A 90 GB day-one patch shipped to two million preloaded clients is 180 petabytes of egress compressed into a rollout window you do not fully control. At hyperscaler list rates near $85 per TB, that single event bills out around $15.3 million. At a high-volume game CDN contract closer to $2 per TB, the same bytes cost roughly $360,000. Nothing about the delivery path changed. Only the contract did. That ratio, not p95 latency, is what decides which game CDN a studio actually ships on.
This rewrite gives you four things: the burst-capacity math for launch day, the transfer-layer behaviours that actually break large file distribution at scale, a 2026 provider comparison with real cost-per-TB anchors, and a workload-profile decision matrix so you can pick without a three-month bake-off. Patch delivery is a throughput and economics problem wearing a latency costume.
Install sizes kept climbing through 2025. Flagship PC and console titles routinely ship 100–200 GB base installs, and seasonal content drops of 15–40 GB are normal rather than exceptional. Console platforms preload before unlock, which concentrates demand into a narrow decrypt-and-verify window instead of spreading it. The result is a traffic shape no web CDN was designed around: near-zero baseline, then a sustained multi-terabit plateau for six to twelve hours, then collapse.
Three properties matter more than everything else combined for a game patching CDN:
Modern patchers do not download files. They download chunk manifests, then fetch content-addressed blocks of roughly 1–4 MB. A CDN that caches only whole objects will fetch a 40 GB archive from origin to satisfy a 2 MB range hit. Confirm three behaviours in a proof of concept before signing anything: partial-object caching with byte-range fill, correct 206 handling on cache miss with concurrent range coalescing, and stable ETag or immutable-content semantics so resumed transfers after a mobile handoff do not restart from zero.
Instrument resume success rate as a first-class metric. In 2026 conditions, mobile and satellite-backed connections in emerging markets still see connection resets during multi-gigabyte transfers at rates that make a broken resume path visible in support tickets long before it appears in your dashboards.
Take peak concurrency, multiply by average sustained per-client throughput, and you have the Tbps figure your provider must absorb on your behalf while also serving everyone else. One million concurrent clients pulling at a modest 20 Mbps is 20 Tbps. That is not a hypothetical number — it is a normal AAA unlock hour, and it is the number to put in front of a sales engineer before discussing price.
Ask for three commitments in writing: guaranteed burst headroom in Tbps for named regions, the escalation path if you exceed it, and whether your traffic is subject to fair-use shaping when a neighbouring tenant bursts simultaneously. Shared-capacity CDNs are fine right up until two publishers unlock on the same Tuesday.
The most effective 2026 launch pattern is publisher-side rate governance: staged unlock windows by region, client-side jittered start times, and a server-driven token that gates manifest access. Distributing the same bytes over ten hours instead of two cuts your peak Tbps requirement by 80 percent and moves you into a cheaper capacity band. Studios that treat throttling as a failure mode overpay for headroom they use twice a year.
Prices below are public list or published entry pricing as of 2026 and shift with commitment. Treat them as order-of-magnitude anchors for cost-per-TB modelling, not quotes.
| Provider | Indicative cost per TB | Strength for patch delivery | Trade-off |
|---|---|---|---|
| BlazingCDN | $5 entry, $2 at 2 PB+ | NVMe SSD edge storage keeps large chunk sets hot under churn; flat volume tiers make launch budgeting deterministic; onboarding measured in about an hour | Smaller brand footprint than incumbents in procurement checklists |
| Akamai | Negotiated, typically well above commodity tiers | Deepest last-mile embedding and long-standing console platform relationships; unmatched burst absorption | Contract complexity and price make it hard to justify below multi-PB volume |
| Amazon CloudFront | ~$85 list, single-digit dollars under private pricing | Tight coupling to S3-backed build pipelines and IAM | List pricing is untenable for patch volume; private terms require committed spend |
| Bunny.net | ~$5–10 depending on region mix | Excellent developer ergonomics and edge scripting; strong price transparency | Regional pricing multipliers complicate global launch forecasting |
| CDN77 | Commit-based, low single digits | Genuinely strong at sustained large-file throughput and dedicated capacity deals | Sales-led onboarding; less self-serve |
| Gcore | Mid single digits at volume | Strong reach into markets where western CDNs are thin; gaming-adjacent compute in the same account | Performance consistency varies more by region than tier-1 incumbents |
| Cloudflare | Enterprise-negotiated | Security posture and platform breadth are best in class | Large-file egress at patch scale requires enterprise terms; not the default choice for pure download workloads |
Model your actual year: base installs, monthly patches, seasonal events, re-download rate from platform reinstalls. A mid-size live-service title distributing 800 TB monthly pays roughly $2,800 at $3.50/TB versus roughly $68,000 at CloudFront list. That delta funds an entire build-and-release team. Run this before you argue about edge compute features.
| Workload profile | Dominant constraint | Best-fit approach |
|---|---|---|
| Indie or early-access, under 50 TB/month | Predictable spend, zero ops headcount | Flat-tier provider at $5/TB entry; skip multi-CDN entirely |
| Live-service, 500 TB–2 PB/month, steady patch cadence | Cost per TB and cache hit ratio under churn | Single primary on volume commit ($2.50–3/TB band) plus origin shield; secondary CDN for failover only |
| AAA launch, 10+ PB in one week | Peak Tbps headroom, regional saturation risk | Multi-CDN with weighted DNS or manifest-level steering; staged unlock throttling; contracted burst guarantees |
| Mobile-first with global emerging-market base | Resume reliability and regional reach | Provider with strong non-western footprint; aggressive delta patching to cut bytes at the source |
| Console-platform distribution | Platform holder controls the pipe | Optimise package layout and delta size; your CDN choice governs PC, launcher and asset side-loads only |
Thundering-herd origin fill. A fresh patch with cold edges triggers thousands of simultaneous misses per region. Without request coalescing and a shield tier, your origin object store gets hit with a self-inflicted denial of service. Pre-warm the top 20 chunk sets per region before unlock, and verify coalescing is enabled rather than assuming it.
Cache key fragmentation. Signed URLs with per-user tokens in the path split one hot object into millions of unique keys and drive hit ratio toward zero. Keep authentication in query-independent token headers or normalise the cache key so signature material is excluded.
Silent range degradation. A configuration change disables partial-object caching. Hit ratio looks fine because full objects are cached; origin egress quietly triples. Alert on origin bytes per delivered byte, not on hit ratio alone.
Rollback with no path back. Immutable content addressing makes rollback trivial: point the manifest at the previous chunk set. Mutable filenames with long TTLs make it a purge race across every edge. Design for the former.
For studios whose deciding variable is cost per terabyte across a predictable patch cadence, BlazingCDN sits in the same league as Bunny.net, CDN77 and Gcore, and competes on published, flat volume tiers rather than negotiated opacity: $100/month up to 25 TB ($0.004 per additional GB), $350 up to 100 TB, $1,500 up to 500 TB, $2,500 up to 1,000 TB, and $4,000 up to 2,000 TB with additional GBs at $0.002. Entry is $5 per TB and at-volume pricing reaches $2 per TB. NVMe SSD edge storage matters specifically for patch workloads, where each release invalidates the previous chunk set and retention under churn determines your origin bill.
The operational claim worth testing rather than believing: stability and fault tolerance comparable to Amazon CloudFront, with 100% uptime, flexible per-property configuration, fast scaling under unlock-hour spikes, and onboarding measured in about an hour. For enterprises moving petabytes, the cost difference is structural, not marginal. Details on the gaming configuration live at BlazingCDN's game distribution setup.
Patch delivery is dominated by sustained multi-gigabyte transfers with heavy byte-range activity and near-total object churn on every release. General web CDNs optimise for small-object request rates and TTFB. The metrics that matter here are aggregate egress capacity, partial-object cache retention, resume success rate and origin bytes per delivered byte.
Multiply projected peak concurrent downloaders by the average sustained per-client throughput you expect, typically 15–30 Mbps depending on region mix. One million concurrent at 20 Mbps is 20 Tbps. Then subtract whatever peak you can shave with staged regional unlocks and jittered client start times, and negotiate contracted burst headroom against the remainder.
Above roughly 5 PB per month or for one-shot AAA launches, yes — it buys regional failover and negotiating leverage. Below that, the engineering cost of manifest-level steering, dual cache warming and reconciled billing usually exceeds the benefit. A single primary with a contractually tested secondary for failover covers most live-service titles.
Steady-state, above 95 percent is achievable with immutable content-addressed chunks and adequate edge storage. In the first hour after a release, expect it to dip sharply as cold edges fill. Pre-warming key regions and enabling request coalescing keeps that dip from becoming an origin incident.
Hyperscaler list pricing sits near $85 per TB, which nobody at patch scale pays. Commodity high-volume CDNs land between $2 and $10 per TB depending on commitment and region mix, with entry tiers around $5 per TB and at-volume contracts reaching $2 per TB. Model your annual byte volume first; the tier you land in matters far more than the headline rate.
Yes, often more than switching providers. Content-defined chunking with binary diffs commonly cuts patch payloads by 60–90 percent versus full-file replacement, and correct package layout prevents small code changes from invalidating large asset blocks. Fix the bytes before you renegotiate the rate.
This week, instrument one metric you probably do not have: origin bytes divided by delivered bytes, sampled per region per hour, over a full patch cycle. If that ratio exceeds 0.05 outside the first fill hour, you have a range-caching or cache-key problem that is costing real money right now. Then take your last twelve months of egress, apply the $5/TB and $2/TB anchors, and see what the delta funds.
What is your current resume success rate on transfers above 20 GB? If you cannot answer that from a dashboard, that is the second thing to instrument.