The opening chapter of this part established a 4.7× spread on H100-hours. The chapter on egress established that the transfer line can consume two weeks of a GPU per month. Put them together and you get the decision most teams at this layer actually face: stay on a hyperscaler, or move the accelerator workload to a specialist.
This chapter prices both sides honestly, including the parts the specialists do not put on their pricing page.
The full-stack comparison
One H100, running continuously for a month (730 hours), with 10 TB of internet egress. All rates published, 2026-08-05:
| Provider | Compute | Egress | Monthly total | vs cheapest |
|---|---|---|---|---|
| RunPod (Community) | $2,183 | $0 | $2,183 | 1.0× |
| Lambda (on-demand) | $2,913 | $0 | $2,913 | 1.3× |
| AWS (on-demand) | $5,022 | $913 | $5,935 | 2.7× |
| GCP (on-demand) | $8,074 | $1,137 | $9,210 | 4.2× |
| Azure (on-demand) | $8,972 | $882 | $9,854 | 4.5× |
A 4.5× spread on the same chip doing the same work.
Note what egress does to the shape. On compute alone AWS is 2.3× RunPod. Add transfer and it becomes 2.7× — because the neoclouds charge nothing for the line that the hyperscalers meter. Egress does not soften the neocloud case; it widens it.
That is the number that makes this a real decision rather than a preference. Now the other side.
What the cheap price is not buying you
A 4.5× gap that survives scrutiny is worth acting on. A 4.5× gap you have not scrutinised is a trap. Here is the itemised list, with prices attached where prices exist.
Capacity tiering — read the tier name. RunPod's $2.99 is the Community rate: capacity rented from third-party hosts, priced accordingly. Its Secure Cloud tier costs more. Comparing a hyperscaler's dedicated on-demand rate against a neocloud's community tier is not a like-for-like comparison, and it is the single most common way these tables get abused — including by vendors quoting them. If you are running production inference, price the secure tier and rebuild the table.
Node failure is a support ticket, not an API call. On AWS a dead instance is replaced programmatically in seconds. On a smaller provider it can be a ticket and a wait. For checkpointed training that is an annoyance costing you the time since the last checkpoint. For production inference serving customers it is an availability incident, and the cost is whatever your SLA says it is.
There is no account team. No committed-spend negotiation, no credits, no architecture review, no escalation path that ends in a named person. For a ten-person company this is mostly fine. At $500k/year of spend it is not, and the hyperscaler discount you can negotiate at that volume erodes some of the gap in the table above.
The compliance list is short. Specific certifications, data residency guarantees, a signed BAA, FedRAMP — these narrow the eligible vendor list faster than any price consideration, and they are the honest reason many teams pay 4.5×. If you are in that position the table is not a decision, it is background.
Everything else you use is on the hyperscaler. Your data, VPC, IAM, secrets, observability, and CI live somewhere. Moving the GPU out means the GPU now reaches back across the public internet to all of it — which reintroduces the egress you just saved, plus latency, plus a security surface. The clean case for a neocloud is a self-contained workload; the messy case is a GPU that needs constant conversation with the rest of your stack.
Managed services do not come with it. No Bedrock, no Vertex, no managed vector store, no IAM-integrated object storage. You are buying a machine, not a platform. Everything above bare metal is now yours to run, which is the subject of the total-cost-of-self-hosting chapter and is not a small line.
The honest framing
The neocloud discount is real, large, and specific: it is a discount on renting a GPU, and it is largest for workloads that are self-contained, tolerant of interruption, and egress-heavy.
It is not a discount on running a company's infrastructure. Teams that migrate everything to chase a 4.5× number and then rebuild identity, networking, storage, and on-call rotations discover they traded a variable cloud bill for a fixed engineering bill, and engineering is the more expensive of the two.
The pattern that actually works for most small teams is split, not switch: control plane, data, and the services that need each other stay where they are; the accelerator workload — batch training, offline inference, embedding jobs, anything that takes an input and returns an output without needing the rest of your stack mid-flight — goes to the cheap machine. That captures most of the 4.5× and touches almost nothing.
Where the numbers get uncomfortable for the neoclouds
Two rows from the earlier ladder are worth restating here, because they cut against the specialists:
AWS spot is $2.62/GPU-hour — cheaper than RunPod's Community rate. If your workload already tolerates interruption, the cheapest H100 on the board is on the hyperscaler you are probably already using, with no migration, no new vendor review, and all your existing services one VPC away. "Leave the hyperscaler to save money" is not automatically correct, and for interruption-tolerant training it is often wrong.
AWS three-year reserved ($2.97) and RunPod Community ($2.99) are the same number. One of those locks three years of your budget to a specific accelerator generation. The other is a credit card. That comparison is not close on any axis except price.
Together's dedicated H100 is $5.49/GPU-hour — a specialist charging within 20% of AWS on-demand. "Specialist" and "cheap" are not synonyms, which is the same lesson the managed endpoints chapter reached from a different direction.
The diagnostic
- Is your accelerator workload self-contained? If it needs your database, your auth, and three internal services mid-run, the egress and latency you reintroduce will eat the discount. If it takes a file and returns a file, it is a clean candidate.
- Which tier did you price? Community versus secure, spot versus on-demand. Compare like for like or the exercise is theatre.
- What does an hour of unavailability cost you? Multiply by realistic failure frequency. If the number is large, you are buying reliability, and you should say so out loud rather than discovering it during an incident.
- Have you priced AWS spot? It is cheaper than the neoclouds and you are already there.
- What is your real monthly egress? It is a bigger lever than the hourly rate for anything media-heavy, and it is the line that most favours moving.
- Would splitting capture most of it? Usually yes. Migration is rarely the right unit of change.
What this chapter is not saying
It is not saying neoclouds are risky and hyperscalers are safe. A 4.5× price difference is not a rounding error you should pay for peace of mind, and plenty of serious workloads run happily on specialist capacity.
It is saying the comparison is total cost of the workload, not the price of the machine — and that the honest version of this comparison has line items on both sides. Most published versions have line items on one.