AWS vs Azure vs GCP:
the bill isn't where you think it is.

I run this site on a residential fiber connection in Kathmandu, and every time I've looked at moving a workload from a hyperscaler to a plain VPS, the compute line item was never the reason. Compute is boring now — AWS, Azure, and GCP all sell a 4 vCPU / 16 GB general-purpose instance for somewhere between thirteen and nineteen cents an hour, and none of them will win or lose your business on that number alone. The line item that actually moves the total is the one that's hardest to find on any pricing page: what it costs to send data back out to the internet.

That matters more, not less, from a place like Nepal, where a meaningful share of a small site's or API's traffic is inherently cross-region — there's no local hyperscaler zone to keep egress cheap or free, so every byte served to a reader in the US, Europe, or India crosses a real, billed boundary. But the underlying math isn't a Kathmandu-specific quirk. It applies to any solo developer, small SaaS, or side project serving images, video, API responses, or model output at meaningful volume, anywhere in the world. This piece exists because I kept seeing the same mistake in cost comparisons: everyone matches compute prices carefully and then hand-waves egress, which is backwards — compute is the part that's already been commoditized.

Here's where the three hyperscalers diverge from each other, and from independent VPS providers, sharply. AWS gives every account 100 GB of free data transfer out to the internet per month, pooled across nearly all its services — not per instance. Past that, standard published tiers charge $0.09/GBfor the next 10 TB, stepping down to $0.085/GB, then $0.07/GB, then $0.05/GB at higher volume tiers. Azure's egress pricing mirrors that structure closely, with a comparable small free allowance and a first-tier rate around $0.087/GB. GCP's general internet-egress rate is commonly cited around $0.12/GBafter a token 1 GB/month free allowance — GCP's free tier for egress is dramatically smaller than AWS's or Azure's.

Compare that to how DigitalOcean and Vultr price the same thing. A DigitalOcean droplet at roughly the 4 vCPU / 16 GB tier runs about $126/month and bundles several terabytes of outbound transfer directly into that price, with overage billed at a flat $0.01/GiB— not per hundred megabytes, per gibibyte. Vultr's comparable High Frequency tier runs about $96/monthwith roughly 5 TB included, at a similarly flat, low overage rate. That is a 9-12x gap in the marginal cost of the exact same gigabyte, depending which hyperscaler you're comparing against.

None of this is an argument that hyperscalers are overpriced or that everyone should run on a VPS. It's an argument that the decision should be made on the line item that actually varies by workload — egress — rather than the one that's converged across providers. I covered the broader architecture side of this same shift, where inference and hybrid deployment decisions are reshaping enterprise cloud spend, in Cloud 3.0: The AI Inference Economics Reckoning, and the broader pullback from all-in public-cloud commitments in my piece on cloud repatriation. This post is the unglamorous cost model underneath both of those bigger stories.