Subsea Cables & Internet Infrastructure

The human mind's energy consumption is far more efficient than large language models. Even a genius like Einstein or Poincaré used 30 watts max. 🙂 In contrast, META's Lama 2.0 front tier model required 700 MWh and later versions much more. AI service providers are using brute force as opposed to finesse in their struggling attempt to outperform human intelligence and creativity. Throwing lots of power and GPUs will not lead to the Singularity. AI advocates calls it 'scaling', but I call it 'diminishing returns'. As a result of the brute force approach, AI service providers are losing money on their $20 a month subscription plans due to the high cost of compute and power. VC money pays not just for infrastructure, but also the AI service provider's operating costs. In New York video rendering companies believe the AI services they buy are definitely below cost, yet are still not easy to use. At some point the introductory pricing will disappear in order to c...

Bulk Infrastructure, the Nordic data center and transport player, wants to build a subsea cable linking Norway to Canada for an estimated $620 million. While I applaud Canada for standing up to Bully Trump, this project is fundamentally flawed 1. Most Canadian Internet traffic peers in the US at Secaucus Equinix or in Chicago. It does not go to Europe and the same holds for voice traffic. Most Canadian international calls are to the US. None of this should surprise anybody. Nearly 50% of Canadians live within 100 kilometers of the US border and the two countries are each other's other largest trading partners. Bulk's project has no firm economic foundation. I worked at Hibernia Atlantic for six years from 2005 to 2011, the only high capacity cable directly linking Canada to Europe. Canadian transport demand to Europe was tiny. European transport demand to Canadian destinations was almost non-existent. Internet traffic between the US and Canada is many magnitudes bigger than be...

Nscale has only one major client, Microsoft, and they just cancelled. So its IPO, desperately needed to finance operations and investment, is unlikely to happen. Last week another neocloud, Australian Firmus, was forced to cancel its IPO when investors decided that the company was not worth $35 billion based on 2026 fiscal year revenues of $51 million. On top of all this, we have Coreweave's long term bonds yielding 16%. As those rates, public financing makes no sense and the company needs private financing to avoid default on its existing debt.  Investors should not confuse big orders with success. Companies must be able to deliver in a timely fashion. In the case of Nscale, they could not meet the delivery date for the data center because they could not secure the required power until the next decade. Moreover, the big and growing order backlogs in the AI data center sector reflect not just strong demand, but the difficulty of building these data centers and equipping then. The c...

Nearly half of AI data center debt issued in 2026 is rated high risk (junk category). These bonds generally yield market rates today between 7% and 16%. So if data center company A's bonds yield 15%, then any new debt will probably be priced at 15% or above unless the debt has special features like collateral or equity conversion options. Yet the stock market is valuing these companies much more generously than their bond market counterparts despite persistent net income or net operating losses. Who do you trust? Well, the bond market is an income and risk calculation machine. It is not susceptible to fads or hero worship. Risk miscalculations are possible, but corrected immediately as new information arrives. In contrast, Warren Buffett has noted that short run the stock market is a voting machine and long term a calculating machine. So popularity dominates short term, but long term the wealth created is the sole consideration (present value of net operating cash flow). Hence a wi...

The International Cable Protection Committee has conducted a statistical study using an 18 year time series that includes 3,000 repairs. A link to the article is attached to this post. A fault is any damage to a cable that affects performance or might in the future (like a shunt fault). 1. Fault rates per kilometer of cable have dived over the last two decades. Total annual repairs range between 150 and 200, yet total subsea cable mileage has risen from 900,000 kilometers in 2008 to 1.4 million today. However, this underestimates the decline as fibre pair mileage is probably 5x greater today than in 2008 as fiber pairs per cable have risen from 4 pairs in 2008 to 16 pairs today. 2. Faults in European waters and the Atlantic have dramatically declined due to better burial of subsea cables, distributing cable maps to commercial ships, more careful cable routing, and less fishing in some waters. For example, Northwest Europe, which includes Ireland, UK, France, Iceland, Belgium, Netherla...

According to Coreweave's CEO, "The market is being irrational". Except it is not. The company's quarterly interest payments are 4x its quarterly operating income. The company has simply too much debt and cannot grow revenues fast enough to internally finance interest payments. Right now Coreweave's liabilities-to-revenue ratio is 4.1. This includes bank loans, bonds, and long term leases. Moreover, revenues are a tiny fraction of capital spending.  Consequently, most investment spending is debt financed, but the debt markets are now closed to Coreweave because it would be forced to pay what its current long debt is yielding, namely 15% to 16%, usurious rates. In addition, its order backlog is growing due to a faltering supply chain, and global interest rates are steadily rising.  The most common CEO lament during the 2001 dotcom crash was, ironically, the market is being irrational. 😉 The current price to insure Coreweave's bonds against default implies a 60%...

This project will deploy the first fibre optic cable to connect connect Ireland to Continental Europe. Pisces provides a unique route that sharply improves the resiliency of Ireland's connectivity to the rest of the world. The EU loves projects that bypass non-EU countries and directly ties together EU member states. As the UK is notorious for surveillance, it is particularly welcomed. Consortium members include the Irish subsea permitting firm, MDM, an equity fund appropriately named Equitix, and Orange. The cable will have 16 fiber pairs. I estimate aggregate throughput will be at least 320 Tbps. I admire the unique routing of this system. What is not clear to me is the hyperscaler demand for its capacity. Microsoft has agreed at least in principle to take a fibre pair. There is certainly a prima facie case for Google and Amazon to take single fibre pair IRUs on the system because it will improve their own network's resiliency. However, that advantage must be balanced against...

WACS has a lower latency than Equiano between Lagos and Lisbon, but you are going to pay a high price in outage time if history proves a good guide. The ACE record is significantly better than WACS, but both systems are light years behind Equiano. For reliability Equiano wins hands down. In fact, if you can only afford one cable and Equiano has your desired route, Equiano is the way to go. See the three cable's outage performance below. Equiano: RFS February, 2023. Wet system cumulative outage hours: 0. Dry System: Under 48 hours. WACS RFS: May 2012. Wet system cumulative outage hours: 4,903. Dry system outages: 140. Downtime/System Operational Life Span: 4%. ACE RFS: December 2012. Wet system cumulative outage hours: 2,472. Dry system outages: 70. Downtime/System Operational Life Span: 2%. What adds salt to the wound is that force majeure protected the capacity owners from any obl...

Layer 1 bandwidth is increasingly tight in the Pacific Triangle consisting of Tokyo, Hong Kong, and Singapore. In particular, many carriers' spare ADC capacity is completely gone. Whole fibre pairs are exhausted. Most older cables like SJC are maxed out. Even SJC2 is mostly gone. A large number of resellers is compounding the bandwidth misery by demanding 100G pricing in the $9K to $12K range. This is absurd and not happening. I know of 3x ADC 100G waves available on the Tokyo/Singapore route. One year contracts only. Non-Chinese carrier and equipment. Demarcation points are SG1 and Maruyama CLS. MRC is an inexpensive $13.7K. Don't be a fool and shop this price because this is the network operator's offer, not a resale by me. I don't resale. I take you to the Source (like the Machine Main Frame in the Matrix film series). There is one SJC2 100G wave available for contract early February. SG3/CC1. Available on a 3 year contract at $13.5K MRC plus a $10K install. One SJC...

I am starting to create subsea cable databases available to clients that include a variety of details ranging from RTDs, fibre pairs on each branch, outage histories, etc. Will probably incorporate a new company in Latvia for all the telecommunications procurement and consulting activities. Be bringing on part time employees. Also creating a new website to create some independence from social media: https://subseacables.blog/blog/. 

Insurance on AI data center provider Coreweave's debt is pricing a 50% chance of bond default during the next 24 months. Current yield on the company's long debt is 15%. AI is a real estate bubble and Coreweave is likely to be the first to crash and burn. Most top notch computer scientists like Turing Prize winner Jann LeCun view the current generation of AI, large language models, as a hyped technology that is fundamentally flawed. LLMs are statistical models whereas human intelligence can generate universal statements, not just probabilities. Statistical models like LLMs simply extrapolate trends and are not capable of true imagination. In the language of mathematics, anything truly original is orthogonal to the estimated statistical model. This is the true origin of AI slop. See https://www.bloomberg.com/news/articles/2026-10-14/coreweave-s-brazen-dream-crashes-into-reality-as-yields-top-15?srnd=homepage-asia. 

Good outage and repair history. Diverse to any subsea cable routes. Enjoys comparable latency to subsea routes. Good standalone solution because repairs have averaged under 15 hours. 

Larger wavelengths reduce overhead loss and result in net usable bandwidth gains. Each 100G wavelength has approximately the same overhead loss as a 800G wavelength so bigger wavelengths reduce the ratio of overhead to data payload. Hence a fibre pair can gnerates more total payload transmission. Moreover, META avoids muxing and demuxing since many of their applications are 400G and 800G. Ciena, of course, wants to sell more expensive and profitable higher performance coherent and DWDM tech. I don't see this announcement as a remarkable breakthrough. It just reaffirms that the initial design capacity of 270 Tbps is achievable. For full details, https://www.ciena.com/insights/blog/2026/ciena-and-meta-achieve-world-record-800-gbs-across-16,608-km-on-transpacific-submarine-cable. 

Firmus wanted to go public valued at $30 billion. How much actual data center capacity does it have booking revenue? 46 Megawatts. About the same power consumption as Vienna's Interxion 2 data center. Now Firmus has contracts for 912 megawatts. But there is no guarantee given supply side chain obstacles that it can meet its obligations. Nor do we know if these 'signed contracts' have any contractual teeth. A good guess is less teeth than grandma. Despite only modest-sized AI data center, Firmus has managed to incur $30 billion in debt and plan to issue more.  For full details, https://thetechcapital.com/firmus-shelves-ipo-valuation-concerns

What is amazing is how the EU consistently rewards European companies that failed to anticipate technological change like Eutelsat and Arianespace. Its French executives were too busy with two hour lunches, Bordeaux wines, and crème brûlée to realize that LEO satellites were the future. See the quotes below. 😉 Eutelsat is one of the prime beneficiaries of the IRIS2 project which involves a 330 LEO and 18 MEO constellation that will only cost $18 billion. We all know that Starlink could have done the project for under a billion dollars. A Starlink V3 satellite costs $1.2 million to manufacture and provides 1 terabit download throughput. The point is that the European satellite industry is grossly inefficient. It lacks the economies of scale in production. Hence a Fortress Europe approach to technology is likely to be super expensive. EU procurement often rewards mediocrity because it is trying to create sovereign infrastructure in areas where the Americans dominate and won fair and...

Ooredoo is a multinational mobile service provider with operations in Algeria, Tunisia, Palestine, Kuwait, Saudi Arabia, Qatar, Iraq, Oman, Maldives, and Indonesia. Its 2025 revenues were $6.4 billion revenues. Like many multinational mobile carriers such as Vodafone and Bharti, Ooredoo is becoming an important subsea cable player. Its first noticeable success was winning the competition to host the 2Africa CLS in Oman. More recently, Ooredoo announced in January 2025 the Fiber In The Gulf (FIG) cable that will link Kuwait, Iraq, Qatar, Bahrain, the UAE, Oman, and Saudi Arabia. Due to the short distances between the countries involved, this 24 fibre pair cable is expected to achieve 720 Tbps or 30 Tbps per pair, a Middle East record. The shorter the distance between DWDM optical nodes, the less optical errors accumulate, and hence the greater the throughput. The higher optical power or intensity required for long subsea spans inevitably creates nonlinear errors or optical distortion th...

The 16 fibre pair MVISTA project has the potential to disrupt the cozy world of Southeast Asian telecommunications. The majority shareholder is the Stavian Group, a Vietnamese industrial conglomerate building one of the first carrier neutral data centers in Vietnam. Stavian is a good example of how cash heavy and well established companies are entering the Southeast Asian service provider market in search of their slice of the Good Life. Another example is Keppel, which began life as a builder of offshore rigs after WW2 and today is a very successful asset manager of real estate, infrastructure (power), and connectivity. Companies outside telecommunications are drawn to subsea cables due to the robust connectivity demand in Southeast Asia. This reflects the region's fast growth, but also a fragmented geography. Many countries are islands like Indonesia, New Guinea, Philippines, Singapore or peninsulas like Thailand and Malaysia. The easiest way to build cost effective communication...

Coreweave is a startup AI Data center play. It makes Oracle look downright healthy and athletic. While Oracle's legacy software business generates a lot of cash flow, Coreweave totally depends on external financing in the form of stock sales and debt. Moreover, Coreweave has exactly the same problem as Oracle, namely high customer risk. Coreweave's main clients are Microsoft, META, and OpenAI. Now Coreweave claims to have a $100 billion backlog. But the Devil is in the Details. No contract is valid if Coreweave cannot deliver power, GPUs, and space to fulfill them. And that is exactly the probem. Demand vastly exceeds supply with no relief in sight for years to come. Secondly, it is likely that OpenAI and Anthropic are issuing more compute requests than they actually need given their soft commitment contracts. So if you add up the backlog across all customers, it probably vastly exceeds real demand.  These key fundamental factors are the same as in the case of Oracle. An AI da...

I am seeing a growing number of requests for subsea capacity linking Thailand to Singapore. Thailand is challenging for many reasons. Many consortium members own no fibre beyond the Thai cable landing stations. This leaves them hostage to the country's two main cable landing station operators, TOT and NT, state owned telecom operators, that have exploited their CLS control to charge high backhaul pricing to the carrier neutral data centers.  The country has three key cable landing stations. Sri Racha, just outside Bangkok, hosts ADC, AAG, and some older cables. In the far South, near Thailand's Malaysian border, a large number of cables land at the Songkhla and Satun cable landing stations. Among the more important high capacity modern cables, AAE1, IAX, and MIST land at Satun. SJC2, AAE1, MVISTA, APG, SEA-H2X, and TIS use the Songkhla CLS. As you can readily infer from the map, backhaul pricing from the South to Bangkok is quite high due to the distance and terrain. The terrai...

AI data centers are extremely capital intensive. Indeed, much more than a standard data center due to the power infrastructure, bare metal servers, and GPUs. But whereas most capital intensive industries have low operating expenses, the AI data center has very high operating expenses as many facilities require hundreds of megawatts or in some cases gigawatts of electricity. So this leads to a high capex, high debt business model combined with high operating expenses. Not to mention, low customer switching costs since AI service providers generally just download their data and models into these bare metal servers and estimate them. Hence, the balance of power generally favors the customer, AI service provider, as opposed to the standard data center model where customers incur very high exit costs. The result are contracts with no penalty termination clauses and 60 to 90 days cancellation notice requirements. In Oracle's case all of its free cash flaw has evaporated due to the immens...