Subsea Cables & Internet Infrastructure

 The human brain runs on 20 to 30 watts. AI (large language models) model estimation consumes gigawatts. A teenager can learn to drive in a couple hours. In contrast, Tesla AI requires 10 billion miles of actual car driving data. And it still makes elementary driving errors like failing to stop for trains: https://www.bloomberg.com/news/articles/2026-08-22/nvidia-customers-notified-about-ai-related-price-hikes-above-15? So is betting the house on a resource-intensive, but second rate technology? Large language models appear too expensive and too mediocre in performance to be the future of AI.

The Echo cable is a 12 fibre network directly linking Singapore to California with with a Guam branching unit. Bifrost, the other cable connecting Singapore and the US, also lands in Bifrost. I believe each cable effectively serves as backup for the other because traffic can be rerouted at Guam. Both cables are only 12 fibre pairs due to the great distances involved. Its design capacity is 260 Tbps. A key feature of both Bifrost and Echo is they bypass the South China Sea even though it is part of the shortest path between Singapore and the US West Coast. Hence Chinese permits are avoided, resiliency is improved since most Intra-Asian traffic traverses the South China Sea, and the cable is more protected from Chinese tapping or attack. However, it does come as a price. The cable must be buried quite deep through the shallow Indonesian waters to minimize fishing boat or anchor damage. And Indonesia is quite slow to issue permits for new subsea cables. So two cables likely...

JAKO is a high capacity cable connecting Korea and Japan via a Busan, Korea, and a Fukuoka, Japan landing. Busan is a major Korean cable landing spot. Nine cables including JAKO land in Busan. The cable's consortium consists of Microsoft, Amazon, Arteria (important Japanese back haul carrier), and the Korean Dreamline company. Dreamline is a Korean carrier that provides tower services, metro and long haul connectivity. It is increasingly common for hyperscalers to team up with competitive carriers in Asia because the regional PTTs are viewed as difficult, slow to make decisions, and too concerned about protecting their home turf. Uncooperative is hyperscaler diplomatic language to describe the telecom incumbents. NTT is an exception to this generalization. Obviously this cable reflects Microsoft and Amazon's cloud and AI ambitions. No public information is available on the number of fibre pairs. Given the fixed costs associated with cable deployment and rapidly growing traffic,...

Apricot is one of the most exciting Southeastern Asian cable projects to be RFS next year. In fact, it is the first subsea network to be truly physically diverse from the South China Sea. The 12 fibre pair main trunk's end points are Singapore and Tokyo, but the path completely bypasses the South China Sea by detouring through Indonesia and up the Philippines' East Coast. This increases latency, but makes Apricot the natural complement for any cable system like ADC or SJC2 that goes through the South China Sea. Obviously the latency is higher than the usual suspects, but most buyers need resilient networks to keep their jobs. Apricot lands in Singapore, Indonesia, the Philippines, Taiwan, Guam, and Japan. The cable system has military value as it links Taiwan to the Guam, the latter being the US military command center for the Western Pacific. Its routing around the South China Sea also makes it less vulnerable to sabotage. Design capacity was originally intended to be 211 Tbps...

Although most subsea cables with Singapore POPs land at Singapore, some do not, and these exceptions are the keys to better uptime for your Layer 1 network. Five fibre pair AAE1 lands at a Penang, Malaysia CLS that has a fully diverse fibre ring for automatic route protection into Singapore. Each side of the fibre ring uses a separate bridge to cross the Malaysian/Singapore border. One side of the ring terminates at the older of the two Global Switch data centers and the other at SG3. So AAE1 neither lands at Singapore nor uses any of the Singapore cable landing stations. This is one reason why AAE1 is a 'must' for network capacity buyers. Other reasons include low latency for the Marseille/Singapore end points and relatively good pricing due to the system's high capacity. Bay of Bengal Gateway also lands in Penang, Malaysia, but my understanding is that it and AAE1 use different cable landing stations. However, I have not seen backhaul maps so I cannot clai...

Building a resilient Layer 1 network where Singapore is a major POP is very difficult. Most subsea cables land in Singapore via designated sea lanes with four to five cables per lane. This increases the risk of a dragged anchor taking many cables. Although over the last ten years there have been few or none such incidents, a network planner is judged on future performance and proper risk management. Moreover, most subsea cables land at subsea cable landing stations (only 4 in total) serving many other cables. So dividing traffic among several cables is not good enough if resiliency is the goal. Tuas CLS: Apricot, Asia Direct Cable, Bifrost, I-2SEA, I2I, IAX, Indigo West, INSICA, MIST, MVISTA, PEACE, SEA-H2X, SMW4, SMW5, SMW6, SJC, & UMO. Tanna Merah: SEAX-1, BJC2, BJC3, & ASC. Changi North: AAG, BSCS, CANDLE, C2C, Echo, JAKABARE, MIC-1, RISING 8, TGN-IA, TGN-INDICOM, & TIS. Changi South: APG, ASE, C2C, MATRIX, & SJC2. Cable landing diversity is extremely hard to achieve...