What US Carriers Actually Sell an AI Compute Center

Ask three US carriers what they would sell a 100 MW AI campus and you will get three answers that are all true and none of which are comparable. One quotes dark fiber on a twenty-year indefeasible right of use. One quotes managed 400G wavelengths with a four-hour MTTR. One quotes hourly metered bandwidth from a portal, turned up in minutes. They are not competing on the same product, and the price per bit differs by more than an order of magnitude between them. That is not carrier confusion. It is the market sorting itself out around four very different jobs, and the job you actually have determines which quote is even relevant.

This article is the companion to How US AI Compute Centers Should Choose a Telecom Carrier, which covered the selection method: traffic classification, product ladder, break-even economics, diversity geometry, SLA arithmetic and the lead-time stack. This one is the other side of the table. It walks through what the wholesale and hyperscale-facing carriers in the United States are actually shipping in 2026, what each product is good at, what the published specifications quietly leave out, and how to map any of it back to your own traffic.

1. What changed on the demand side

Four things about AI traffic break the assumptions that carrier product catalogs were built on, and each one maps to a specific product response.

Capacity per site went up by roughly an order of magnitude. Conventional data center interconnect was designed for replication, backup and disaster-recovery failover. Distributed training is a different animal: Cisco's modeling of scale-across puts the requirement at roughly fourteen times conventional DCI, with two 100 MW sites needing somewhere between 12,000 and 32,000 coherent 400G or 800G ports where a conventional design needed 1,000 to 2,000. Carriers responded by pushing channel speeds up the coherent ladder. Zayo completed full 400G enablement of its North American core and has publicly targeted commercial 800G wavelength service by the end of 2026 [7][14]. AT&T has extended 400G wavelength connectivity to more than 40 US metros, with 400G handoff enabled across 440,000 properties serving more than 2.3 million business tenants, and is extending capacity to 1.6 Tbps across key metro, regional and long-haul routes [13]. Verizon has validated 1.6 Tbps transmission over a live metro network using Ciena WaveLogic 6 Extreme coherent optics, a single 1.6 Tbps wavelength operating through a dense ROADM environment [13].

Determinism became a saleable attribute. In Ciena's most recent global service provider survey, 35 percent of respondents identified network consistency — guaranteed performance stability and low jitter — as the leading opportunity for monetizing premium connectivity, ahead of raw capacity [13]. That is a change in what carriers think they are selling. Synchronous training converts latency variance directly into idle accelerators, so a route that is fast on average and inconsistent in the tail is worth less than a slower route that is stable. This is why purpose-built low-latency routes still command a premium, and why route transparency (latency figures, route maps, KMZ files) has become a standard ask rather than a favor.

Deployment speed became a product feature. Lumen's Internet On-Demand activates in about five minutes and reaches more than 10 million locations [3]. Zayo's Quick Connect data center program — now more than 100 facilities — turns up IP and Ethernet services within ten days of order [7]. Neither of those replaces a fiber build, and both are worthless if your campus is not already on-net. But they change what you can promise a tenant during the eighteen months while the real build is in permitting.

The physical layer became a supply-chain story again. Lumen agreed to reserve 10 percent of Corning's global fiber capacity for each of two years, its largest cable purchase ever, and the point of the deal was not price — it was queue position [1]. Corning's new generation-AI fiber and cable system lets Lumen fit two to four times the fiber count into existing conduit, which is the only economically sane way to add capacity in a corridor where the trenching was done decades ago [1]. Lumen has stated it will install more than 22,000 route miles of Corning fiber across both upgrades and more than 20 new routes, and will more than double intercity fiber miles over five years [5]. At OFC 2026 Lumen described pulling cables with as many as 1,724 individual fibers into existing conduit, and building in-line amplifier sites in remote areas as precast "micro data centers" to handle the resulting fiber counts [14].

Two of the three hyperscale-side design philosophies aired at OFC 2026 are worth internalizing, because they explain why carriers are re-architecting rather than just adding channels. Google's position is that "fiber is the new wavelength": with single-fiber capacity approaching its limit, deployment granularity moves from a wavelength to a fiber or an entire system, and the company is deliberately reducing reliance on ROADM and FlexGrid in favor of simpler point-to-point designs to speed deployment [14]. Meta's position is that beyond roughly 10 km, intensity-modulated direct detection runs out and high-capacity transport has to move to DWDM, which means the 800ZR generation must land on standardized interfaces and open line systems so that sites can be upgraded one at a time across vendors [14]. If you are writing an RFP, "open line system, standards-based pluggables, upgradable per site" is the ask those two positions justify.

2. The eight-product map

Here is the catalog, ordered by how much of the stack you end up owning. Every US carrier plays in several layers at once, and the same company will happily quote you layer 1 and layer 5 for the same campus.

Layer	What you get	Who sells it in the US	When it is right
1. Dark fiber lease / IRU	Unlit strands, your own optics, your own line system	Zayo, Lumen, Verizon, AT&T, Crown Castle fiber	Long routes, high channel counts, capacity you can forecast. Survives carrier bankruptcy because the asset is yours.
2. Managed wavelength	A lit 10G/100G/400G/800G lambda with route transparency	Zayo, Cogent, Lumen, AT&T, Verizon	Capacity plan uncertain, ops team small, need it in weeks not quarters
3. Carrier Ethernet (EPL/EVPL)	Layer-2 private line with a committed information rate	Lumen, Zayo, AT&T, Verizon, Comcast Business	Building-to-building on a campus, metro point-to-point, MEF-certified handoff
4. IP transit / DIA	Public internet with a commit tier, burstable	Cogent, Lumen (AS3356), Zayo (AS6461), AT&T	North-south traffic only. Never for GPU-to-GPU or synchronous storage replication.
5. NaaS / on-demand	Hourly metered bandwidth, self-served from a portal or API	Lumen Internet and Ethernet On-Demand, Megaport, Console Connect	Bursty edges, temporary sites, capacity you cannot forecast, testing before committing
6. Private AI fabric / dedicated build	New fiber on new routes, engineered for one tenant	Lumen Private Connectivity Fabric, Zayo AI corridors	Multi-campus training at scale, where route geometry and latency are the product
7. Cloud on-ramp / interconnection fabric	Virtual circuits to clouds and to other tenants in the same building	Equinix Fabric, Megaport, Console Connect, Zayo CloudLink	Multi-cloud egress, partner interconnects, anything measured in minutes
8. Managed services	Somebody else monitors, troubleshoots and escalates	Lumen Managed Services, Zayo fiber monitoring across 14,000+ route miles [7]	You took dark fiber and do not want to build a NOC

Two structural warnings about this table. First, layers 1 and 6 are the only ones where you control the physical route, which means they are the only ones where a diversity claim can be independently verified rather than believed. Second, layers 5 and 7 are the only ones where lead time is measured in minutes — and both are footprint-limited. A NaaS port you can activate in five minutes is worth nothing at a greenfield site that no carrier has on-net.

3. Lumen: fabric, fiber supply, and consumption pricing

Lumen is the clearest example of a carrier that rebuilt its product line around AI demand rather than bolting a marketing layer onto wavelength services. Three things it sells are genuinely differentiated.

Private Connectivity Fabric

Private Connectivity Fabric (PCF) is the flagship: a custom network that bundles dedicated access to existing Lumen fiber, installation of new fiber on existing and new routes, and access to Lumen's digital services [1][5]. The company reported reaching $5 billion in AI-related sales, a figure it attributes in part to the Microsoft partnership disclosed weeks earlier [2][20]. Frost & Sullivan's assessment of the wavelength business describes the network at approximately 78,000 route miles with dense fiber routes supporting high-capacity 400G service, and notes Lumen's plan to implement 800G channels in its network by 2025 [5]. The same assessment catalogs the hyperscaler relationships — Microsoft, Meta, Google Cloud and AWS — and notes that Lumen's modernization had outfitted more than 525 data centers for NaaS products at the time of writing [5][20].

The honest caveat sits in the same analysis: Lumen has divested much of its international footprint, including the EMEA business sold to Colt in 2022, and will rely on partners to extend PCF into some international locations [20]. If your campus map is multinational, ask where the partner segments are and who holds the SLA across them.

Fiber supply as a product

The Corning agreement is worth understanding as a procurement signal rather than a press release. Reserving 10 percent of a manufacturer's global capacity for two years is a statement about how tight fiber supply got, and about which customers a carrier expects to serve with it [1]. For a compute operator the practical implication is straightforward: if you are asking a carrier to build to a new site in 2027, ask whether that build depends on fiber the carrier has already secured, because a corridor that is supply-constrained is a schedule risk wearing a connectivity costume.

Consumption pricing — and the 240-hour month

Lumen's on-demand catalog is the most transparent published pricing in the US wholesale market, which makes it useful as a reference even if you buy elsewhere. Internet On-Demand runs 1 Mbps to 100 Gbps on existing on-net ports, activates in about five minutes, carries a 99.99 percent service availability SLA, and is billed hourly: $0.37 per hour at 50 Mbps, $0.47 at 100 Mbps, $1.06 at 1 Gbps and $3.93 at 10 Gbps [3]. Ethernet On-Demand is listed at $1.77 per hour or $424.80 per month, with on-demand bandwidth in the 50–100 Gbps range and custom configurations from 2 to 400 Gbps, MEF CE 3.0 certified [4].

Now read the unit economics, because this is where a self-serve catalog quietly misleads.

Speed	Published hourly	Published "per month"	True 730-hour month	Effective $/Mbps/month
50 Mbps	$0.37	$89	$270	$5.40
100 Mbps	$0.47	$113	$343	$3.43
1 Gbps	$1.06	$254	$774	$0.77
10 Gbps	$3.93	$943	$2,869	$0.29

$424.80 divided by $1.77 is exactly 240.0 hours. The printed monthly figure describes ten days of a port, not thirty, and the arithmetic is confirmed by the Service Guide, which states that hourly charges accrue for each consecutive hour until disconnection, rounded up to the nearest minute, and that a term-commitment model is billed as a monthly recurring charge with any increase above the committed level invoiced at an incremental hourly rate [6]. This is not a Lumen-specific trick — it is what happens whenever a consumption price is restated as a monthly one. Always divide the monthly by the hourly before you put it in a comparison sheet. The second thing the table shows is that ten gigabits costs about 3.7 times one gigabit per hour and delivers ten times the capacity, so the honest comparison is never "cheapest port." It is cost per Gbps of what you will actually turn up by month 36.

4. Zayo: corridors, 400G everywhere, and the latency product

Zayo's AI pitch is physical and specific, which makes it the easiest to audit. Three components matter.

New long-haul corridors built for AI demand

Zayo broke ground on three purpose-built long-haul dark fiber routes: Chicago to Columbus at 385 miles, engineered as a low-latency route with connectivity to all major data centers in both metros; Chicago to Minneapolis at 521 miles with ultra-high fiber counts tying into the Chicago–Columbus corridor and the SPREAD networks; and Phoenix to Tucson at 123 miles, connecting into three planned further builds toward El Paso, Dallas and Mexico [7]. These sit inside a commitment to construct more than 5,000 new route miles of long-haul fiber, on a stated expectation that AI demand grows two to six times by 2030 [7][19].

The reason that list should interest a compute operator is not the mileage. It is that these are corridors chosen by market modeling rather than by where existing conduit happened to be, which means they are the routes where a new campus can plausibly get diverse fiber without a two-year special construction. If your site is inside one of them, say so in your RFP and ask for the route map as a contract exhibit.

Full 400G enablement, and 800G next

Zayo completed the full 400G upgrade of its North American core network, added 393.6 terabits of optical capacity and ten additional 400G-enabled points of presence in a single quarter, expanded fiber monitoring to more than 14,000 route miles, and added six IP PoPs including two 400G-enabled ones in Montreal and Chicago [7]. It describes itself as operating the largest 400G-enabled wavelength network in North America, with long-haul wavelengths from 10G to 400G and a network spanning over 18 million fiber miles, 51,000 buildings and 1,500 data centers [10]. At OFC 2026 Zayo stated a plan for commercial deployment of 800G wavelength service by the end of 2026 [14]. The same panel produced the number that best conveys where this is heading: fill a single route with 11,000 fibers and run C plus L band across it, and total capacity approaches half an exabyte per second [14].

The latency product: SPREAD and the thirteen routes

Zayo acquired Spread Networks in 2017 for $127 million — an 825-mile, high-fiber-count long-haul route between New York and Chicago built explicitly to be the shortest, lowest-latency path between the two, at a construction cost reported at $300 million [9]. The route had 432 fibers with less than 25 percent utilized at acquisition, plus overpull rights and additional conduit [9]. Zayo now markets 13 ultra-low-latency routes including it, guarantees primary, secondary and tertiary diversity on the New York–Chicago path, claims 99.999 percent uptime, states that 99 percent of the route is buried, and offers five tiers of latency all below 17 ms [8].

Here is why that matters more than the marketing suggests. Run the arithmetic independently. 825 miles is 1,328 kilometres. At 4.90 microseconds per kilometre of propagation delay in standard single-mode fiber — the figure derived from a group index of about 1.468 — the theoretical round-trip floor for that route is 13.00 milliseconds. Zayo publishes 13.0 ms. The published number is the physical number, which means the route tax is effectively zero: this path is within measurement error of the shortest possible fiber path, against the 15 to 25 percent route tax typical of carrier networks and the 1.25x straight-line penalty that would put a conventional route at roughly 14.1 ms [8]. That is the entire argument for paying for a purpose-built route, and it is checkable in about thirty seconds with a calculator and a route length.

5. AT&T and Verizon: hyperscale contracts and metro density

The two largest US carriers approach this market from assets they already own, and their AI narratives are built on scale rather than on new product invention.

AT&T positions wavelength services explicitly at hyperscalers, advertising 99.999 percent availability with SLAs, secure direct point-to-point connectivity nationwide without transiting the public internet, and capacity scaling to 1.6 Tbps across key metro, regional and long-haul routes [12][13]. Its 400G wavelength footprint reached more than 40 US metros with 400G handoff enabled across 440,000 properties serving 2.3 million business tenants — the expansion framed explicitly around distributed AI, cloud and analytics workloads [13]. The AWS Interconnect last-mile offering extends AT&T fiber and 5G from enterprise locations directly into AWS environments, which is the clearest example yet of a carrier product designed around inference traffic rather than training traffic [13]. AT&T completed its $5.75 billion acquisition of Lumen's mass-market fiber business in February, adding more than four million fiber locations and construction capability across 11 states, and expects its footprint to exceed 60 million locations by 2030 [13].

Verizon signed an agreement valued at more than $1 billion to provide dark fiber connectivity for Google data centers, and management said at the time that it expected to sign more agreements of that type before year end, with several billion dollars of new revenue possible over the following years [17]. Verizon has also validated 1.6 Tbps transmission over a live metro network using Ciena's WaveLogic 6 Extreme, demonstrating a single 1.6 Tbps wavelength through a dense ROADM environment, with the technology also designed to extend 800G across substantially longer distances [13]. Ciena's claim for that generation is that it roughly doubles capacity within the same space and power envelope, cutting space and power per bit by about half [13]. Verizon's Frontier acquisition expanded its reach to roughly 30 million fiber passings against a 40 to 50 million target [13].

The strategic read for a compute operator: neither transaction was an AI acquisition. Both were broadband and convergence plays [13]. But distributed inference changes what those access assets are worth. Training concentrated network value inside a handful of massive facilities and the optical links between them; inference distributes it outward toward metro networks, enterprise premises, wireless aggregation and the edge. Ciena's survey found 88 percent of respondents expect edge-hosted services — localized inference, edge compute, GPU-as-a-service — to account for at least 10 percent of enterprise revenue within three years [13]. If your business includes inference delivered to enterprises, the metro density question matters more to you than the long-haul corridor question, and the two carriers with the deepest metro footprints are the ones to anchor the conversation with.

6. The wholesale specialists: what Cogent tells you about the floor

Cogent is useful as a benchmark because its published specification is unusually blunt. Optical Wavelengths are offered at 10G, 100G and 400G (10GE LAN-PHY, 10G OTU2e, 10G WAN PHY, 100GE LAN-PHY, 100G OTU4, 400GE LAN-PHY), available in more than 1,000 carrier-neutral data centers across 150 North American markets, on a wholly owned fiber network with diverse coast-to-coast routing options, multiple unique east-west and north-south routes, full signal transparency, an installation guarantee within 30 business days, a 100 percent network availability SLA, and a stated willingness to share route metrics including latency figures, route maps and KMZ files [11]. Its brochure makes the positioning explicit: for hyperscale or high-volume transport, wavelengths are "more cost-effective and flexible than dark fiber" [11].

Take that claim seriously and then test it rather than accepting it. It is true at short distances and low channel counts, and false at long distances and high channel counts — which is exactly what the break-even model in the companion article quantifies. A carrier telling you wavelengths beat dark fiber is usually right when you have four channels and forty route miles, and usually wrong when you have twenty channels and four hundred. The value of a published specification like Cogent's is that it gives you the inputs — market count, speed tiers, installation interval, SLA structure — to run the comparison yourself instead of arguing from brochures.

7. NaaS and neutral fabrics: how they differ from carrier NaaS

There are two different things called network-as-a-service, and mixing them up causes bad decisions.

Carrier NaaS (Lumen's Internet and Ethernet On-Demand) is metered capacity over a port on the carrier's own network. Its strengths are reach and speed: Lumen's on-demand internet is available at more than 10 million locations and activates in about five minutes [3]. Its limitation is the same as its strength — it is only as good as the carrier's footprint, and off-net segments of an on-demand connection cannot carry multiple logical connections over the same physical port [6].

Neutral fabric NaaS (Megaport, Console Connect, Equinix Fabric, and carrier equivalents such as Zayo CloudLink) sells virtual circuits between ports that are already inside interconnected data centers, typically with cloud on-ramps attached [18]. Its strength is neutrality and granularity: you can stand up a private connection to a cloud or to another tenant without ordering anything physical. Its limitation is harder-edged than it looks — a fabric circuit is only useful if both endpoints are already in the fabric's footprint, and the fabric does not own the glass between buildings. For a greenfield campus that nobody has on-net, a fabric is a month-24 product, not a day-one product.

The rule of thumb: use fabric and carrier on-demand for the edges of your network — tenant onboarding, cloud egress, burst capacity, partner interconnects, and pre-commitment testing — and use layers 1, 2 and 6 for anything carrying synchronous training traffic between your own halls.

8. Hollow-core fiber: the next differentiator, priced accordingly

Hollow-core fiber is the one technology in this market that changes a number physics had previously fixed. Light travels about 47 percent faster in an air core than in glass, cutting propagation latency by more than 30 percent, with reported signal loss below 0.09 dB/km, nonlinear effects reduced by three to four orders of magnitude, and a claimed 50 percent extension of unrepeatered reach [15][16]. The practical translation: the 4.90 microseconds per kilometre you have been budgeting becomes roughly 3.33, so a 100 km round trip drops from 0.979 ms to 0.666 ms — 0.313 ms saved, and at 800 Gbps the bytes in flight on that round trip fall from 97.9 MB to 66.6 MB. That second number is the one that matters for distributed training: less data committed to the wire is less buffering you have to provision, which is a capacity and cost argument rather than a latency argument.

On the New York–Chicago corridor, hollow core would take the same 825-mile route from 13.00 ms to 8.85 ms — within half a millisecond of the 8.2 ms that McKay Brothers achieves by microwave over the same corridor, and without microwave's weather exposure [8][16]. That is a genuinely striking result: fiber catching air.

Microsoft, through its Lumenisity acquisition, is the demand-side anchor here, with plans reported to deploy on the order of 15,000 kilometres across the Azure global network by the end of 2026 [15]. Lumen is reported to have moved hollow core from trial into small-volume production inside 2026 AI training cluster contracts, with early customers in high-frequency trading and large model development [16]. LightCounting's 2026 projection has hollow core taking 15 percent of the high-speed optical transceiver market by 2027 and passing 35 percent by 2030 [16]. Read those as directional; the deployment constraints above are real and the supply chain is the gating factor.

9. Mapping the catalog back to your requirement

Every product above collapses into a decision on four variables: how much capacity, over what distance, by when, and with what consequence if it fails. This matrix is the short form.

Situation	Distance	Capacity	Lead time available	Start here
One hall, no inter-building training	Intra-building	Any	Any	No carrier product. This is a structured cabling and optics problem.
Two halls on one campus	0.5–5 km	400G–800G, growing	6–12 months	Dark fiber on a short metro lease, or your own conduit if you own the land
Two campuses in one metro	20–80 km	Multiple 400G	3–6 months	Run the break-even. Below about 40 route-miles and four channels, managed wavelengths usually win.
Regional training cluster	200–800 km	1–2 Tbps	12–24 months	Dark fiber or IRU on new-build corridors; expect 10–22 months and start before you buy land
Latency-sensitive coordination	Any	Moderate	Any	Purpose-built low-latency route, with published RTT checked against physics
Inference delivery to enterprises	Metro and last mile	10G–400G	Days	Metro Ethernet, cloud interconnect, carrier on-demand at the edges
Unknown capacity shape	Any	Unforecastable	None	NaaS on-demand while you learn the curve, then commit on layer 2 or 3

Notice what the matrix does not contain: a single row where transit is the answer for training traffic, and a single row where dark fiber is the answer when you have three months. Products are constrained by physics and by construction schedules long before they are constrained by price.

10. Ten ways the specification misleads

1. The printed monthly figure is not a month. Divide every "per month" number by the hourly rate. Lumen's $424.80 is 240 hours [4].
2. "400G-enabled network" is not "400G to your building." Core enablement and handoff availability are different facts. Ask for 400G handoff at your specific address, in writing.
3. On-net claims are per-port, not per-campus. A carrier can be on-net at the meet-me-room and off-net at your hall, which changes both price and lead time by an order of magnitude.
4. Off-net segments limit what you can do with a port. Lumen's Service Guide states that off-net Internet On-Demand connections cannot support multiple logical connections over the same physical port [6].
5. "100 percent network availability" SLA describes credits, not uptime. Read the credit cap and the exclusions before treating it as a reliability claim [11].
6. Latency tiers are sold as tiers. Five tiers all below 17 ms is a meaningful offer on a 1,328 km route, and meaningless on a 12 km one. Always convert to microseconds per kilometre and compare against the 4.90 physical floor.
7. Quick Connect and on-net programs are footprint-limited. Ten-day turn-up applies where the capacity already exists [7]. Verify at your address, not in the brochure.
8. Capacity announcements are network-wide, not route-specific. 393.6 terabits added is a network fact; which routes got it is a contract question.
9. Reference customers are not a route map. A carrier serving ten of the top digital companies may be serving them in Ashburn and nowhere else.
10. Hollow-core latency claims are real and the cost is too. Minus 30 percent latency at roughly 1,800 times fiber cost per kilometre [15][16]. Both numbers go in the model.

11. Twelve questions for the RFP

1. Is the service on-net or off-net at each of my addresses, and what is the quoted lead time for each?
2. Give me the route length in route-miles and in kilometres, and the published one-way latency. I will compare it to 4.90 µs/km myself.
3. Provide route maps as KMZ or GIS files as a contract exhibit, including the location of every shared conduit, bridge and tunnel crossing [11].
4. Do my primary and diverse paths share any conduit, bridge, tunnel, pole line or ILA site? Identify each one.
5. What are the four SLA numbers: availability percentage, MTTR, credit cap as a percentage of monthly recurring charges, and the measurement window?
6. At 400G and at 800G, what is the handoff interface, and is it standards-based pluggable on an open line system so I can upgrade site by site [14]?
7. If I take dark fiber, what is the maintenance obligation per route-mile per year, and who owns splicing, permits and locates?
8. If I take an IRU, what is the tenor, is it recorded, and what happens in an insolvency?
9. What is the incremental cost of the next eight channels, quoted as a per-channel price rather than a discount schedule?
10. What is your fiber supply position for a build to this address in the next 24 months, and does it depend on fiber you have already secured [1]?
11. Where are you trialing hollow-core fiber, and at what price per fibre-kilometre would it enter a commercial quote?
12. Provide three reference customers with a route of similar length and channel count, and the date each was turned up.

Where this leaves you

The US carrier market for AI compute is not short of products. It is short of operators who know which product they are buying. The catalog spans eight layers, the price per bit spans more than an order of magnitude between the bottom and the top of it, and the honest answer for a serious campus is almost always three layers held at once: glass you control on the routes that carry training traffic, managed wavelengths on the routes you cannot forecast, and on-demand capacity at the edges where the shape of demand is genuinely unknown.

Two habits will carry most of the value. First, convert every claim into a physical number before you believe it — microseconds per kilometre, dollars per strand-mile per month, hours per month, megabytes in flight. Carriers publish the inputs more often than they publish the conclusions, and the arithmetic takes minutes. Second, treat route geometry as the deliverable rather than the service. A diversity exhibit with KMZ attachments is worth more than any SLA, because the SLA pays you after the outage and the geometry is what prevents it.

If you are scoping connectivity for a new AI hall or campus, bring the route lengths, channel counts and go-live dates and we will run the break-even and diversity analysis against real quotes.

Request a connectivity scope review

References

1. Lumen Technologies, "Corning and Lumen Reach Supply Agreement on Next-Generation Fiber-Optic Cable to Support Data Center AI Demands," news release, 1 August 2024. news.lumen.com
2. Lumen Technologies, "Microsoft and Lumen Technologies Partner to Power the Future of AI," news release, 24 July 2024. news.lumen.com
3. Lumen Technologies, "Internet On-Demand" service page (hourly rates, activation time, 99.99% availability SLA, 1 Mbps–100 Gbps). lumen.com
4. Lumen Technologies, "Ethernet Services" page (Ethernet On-Demand at $1.77/hour or $424.80/month; 50–100 Gbps on-demand; 2–400 Gbps custom; MEF CE 3.0). lumen.com
5. Frost & Sullivan, "2024 Lumen Technologies: Best Practices Leadership in Wavelength Services, North America" (approximately 78,000 route miles; 800G channel plans; Corning partnership; hyperscaler relationships). assets.lumen.com
6. Lumen Technologies, "Internet On-Demand Service Guide," version dated 22 July 2026 (hourly metered billing, term-commitment model, incremental hourly charges, off-net port limitations). assets.lumen.com
7. Zayo, "Zayo Levels Up Its AI-Ready Infrastructure with New Routes and 100% 400G Enablement" (Chicago–Columbus 385 mi, Chicago–Minneapolis 521 mi, Phoenix–Tucson 123 mi; 393.6 Tb added; 14,000+ route miles monitored; 100+ Quick Connect data centers at 10-day turn-up). zayo.com
8. Zayo, "New York to Chicago Route Overview" (SPREAD Networks: 825 route miles, 99% buried, 99.999% uptime, five latency tiers all below 17 ms, primary/secondary/tertiary diversity, dedicated NOC). zayo.com
9. Zayo Group Holdings, "Zayo to Acquire Spread Networks," 27 November 2017 ($127 million; 825-mile route; 432 fibers with less than 25% utilized; overpull rights). zayo.com
10. Zayo, "Data Centers" solutions page (18 million fiber miles, 51,000 buildings, 1,500 data centers; largest 400G-enabled wavelength network in North America). zayo.com
11. Cogent Communications, "Optical Wavelengths" brochure (10G/100G/400G interfaces; 1,000+ carrier-neutral data centers across 150 markets; 30-business-day installation guarantee; 100% network availability SLA; route metrics including KMZ shared). cogentco.com
12. AT&T, "Business networking services" (wavelength services positioned at hyperscalers; 99.999% availability with SLAs; secure direct point-to-point nationwide). att.com
13. RCRTech, "The AI traffic effect: A new investment cycle for fiber" (AT&T 400G in 40+ metros across 440,000 properties and 2.3 million business tenants; 1.6 Tbps capacity; Verizon 1.6 Tbps WaveLogic 6 Extreme metro trial; AT&T–Lumen mass-market transaction; Verizon–Frontier; Ciena survey findings of 88%, 35% and 96%). rcrtech.com
14. Simple Tech Trend, "OFC 2026 — DCI and Scale Across in the AI Era" panel summary (Meta on the 10 km DWDM boundary and open line systems; Google on fiber as the new wavelength; Zayo on 800G commercial deployment by end of 2026 and 11,000-fiber C+L capacity; Lumen on 17 million fiber miles, 1,724-fiber cables and ILA micro-sites; Nokia on 28 W 800G modules). simpletechtrend.com
15. Communications World Weekly (CWW), "空芯光纤规模化商用须迈过五道坎" (propagation speed +47%; latency reduction above 30%; loss below 0.09 dB/km; nonlinear effects reduced by 3–4 orders of magnitude; CNY 36,000 per fibre-km against CNY 20 for G.652D; deployment, standards and ecosystem gaps; Microsoft's planned Azure deployment). cww.net.cn
16. Internet Pros, "Hollow-Core Fiber in 2026" (conventional versus hollow-core comparison table; Microsoft and Lumenisity; Corning, OFS, Sumitomo, Nokia and NKT Photonics; Lumen small-volume production in 2026 AI cluster contracts; LightCounting market projections). internet-pros.com
17. Cailianshe (财联社), "威瑞森上调全年业绩指引 与谷歌签署逾10亿美元AI光纤协议," 24 July 2026 (Verizon–Google agreement valued above $1 billion for dark fiber connectivity to Google data centers; management expectation of further agreements). cls.cn
18. Megaport, services and platform overview (software-defined NaaS; virtual circuits to cloud providers and data centers). megaport.com
19. Zayo, "Global Network Capabilities." zayo.com
20. 451 Research / S&P Global Market Intelligence, "Lumen launches Private Connectivity Fabric, reaches $5B in AI-related sales" (PCF scope; $5 billion AI-related sales; Microsoft partnership; 22,000 route miles of Corning fiber and 20+ new routes; 525+ data centers outfitted; international divestitures). assets.lumen.com