Published list prices, read from the vendor's own rate card and pricing API on a stated date, with the arithmetic shown. The conclusion is not that public cloud is expensive. It is that two of the four meters are priced in a way that punishes exactly the workload video generates, and that South Asian delivery does not receive the volume discount that makes the model work elsewhere.
Three things, in the order they matter to whoever signs the invoice. The arithmetic behind every one of them is further down this page, with the vendor's own rate card and the date it was read.
Three meters, and only one of them is about your catalogue. Encoding bills per output minute per rendition, so a ladder change re-bills the whole library. Storage bills per gigabyte-month against an archive that only grows. Delivery bills per gigabyte against an audience you are trying to grow — and it is the largest of the three by a distance.
Rented delivery is sold as though volume earns a discount. On the North American card it does: the rate falls 4.25× across the tiers. On the South Asian card it falls 1.51×. Tripling delivered volume multiplies the bill 2.89× and improves the blended rate by 3.8%, so scale is not a lever a regional operator can pull.
On an owned deployment: transit or peering inside a commitment, power and rack, hardware amortisation, the cloud half if you chose hybrid, gateway fees on your own merchant account, and the agreement with us. None of those is per gigabyte delivered or per viewer watching, which is the whole difference.
What this page will never give you is a percentage saving. That number needs the owned side, which is set by your volume, your utilisation and the amortisation period your finance side accepts — five inputs listed at the bottom of this page. A headline saving would be the one invented figure on a page built entirely from published ones.
Four figures from the published cards below. They are the reason this architecture exists, and none of them is our opinion.
2.05×
What the same petabyte costs
$83,407.87
One petabyte, India tiers
1.51×
How far the discount falls
$0.072
The floor you never get under
Rented delivery is priced so that growth buys a discount. On the India card that discount is almost absent, so ten times the audience is close to ten times the bill. Owned origin and owned edge change the shape rather than the rate: the cost becomes a commitment you size once, and inside it the next gigabyte is free. What that saves is a function of your volume, your hardware and your amortisation period, so this page gives you the model and the published rates to run it against rather than a percentage that would be true of somebody else.
A streaming service of this shape normally pays three vendors that all bill by usage: one to transcode, one to deliver, one to measure. Every additional viewer increases all three. We removed all three.
This is not a discount. It is a different cost structure. Owning the pipeline converts a variable cost that grows with success into a fixed cost that does not — which is precisely the curve a subscription business needs.
Three meters carry almost all of a streaming bill. Each is quoted below at its vendor list price, for a stated region, on a stated date. No negotiated rate, private pricing or commitment discount is assumed anywhere on this page, because none of those are published and a number that cannot be checked is not worth printing.
Billed per minute of OUTPUT, not of source. A four-rendition ladder bills four times the runtime of the file you fed it, once per rendition, every time you re-encode. AWS defines the tiers by output height: SD is below 720, HD is 720 through 1080, and anything above that is 4K.
| Output tier | Basic · AVC | Professional · AVC | Professional · HEVC |
|---|---|---|---|
| SD — under 720 | $0.0085 | $0.0136 | $0.0272 |
| HD — 720 to 1080 | $0.0170 | $0.0272 | $0.0544 |
| 4K — above 1080 | $0.0340 | $0.0544 | $0.1088 |
The meter that only moves in one direction. A catalogue is never deleted, so this line grows every month a title is added and never shrinks on its own. Requests are billed separately from bytes, which matters for segmented video because one hour of playback is several hundred object reads.
| Meter | Rate | Unit |
|---|---|---|
| S3 Standard — first 50 TB | $0.025 | per GB-month |
| S3 Standard — next 450 TB | $0.024 | per GB-month |
| S3 Standard — over 500 TB | $0.023 | per GB-month |
| GET and all other requests | $0.0004 | per 1,000 requests |
| PUT, COPY, POST, LIST | $0.005 | per 1,000 requests |
The meter that decides the architecture. The rate is set by the geography of the edge location that serves the request, not by where the origin sits — so an audience in South Asia is priced against the South Asian card no matter which region the bucket is in.
| Monthly tier | US, Mexico, Canada | India | Singapore, HK, and SEA |
|---|---|---|---|
| First 1 TB | Free | Free | Free |
| Next 9 TB | $0.085 | $0.109 | $0.120 |
| Next 40 TB | $0.080 | $0.085 | $0.100 |
| Next 100 TB | $0.060 | $0.082 | $0.095 |
| Next 350 TB | $0.040 | $0.080 | $0.090 |
| Next 524 TB | $0.030 | $0.078 | $0.080 |
| Next 4 PB | $0.025 | $0.075 | $0.070 |
| Over 5 PB | $0.020 | $0.072 | $0.060 |
Nothing here is modelled or estimated. Each figure is the rate card above applied to a stated quantity, with the quantity chosen so the tier boundaries land cleanly rather than to flatter a conclusion.
1 TB free, then 9, 40, 100, 350 and 524 TB billed at the published tier rates. 1 TB is taken as 1,024 GB. No discount, commitment or private pricing is assumed, because none is published.
| Price group | 1 PB per month | Blended per GB | Against the US price |
|---|---|---|---|
| US, Mexico, Canada | $40,637.44 | $0.0388 | 1.00× |
| India | $83,407.87 | $0.0795 | 2.05× |
| Singapore, HK, and SEA | $90,112.00 | $0.0859 | 2.22× |
360p and 480p bill as SD, 720p and 1080p bill as HD, at the Basic AVC rates above. This is per minute of SOURCE, with the ladder cost already summed across renditions.
| Rendition | Tier | Rate per output minute |
|---|---|---|
| 360p | SD | $0.0085 |
| 480p | SD | $0.0085 |
| 720p | HD | $0.0170 |
| 1080p | HD | $0.0170 |
| Ladder total | — | $0.0510 per source minute |
Four readings of the numbers above. Each one is checkable against a table on this page, and where a claim depends on something not published, that is said rather than filled in.
The published US card falls from $0.085 to $0.020 per GB across its tiers — a factor of 4.25 between the first billed byte and the cheapest one. That decline is the entire reason a large service can run on rented delivery and still have a business at the bottom of the price list.
The India card falls from $0.109 to $0.072, a factor of 1.51. The Singapore and South East Asia card falls from $0.120 to $0.060, a factor of 2.00, and only reaches that floor above 5 PB per month. For a service whose audience is in Bangladesh, the practical consequence is that scale stops being a lever. Growing ten times larger does not move the unit price meaningfully; it multiplies the bill by roughly ten.
That is the number that decides the architecture, and it is not an opinion about cloud. It is the vendor's own published card, read on the date stated under each table.
The coupling is not the origin fetch. AWS states that cacheable data transferred from its own resources into CloudFront edge locations carries no additional charge, so the S3-to-edge hop is genuinely free and it is worth saying so rather than implying otherwise.
The coupling is that the two meters are sold separately and consumed as a pair, and only one of them has a ceiling you control. Storage is billed per GB-month against a catalogue that only grows. Delivery is billed per GB against an audience you are trying to grow. An object earns nothing while it sits in storage and only earns when it is delivered, so every commercial success on the catalogue side arrives as a bill on the delivery side. There is no configuration that separates them, because the dependency is in the business, not in the account.
Segmented video adds a third meter to the same pair. One hour of playback at six-second segments is roughly 600 object reads before manifest refreshes, so request charges scale with watch time rather than with catalogue size. At the published CloudFront HTTPS request rate for India, a million requests is $1.20 — small next to the bytes, and not zero.
Owned infrastructure does not make delivery free. It changes the shape of the cost from a meter to a commit, and that is a different thing that is worth stating precisely.
Encoding becomes capital. A GPU is bought once and its cost per minute falls every time it is used, approaching zero as utilisation rises. The rented equivalent charges the same $0.0170 for the ten-thousandth HD output minute as for the first, because there is no volume tier on MediaConvert output minutes at all.
Origin storage becomes capital on the same terms, and reads stop being billable events. A cache miss against owned storage costs the electricity and the disk time and nothing else, which is what makes a full catalogue re-crawl or a cache warm a scheduling decision rather than a budget decision.
Delivery becomes a committed rate rather than a per-GB meter. Inside a transit or peering commit the marginal cost of the next gigabyte is zero, and the cost only steps when the commit is renegotiated. Against the published India card the marginal gigabyte never falls below $0.072 no matter how large the service grows.
What owned infrastructure takes on is real and belongs in the same paragraph. Capacity has to be bought before it is needed rather than after. Utilisation below the planning assumption is money already spent. Hardware fails and somebody has to be on call for it. The trade is a fixed, plannable cost with a floor you own, against a variable cost with no ceiling, and it is only the right trade when the volume is high enough and steady enough to fill what you bought.
There is no published price for the owned side because there cannot be one. What there is instead is a model, and it takes five inputs that are specific to a deployment. Delivered petabytes per month, taken from CDN logs rather than estimated. Catalogue size in TB and its monthly growth. Peak concurrency, which sizes the edge rather than the origin. Encode hours per month, including re-encodes after a ladder change. And the amortisation period the finance side will accept on hardware, which is usually the input that moves the result most and is almost never the one people bring to the first conversation.
Given those five, the owned side resolves to a number, and it can be put next to the tables above and compared honestly. Given fewer than those five, any number quoted at you — by us or by anyone — is a number about a different service.
Against the published pay-as-you-go card read on 25 August 2026, the India price group starts at $0.109 per GB after the free 1 TB and falls to $0.072 per GB above 5 PB per month. The Singapore and South East Asia group starts at $0.120 and falls to $0.060. One petabyte per month priced against the India tiers is $83,407.87, a blended $0.0795 per GB.
Because the volume tiers decline far less steeply. The US, Mexico and Canada card falls from $0.085 to $0.020 per GB across its tiers, a factor of 4.25. The India card falls from $0.109 to $0.072, a factor of 1.51. Priced against the same 1 PB month, that is $40,637.44 in the US group against $83,407.87 in the India group — 2.05 times the cost for the identical traffic.
Billing is per output minute, per rendition. At the Asia Pacific Singapore Basic AVC rates, SD is $0.0085 per output minute, HD is $0.0170 and 4K is $0.0340. A four-rendition ladder of 360p, 480p, 720p and 1080p therefore bills $0.0510 for each minute of source, so a 1,000-hour catalogue costs $3,060.00 for one full pass, and bills again in full after any ladder or codec change.
No. AWS states that cacheable data transferred to CloudFront edge locations from AWS resources incurs no additional charge. The coupling between storage and delivery is commercial rather than technical: storage is billed per GB-month against a catalogue that only grows, delivery is billed per GB against an audience you are trying to grow, and an object only earns when it is delivered.
That no per-gigabyte meter sits in the delivery path. Origin storage and the delivery edge both run on owned capacity, and transit is bought as a committed rate, so the marginal cost of the next gigabyte inside the commit is zero. It does not mean delivery is free — it means the cost is fixed and plannable rather than proportional to how much the audience watches.
When volume is low enough to sit inside the cheap end of the card, when traffic is too spiky or too uncertain to size hardware against, or when the audience is in a price group whose tiers actually decline. Below roughly the point where a commit would sit idle, the meter is cheaper than the floor, and the assessment is expected to return that answer where the arithmetic gives it.
Delivered petabytes from CDN logs, catalogue size and growth, peak concurrency, encode hours including re-encodes, and the amortisation period your finance side will accept. With those the owned side resolves to a number that can sit next to the tables above.