Gigabyte to Exabyte

Hard drive manufacturers measure 1 TB as 1,000,000,000,000 bytes (decimal). Windows displays storage in gibibytes (binary) but historically labelled them as "GB" — so 1,000,000,000,000 bytes ÷ 1,073,741,824 ≈ 931 GiB, which Windows displayed as "931 GB". macOS (since 10.6) correctly reports the same drive as "1 TB" using decimal GB. The drive is not lying; the OS was using a binary unit with a decimal label.

8 GB RAM is the current minimum for gaming; 16 GB is the recommended standard for most modern games at 1080p and 1440p; 32 GB benefits heavily multitasking systems or games with large open worlds. Memory-intensive tasks like video editing, 3D rendering, and running large language models locally typically require 32–64 GB or more.

A 4K movie in H.264 or H.265 encoding is typically 50–100 GB on Blu-ray; streaming services compress aggressively to 15–25 GB for 4K HDR content. Netflix's 4K streams average about 7 GB per hour; the downloaded version via the Netflix app for offline viewing is roughly 3–6 GB per hour at high quality settings.

1 GB of mobile data supports roughly: 2–3 hours of music streaming, 1 hour of HD video streaming, 2–3 hours of web browsing, or 30–60 minutes of video calling. Social media apps with autoplay video are heavy consumers — TikTok and Instagram Reels can use 300–600 MB per hour of active use.

AI model sizes vary enormously. GPT-2 (2019) is about 1.5 GB; Llama 2 7B is roughly 13 GB in float16 precision; Llama 2 70B is about 130 GB. GPT-4-class models are estimated at 500+ GB. Quantised (compressed) versions are smaller: a 4-bit quantised 7B model fits in about 4 GB, runnable on a modern laptop. Training requires far more — the training dataset, gradients, and optimizer states for a 70B model can occupy 1–2 TB of GPU memory across a cluster. The trend toward larger models is driving consumer GPU memory from 8 GB to 16–24 GB as a baseline for local AI inference.

One exabyte = 1,000,000 terabytes = 1,000 petabytes. If you filled 1 TB external hard drives and stacked them end to end, 1 EB worth would stretch roughly 200 km. In content terms: 1 EB can store about 250,000 years of HD video, or about 100 billion hours of music at 128 kbps. All the data produced by the Large Hadron Collider per year is about 15 petabytes — still 67× less than one exabyte.

Global data creation, capture, copy, and consumption is estimated at roughly 2.5 exabytes per day (IDC 2023 estimate), growing roughly 23% annually. This includes IoT sensor readings, financial transactions, social media posts, surveillance camera footage, scientific instrument output, and all other digital activity. Most of this data is transient and never stored long-term.

Amazon Web Services, Microsoft Azure, and Google Cloud each store estimated tens to hundreds of exabytes of customer data in their cloud platforms. Meta (Facebook/Instagram) stores an estimated 100+ exabytes across all data types. The NSA's Utah Data Center is estimated to hold yottabytes in capability, though actual stored volumes are classified. Collectively, global cloud storage is in the hundreds-of-exabytes range.

An exabyte (EB) = 10¹⁸ bytes (SI decimal). An exbibyte (EiB) = 2⁶⁰ bytes = 1,152,921,504,606,846,976 bytes — about 15.3% larger. This is the largest practically relevant gap between SI and IEC units in storage contexts. For a data center procuring 10 EB of storage, the SI vs IEC difference represents about 1.5 EB of capacity discrepancy in the contract.

Data archaeology is the practice of recovering information from obsolete storage media and formats — 9-track magnetic tapes, 8-inch floppy disks, MiniDiscs, Zip drives, and early optical formats. The challenge is threefold: hardware to read the media no longer exists or is failing, file formats and encoding schemes are undocumented, and magnetic media degrade over time (tape has a 10–30 year shelf life). At exabyte scale, organisations like national archives face the prospect of vast digital collections becoming unreadable within decades. Active migration strategies — periodically copying data to current formats and media — are the only reliable defense, but the cost scales linearly with data volume.