Subtitle: Anthropic's IPO prospectus leaked this week and put real numbers on the AI economy, and this article answers the question those numbers actually raise for a small business owner: what happens to your work if the plan, tier, or usage limit you built on changes shape with thirty days notice?
Five hundred eighteen billion dollars.
That is the figure for future cloud, computing, and infrastructure obligations disclosed in Anthropic's IPO prospectus, reported by CNBC on September 28, 2026. It is a number with no useful comparison in small business life. It is not a mortgage. It is not a truck payment. It is a commitment larger than the annual economy of most countries.
And it has almost nothing to do with you, except for one thing: numbers that size mean the pricing, tiers, and usage limits on every AI product you use are still being figured out in public.
So here is the direct answer to the question in the headline. You protect your business by knowing exactly which of your workflows stops working if your provider changes its limits tomorrow, and by having a second provider already configured and tested on the one that matters most. Not researched. Not bookmarked. Configured, with your prompts loaded and one real job run through it.
That is the whole thing. Everything after this is how to actually do it.
Because here is the thing most people get wrong about this story. They read the headlines and start worrying that their AI company is going to collapse. That is the wrong worry. Anthropic is growing fast and the prospectus shows it. The real risk to your business is smaller, more boring, and far more likely: the specific plan you built your Tuesday morning on gets restructured, and you find out by email.
Down here in Alabama, everybody I know owns a generator they hope never to use. Nobody buys one during the storm.
Key Takeaways
- The risk to your business is not that your AI provider fails; it is that the specific plan, tier, or usage limit your workflow depends on changes shape with short notice.
- Anthropic's own documentation commits to at least 60 days notice before retiring a publicly released model, and its published deprecation history shows that roughly 60 days is what you actually get.
- A Zapier survey of 542 U.S. executives found 89 percent believed they could switch AI vendors within four weeks, but among the 66 percent who had actually attempted a migration, 58 percent said it failed or took significantly more effort than expected.
- Portability is cheap to build in advance and expensive to retrofit, which is why you configure the backup before you need it, not during the outage.
- You do not need a second subscription for everything. You need one for the two or three workflows that would genuinely hurt to lose.
The Problem: You Probably Do Not Know What Stops
Most small business owners I talk to cannot answer a simple question: if your main AI tool changed its usage limits next month, what specifically would break?
Not "I'd be annoyed." What would stop.
That gap is the actual problem. AI crept into small businesses sideways. You started using a chatbot for email drafts, then it became the thing that writes your proposals, then your VA built a process around it, and somewhere in there a subscription that felt like a toy became load-bearing.
Load-bearing without documentation is the definition of fragile.
And the people who should know better do not know either. Zapier surveyed 542 U.S. C-level executives with active paid AI contracts in early 2026 and found that only 6 percent said they could stop using their primary AI vendor without interruption. Those are companies with procurement departments. You probably do not have one of those.
Candidly, I have a personal reason this subject gets my attention. I have lost a business before. Not to a dramatic market collapse, but to a set of dependencies I had not mapped and could not replace quickly enough when they shifted. The lesson stuck. Dependency is not the problem. Unexamined dependency is.
Now let me acknowledge the honest difficulty here, because a lot of advice on this topic pretends it is easy.
Switching AI providers is genuinely annoying. Your prompts are tuned to one model's quirks. Your integrations point at one API. Your team learned one interface. Your outputs have a voice you spent months getting right. Anyone who tells you that you can just paste your prompts into a different tool and get identical results has not done it.
That difficulty is exactly why the work has to happen before the deadline, not after. The reframe is simple: you are not trying to be provider-neutral. You are trying to be provider-replaceable on the handful of workflows that actually carry weight.
The Evidence: Why the Ground Is Still Moving
Let me lay out what is actually verifiable, because the narrative in most coverage this week was incomplete.
One. The leaked prospectus shows growth, not distress, and the loss number is widely misread. Anthropic's 2025 revenue came in near $4.6 billion, up roughly twelvefold, according to reporting from TechCrunch and Fortune on September 28 and 29, 2026. The headline net loss of $42 billion sounds catastrophic until you read the composition: roughly $34 billion of it is a non-cash accounting charge reflecting an increase in the estimated value of convertible financing. That is not money spent running the company. The operating picture is different again. Reporting indicates Q2 2026 revenue of $11.5 billion and a second consecutive quarter of adjusted operating profit. I want to be fair about that, because fear-mongering about AI companies collapsing is lazy and wrong.
Two. The real number to watch is the cost side. Per the same prospectus reporting, compute and infrastructure spend was $7.33 billion in 2025, roughly triple the prior year, against that $518 billion in committed future obligations. That is the fact that touches your invoice. When a provider's cost base scales that fast, plan structures get revisited. Not out of malice. Out of arithmetic.
Three. The notice period you get is short, and it is published. Anthropic's own model deprecation documentation commits to at least 60 days notice before retiring a publicly released model. Look at what that means in practice on their published history: Claude Sonnet 4 and Claude Opus 4 were announced for retirement on April 14, 2026 and retired June 15, 2026. Claude Opus 4.1 was announced June 5, 2026 and retired August 5, 2026. Roughly two months, both times. That is a clearly documented, honestly disclosed policy, and it is still two months.
Four. Everybody thinks switching is easy until they try it. In that same Zapier survey, conducted with Centiment between January 30 and February 6, 2026, 89 percent of executives said they could migrate to a new AI vendor within four weeks. Two-thirds had actually attempted a migration. Among those who tried, 58 percent reported the process either failed outright or required significantly more effort than expected. That is the single most useful statistic in this entire article. Confidence and capability are not the same thing.
Five. This is not a new problem, and the research predates the AI boom. A study published in the Journal of Cloud Computing in 2016 surveyed 114 participants across large corporations and small and medium enterprises in the UK on vendor lock-in. It found that 76.6 percent of respondents were unaware of existing standards to mitigate lock-in risk, and that 35.1 percent named loss of control and over-dependence on a single provider as a core barrier to further cloud adoption. Ten years ago, with different technology, the same shape.
And just to show how unsettled the pricing conversation is right now, look at one week in September 2026. Cognition cut Devin pricing by 30 to 40 percent. Google restructured its consumer AI tiers. Anthropic held Claude Sonnet 5.5 at Sonnet 5 pricing, $2 per million input tokens and $10 per million output tokens, while making the model roughly 30 percent faster.
Three providers. One week. Three completely different directions. That is not a market that has settled.
The Solution: The Stop List and the Second Option
Here is what changed how I run this at White Beard Strategies.
I stopped thinking about AI tools as subscriptions and started thinking about them as suppliers. That one word shift fixed my thinking. Nobody sane runs a business with a single supplier for a critical input and no phone number for anybody else.
The system has two parts, and it is deliberately small so you will actually finish it.
Part one is the Stop List. Write down every workflow where AI does real work in your business. Not "I sometimes ask it questions." Real work with a deadline attached. For most small businesses this is four to twelve items: client proposals, weekly content, customer service replies, meeting summaries, invoice descriptions.
Then mark each one with what happens if it stops on Monday. Three categories only: Revenue Stops, Quality Drops, Mild Inconvenience. Be honest. Most items land in the third bucket, and that is fine. The point is to find the one or two that do not.
Part two is the Second Option. For every Revenue Stops item, and only those, you build a working alternate path on a different provider. Working means you ran a real job through it, start to finish, and kept the output.
This is the part people skip. They read a comparison article, decide Gemini or ChatGPT or Claude would be their fallback, and consider the job done. That is not a fallback. That is a preference. A fallback is an account that exists, with your prompts already in it, that you have personally used to produce one finished deliverable.
The reason this works is that it forces you to discover the friction while the stakes are zero. Your prompt will need rewording. Your formatting will break. One model will handle your industry jargon better than the other. Finding all of that on a random Thursday costs you two hours. Finding it the week your workflow stops costs you a client.
There is a third piece that costs nothing and matters more than either: get your prompts out of the tool. Your prompt library, your brand voice document, your standard operating procedures, your reference materials. Those are your assets, not the vendor's. Keep them in plain text files somewhere you own, whether that is Google Drive, Notion, or a folder on your laptop. A prompt saved only inside one platform's custom instructions is a prompt you are renting.
Here is a prompt to run your own audit. Use it with whatever tool you already pay for.
[The Job]
Help me map my business's dependency on AI tools and identify where I am exposed if a provider changes its pricing, plan structure, or usage limits.
[The Background]
I will describe the AI tools I currently pay for and the specific tasks I use them for. My business is small and I do not have a technical team. I care about which workflows would actually stop producing revenue, not about theoretical risk. I want to spend the minimum amount of effort for the maximum reduction in exposure.
[The Deliverable]
A table of my AI-dependent workflows, each rated Revenue Stops, Quality Drops, or Mild Inconvenience, followed by a short recommendation for each Revenue Stops item covering which alternate provider to configure and which specific assets I should move out of the vendor's platform and into storage I control. Written from the perspective of an operations consultant speaking to a non-technical small business owner.
[The Questions]
Ask me any questions you have.
Practical Steps: Do This in One Afternoon
-
Build the Stop List first, on paper, before you open any tool. Give yourself twenty minutes and write down every place AI does work with a deadline attached. Paper matters here because it keeps you from wandering into the software and getting distracted by features.
-
Rate each item honestly using the three categories. Revenue Stops, Quality Drops, Mild Inconvenience. If everything you wrote down lands in Revenue Stops, you are being dramatic. If nothing does, you are not looking hard enough. Most owners find one or two.
-
Export your prompt library to plain text today. Copy every prompt, custom instruction, and brand voice document into a folder you control. Name the files clearly. This takes under an hour and is the single highest-return item on this list because it is the piece that transfers to any provider.
-
Open one free account on a second provider and run a real job through it. Not a test question. An actual deliverable you needed anyway. Keep the output next to the original so you know exactly how much quality difference you are dealing with and where the prompt needs adjusting.
-
Write down the adjustments in a one-page switching note. What you changed in the prompt, what broke, what you had to do differently. One page. This document is what turns a four-week panic migration into a two-hour Tuesday.
-
Put a calendar reminder on the first business day of each quarter. Fifteen minutes to check whether your providers have announced plan changes, and to re-run one job through the backup so the account stays warm. Four times a year, an hour total.
-
Check the deprecation page for any model you depend on by name. Anthropic, OpenAI, and Google all publish model lifecycle documentation. If a workflow of yours is pinned to a specific model version, that page tells you your actual runway.
Frequently Asked Questions
How much notice do AI companies actually give before changing a model or plan?
It varies by provider and by what is changing. Anthropic's published documentation commits to at least 60 days notice before retiring a publicly released model, and its deprecation history shows announcements landing about two months ahead of retirement. Pricing and plan structure changes are a separate matter from model retirement and typically move faster. Check each provider's own policy page rather than assuming.
Should I pay for two AI subscriptions just to have a backup?
Usually not. Most second providers offer a free tier that is sufficient to keep an account configured, prompts loaded, and one workflow tested. Pay for a second subscription only if a single workflow genuinely stops your revenue and you cannot tolerate even a day of degraded output. For everyone else, a warm free account is enough insurance.
Is my usage limit or my price the bigger risk?
For most heavy users, the limit. Headline per-token pricing has been stable or falling, but the plan-level constraints, including usage windows and message caps on consumer tiers, are where workflows actually break. Community reporting consistently points to usage windows rather than price as the binding constraint. Track what your plan allows, not just what it costs.
Will my prompts work the same on a different AI provider?
Partially. Structure, context, and instructions transfer well. Tone, formatting habits, and length calibration usually need adjustment, and model-specific tricks often do not carry over at all. Expect to spend fifteen to thirty minutes tuning each important prompt on a new provider. That is exactly why you do it before you need it.
Does this mean AI companies are financially unstable?
No, and it would be a mistake to read it that way. Anthropic's leaked prospectus showed 2025 revenue near $4.6 billion, up roughly twelvefold, plus a second consecutive quarter of adjusted operating profit per reporting on the filing. The instability is in pricing and packaging models across the whole industry, not in the survival of the major providers.
The Close
Five hundred eighteen billion dollars in committed infrastructure obligations is not a number you can do anything about.
Knowing which two workflows in your business would stop on Monday is a number you can do something about, and it takes an afternoon.
That is the whole trade. You cannot control what any provider decides about tiers, limits, or packaging. You have zero leverage there and you always will. What you control is whether their decision arrives as an inconvenience or as a crisis, and that is determined entirely by work you either did or did not do beforehand.
So let me be as plain as I know how to be.
If you are reading this and you cannot name the workflow that stops, you do not have an AI strategy. You have an AI habit. There is no shame in that, most people are in exactly the same place, and the fix is an afternoon of unglamorous work with a legal pad. But the fix does not happen on its own, and it does not happen after the email arrives.
Build the Stop List. Export the prompts. Warm up one backup account. Then go back to running your business and forget about it until the calendar reminder pops in January.
Nobody buys the generator during the storm.
About the Author
Jonathan Mast is the founder of White Beard Strategies, where he teaches non-technical entrepreneurs how to actually use AI in their businesses. He runs a Facebook community of more than 500,000 members and the AI Insiders membership, and speaks regularly on practical AI adoption for small business. He has built a business, lost a business, and built another one, which is why he writes about dependency and risk more often than most people who write about AI.
Ready to build the systems instead of guessing at them? AI Insiders is White Beard Strategies' membership for entrepreneurs who want the frameworks, the prompt libraries, and the live support to implement them, at $247 per month. If you would rather start smaller, our Express AI Training sessions are $10 live or $20 for the replay.
