out to the internet and back. In an AI cluster, that flips. Over 70 to 90 percent of the traffic stays inside the data center, moving between GPUs and between servers. Mulcas east west traffic analysis. The network fabric that carries that traffic includes the internal spine leaf switches and the cross connect links that attach the tenant’s AI fabric to outside carriers and cloud on ramps. Even a small bottleneck in those links can cascade.

To handle these workloads, operators pick between two main networking technologies.

InfiniBand uses credit based flow control and a centralized subnet manager to guarantee lossless delivery. It supports native remote direct memory access (RDMA), which lets one GPU write directly into another’s memory without involving the CPU. RoCEv2 (RDMA over Converged Ethernet version 2) runs RDMA over standard Ethernet switches, using priority flow control and explicit congestion notification to approximate the same lossless behavior.

InfiniBand holds a slight raw latency edge, but a well tuned Ethernet fabric now comes very close. Meta’s Llama 3 training ran on twin clusters, one on InfiniBand and one on Arista Ethernet, and Meta reported no network bottlenecks on either fabric. Intelligent Visibility Ethernet vs. InfiniBand guide. NVIDIA’s reference architecture recommends a dual fabric design. A lossless east west compute fabric (InfiniBand or Spectrum-X Ethernet) handles GPU collectives. A separate north south fabric with BlueField DPUs handles data ingestion, storage, and external connections. NVIDIA AI Enterprise Reference Architecture. In a colocation cage, that second fabric often connects through cross connects to cloud on ramps or storage arrays.

The financial stakes are huge. If the network fabric loses just a few percent of throughput, the extra training time can cost millions. A growing number of AI tenants now demand virtual cross connects as a complement to physical ones. A virtual cross connect is a software defined link provisioned through a portal in minutes, adjustable on demand, without waiting days or weeks for a physical cable. Flexential cross connect blog. That flexibility matters for dynamic multi cloud access during training.

What do cross connects cost and how are they priced?

Cross connect pricing is not regulated, and it varies sharply from one AI data center to the next.

A typical monthly recurring charge for a single cross connect at a major U.S. colocation provider runs between $100 and $300. Datacenters.com cross connect guide. CoreSite charges $250 per month per fiber cross connect. Spiceworks community report. At those rates, 30 cross connects cost a tenant $90,000 per year, a sum that often exceeds what the same tenant pays for power. LinkedIn industry commentary.

Some facilities have taken the opposite approach.

165 Halsey Street, a carrier hotel in Newark, charges only a one time setup fee of $500 for single mode fiber or $750 for copper. There is no monthly recurring charge. 165 Halsey Street. Stream Data Centers provides complimentary cross connects as part of its wholesale connectivity package. Stream Data Centers glossary. TRG Datacenters in Houston offers free cross connects as a baseline service. LinkedIn industry commentary. The stark range has prompted calls from industry voices for a Free The Cross Connects movement, arguing that interconnection should be treated as essential core infrastructure, the same as cooling or backup power.

For the AI data center operator, cross connect fees are a high margin revenue stream. Once a cable is installed, there is effectively no ongoing operating cost. Yet the monthly billing continues for the full term of the customer’s license. That economic structure is a big reason operators work to keep cross connect pricing opaque and sticky.

Table, Cross connect pricing comparison

What legal risks exist around cross connect pricing and carrier access?

The biggest risk for an AI tenant is anti competitive lock in. In a carrier neutral facility, the tenant should be free to choose any carrier. But some operators, according to industry reports, have pressured carriers to refuse service at rival buildings. That practice, called anti tethering, can show up as threats of eviction, refusal to renew a license, placement on a blacklist, or sudden sharp increases in cross connect fees. LinkedIn industry analysis.

Such conduct may run afoul of the Sherman Act’s ban on contracts in restraint of trade and monopolization. 15 U.S.C. §§ 1-2. A court could treat a dominant AI data center as controlling an essential facility and find that denying access or charging discriminatory fees is an antitrust violation. No definitive U.S. court decision has applied the essential facility doctrine to colocation cross connect practices, however. The path to an antitrust win is therefore uncertain and expensive.

A second, quieter risk is the weak legal position of a colocation customer. In most U.S. jurisdictions, a colocation customer is not a leasehold tenant in the traditional real property sense. The relationship rests on a license agreement or master service agreement, not a lease. That means the customer does not have the same statutory protections against arbitrary fee increases or non renewal that a commercial office tenant would enjoy. LinkedIn industry analysis. An AI data center operator could theoretically triple the cross connect fee at renewal, and the tenant’s only practical fix would be to move out, an expensive and disruptive task for a built out AI cluster.

The FCC’s multiple tenant environments rule gives a tenant some leverage, but enforcement against AI data centers remains untested. The 2022 order was aimed mainly at residential apartment buildings and office towers. The rule’s definition of multiunit premises appears to cover a colocation facility, yet no reported FCC action has applied § 64.2500 to an AI data center carrier access dispute. Tenants and their counsel should treat the rule as a strong argument, not a guaranteed shield.

Practical steps for negotiating cross connect terms in an AI colocation deal

Investigate the carrier mix before signing. Ask the AI data center operator for a written list of on net carriers, not just a count. Confirm that the cloud on ramps you need, such as AWS Direct Connect, Azure ExpressRoute, or Google Cloud Interconnect, are already present and cross connectable at standard rates. If you can, visit the meet me room.

Get cross connect pricing in writing with a clear cap. The base monthly recurring charge should be fixed or capped for the full term. Ask for a schedule of one time installation fees. If you expect to add links as the cluster scales, negotiate a volume discount or a flat monthly fee for a block of cross connects.

Look for facilities that charge no monthly recurring fee. As the examples of 165 Halsey Street, Stream Data Centers, and TRG Datacenters show, zero monthly cross connect facilities exist in major markets. For an AI tenant deploying dozens of links, eliminating $250 per month per cross connect can save a six figure sum each year.

Demand low latency, lossless fabric options. Confirm the facility can support both InfiniBand and high speed Ethernet in your cage. Ask whether the cross connect pathways can handle the fiber types you need and whether the operator plans to support 800 Gbps or 1.6 Tbps links in the near term.

Check for exclusive marketing disclosures. A common carrier that has the exclusive right to market its service to tenants of a multiunit premise must disclose the arrangement on all written marketing materials, whether electronic or in print, directed at tenants or prospective tenants in a manner that is clear, conspicuous, and legible. 47 C.F.R. § 64.2500(e). If you see a carrier that fails to make that disclosure, it may signal an unlawful exclusive deal.

Review the contract for anti tethering red flags. A clause that limits your ability to connect to carriers at other AI data centers, or that gives the operator broad discretion to raise fees or block a carrier, should be escalated and pushed back. While federal antitrust law provides a backstop, it is far cheaper to negotiate these points up front.

Key takeaways

• A cross connect is a dedicated point to point cable inside an AI data center that links a tenant’s equipment to a carrier’s gear, keeping traffic private and off the public internet.
• Carrier neutral AI data centers give tenants equal access to many competing networks, a setup that federal rules and FCC orders protect by banning exclusive carrier contracts and undisclosed exclusive marketing.
• AI training workloads flip the traditional traffic pattern. East west GPU to GPU data now makes up 70 to 90 percent of traffic and demands lossless, 800 Gbps plus links. A single dropped packet can cost hours of training time.
• Cross connect pricing is unregulated and ranges from $100 to $300 per month at many facilities, supporting high margin revenue for operators. A growing number of AI data centers are offering zero monthly fee cross connects as a competitive differentiator.
• The legal position of a colocation tenant is weaker than a traditional office tenant. Tenants lack statutory protections against arbitrary fee increases and non renewal, making aggressive contract negotiation critical.
• Anti tethering, where an operator pressures carriers not to serve rival buildings, raises serious Sherman Act concerns but has not yet produced a definitive U.S. court ruling applying the essential facility doctrine to AI data center cross connects.
• Practical steps include demanding written cross connect fee caps, verifying on net carrier counts, checking for exclusive marketing disclosures, and confirming the facility can support the low latency fabric an AI cluster needs.

Frequently asked questions

Q:What is a cross connect in an AI data center?

A:A cross connect is a physical cable (usually fiber optic) that directly links two termination points inside a single AI data center, such as a tenant’s server cabinet and a carrier’s equipment in a meet me room, without routing through the public internet.

Q:What does carrier neutral mean?

A:A carrier neutral AI data center does not favor one telecom provider. It gives every carrier the same physical access and pricing, and it does not require tenants to use a specific network. The facility earns its money from space, power, and cooling, not from selling network services.

Q:Which federal rules stop an AI data center from locking in one carrier?

A:The key rule is 47 C.F.R. § 64.2500. It bans common carriers from signing contracts that restrict a building owner’s right to let other carriers serve tenants. It also bans exclusive revenue sharing deals and requires disclosure of exclusive marketing arrangements. The FCC reinforced these protections in its 2022 Multiple Tenant Environments Order.

Q:Do those FCC rules apply to a private multi tenant AI data center?

A:The definition of multiunit premises in the rule covers any contiguous area under common control with two or more distinct units, which logically includes a colocation facility. However, no published FCC enforcement action has applied § 64.2500 specifically to a data center cross connect dispute, so the rule remains an untested shield.

Q:How much does a cross connect cost?

A:Monthly recurring charges for a single cross connect at a large U.S. colocation provider typically run between $100 and $300. Some facilities charge nothing monthly, only a one time setup fee, while others charge $250 per month or more. At scale, 30 cross connects can cost $90,000 per year.

Q:Why do AI training clusters need such fast cross connects?

A:AI training uses thousands of GPUs that exchange massive data blocks in synchronized bursts. A single lost packet can force the job to restart from the last checkpoint. To keep GPUs fully utilized, the network fabric, including the cross connects, must sustain 800 Gbps to 1.6 Tbps with under a microsecond of latency and no dropped packets.

Q:What is anti tethering?

A:Anti tethering describes conduct where a dominant AI data center operator pressures carriers not to serve competing facilities, often through threats of eviction, non renewal, or sudden fee increases. It is a competition law concern and could violate the Sherman Act, though litigating such a claim is costly.

Q:What are the main networking fabrics used in AI clusters?

A:The two leading technologies are InfiniBand (credit based flow control, native RDMA, slight latency edge) and RoCEv2 Ethernet (RDMA over Ethernet with congestion control). Meta demonstrated that a well tuned Ethernet fabric can match InfiniBand performance for large model training. NVIDIA recommends a dual fabric design with a lossless east west compute fabric and a separate north south fabric for data ingestion and storage.

Q:Can a colocation tenant challenge a sudden fee increase?

A:A colocation customer is usually not a leasehold tenant and lacks the statutory protections that a commercial lessee would have. The best defense is a contract that caps cross connect fees for the license term and limits the operator’s discretion to impose new charges.

Q:Are there AI data centers that offer free cross connects?

A:Yes. Several providers, including Stream Data Centers, TRG Datacenters, and 165 Halsey Street, offer cross connects with no monthly recurring charge. Others are beginning to include cross connects as a standard feature of colocation packages.

Subscribe to The Compute Law Brief

The Compute Law Brief is a free weekday newsletter on the law of AI infrastructure across tax, real estate, construction, power, and deals. The big US build markets and federal law. Three minutes a morning. No paywall, and no email gate to read the blog. Subscribe if you want it in your inbox.

Junde Liu, JD, LL.M. (Taxation) candidate at UF Law. Originally published on Compute Law Blog. This article is general information and does not constitute legal advice. Reading it does not create an attorney client relationship. The reader should not act on the basis of any content here without first consulting a licensed attorney in the relevant state. Last reviewed for accuracy May 23, 2026.