Your AI stack needs a second supplier

You picked one AI tool, wired it into how you work, and it has been fine for a year. Then something moves that you did not choose. The model name written into a Zapier step stops existing. The price you bookmarked turns out to have had an end date on it the whole time. The API returns errors on the morning a client deliverable is due. None of these is an emergency for the vendor. Each one is an emergency for you, because you have one supplier and nothing behind it.

The companies selling you the models have already bought themselves second and third suppliers, at a scale measured in gigawatts and billions of dollars [1][2]. This guide is about copying the shape of that decision at a size that fits a one-person business. It is not for you if AI is a convenience in your week, something you paste into a few times where nothing downstream depends on the answer. In that case the correct amount of preparation is none. It is for you if a day without your main model means work does not ship.

The companies selling you AI already bought a second supplier

Anthropic trains and runs Claude on AWS Trainium, Google TPUs and NVIDIA GPUs, and in April 2026 it committed to multiple gigawatts of additional TPU capacity that it expects to come online starting in 2027 [2]. Its own stated reason is worth reading twice, because it is the argument of this guide in one sentence: “This diversity of platforms translates to better performance and greater resilience for customers who depend on Claude for critical work” [2].

Then, on 22 July 2026, it widened one of those paths. AMD and Anthropic announced a deal for up to 2 gigawatts of AMD Instinct MI450 Series GPUs, with the first gigawatt beginning to deploy in the first half of 2027, alongside an AMD equity investment in Anthropic of up to $5 billion [1]. The announcement notes that this “builds on Anthropic’s existing use of AMD Instinct MI355X GPUs” [1], and that clause is the part worth stealing. The second supplier was already carrying real work before the large commitment was signed. Tom Brown, Anthropic’s co-founder and chief compute officer, framed the deal as capacity rather than novelty: “By partnering with AMD across the stack, we are securing the capacity we need and optimizing it for training and serving Claude” [1].

Nothing in that deal changes your bill or your output. The first gigawatt does not begin deploying until the first half of 2027 [1]. What is worth copying is the shape of the decision, which is smaller and more boring than the headline numbers suggest. Anthropic did not move off NVIDIA [2]. It ran work on a second vendor’s chips first, then scaled that path up once it knew the path held [1]. Your version is not a second stack. It is one afternoon spent proving that your three most important AI jobs still work somewhere else.

Every model you use has a published retirement date

This is the failure that catches people who thought they had nothing to prepare for. Anthropic’s deprecation policy commits to “at least 60 days’ notice before model retirement for publicly released models”, and states plainly that “requests to models past the retirement date will fail” [3]. The page carries a live table. As of today, claude-opus-4-1-20250805 is listed as retired, with a retirement date of 5 August 2026, and claude-haiku-4-5-20251001 is still active with a tentative retirement date of “not sooner than October 15, 2026” [3].

OpenAI publishes the same thing on different terms: at least 6 months of notice for generally available models, at least 3 months for specialised variants such as Codex and deep research models, and much shorter notice, such as 2 weeks, for preview models [4]. Its shutdown list has dates on it right now. The gpt-5-2025-08-07 and gpt-5-pro-2025-10-06 snapshots are both scheduled to shut down on 11 December 2026 [4].

So spend twenty minutes doing the unglamorous search. Grep your scripts, open your n8n and Zapier and Make steps, and find every place a model name is written down. Dated snapshot names such as gpt-5-2025-08-07 are the ones with a clock on them [4]. Aliases without a date have the opposite problem: no dated entry appears against them, and one week your outputs quietly change character because the alias now points at a newer model. Both are worth knowing about. Neither is worth finding out from a client.

The price you are on has an expiry as well

Prices are not fixed points either, and the expiry is sometimes printed on the page you are reading. Gemini 3.8 Flash is listed at $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026, then $1.50 and $7.50 starting 1 January 2027 [6]. That is a doubling, announced in advance, in a footnote. If your margin on a productised service depends on that rate, your margin has a date on it.

Within a single vendor the spread is wider than most people assume. On Anthropic’s published prices, Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output, Claude Sonnet 5 costs $2 and $10, Claude Opus 5 costs $5 and $25, and Claude Fable 5.1 costs $10 and $50 [5]. Newer is not automatically dearer: Sonnet 5 at $2 and $10 is cheaper than the older Sonnet 4.6 at $3 and $15 [5]. The Batch API takes 50% off both input and output tokens, and for most models a prompt cache hit is charged at 0.1x the base input price [5].

The practical move is to know your own monthly number well enough that a price change is arithmetic rather than a feeling. Until you have that number, a doubling reads as a catastrophe or a shrug depending on your mood rather than on your money.

A second provider is an afternoon, not a rebuild

The plumbing is the easy part, and it is easier than it was two years ago because the vendors started publishing endpoints that speak one another’s dialect. Google offers an OpenAI-compatible endpoint for Gemini: you change three things, the API key, the base URL to https://generativelanguage.googleapis.com/v1beta/openai/, and the model name [7]. If your code already talks to a provider through a standard OpenAI client, that is a configuration change, not a rewrite.

If you do not write code, the equivalent is shorter still. Find the step or node where the model is named, and point it at a key from the other provider. The reason to do this today, while nothing is broken, is that the swap always looks trivial until you try it and find the one workflow that quietly depends on something the other provider does not do.

So run the drill once. Create an account and a key with a second provider, take your three highest-volume jobs, run them through the alternative, and read the outputs properly rather than checking that no error appeared. Write down what changed. Then go back to whatever you were using. You are not switching. You are buying the knowledge that switching is possible and what it would cost you, which is the only thing that makes it a decision rather than a panic.

What locks you in is the prompt, not the plumbing

Once you have done the drill you find the switching cost is somewhere other than where you looked. The API call ports in minutes. What does not port cleanly is everything you built on top of one model’s habits: prompts tuned by trial and error to its particular literalism, tool schemas it happened to follow, the structured output you stopped validating because it was always right, and your own unwritten sense of what a good answer looks like from this model.

Two habits fix most of that, and both cost nothing. First, keep prompts in files you own, not only inside the tool that runs them. A prompt that lives exclusively in a Zapier step is a prompt you will rewrite from memory under pressure. Second, keep a spot-check set: five to ten real inputs with outputs you have already judged as acceptable. When a model retires, a price doubles or a provider goes down, that set is what turns a switch into a thirty-minute evaluation instead of a week of vague unease about whether the new thing is worse.

The compatibility layers reward this too, because they are shallower than they look. Google’s own documentation says support for the OpenAI libraries is “still in beta while we extend feature support”, and its image generation section warns that parameters it does not list “will be silently ignored by the compatibility layer” [7]. Silently is the operative word. You will not get an error. You will get a plausible response produced without the setting you thought you had applied, and only a spot-check set will catch it.

How much redundancy is actually worth paying for

Now the honest sizing, because most guidance on this subject is written for companies with a platform team. Uptime is good and not perfect. On the day this guide was checked, the Claude status page showed 90-day uptime of 99.5% for the Claude API, 99.4% for claude.ai and 99.44% for Claude Code, with incidents logged on 1, 2 and 3 September 2026 [8]. Take the API figure at face value and 0.5% of a 90-day window is roughly 11 hours. Spread across three months that is invisible. Concentrated on a Thursday afternoon it is not.

So sort the work rather than the tools. Put every AI job you run into two piles: the ones that can wait a day, and the ones that cannot without costing you a client, a deadline or a refund. For most solo operators and small teams, nearly everything lands in the first pile, and the correct spend on redundancy for that pile is zero. No second subscription, no running two providers in parallel, no abstraction layer. Just a tested escape route and prompts you can carry.

For the second pile, and it is usually one or two jobs rather than ten, pay a little. That means a live account and key with a second provider, the fallback path already written and tried once, and a note of which model you would move to. Anything more than that, for a business your size, is buying insurance against a risk you have already reduced to a nuisance.

What still goes wrong

The escape route rots quietly. You test a fallback in March, and by August the model you were going to fall back to has passed its published retirement date [3][4], so the route you are relying on points at nothing. This is why the drill belongs on a quarterly reminder rather than in your memory. The same applies to prices: the Gemini 3.8 Flash rate you costed against runs only through 31 December 2026 [6], and a plan built on it needs revisiting before then, not after.

A second provider also does not protect you from correlated failure. If both providers you chose sit behind the same cloud region, or the thing that actually broke is your own connection, your redundancy is decorative. And in wrapper products the lever does not exist at all. When AI is embedded in a tool that never shows you a model name, the vendor chooses the model and changes it when it suits them, and your only redundancy is whatever that vendor arranged with its own suppliers.

Finally, be careful about reading too much into deals like the AMD one. The first gigawatt of MI450 capacity does not begin deploying until the first half of 2027 [1], so none of it touches anything you run this year, and there is nothing in the announcement about what it will mean for prices or availability. What it is is a signal: a company far closer to the supply chain than you are takes concentration risk seriously enough to sign for up to 2 gigawatts of a second vendor’s chips and to keep three other hardware platforms running alongside them [1][2]. That is not a prediction, and it is not a reason to change anything in your stack this week beyond spending the afternoon.