The case for renting infrastructure was always about focus. A small team could buy compute by the hour, storage by the gigabyte, and inference by the token, then spend its attention on the product. That trade worked while the rented layer was a commodity and the product was not.

It stops working when the rented layer becomes the product.

Most stacks now draw from the same short list of model APIs, object stores, managed databases, and auth and observability vendors. The architecture diagram is a shopping list, and anyone with a credit card can buy it. The floor is level. The advantage sits in the parts a competitor cannot rent: the box that runs your inference, the weights you can run without permission, the data you accumulated, the domain and mail reputation you built. None of those appear on a vendor's price page.

Same parts, same price

Price convergence is visible in the public tables. When checked on 19 October 2026, Anthropic listed Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens. OpenAI listed GPT-6.1 Sol at the same $2 and $10. Two vendors, one price point, and a market of wrappers selling the difference as a feature.

A stack assembled from those parts is competent on day one and identical to its competitors by lunch. The product can still win on distribution, taste, or service. It cannot win on the infrastructure it rented, because the competitor rented the same infrastructure.

The invoice is not the cost

Managed inference is cheap at the margin and less cheap at volume. The break-even is easy to calculate.

Vantage's AWS instance tracker lists a g6.xlarge, one L4 GPU with 24 GB of memory, starting at $0.8048 per hour on demand, or about $588 for 730 hours. At $10 per million output tokens, a workload that emits 58 million of them in a month pays roughly that much in output tokens alone, before input tokens. The rent buys the whole card for the month, power and cooling included, with operations still yours.

Quality does not match at that price. A 24 GB card runs a smaller model than a frontier API. Plenty of production traffic, from classification to routing and first drafts, does not need the frontier. A team below the line should rent. A team pushing past it, with retries, evals, and batch jobs counted in full, should look hard at the box.

The split matters more than the verdict. Own the part with predictable utilization. Rent the burst. A single server you control can carry the steady load while the provider absorbs the spikes. That puts the common case on a fixed cost and leaves the spike rented.

Own the part with predictable utilization. Rent the burst.

Storage repeats the pattern with a sharper edge. Cloudflare's R2 pricing page, last updated on 1 October 2026, lists standard storage at $0.015 per GB-month and no egress charge. AWS lists $0.09 per GB for data transfer out from S3 in its Ireland region. The exit carries the cost. On S3, a product serving video, model weights, or customer exports pays the fee every month, and the fee grows with success. A server with a flat bandwidth line turns that variable into a fixed cost.

Latency and privacy are paid tiers

A hosted call leaves your network, crosses to a provider region, waits in a queue behind other tenants, and streams tokens back over the same path. A model on your own rack answers over a switch. The size of the difference depends on distance, model, and load. The direction is fixed; only the size varies.

Providers have started charging for the geography. Anthropic's pricing documentation lists a 10% premium for regional endpoints on its cloud partners and a 1.1x multiplier for US-only inference on its own API. OpenAI's pricing page applies a 10% uplift to regional processing endpoints for models released on or after 5 March 2026. Keeping data in a jurisdiction is a product tier, not a default.

Retention follows the same logic. OpenAI's data controls page states that API data has not been used for training since 1 March 2023, but abuse-monitoring logs are retained for up to 30 days by default. Zero Data Retention requires approval and extra controls. A team that needs data to stay on its own hardware has to run the inference; a better terms page cannot substitute for it.

The roadmap is not yours

Every managed service ships on the vendor's schedule. That works while the schedules match, and becomes a release risk when the vendor changes a license, retires a model, or repositions the product.

Redis moved from BSD to dual RSALv2 and SSPLv1 licensing on 20 March 2024, starting with version 7.4. Existing deployments kept working. The next version did not carry the same permissions, and every architecture decision that depended on Redis inherited a new variable. HashiCorp moved Terraform to the Business Source License in August 2023, and OpenTofu was forked within weeks. Neither company broke a contract, and both showed who controlled the roadmap.

Model providers retire products with less notice. Anthropic's pricing page marks Claude Opus 4.1 as retired except on Bedrock and Google Cloud. A product built on a pinned model carries an expiry date it did not set. Owning weights you can serve locally gives you a version that keeps answering after the provider moves on, though quality can still change.

What to own

Ownership is a set of decisions. The test: what still works if a vendor disappears tomorrow?

  • The data comes first. It compounds, and no competitor can buy the years it took to collect.
  • The weights you depend on belong in a form you can run.
  • The mail domain and its reputation stay yours; deliverability is built slowly and lost quickly.
  • The server that carries the steady load stays, even when you rent it by the month.

Rent the burst, along with payments and compliance, where the vendor carries a license and a liability you do not want. Rent the frontier model for work that needs it. Bring home what makes the product yours and rent the rest.

A stack made entirely of rented parts is a demo of someone else's product. It can be assembled in an afternoon, and it can be repriced, deprecated, or terminated by a company that has never heard of you. Sovereignty used to be an ideology. It is now a latency budget, a privacy tier, and a line on the invoice that does not move when someone else changes their terms.

Sources

  • Claude platform pricing (Anthropic, checked 19 October 2026): https://platform.claude.com/docs/en/about-claude/pricing
  • API pricing (OpenAI, checked 19 October 2026): https://developers.openai.com/api/docs/pricing
  • R2 pricing (Cloudflare, last updated 1 October 2026): https://developers.cloudflare.com/r2/pricing/
  • Amazon S3 pricing (AWS, checked 19 October 2026): https://aws.amazon.com/s3/pricing/
  • g6.xlarge pricing and specs (Vantage, checked 19 October 2026): https://instances.vantage.sh/aws/ec2/g6.xlarge
  • Amazon EC2 G6 instances (AWS, checked 19 October 2026): https://aws.amazon.com/ec2/instance-types/g6/
  • Data controls in the OpenAI platform (OpenAI, checked 19 October 2026): https://developers.openai.com/api/docs/guides/your-data
  • Redis adopts dual source-available licensing (Redis, 20 March 2024): https://redis.io/blog/redis-adopts-dual-source-available-licensing/
  • OpenTofu announces fork of Terraform (OpenTofu, August 2023): https://opentofu.org/blog/opentofu-announces-fork-of-terraform/