publish date 2026-08-06
Welcome back! This is Canaan's weekly newsletter on Bitcoin mining, energy, and compute infrastructure.
For bitcoin miners, siting industrial-scale compute has historically been a one-variable problem: find the lowest-cost, reliable megawatt. That strategy worked because bitcoin mining is one of the few compute workloads at scale that is largely indifferent to network bandwidth. An ASIC exchanges only a few kilobytes of data with its mining pool. Even a thousand-machine facility can operate comfortably on a 50 Mbps connection. What matters is uptime and low round-trip latency, not throughput, because a share submitted too late is unpaid. Connectivity rarely disqualified a site. Power economics decided almost everything.
Satellite connectivity reinforced that. Starlink's business tier delivers speeds exceeding 400 Mbps, and its Community Gateway product delivers multi-gigabit capacity in locations where no terrestrial fiber exists. For bitcoin mining, network access is no longer a meaningful siting constraint almost anywhere in the world.
AI is different.
Modern AI campuses require not only abundant power, but also massive internal and external networking capacity. Training clusters depend on hundreds of gigabits (if not terabit-class connections) of bandwidth between accelerators, servers, and storage, while also requiring high-capacity fiber connections to move enormous datasets. For many AI workloads, bandwidth becomes just as important as power.
Even within AI, there are two distinct siting models.
AI does not have one siting profile. It has two, and they pull in opposite directions. McKinsey's analysis of hyperscaler strategy puts a number on it: training is insensitive to latency and tolerates delays of up to 100 milliseconds between adjacent regions, which is precisely what lets operators site it in remote, power-rich areas where grid capacity, land, and water are still available. AI training increasingly resembles bitcoin mining: both prioritize access to inexpensive, scalable power over proximity to end users.
Inference has the opposite requirement. Its build-outs are landing in metro and near-metro locations chosen for low round-trip time and dense interconnection, and McKinsey expects inference demand to overtake training by 2030, growing at an annual rate of roughly 35% to exceed 90 GW of capacity. A site that is power-rich and fiber-poor is not an inference platform, however many megawatts sit behind the fence.
Training pays a one-time freight cost to move data to the compute. Inference pays a continuous latency tax to move compute to the users. The market has already repriced it: long-haul dark fiber demand doubled between 2024 and 2025, and optical analysts now describe AI, rather than telecom, as the primary growth engine for fiber. In the AI era, power determines where compute can be built, but bandwidth increasingly determines what that compute can do.
The bottom line is that a megawatt is no longer just a megawatt. The same 100 MW behind the same substation can generate hashrate, training rent, or inference rent, and it’s the network position that decides which. For miners weighing a pivot to AI infrastructure, the critical question is no longer just how many megawatts a site can support, but what latency class the site can credibly serve. In the next phase of digital infrastructure, power determines where compute can be built. Connectivity determines what that compute is worth.
In the News
Network at a Glance
Project Spotlight
The AvalonMiner 1246, launched in January 2021, was the first Avalon unit to reach 90 TH/s. It drew 3,420W for roughly 38 J/TH and shipped in binned configurations from 81 to 96 TH/s, letting operators match machines to their power cost. Unlike earlier Avalon generations that relied on external controllers, the A1246 incorporated its own network management interface, bringing the line closer to the self-contained architecture used across the industry.
The more interesting part of the generation was the variant. Canaan introduced the Avalon Immersion Miner in 2020, described at the time as the first liquid-cooling immersion device built for blockchain computing, with the A12 series carrying it into volume as the 1246I. Immersion was a curiosity then. Five years later, it is a product category, and Canaan's third-generation immersion unit, the A1566I, runs in commercial fleets at 249 to 267 TH/s.
The hydro and immersion machines shipping today did not appear with the AI thermal-density conversation. Canaan was building them when the industry still thought air was good enough.
Follow and Contact Us
Disclaimer: This newsletter shares industry commentary and third-party news for informational purposes only. The views and opinions expressed by third-party sources are those of their respective authors and do not necessarily reflect the views of Canaan Inc. For official news, please refer to Canaan’s press releases and SEC filings at https://investor.canaan-creative.com/.
For bitcoin miners, siting industrial-scale compute has historically been a one-variable problem: find the lowest-cost, reliable megawatt. That strategy worked because bitcoin mining is one of the few compute workloads at scale that is largely indifferent to network bandwidth. An ASIC exchanges only a few kilobytes of data with its mining pool. Even a thousand-machine facility can operate comfortably on a 50 Mbps connection. What matters is uptime and low round-trip latency, not throughput, because a share submitted too late is unpaid. Connectivity rarely disqualified a site. Power economics decided almost everything.
Satellite connectivity reinforced that. Starlink's business tier delivers speeds exceeding 400 Mbps, and its Community Gateway product delivers multi-gigabit capacity in locations where no terrestrial fiber exists. For bitcoin mining, network access is no longer a meaningful siting constraint almost anywhere in the world.
AI is different.
Modern AI campuses require not only abundant power, but also massive internal and external networking capacity. Training clusters depend on hundreds of gigabits (if not terabit-class connections) of bandwidth between accelerators, servers, and storage, while also requiring high-capacity fiber connections to move enormous datasets. For many AI workloads, bandwidth becomes just as important as power.
Even within AI, there are two distinct siting models.
AI does not have one siting profile. It has two, and they pull in opposite directions. McKinsey's analysis of hyperscaler strategy puts a number on it: training is insensitive to latency and tolerates delays of up to 100 milliseconds between adjacent regions, which is precisely what lets operators site it in remote, power-rich areas where grid capacity, land, and water are still available. AI training increasingly resembles bitcoin mining: both prioritize access to inexpensive, scalable power over proximity to end users.
Inference has the opposite requirement. Its build-outs are landing in metro and near-metro locations chosen for low round-trip time and dense interconnection, and McKinsey expects inference demand to overtake training by 2030, growing at an annual rate of roughly 35% to exceed 90 GW of capacity. A site that is power-rich and fiber-poor is not an inference platform, however many megawatts sit behind the fence.
Training pays a one-time freight cost to move data to the compute. Inference pays a continuous latency tax to move compute to the users. The market has already repriced it: long-haul dark fiber demand doubled between 2024 and 2025, and optical analysts now describe AI, rather than telecom, as the primary growth engine for fiber. In the AI era, power determines where compute can be built, but bandwidth increasingly determines what that compute can do.
https://x.com/matthew_sigel/status/2080649280356847968?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E2080649280356847968%7Ctwgr%5E40632522bec80a3114a0819a2a089cddb3e0dac8%7Ctwcon%5Es1_&ref_url=https%3A%2F%2Fcanaan-newsletter.beehiiv.com%2Fp%2Fpower-says-where-bandwidth-says-what
The bottom line is that a megawatt is no longer just a megawatt. The same 100 MW behind the same substation can generate hashrate, training rent, or inference rent, and it’s the network position that decides which. For miners weighing a pivot to AI infrastructure, the critical question is no longer just how many megawatts a site can support, but what latency class the site can credibly serve. In the next phase of digital infrastructure, power determines where compute can be built. Connectivity determines what that compute is worth.
In the News
- Core Scientific Pays $42M to Exit Block's Bitcoin Mining Deal as AI Revenue Surges
- MARA Adds Power, Data Center Veterans to Board Amid AI Push
- Cipher Secures 900 MW Texas Site Option and Accelerates Black Pearl AI Data Center
- Crusoe Plans $1B Texas AI Data Center Expansion, Pursues Nuclear Power
- CoreWeave Plans 360MW Indonesia Expansion in First Asia AI Data Center Push
- DeepSeek Plans 1-Gigawatt AI Data Center in Inner Mongolia
- Southern Company Plans $2.15B Raise via Convertible Bond
- Meta and BlackRock form $14 billion venture for 1-gigawatt El Paso AI campus
- Bitcoin difficulty falls year over year for only second time: Hashrate Index
- Morgan Stanley Sees CIFR, GLXY Upside From ERCOT's 65 GW Batch Zero
Network at a Glance
- BTC price (USD): ~$64,583
- Network hashrate: 914.1 EH/s
- Difficulty: 126.2T
- Hashprice: ~$32.46 / PH / day
Project Spotlight
The AvalonMiner 1246, launched in January 2021, was the first Avalon unit to reach 90 TH/s. It drew 3,420W for roughly 38 J/TH and shipped in binned configurations from 81 to 96 TH/s, letting operators match machines to their power cost. Unlike earlier Avalon generations that relied on external controllers, the A1246 incorporated its own network management interface, bringing the line closer to the self-contained architecture used across the industry.
The more interesting part of the generation was the variant. Canaan introduced the Avalon Immersion Miner in 2020, described at the time as the first liquid-cooling immersion device built for blockchain computing, with the A12 series carrying it into volume as the 1246I. Immersion was a curiosity then. Five years later, it is a product category, and Canaan's third-generation immersion unit, the A1566I, runs in commercial fleets at 249 to 267 TH/s.
The hydro and immersion machines shipping today did not appear with the AI thermal-density conversation. Canaan was building them when the industry still thought air was good enough.
Follow and Contact Us
- X: @canaanio
- Youtube: Canaanmining
- LinkedIn: Canaan Inc.
- Website: canaan.io
- Support: Customer Care
Disclaimer: This newsletter shares industry commentary and third-party news for informational purposes only. The views and opinions expressed by third-party sources are those of their respective authors and do not necessarily reflect the views of Canaan Inc. For official news, please refer to Canaan’s press releases and SEC filings at https://investor.canaan-creative.com/.
