Amazon Just Put Another $25 Billion Into Anthropic, And Anthropic Promised $100 Billion Back To AWS: What Is The Reciprocal Lockup, And Why Should Business Owners Care?
The biggest AI deals of 2026 aren't acquisitions.
They're reciprocal contracts.
Amazon just announced a new commitment of up to $25 billion in Anthropic, on top of the $8 billion Amazon has already poured into the company (CNBC).
In return, Anthropic committed to spending more than $100 billion on Amazon Web Services over the next 10 years.
Read that twice.
Amazon gives Anthropic $25 billion.
Anthropic gives Amazon $100 billion.
It's the same money, going in a circle, locking each side into a 10-year dance.
Less than 48 hours earlier, Bloomberg confirmed Google is preparing a similar move, with up to $40 billion of new investment into Anthropic (Mark McNeilly).
OpenAI signed an agreement to spend more than $20 billion on Cerebras chips while Cerebras gets a stake-style backing (Mark McNeilly).
Meta committed to $115 billion in 2026 capex, the bulk going to its own AI infrastructure (Mark McNeilly).
A pattern is forming.
Hyperscalers are locking up AI labs.
AI labs are locking in hyperscaler compute.
And the small businesses using their products are about to find out what that means for them.
What Just Happened With Amazon, Anthropic, And AWS?
The Amazon-Anthropic deal has three moving parts.
Part 1. Amazon agreed to invest up to $25 billion in Anthropic. That's on top of the $8 billion already invested in recent years (CNBC).
Part 2. Anthropic agreed to spend more than $100 billion on Amazon Web Services over the next 10 years.
Part 3. Amazon expects to spend roughly $200 billion in 2026 on capital expenditures, with most of it going toward AI infrastructure (CNBC).
The Cerebras-OpenAI deal is identical in structure.
OpenAI commits more than $20 billion in chip purchases.
Cerebras gives OpenAI an equity stake.
OpenAI helps fund Cerebras' data centers with around $1 billion (Radical Data Science).
Same structure. Different parties.
This is the new shape of AI.
Money flows in circles.
Lockups extend over a decade.
And the result is the consolidation of the AI economy around three or four mega-stacks before the next decade even begins.
What Is The Reciprocal Lockup, And Why Should A Business Owner Care?
Here is a framework.
The Reciprocal Lockup is a deal where the buyer becomes the supplier, the supplier becomes the buyer, and both sides commit to multi-year terms that make switching nearly impossible.
In normal business, when you buy from a vendor, you can leave them.
In the Reciprocal Lockup, your vendor's biggest customer is your vendor's biggest investor. Their biggest investor is their biggest customer.
The only way to leave is to walk away from both at once.
For a business owner, three things change.
Thing 1: Your AI vendor isn't just a vendor anymore.
When Anthropic owes Amazon $100 billion in usage commitments, Anthropic is no longer free to build its products in a vacuum. Every roadmap decision is shaped by what AWS wants. Every cost structure is tied to AWS pricing. The same is true for OpenAI and Microsoft, OpenAI and Cerebras, and increasingly Google and Anthropic.
If you're buying AI from any of these labs, you're buying into a multi-party stack that you didn't sign up for and can't easily exit.
Thing 2: Pricing is now strategic, not market-driven.
In a normal market, AI pricing would be falling fast (because compute is cheap and getting cheaper). In a Reciprocal Lockup market, pricing is set to satisfy the cloud commitments, not the customer.
You'll see weird pricing patterns this year. Tiers that don't map to feature differences. Volume floors. Bundled discounts. Long-term contracts in exchange for "savings."
That's not a glitch. That's the lockup showing up in your bill.
Thing 3: Your data follows the lockup.
When you choose Claude on AWS, your data lives on AWS. When you choose ChatGPT through Microsoft, your data lives on Azure. Migrating off later is technically possible, but the lockups make it commercially uncomfortable for the lab to make migration easy.
Your data inertia becomes part of someone else's lockup.
How Many Cloud-AI Stacks Will There Actually Be By 2027?
Three.
Maybe four.
The shape is already visible.
Stack 1: Microsoft + OpenAI + Cerebras. Compute, frontier models, Workspace agents inside the Microsoft ecosystem.
Stack 2: AWS + Anthropic. Compute, enterprise-grade models, the bulk of Fortune 500 cloud spending.
Stack 3: Google Cloud + Gemini + Anthropic investment + Apple integration. The agentic and consumer story.
Stack 4 (maybe): Meta + Llama + Meta's own infrastructure. Vertical integration, internal data, possible third-party offering.
If you're a small business in 2026, you're going to live inside one of these stacks. The question is whether you choose deliberately or get drifted into one by accident.
The Reciprocal Lockup means the choice is more permanent than it used to be.
What About The Meta Keystroke Story Should I Worry About?
This week, Reuters reported that Meta is installing tracking software on US employees' work computers, capturing mouse movements, clicks, keystrokes, and screenshots, then feeding that data into its AI training pipeline (Mark McNeilly).
The Meta spokesperson framed it as: "If we're building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them."
Translation: AI agents need watch-and-learn data, and Meta is starting with its own employees.
Two implications for business owners.
Implication 1: Your team's data may already be in someone's training pipeline.
If a SaaS tool you use has a quiet "model improvement" toggle that defaults to on, your team's keystrokes, prompts, and outputs may already be feeding a model. Most owners have never checked.
Implication 2: The agents that show up in your business in 2027 will be more capable than you expected.
Models trained on real human computer behavior will run customer service better, fill forms faster, and click through interfaces with less hand-holding. That's the upside. The downside is that those same agents will be very good at outcompeting human employees on the exact tasks the data was harvested from.
Either way, your AI Supply Chain Audit (we covered this Thursday) just got a new column.
What does each vendor train on, and where does my team's data fit in?
How Should A Business Owner Respond To The Reciprocal Lockup This Week?
Three moves.
Move 1: Pick your stack on purpose.
Stop letting individual managers each pick their own AI tool. By the end of 2026, you should be deliberately leaning into one or two of the three stacks, not all three. Pick on capability fit, not on whoever made the most noise this week.
Move 2: Demand portability clauses.
In any new AI contract you sign, ask for two things. First, the right to export all of your data and prompts in standard formats within 30 days of any termination. Second, model and pricing transparency, in writing, before each renewal.
Most vendors will say yes if you ask. Almost no business owner asks.
Move 3: Build at least one workflow on the open-source frontier.
Open-weight models like Gemma 4, Qwen 3.6, and Kimi K2.6 are now competitive with frontier closed models on many enterprise tasks (Mark McNeilly). Even if your primary stack is Microsoft or AWS, build at least one production workflow on an open-weight model. It keeps your skills sharp, your costs down, and your portability intact.
The open-source layer is the only part of the AI economy that isn't subject to a Reciprocal Lockup.
That makes it the most valuable insurance policy a small business can hold this year.
Is There A Smarter Way To Buy AI In A Reciprocal Lockup Market?
Yes.
Three principles.
Principle 1: Buy outcomes, not seats.
Outcome-based pricing (which Google announced at Cloud Next this week) makes lockups easier to escape, because each task has its own discrete cost. If your AI vendor still charges per seat or per token only, push for an outcome-based addendum.
Principle 2: Spread your data, not just your spend.
Avoid putting all of your operational data behind one vendor. Customer data in one stack. Internal documents in another. Your CRM somewhere else. Diversification is your hedge against the day a lockup turns into a price hike.
Principle 3: Run a quarterly portability test.
Once a quarter, export a sample of your AI prompts, outputs, and data from each stack. Confirm you can read it elsewhere. If the export is broken or incomplete, you've found a lockup risk before it bites you in a renewal.
These three principles don't break the Reciprocal Lockup.
They turn it into a manageable risk instead of a hidden cost.
If you'd like a partner running these audits with you, we do a stack review and portability assessment in our complimentary 1-on-1 AI Implementation Session. We'll map your current AI vendors, identify Reciprocal Lockup risk, and build a 90-day plan to keep you optionable while still capturing the upside. Book your session here.
TL;DR
- Amazon committed up to $25 billion more to Anthropic, while Anthropic committed more than $100 billion back to AWS over 10 years (CNBC)
- Google is preparing up to $40 billion of new investment into Anthropic, and OpenAI signed a $20 billion+ Cerebras chip deal with reciprocal funding (Mark McNeilly)
- Meta is harvesting employee keystrokes, mouse moves, and screenshots to train its agent models, signaling where AI training data is heading (Mark McNeilly)
- The AI market is consolidating into 3 or 4 mega-stacks: Microsoft+OpenAI, AWS+Anthropic, Google Cloud+Gemini+Anthropic-investment, and Meta+Llama
- The Reciprocal Lockup framework: when your vendor is your supplier's biggest customer, switching becomes nearly impossible
- Three moves: pick your stack on purpose, demand portability clauses, build one workflow on open-source models
Frequently Asked Questions
What does Amazon's $25 billion investment in Anthropic actually mean for my business?
It means the AI tool you use through AWS Bedrock or directly via Anthropic now has runway to ship for the next decade. The downside is your contract terms and pricing are increasingly shaped by AWS commercial priorities rather than open-market competition. Watch for changes in your renewal pricing.
Should I avoid AWS-based AI tools because of the lockup?
No. AWS plus Anthropic is one of the strongest enterprise AI stacks available right now. The point isn't to avoid lockups. The point is to enter them deliberately, with portability clauses in your contract.
What is open-weight AI, and is it really competitive with Claude or GPT?
Open-weight models like Gemma 4, Qwen 3.6, and Kimi K2.6 are AI models whose weights are published publicly, meaning you can run them on your own infrastructure or any cloud you choose. On many enterprise tasks, they now match closed frontier models, which makes them a strong portability hedge.
How worried should I be about the Meta keystroke surveillance story?
For your customers, low. For your employees, moderate. The bigger lesson is that AI training data is moving from "scraped from the internet" to "captured from real human work." Verify what your SaaS tools are training on and toggle off any data sharing you can't justify.
Is there a way to run AI in 2026 without a Reciprocal Lockup at all?
Yes, but it takes more engineering work. You can self-host open-weight models, route through neutral inference providers like Together AI or Fireworks, and own your own data layer. Most small businesses won't go fully self-hosted, but having one production workflow there is good insurance.
Ready To Audit Your AI Stack Before The Lockup Tightens?
The Reciprocal Lockup is happening with or without your input. The only choice is whether you live inside it deliberately or by accident.
Book a complimentary 1-on-1 AI Implementation Session with our team. We'll map your current AI stack, flag the highest lockup risks, and build a 90-day plan to keep you optionable while still capturing the upside of the new agentic era.
