How US AI Compute Centers Should Choose a Telecom Carrier

Picking a telecom carrier for an AI compute center looks like a procurement exercise and is really a site-selection and schedule exercise wearing a procurement costume. Power still decides where a campus can go, but connectivity increasingly decides when it can earn. The failure mode is now familiar across the US market: a hall is built, energized, racked and staffed, and then sits at a fraction of intended capacity because the diverse fiber it needed was scoped after the land was bought. Operators have started calling the result stranded compute, and the arithmetic behind it is ugly. A 10 MW hall of GB200-class racks holds roughly 6,000 GPUs; at a modest $2.00 per GPU-hour — the mid-market rate examined in our GPU rental ROI analysis — every month that fiber is late costs about $8.6 million in deferred revenue against capex that is already depreciating [see the worked outage math below].

This article is a selection method rather than a vendor list. It covers seven things, in the order they actually matter: what traffic you are really carrying, which product on the carrier ladder matches that traffic, when dark fiber mathematically beats a managed wave, what physics no contract can change, why two carriers is not two paths, which four SLA numbers carry all the weight, and how to keep fiber off your critical path. Every number here is either sourced or calculated from sourced inputs, and the two models are simple enough to re-run with your own quotes.

1. Start with the traffic, not the tariff

Carriers price a circuit. You are buying a traffic pattern. Those are different objects, and the gap between them is where most bad contracts get signed. Sort your traffic into four buckets before you issue an RFP, because each one has a different cost driver, a different latency sensitivity and a different answer.

Traffic class	Direction and shape	What actually drives cost	Product that usually fits
Scale-up (inside the rack)	NVLink domain, tens of TB/s, sub-metre	Copper vs on-board optics; not a carrier product at all	None — this is a server bill of materials
Scale-out (inside the hall)	East-west, all-reduce bursts, 400G–800G per port, microseconds	Switch radix, transceiver count, structured cabling	None — owned fabric, possibly managed
Scale-across (campus to campus)	East-west, synchronous and coordinated; incast when thousands of GPUs finish a step together	Fiber pairs, coherent optics, route miles, route diversity	Dark fiber or managed wavelength
North-south (inference, storage, DR, cloud on-ramps)	Bursty, transactional, latency-tolerant by comparison	Committed rate, 95th percentile, peering vs transit	IP transit, DIA, EPL, cloud direct connect

The distinction that catches people out is scale-across. Conventional DCI was designed for asynchronous replication and disaster-recovery failover, traffic that is patient and can be rescheduled. Distributed training is neither: thousands of accelerators finish a compute phase at the same moment and flood the inter-site pipe, and in synchronous training the slowest group sets the pace, so network delay converts directly into idle GPUs [17]. That is why the bandwidth step is so large, and why the interconnect becomes part of the computation rather than plumbing around it.

Practical consequence: if your roadmap includes multi-building training, the carrier conversation belongs in the earliest site-diligence window, alongside substation availability. If your workload is inference and content delivery, fiber still matters but the binding product is transit and peering, and you should spend your effort on commit tiers and 95th-percentile mechanics instead.

2. The product ladder: know which rung you are standing on

Every carrier sells some subset of the same six rungs. The rung determines who owns the optics, who can upgrade capacity, what your lead time looks like, and how much of the failure domain you control. Confusing rung four with rung five is the single most expensive category error in this market.

Rung	What you receive	Who owns the optics	Capacity ceiling	Typical turn-up	Best fit
1. IP transit / DIA	Internet access with routing, peering, DDoS	Carrier	Your commit tier; 95th percentile billing	Days to weeks where on-net	Inference egress, SaaS, general internet
2. Carrier Ethernet / EPL	Layer 2 point-to-point with committed rate	Carrier	Port speed, 10M–100G	Weeks	Branch, cloud on-ramp, moderate site-to-site
3. Managed wavelength	A dedicated lambda, 10G/100G/400G clean handoff	Carrier	Contracted wave size	Weeks on-net, longer off-net	Predictable capacity, lean optical team
4. Dark fiber lease	Unlit strands; you light them	You	Whatever your optics support	Weeks on-net; 120–150 days off-net; 6–18 months if new build	High and growing capacity on stable routes
5. IRU	Long-horizon exclusive right, 10–30 years, 20 most common	You	Whatever your optics support	Same civil work, longer paper	Capital-grade control; survives the grantor's insolvency
6. Build your own	Your conduit, your cable, your permits	You	Whatever you pull	12–36 months for a diverse lateral	Campuses where no carrier will serve you

Two mechanics on the lease-versus-IRU choice deserve explicit attention, because they are accounting decisions as much as engineering ones. Under a lease the fiber is an operating expense and the whole remaining contract sits as a liability; under a prepaid IRU the cost is booked as an asset and amortized over the contract benefit period, commonly 20 years [14]. The IRU buyer pays a lump sum calculated on strand-miles plus an annual maintenance charge calculated on route-miles — note the different denominators, because it means adding strands later is priced differently from adding route length [14]. An IRU is also the structure that has historically survived the insolvency of the fiber owner, which matters when your supplier is leveraged [15]. In exchange you take the traffic risk, the routing risk, the electronics obsolescence risk and the maintenance obligation, and you should price all four before you sign [14][15].

On rung 3 versus rung 4, the honest comparison is not monthly fiber rent against a monthly circuit price. You have to load in the optics at both ends, spares, amplification or regeneration over distance, a monitoring platform, and the engineering time to run it — or accept that a managed wave has bundled all of that into the monthly rate.

3. Distance decides the economics: compute your break-even rate

Here is the question to put to every carrier quoting you dark fiber: what is your rate per strand-mile per month? Then compare it to your break-even rate, which you can compute in about thirty seconds.

Set up the comparison at a given channel count N of 100G circuits:

• Dark fiber monthly = 2 × route miles × rate + $3,500/mo + $50/mo × (N − 1)
• Managed wave monthly = N × $6,000/mo

The $3,500 bundles $2,000/mo of operations and monitoring, $1,000/mo of amortized optics against a $60,000 one-time line system over a 60-month term, and $500/mo of cross-connects at two sites. The $6,000 is the midpoint of the published metro 100G wave band of roughly $3,500 to $12,000 [20]. Solving for the rate gives:

break-even rate = (5,950 × N − 3,450) / (2 × route miles)

Route length	N = 1 (100G)	N = 4 (400G)	N = 20 (2 Tbps)	Reads as
10 miles	$125	$1,018	$5,778	Dark fiber almost never wins on a short metro hop
40 miles	$31	$254	$1,444	Wins only at high channel counts
100 miles	$12	$102	$578	Contested ground; run your own numbers
400 miles	$3	$25	$144	Dark fiber usually wins
760 miles	$2	$13	$76	Dark fiber wins decisively

Break-even dark fiber lease rate in US$ per strand-mile per month. Below the number, dark fiber is cheaper; above it, buy the managed wave.

Sanity-check the model against a real published deal: ten strands between Phoenix and Los Angeles, roughly 760 route miles, leased for $12,000 per month and carrying about 2 Tbps of owned DWDM capacity. That is $1.58 per strand-mile per month, against a break-even rate of $76 at twenty channels — the deal clears break-even by roughly 48 times [11]. The same $1.58 on a 40-mile metro route would be irrelevant, because break-even there is $1,444 at 2 Tbps. Distance, not price, is what makes dark fiber work.

Worked through over five years at 760 miles and 2 Tbps, dark fiber totals about $1.24 million against roughly $7.23 million for managed waves — 5.8 times cheaper, with crossover in month three, and an effective $10.33 per Gbps per month against $60.15. For calibration on the transit end of the ladder, published on-net transit rate cards show the same shape: about $1.70 per Mbps at a 100M commit falling to $0.11 per Mbps at 100G, which is $110.50 per Gbps per month [18]. Carrier-neutral cross-connects sit in the $50 to $350 per month band, and higher in interconnection-rich facilities [18][19].

4. The physics you cannot negotiate

Latency in fiber is a distance problem with an exact answer. Light in standard single-mode fiber travels at the group velocity c/n_g; with a group index of 1.4682 at 1550 nm, every kilometer adds 4.90 microseconds one way and 9.79 microseconds round trip [6]. One hundred kilometers of fiber is 0.979 ms of round trip before any equipment, amplifier, switch or forward-error-correction step contributes anything.

Fiber route length	One-way delay	Round trip	What it looks like in practice
5 km	0.025 ms	0.049 ms	Campus building to building
50 km	0.245 ms	0.490 ms	Metro, same market
100 km	0.490 ms	0.980 ms	Edge of 400G ZR without amplification
300 km	1.470 ms	2.940 ms	Regional; still coherent-pluggable territory
1,357 km	6.65 ms	13.29 ms	Chicago to northern New Jersey, premium route

The 1,357 km row is not a guess. A published premium fiber route between Chicago and Equinix NY4 in New Jersey measures 6.65 ms one way, which back-solves to 1,357 km of actual glass against a great-circle distance of about 1,200 km [7]. That is a 13 percent route tax. Real fiber follows roads, rail corridors and rights-of-way, so expect 15 to 25 percent over the straight line on most regional paths, and treat any quote that quotes you the geodesic distance as a red flag.

Now the number that changes how you specify equipment. At 800 Gbps across 100 km, one-way propagation is 490 microseconds, which puts 49 MB on the fiber in a single direction; by the time an acknowledgement can return, 98 MB is in flight [17]. That is why long AI links need deep shared buffer pools rather than the shallow-buffer switches optimized for intra-data-center latency — if you cannot absorb the burst while waiting for congestion signals to come back, you drop and retransmit, and the retransmission costs far more than the buffering did [17].

On optics selection, the coherent pluggable generation has made this considerably easier. A 400G ZR module reaches about 120 km with no external transponder and needs greater than 26 dB OSNR; ZR+ modes trade reach for capacity, with 200G QPSK reaching 800 to 1,000 km and requirements in the 22.5 to 24 dB range, and dispersion compensation to roughly 24,000 ps/nm handled electronically in the DSP [16]. Coherent DSP adds well under a microsecond of latency, which is free for training workloads. The caveat is power and thermal: confirm your switch platform supports the module's power class and CMIS version before you order, and remember that clean end-faces matter more with coherent receivers than with direct-detect because they are more sensitive to reflections [16].

5. Two carriers is not two paths

The most common failure in carrier selection is buying redundancy that is not redundant. Facilities advertise carrier counts; what you need is path count. In several US metro markets, two or more providers lease adjacent conduit in the same bridge or share a railroad right-of-way, because those corridors are the only feasible crossing. When a cut happens there, the advertised diversity collapses [9]. One review of a facility advertising twenty carrier options found fifteen of them were subleasing last-mile fiber from the same underlying provider [11].

Do the arithmetic and the point lands hard. Two physically independent paths each at 99.99 percent give a joint availability with 0.32 seconds of downtime per year. The same two contracts with a correlation factor of 0.2 — meaning twenty percent of failures hit both — give 10.5 minutes per year [20]. That is a roughly 2,000 times worse outcome from identical paper. Correlation is the whole game.

So specify the geometry, not the vendor list:

• Two building entrances, on different sides of the building, at least 66 feet apart, so a localized event cannot take both [8].
• Separate conduit and manholes, with no shared handhole anywhere on the route.
• Different bridge, tunnel and river crossings. Where a chokepoint is physically forced, document it and accept it explicitly rather than discovering it during an outage.
• A route exhibit in KML or GIS, showing both paths and every point where they come within a few hundred feet of each other. A sales slide is not an exhibit.
• Separate maintenance windows written into both contracts, so a planned upgrade on one path does not coincide with the other.

Budget for it honestly: in one published case, the genuinely diverse path added 85 percent over the primary, taking the pair from $12,000 to $22,000 per month — and then paid for itself the first time a railroad maintenance crew damaged the primary route [11].

6. The SLA: four numbers, and only one that maps to money

Uptime percentages are marketing. Convert them before you compare anything.

Stated availability	Downtime allowed per year	What a credit is typically worth
99.9%	8.76 hours	25–100% of one month's recurring charge [13]
99.99%	52.56 minutes	Same cap, much higher price
99.999%	5.26 minutes	Same cap, usually requires protected or diverse design

Now put the credit next to the loss. Take a 10 MW hall: at roughly 120 kW per GB200-class rack with 72 GPUs each, that is about 83 racks and roughly 6,000 GPUs. At $2.00 per GPU-hour the hall earns $12,000 per hour. A four-hour MTTR costs $48,000. A credit equal to 100 percent of one month's recurring charge on a $15,000 per month circuit pays $15,000, or 31 percent of the loss — and that is the best case, since credits are frequently capped well below a full month [13]. At $1.50 per GPU-hour the same outage costs $36,000 and the credit covers 42 percent; the gap is structural, not negotiable, because carriers exclude consequential damages as standard [13].

The four numbers to extract, in priority order:

1. MTTR, by region and by failure class. A published standard is a four-hour mean time to repair on fiber with a two-hour premium tier that requires pre-positioned spare fiber and prioritized dispatch [11]. MTTR is the only SLA term that converts directly into downtime hours and dollars.
2. What counts as down. Many SLAs count only complete loss of service, so a circuit that is up but dropping packets may never register. Check that packet delivery and latency carry their own thresholds and their own credits.
3. The credit schedule, cap and claim window. Credits are usually a percentage of the affected circuit's monthly charge, claimed in writing within a contractual window, and they require you to have opened a ticket at the time — the clock usually starts at ticket open, not at failure [13].
4. Maintenance windows and exclusions. Ask how many hours of scheduled maintenance to expect per year and how much notice you get. Force majeure, your own equipment, and anything on your side of the demarcation are excluded everywhere.

7. Contract terms that bite eighteen months later

The network terms get the attention; the commercial terms cost the money. Insist on these before signature:

• Term and escalator. 12, 24 and 36 months are standard, with longer terms buying 15 to 30 percent off the monthly rate [13]. Model the downside: if you sign 36 months and leave at month 12 on a $5,000 per month circuit, termination fees computed at 50 to 100 percent of remaining charges land between $60,000 and $120,000 [13].
• Minimum annual revenue commitments. A MARC buried in the pricing schedule obligates you to spend a floor across all services whether you use them or not. Smaller carriers will often waive it entirely; use that as leverage with larger ones [13].
• Auto-renewal traps. Many agreements renew into a full new term unless written notice arrives inside a narrow window, sometimes 60 or 90 days before expiry. Start renewal talks 90 to 120 days out and push for month-to-month conversion instead [13].
• Chronic outage clause. This is your exit. Without it, repeated small failures generate small credits and no right to terminate.
• Named escalation path with names and phone numbers, not a queue. Ask for it before you need it.
• Move, add, change pricing pre-agreed. On a campus that will grow for five years, MAC charges are a recurring tax you can price now or discover later.
• Relocation and eminent domain handling on IRUs: who pays when a highway project forces the route to move [15].

8. Due diligence past the network map

Carriers are counterparties with balance sheets, and the US fiber map has been redrawn recently. Crown Castle sold its entire Fiber Solutions business to Zayo for $4.25 billion, closing on 1 May 2026, taking Zayo to about 224,000 route miles in North America and returning Crown Castle to a pure-play tower company [3][4]. Read the price against the cost basis: Crown Castle's cumulative fiber investment was roughly $17.5 billion including the $7.1 billion Lightower purchase in 2017, and it sold for less than half [4]. Two companies looked at the same asset class and reached opposite conclusions — worth knowing when you are signing a twenty-year agreement.

On the other side, Lumen's credit profile improved sharply in 2026: following the $5.75 billion sale of its Mass Markets fiber business to AT&T and the retirement of $4.8 billion of super-priority debt, leverage fell from about 5.9 times to the high-4 times range, and Moody's, Fitch and S&P all raised ratings — to B2, B and B respectively, with S&P affirming the issuer rating at B- [5]. Note what that is and is not: those are still speculative-grade ratings, several notches below investment grade, and the company still carries a large debt load while it funds hyperscaler builds [5]. None of that argues against buying from Lumen, which operates a genuinely large network — on the order of 163,000 on-net buildings, roughly 340,000 global fiber route miles, more than 100,000 route miles of 400G-capable fiber and over 2,200 on-net third-party data centers [1][2]. It argues for knowing what you are signing and for structuring long-horizon rights so they survive a restructuring.

Three more checks: ask whether the carrier owns the conduit or is reselling another party's last mile; ask where spare reels and repair depots physically are, by region; and ask for reference calls on two customers who have actually had a cut, not two who have not.

9. Keep fiber off your critical path

Fiber construction is a strictly sequential process, and the sequence is long. Permitting alone — municipal, state and federal rights-of-way, environmental assessment, traffic control plans — commonly consumes six to twelve months even when expedited. Trenching a typical five to ten mile lateral then takes three to six months, and it cannot start before permits are in hand. After the cable is placed, high-count cable requires fusion splicing of every strand at a rate of roughly 20 to 40 splices per day in field conditions, followed by OTDR traces, chromatic dispersion measurement and end-to-end bit error rate testing [9].

Phase	Typical duration	Parallelizable?
Permitting and rights-of-way	6–12 months	No — sequential by jurisdiction
Trenching and conduit placement	3–6 months	No — must follow permits
Cable placement and splicing	1–3 months	Only by adding crews to different segments
Testing and acceptance	0.5–1 month	No — needs the same physical segments
Total, new diverse lateral	10–22 months (12–36 in published worst cases)	—

Sources: [9] for the 12–36 month envelope and per-phase ranges; [12] for the 6–18 month figure for new dark fiber routes in secondary markets.

Contrast that with what is available if the plant already exists: on-net data center interconnects can turn up in 30 days or less, and even custom off-net builds in mature corridors run 120 to 150 days [10]. The spread between 30 days and two years is the entire argument for parallel development. Run route engineering concurrently with white-space construction, secure permits during initial civil planning, and have the fiber lit at server turn-up rather than after it [10].

One more scale check before you size the order: a single 864-strand cable needs 864 splices at each end. At 20 to 40 splices per day that is 22 to 43 crew-days per end, before testing, and re-splicing any anomaly restarts part of the test cycle [9]. Ribbon fusion splicers that join twelve fibers at once collapse that substantially, so ask which method the crew is bringing.

10. Acceptance testing: what to demand at turn-up

Do not accept a dark fiber pair on a ping. The characterization package is what predicts your MTTR for the next decade, and it is cheap to demand at turn-up and expensive to obtain later:

• OTDR trace in both directions at both wavelengths, with span loss and event table.
• Chromatic dispersion and PMD measurements, with envelopes set for the 400G and 800G coherent modes you intend to run.
• Optical return loss, because coherent receivers are sensitive to reflections.
• End-face images for every connector, graded to IEC 61300-3-35.
• End-to-end bit error rate at the intended line rate, and a remediation clause if anything is out of spec.

This is also where you confirm the diversity exhibit against reality: photograph both entrance facilities, confirm the separation, and file the KML with the as-builts. If the as-built route differs from the route you were sold, you have a contract problem that is much cheaper to resolve before acceptance than after.

11. Ten questions for the RFP

1. Is the fiber on-net at both endpoints today, and if not, how many route miles of new construction are required and who pays for it?
2. What is your rate per strand-mile per month, and what is the per-strand-mile rate for additional strands on the same route?
3. Provide the route in KML or GIS for both the primary and diverse path, including every shared bridge, tunnel, conduit and handhole.
4. What is the physical separation between the two building entrances, in feet?
5. Do you own the conduit, or are you reselling another carrier's facilities? Name the underlying owner.
6. State MTTR by region and by failure class, and whether nights and weekends differ.
7. Provide the full SLA document including exclusions, credit schedule, cap, and the claim window.
8. Where are your spare reels and repair depots for this route, and what is the escalation path with names?
9. What are the move, add, change charges, and the scheduled-maintenance hours and notice period per year?
10. Is there a chronic outage clause, and what is the termination right if it triggers?

Common mistakes

1. Counting carriers instead of paths. Twenty logos can be one conduit. Buy geometry.
2. Treating the SLA as risk transfer. Credits run 25 to 100 percent of one month's charge; a four-hour outage on a mid-size GPU hall is a five-figure loss against a four-figure credit [13].
3. Comparing fiber rent to circuit price without the optics. Add optics, spares, amplification, monitoring and engineering time, or compare against a managed wave that already includes them [20].
4. Scoping fiber after the land is bought. Permitting and trenching are sequential and slow; the fiber schedule will set your go-live date [9][10].
5. Quoting the geodesic distance. Real routes run 15 to 25 percent longer; budget the route tax into every latency target [7].
6. Signing a lease when you need an IRU. If you want twenty-year control and insolvency protection, the lease does not give you either [14][15].
7. Ignoring the counterparty's balance sheet. The US fiber map changed ownership materially in 2026 [3][4][5].
8. Accepting turn-up without characterization. No OTDR, no CD/PMD, no end-face images means no baseline when performance drifts.
9. Optimizing transit price on a training campus. If your traffic is scale-across, transit tiers are the wrong lever entirely [17].
10. Letting both paths share a maintenance window. Diversity you have disabled on purpose is not diversity.

References

1. Lumen Technologies, "Infrastructure Services" — network scale, wavelengths to 400G, on-net footprint. lumen.com
2. Lumen Technologies, "Your GPUs deserve a fast network" — 100,000+ 400G route miles, 2,200+ on-net third-party data centers. lumen.com
3. Crown Castle, "Strategic Review Results" — sale of the Fiber Solutions business to Zayo. crowncastle.com
4. "Zayo Closes $4.25B Crown Castle Fiber Deal, Redrawing the US Long-Haul Map" — 224,000 route miles, Zayo's 50th acquisition, Crown Castle's ~$17.5B cost basis. jain.com
5. "Lumen Ratings Upgraded by Top 3 Global Ratings Agencies", Telecom Ramblings Newswire, February 2026 — Moody's B2, Fitch B, S&P B; $5.75B AT&T transaction; $4.8B debt retirement; leverage 5.9x to high-4x. telecomramblings.com
6. Photonica, "Speed of light in fiber" — group index 1.4682, 4.90 microseconds per kilometer one way. photonica.io
7. MarketsFN, "Microwave vs. Fiber: The Physics Arms Race" — published Chicago to Equinix NY4 fiber latency of 6.65 ms one way against a 1,200 km great-circle distance. marketsfn.com
8. Lynx Planning and Engineering, "Designing Fiber Routes for the AI Data-Center Boom: Diversity, Latency and Permit Shortcuts" — distance-to-latency table, dual entrance facilities at least 66 feet apart, separate conduit and crossings. lynxplanning.com
9. "Fiber Scarcity Now Dictates AI Data Center Locations", Research.io — 12 to 36 months for a diverse lateral; permitting 6–12 months; trenching 3–6 months; 20–40 splices per day; shared bridge and railroad corridors. research.io
10. Crosstown Fiber, "How Does Parallel Subsurface Fiber Engineering De-Risk AI Data Center Deployments in Illinois?" — on-net turn-up in 30 days or less; custom off-net builds 120–150 days. crosstownfiber.com
11. DC Atlas, "Fiber Providers and Carriers" — MTTR commitments, SLA credit economics, the Phoenix–Los Angeles dark fiber case (10 strands, $12,000/mo, ~2 Tbps), diverse path premium of 85 percent, and the twenty-carrier-one-conduit problem. dcatlas.io
12. DC Atlas, "Why Your Data Center Has Been 6 Months Away for 3 Years" — 6–18 months for new dark fiber route construction in secondary markets; fiber as the silent killer for site viability. dcatlas.io
13. LegalClarity, "Circuit Leasing: Contract Terms, SLAs, and Pricing" — term lengths, early termination fees of 50–100 percent of remaining charges, MARCs, auto-renewal windows, credit caps of 25–100 percent of one month's charge, exclusion of consequential damages. legalclarity.org
14. CTC Technology and Energy, "Broadband Factors for Last-Mile Connectivity", report to the California Public Utilities Commission, December 2021 — dark fiber lease and IRU structures, 10–30 year terms with 20 most common, upfront payment on strand miles and recurring maintenance on route miles. cpuc.ca.gov (PDF)
15. EducationSuperHighway, "What is Dark Fiber?" — lease versus IRU mechanics, two-installment payment, survivability across sale or insolvency of the fiber owner, maintenance typically excluded. educationsuperhighway.org
16. AscentOptics, "OSFP ZR Coherent: 400G and 800G DCI Guide" — 400G ZR reach around 120 km, OSNR above 26 dB, ZR+ 22.5–24 dB, 200G QPSK reaching 800–1,000 km, dispersion compensation around 24,000 ps/nm. ascentoptics.com
17. LavX News, "Distributed AI training pushes datacenter networks beyond campus boundaries" — Cisco modeling: scale-across needs roughly 14 times conventional DCI bandwidth; two 100 MW sites may need 12,000–32,000 coherent 400G/800G ports against 1,000–2,000 conventional; about 100 MB in flight on an 800 Gbps, 100 km link; deep shared buffers required. lavx.hu
18. IDACORE, published colocation and IP transit pricing — transit tiers from $1.70 per Mbps at 100M to $0.111 per Mbps at 100G; cross-connects at $150/mo. idacore.com
19. ColoPrice, "Colocation Pricing in 2026" — cross-connects $50–350/mo, IP transit $0.10–$1.00 per Mbps, CBRE H2 2025 North America wholesale rates. coloprice.com
20. FiberQuotes, "Dark Fiber vs Lit Waves: The 5-Year TCO, Control, and Risk Showdown" and "Dark Fiber SLA Math" — managed wave MRC of $3,500–$12,000 per 100G metro, fiber MRC of $0.50–$3 per strand-mile, optics and line system $60k–$250k, O&M $1k–$4k/mo, MTTR of 4–24 hours, correlation-adjusted dual-path modeling. fiberquotes.com