They Should Have Bought A Hard Drive - WAN Show August 14, 2026

Episode Summary

Executive Summary: The episode centered on a broad critique of cloud dependence, rising infrastructure costs, and the value of owning hardware, sparked by a PBS archive loss after a cloud reseller shut down. The hosts discussed moving more systems in-house/colo, becoming an ISP/VPN provider, and the financial logic of CapEx versus OpEx. They also covered creator monetization, AI misuse, app store competition, space data centers, surveillance-style loyalty programs, and several quirky personal updates and upcoming LTT projects.

Main Topics: Cloud fragility and the PBS archive loss (Priority: 5/5): A public broadcaster lost access to 50 TB of archives after its cloud storage reseller shut down, reinforcing the hosts’ long-running warning that data you don’t control can disappear. The discussion emphasized the risks of relying on intermediaries and the importance of backups and owned infrastructure. LMG infrastructure reshaping: colo, ISP, and VPN plans (Priority: 5/5): The hosts explained a cost-saving plan to move some services from OVH to colocated hardware, potentially becoming their own ISP and even a VPN provider. They discussed buying switches, IP space, and leasing fiber to reduce long-term costs while adding redundancy and content opportunities. Rising hardware/service costs and business pressure (Priority: 5/5): RAM, storage, networking, and service prices are rising across the industry, driven by AI demand and supply constraints. The team framed this as a macroeconomic squeeze affecting everything from creator businesses to enterprise hosting and even their battery-bank project. YouTube creator strategy and monetization changes (Priority: 4/5): The hosts reacted to YouTube raising the bar for partner eligibility and used the topic to discuss what actually works in modern content creation: packaging, title/thumbnail testing, and converting short-form discovery into durable long-form audiences. Platform power, AI slop, and data exploitation (Priority: 4/5): They covered Twitch’s default AI-training setting, McDonald’s loyalty-data collection, and YouTube’s mistaken AI slop detection against Kurzgesagt. The recurring theme was that large platforms increasingly extract user data and automate decisions with limited transparency or appeal. Space data centers and satellite compute skepticism (Priority: 4/5): Intel’s satellite-data-center patent spurred a lengthy skepticism session about the practicality of orbiting huge compute clusters. The hosts argued that cooling, launch cost, maintenance, and space junk make terrestrial options vastly more sensible, though they acknowledged smaller distributed satellite compute could make sense. Personal and creative pivots at LTT (Priority: 3/5): The show also covered Linus’s birthday content plan, the badminton training machine, the copper-pipe PC build with Linus’s father, braces updates, and the shift toward more family- and interest-driven content on LTT.

Key Arguments: Owning infrastructure matters because cloud services can disappear when a reseller goes offline or becomes inaccessible. For many businesses, especially media organizations, low monthly cloud fees are tempting because upfront hardware costs are hard to absorb. Rising costs across chips, servers, and services are cascading through the economy and pushing companies to reconsider how much they outsource. Moving selective workloads in-house can save large sums over time; LMG’s plan was framed as roughly $400k in five years. Owning IPv4 and colocated network gear can create optionality: ISP service, redundancy, and possible resale value. YouTube success increasingly depends on packaging and audience conversion, not just content quality. Short-form views are cheap; long-form and live views are more valuable and more monetizable. Platforms are increasingly using AI to classify, target, and influence users, often with weak user control or appeal mechanisms. Space data centers are technically interesting but economically and operationally implausible compared with terrestrial alternatives. Some compute in space or on satellites could still make sense if treated as distributed edge infrastructure, not giant orbital data centers.

Data Points: Archived data lost: 50 terabytes - 9PBS reportedly lost access to archival material after its cloud reseller shut down. Cloud price increase warning: 87% potential increase - OVH signaled a major price hike on some services. Service bill increase: 20% - One external company wanted to raise LMG’s annual bill. Lead time for a component: 44 weeks - A battery-bank IC/buck-related component was backordered for many months. Estimated savings: $400,000 over 5 years - Projected savings from the colocated/ISP/networking redesign versus doing nothing. Switch cost: $65,000 CAD total - Two networking switches for the colo/ISP plan. IPv4 purchase: 256 addresses - LMG bought an IPv4 range for its network plans. IPv4 purchase price: $7,000 - Cost of the IPv4 block that was purchased. Backup/storage example: 4 drives - Luke noted that 50 TB could be backed up on consumer gear with a handful of drives. YouTube Partner Program requirement: 8,000 watch hours / 20 million Shorts views - New requirements for new creators, up from 4,000 hours / 10 million Shorts views. WAN show live peak: Over 15,000 viewers - A recent WAN show outperformed the prior weeks significantly. Another WAN show peak: Around 5,000-6,500 viewers - Comparison point for typical recent live performance. McDonald’s loyalty archive: 515 pages - A data request returned a large dossier of loyalty-program tracking. McDonald’s customer pool: 220 million customers across 70 markets - Reported global AI system built from loyalty data. McDonald’s attrition score: 0 - The system predicted the reporter would essentially never stop eating McDonald’s. New Mexico judgment: Nearly $1 billion - Meta was ordered to pay a large judgment in a public nuisance case. Meta penalties breakdown: $567 million + $375 million - Funds for mental health care plus civil penalties. Meta Q2 2026 revenue: $60.8 billion - Used to contextualize the size of the penalty. Meta Q2 2026 net income: $15.85 billion - Used to contextualize the size of the penalty. GeForce Now machine specs: 32-core AMD EPYC, 56 GB RAM, RTX Pro 6000D with 48 GB VRAM - Cloud desktop escape exploit ran on high-end NVIDIA hardware. Local AI throughput: 41 tokens/sec - A 31B model ran locally after escaping GeForce Now sandbox. Tor Snowflake growth: 1300 to 1700 daily unique proxy IPs - Tor said volunteer proxies rose by 29% in June. Tor Snowflake increase: 29% - Month-over-month growth in daily unique proxy IPs. XSquare throughput: 1,816 parcels per hour - Robot logistics system performance in a recent livestream. XSquare accuracy: Over 98% - Robot system accuracy rate. Competitor throughput: 1,248 parcels per hour - Major competitors’ average throughput cited by the sponsor copy. Brain tissue experiment: Post-mortem adult brain explant - Researchers linked brain slices to a piano-playing robot. Human behavior example: One million views - A video about replacing a battery with Linus’s daughter performed strongly. Global RAM sold out prediction: 2027 - A listener asked when RAM shortages would end; the answer suggested before 2028 would be surprising. Podcast company context: 17 grand a year - A recent internal move to floatplane infrastructure reportedly saved this amount. Badminton machine price: $8,000-$9,000 - The automatic shuttle-feeding machine Linus bought for his birthday. Shuttle budget: $2,000 CAD - Value of shuttles used in testing the machine. Motherboard history: 50% RMA rate - The original ROG motherboard line was recalled as notoriously unreliable. AMD/space heat estimate: ~2,000 watts each - Estimated thermal load for newer AMD Instinct GPUs. ISS thermal dissipation: ~70,000 watts - Used to compare against AI GPU power draw. YouTube Shorts monetization note: 10 million to 20 million views - The Shorts requirement threshold was doubled.

Pivotal Quotes: "The cloud is just someone else's computer." — Luke: Used to summarize the risk of outsourcing data storage and infrastructure. "If it was opt-in, nobody would opt in." — Twitch chief product officer Mike Minton (quoted by the hosts): The hosts cited this as an unusually honest explanation for default AI-training settings. "What are we even talking about, man?" — Luke: Repeated during the debate over data centers in space and the practicality of orbital compute.

Implications: Listeners are being pushed to think harder about ownership, redundancy, and platform power. For creators and businesses, cost control, packaging, and infrastructure independence are becoming increasingly important as cloud, AI, and platform policies reshape the economics of tech.