Subsea Cables & Internet Infrastructure

The SLTE footprint has dramatically shrunk over the last quarter century due to Moore's Law and better optics. As the footprint and power requirements fell, the incentive to migrate the SLTEs to carrier neutral data centers increased because colocation costs fell.

 Microsoft has declined to renew a compute agreement with Oracle, partly on the grounds that the Nvidia chips are too old. This supports Michael Burry's contention that GPUs depreciate faster than the five or six years life span that the AI industry uses in accounting treatments. This in turn suggests that AI data center profits are exaggerated. "Microsoft Corp. is winding down an arrangement to rent AI computing from Oracle Corp., part of a larger reassessment of billions of dollars in cloud partnerships. In recent months, Microsoft has been turning off rented computing capacity from Oracle after declining to renew the contract, according to people familiar with the matter. Over time, the computing Microsoft was renting became less useful because it no longer fit the company’s specifications, said the people, who requested anonymity to discuss a private matter. The Oracle hardware used older-generation Nvidia Corp. chips, the people said." Click here for more details: ht...

Recently a client wanted a port and burst Layer 2 service, similar to what Megaport offers in many parts of the world. The customer wanted to pay a fixed port fee and a variable fee for bursting traffic. In the past this setup used Layer 2 Switched Ethernet. The drawback was the cost of Ethernet switching, network management, and the possibility that the provider would severely overbook the service. Making money on such a service is very hard without overbooking. However, most large capacity providers just sell dumb, Layer 1 pipes. So until recently smaller ISPs had few options. Either Megaport or standard chunky Layer 1 waves. The challenge is that many ISPs has peak traffic well below standard 100G wave services, but above 10G. So rightsizing the network is extremely hard. The good news is that the Optical Internetworking Forum developed the FlexE interface standard, which I view as a Layer 1.5 option. It enjoys all the advantages of layer 1 in terms of hard traffic separation, optic...

1. Massive capital expenditures. Much higher capital intensity per rack than a standard data center.  2. Extremely high operating costs due to the power intensity of AI infrastructure. 3. Low customer switching costs shift negotiating power from supplier to customer.  4. Huge gap between capex and revenue flows: 2 to 5 years. 5. Heavy dependence on debt financing and hence vulnerability to rising risk premiums and interest rates. 6. Underlying technology, large language models, is fundamentally flawed according to many computer scientists. This is corroborated by persistent hallucinations and weak problem solving abilities. 7. Large language models show sharply diminishing returns to scale. 8. Luck is required because there are only a few, very large customers for these projects and loss of a single customer can kill a project. Luck is also important because any of these data center projects requires upgrades to the power transmission grid and also generator capacity. 

What a deal!!! Get yourself a promotion, buy now. 🙂 A point: SG1. Z point: Maruyama CLS. Service: Layer 100G Wave. Interface: 100GE. Term: 1 Year. Delivery: Under 4 weeks.

Some claim that Pangram, an AI program, is good at detecting AI written text. So I tested it on Shakespeare. Apparently the Bard was using AI in the 16th century ... Large language models (LLMs) are not logic engines. They are huge nonlinear statistical models with huge number of parameters, in some cases as many as three trillion. LLMs are black boxes. They spit out answers, but it is not clear how they reach those conclusions. Moreover, complicated statistical models are fragile. This means they provide reasonable answers the majority of the time, but will from time to time produce bizarre responses. In contrast, human intelligence is quite robust. Outliers are extremely rare. 

 https://drive.google.com/file/d/1DgndmDXSnGOa5BiWZsVv0pQ_bEx45bhw/view?usp=sharing

Term: three years. NRC: $8,500. Delivery: 4-6 weeks. Remark: Equiano has the lowest latency between Capetown and Lisbon. A point: LS1, Lisbon Equinix. Z point: CT1, Capetown, SA. Customer responsible for cross connects. 

Without an investment grade rating, Anthropic will be unable to tap the funds of most mutual and bond funds despite a trillion dollars in compute commitments over the next three years. This dramatically increases the likelihood of a liquidity crisis and possible bankruptcy. "The FT says the investment giant Blackstone, the anchor investor in a recent $63 billion deal to finance Anthropic’s chips, told investors in the deal that it expects the AI lab to receive a “high” junk rating after its IPO. Previously, the FT notes, Blackstone had said the opposite: that the Claude maker would receive an investment-grade rating (indicating low risk for lenders) before 2027 rolls around." Click here for more details: https://techcrunch.com/2026/10/30/anthropic-expected-to-receive-high-junk-rating-from-credit-agencies-following-ipo-report-says/. 

Free cash flow equals operating cash flow minus capital expenditures. If capex exceeds economic or real depreciation, then that cash could be returned to the owners via dividend or stock buybacks without sacrificing or reducing the company's ability to generate future profits and cash flow. It is a sustainable payout to the owners. Negative free cash flow increases wealth if marginal capex returns are high enough relative to the risk hurdle. Amazon is still quite profitable today, but investors are concerned future returns are inadequate to justify current negative free cash flow. After all, current wealth is being sacrificed in pursuit of future wealth that may be an illusion. The crux of investor angst is that end user AI revenues are tiny relative to investments. Goldman Sachs forecasts global AI capex to reach $1 trillion in 2026 and $3 trillion in 2030. Gartner, the consulting firm estimates $64 billion in 2026 AI end user revenues. The gap between return as measured by end us...

Musk confuses cooling a hot object in space with a cold object in space. An object in orbit cools more slowly than on Earth because a vacuum eliminates cooling via convection or conduction. Heat only escapes via thermal radiation. Hence the ISS has thermal radiators and controlling heat is a major priority in space vehicle engineering. Musk is right that an object without internal heating will settle down to 3 Centigrade degrees above absolute zero (-273.15). But an orbital data center generates heat continuously and will overheat if the radiators are not sufficiently large. What is scary is that Musk has the title of SpaceX CTO, yet he is clearly unqualified given his physics illiteracy. Click to enjoy the video: https://www.youtube.com/watch?v=S1umCJfbJ0w. 

WACS cumulative outage time totals 4% of operating life and ACE 2%. In contrast, Equiano has had zero wet segment faults and under 48 hours of total down time. Equiano's minor outages have been due to power failures at Lisbon Equinix (LS1) and OADC Lagos (power feed equipment failure) plus backhaul fibre cuts between CLS and CT1 (Capetown Teraco 1). WACS and ACE defenders will argue that any reliability comparison is unfair due to the older cable's longer outage history. But their outages are concentrated in subsea caverns that Equiano avoids. So Equiano's reliability edge is probably real.

1. Facility currently consumes only 720kw. Current power capacity is three megawatts. To me this suggests the facility is only 25% full. This is consistent with the opinions of other members of my tour group (Nigerian Peering & Interconnection Forum). 2. Facility design is standard. The Equiano subsea cable uses the sea facing fibre entrance. There are fibre entrances on the building's other three side. So good fibre entrance diversity. 3. Four diesel generators for power back up. Only one is needed to cover current consumption. So good backup power redundancy. 4. The Google CLS is a room in the data center. It was the anchor tenant that justified the project. 5. I suspect that Starlink will be installing its satellite dishes next to the building on the seaside. It has been a nightmare for Starlink to get its Equinix landlord earth stations up and running due to the cost of long haul fibre. OADC has affordable fibre optic connectivity and the sea and absence of tall buildings m...

 

A Chinese team achieved this new world record of 3.2 petabits per second on a single fibre strand using the C, L, and S bands combined with 24 optical cores. So they tripled the usable spectrum on each of the 24 optical highways to get aggregate throughput of 3.2 Tbps. Each core achieved 133 Tbps. It is increasingly common for terrestrial networks to use both C and L bands. For example, Arelion has used the C and L bands because their network routes only consists of one or two fiber pairs. I call it a low fibre count, DWDM-on-steroids approach. The C band is defined as 1530 to 1565 nanometer wavelengths whereas the L band ranges from 1565 to 1625. L band attenuation in standard fiber is only slightly higher than the work horse C band. So L is the natural place to start when expanding the spectrum range to gain more bandwidth. In contrast, the S band has significantly higher attenuation than either C and L band, and consequently, has not been implemented in commercial hardware. S band ...

 Nebius is one of the healthier publicly traded AI data centers companies. Nonetheless, its 2Q2026 income statement illustrates the industry's depreciation challenge. Despite Nebius linearly depreciating GPUs and other server hardware over a generous five years, total depreciation equals 48% of revenues. This figure underestimates true depreciation because technological change and physical burn out means the true average life span is closer to three years. Facebook has provided detailed data on equipment failure during its 56 day estimation of the Llama 3.0 frontier model: https://ai.meta.com/research/publications/the-llama-3-herd-of-models/ . For a 100,000 GPU cluster, equipment failure happens every 30 to 40 minutes. This halts estimation and requires starting over from the previous iteration. Almost all publicly AI data centers exaggerate their profitability and net cash flows by pretending the server hardware lasts 60% longer than it is really does. Michael Burry, who success...

This offer is for early stage ISPs that find 100G waves too expensive and worry about potential down time on an unprotected point-to-point optical service. This route protected 50G point-to-point Ethernet circuit rides Africa's most reliable cable, Equiano, which has had zero wet segment outages since it went live in February, 2023. The 2Africa cable provides the protect path. End points are OADC Lagos and LS1 Lisbon. Term is 2 years. MPLS restoration. Customer is responsible for cross connects. 

4% O&M; MRS1/SG1.  Fast Path: 135 ms RTD. 

Peace Cable Layer 1 100G floor pricing is approximately $17K per month on 1 year contracts for Chinese carriers. There is a $4K MRC premium for non-Chinese carrier capacity using Western network gear like Ciena or Nokia.  SMW5 MRCs range from upper 20Ks to mid-30K range. Mostly sold out; real offers extremely rare. AAE1 MRCs for the standard latency route range from $26K to lower $30Ks. 

You pay $40K for a 100G wave to the CLS, but its operator hits you with another $60K for the cross connect. 🙂 Much like the Hotel California in the Eagles' song. 😉