---
title: "What Can't Be Bundled"
author: "Simon Gu"
date: 2026-11-13
category: "Product & Tech"
excerpt: "Anything that works the same at every company will end up in somebody's $20 plan."
summary: "A strategy guide for founders and incumbents on which AI products the labs will fold into their $20 bundles, using Amazon Basics, Microsoft Teams and the IBM PC as precedents, and which layers stay defensible, such as entitlements, systems of record, licenses, distribution and a company's own decision data."
url: https://gufamily.org/writing/what-cant-be-bundled
---

# What Can't Be Bundled

There's something very interesting happening right now in the land of tech. With AI, new categories with very large addressable markets are being created overnight, yet a lot of the blue oceans have become much bloodier in the past 6 months. If you are a VC-backed startup, unless you quickly reach profitability or get to $100M ARR in roughly 2–3 years' time, you end up stuck between rounds fighting for survival.

![Line chart of quarters it took software companies to grow from $1M to $100M in ARR. Legora, Sierra, ElevenLabs, and CodeRabbit got there in roughly 6 to 9 quarters, Glean in about 12, top-quartile AI-forward "pacesetters" in about 14, and other software companies in about 22.](https://gufamily.org/writing/what-cant-be-bundled-arr.jpg "Quarters from $1M to $100M ARR. Chart by CodeRabbit, using data from ICONIQ's State of Scaling 2026.")

Traditional exits are becoming less and less common, and I have heard stories within the founder network where acquisition offers appear to turn up at exactly the right time. What would normally be a win for everyone involved becomes a dragged-out diligence process...just to find out that the buyer gets cold feet and, a few months later, launches a competitive product built by agent swarms fed with context from the diligence process.

Shocking to hear at first, but actually believable if you look at the broader ecosystem.

## The AmazonBasics Strategy

Earlier this week, there have been a lot of memes flying around about the death of startups after Anthropic [launched a new set of features](https://claude.com/resources/articles/more-ways-to-work-with-claude-sheets-whiteboards-and-interactive-connectors) that directly competed against Figma, Miro, and Google Sheets.

I also saw a LinkedIn post where someone called the $20 Claude bundle plan, which includes Sheets, Docs, Slides, Meetings, Data, Design, and Whiteboards, the AmazonBasics strategy. I love this super clear framing since it puts into perspective what's happening across the business spectrum.

If you haven't bought a phone charger in a while, AmazonBasics is Amazon's house brand. It started with cables and batteries and grew into just about every category where third-party sellers had already proven people were buying. Though it was typically a low-margin business, Amazon made this work quite well because no one had to guess where the demand was, since Amazon had the data behind every product that got sold on the platform.

In 2020, the *Wall Street Journal* [reported](https://www.wsj.com/articles/amazon-scooped-up-data-from-its-own-sellers-to-launch-competing-products-11587650015) that Amazon employees used that third-party seller data to decide which of its own products to make, which Amazon disputed. Copying product labels that sold in huge volumes with decent gross margins while promoting the Basics brand in search results printed money for Amazon.

The labs are now in the same position. Every connector you turn on and every file you export tells the lab which tools people use and how often, and a lab with that view doesn't need a market research team to know what to build next.

The market is directly measurable. Which raises a few questions: how defensible are you as a company if you enable Claude or ChatGPT to connect directly to your product? Are the distribution gains from integrating with chat going to be worth the potential competition and margin compression in the future?

Most of my mental calories have been spent on the tech side of things over the last few months. It's exciting to switch gears and think through this old pattern, and why the labs more or less have to follow it if they want to become good businesses. I'll also share some thoughts on what I think stays defensible.

## The Platform Always Sees Demand First

Platforms that see demand eventually compete with the customers who showed it to them. Amazon wasn't the first company to use this strategy. Microsoft bundled Teams into Office at an apparent price of zero, which was a brilliant move. It stopped the upstart Slack dead in its tracks when it came to selling deep into enterprises even though Slack clearly offers a much better user experience and wins on a feature level.

By bundling, Microsoft switched the decision-maker from individual users to the CFO, who loved the single line item, which made it awfully hard for Slack to win on price with the people signing the Office renewal.

A separate example comes back to Amazon and is arguably a lot more cutthroat. When AWS offered a managed version of Elasticsearch, Elastic's management team made the decision to change its license to stop AWS from doing so. In response, AWS straight up forked the code and kept going under the name OpenSearch. Lol.

Now fast forward to the last two years. The labs see demand through both usage numbers and exact use cases via the cached prompts, which is a sharper view than any SaaS company ever had. By June, [according to PYMNTS](https://www.pymnts.com/news/artificial-intelligence/2026/anthropic-vertical-software-rollout-threatens-enterprise-ai-builders/), Anthropic had launched 13 industry-specific tools, starting with Claude Code in February 2025, and it no longer always warns customers before shipping a product that competes with theirs.

The immediate perceived revenue impact against the copied companies is very real. For example, when Anthropic released 10 financial services agents on May 5, FactSet's stock [fell as much as 8%](https://www.bloomberg.com/news/articles/2026-05-05/anthropic-unveils-ai-agents-to-field-financial-services-tasks) during the day.

I use Claude Code nearly every day and have praised it more than once on this blog, so I don't read any of this as villainy. I believe that competition is good, and any platform with this view of demand would do the same thing.

However, I think it is worth calling out the elephant in the room, which is when the company competing with you also supplies the model your product runs on and sees how your customers use it. Brookings has a good read on [what happens when AI companies compete with their customers](https://www.brookings.edu/articles/what-happens-when-ai-companies-compete-with-their-customers) that walks through that conflict of interest from the policy side.

The tech-overlord-stealing-your-lunch analogy does break down (thankfully?) in one important place, though. An Amazon seller had nowhere else to find the traffic and distribution the platform provided, so leaving Amazon was rarely an option.

An AI app, on the other hand, can often switch to a different model with a one-line config change, which helps builders. I'll get into this more in the open-weights section, but a consequence of this jailbreak is that it also pushes the labs even harder toward the bundle.

## Why the Labs Have to Bundle

Selling tokens is a strong business only while the lab stays ahead. Anthropic's [S-1](https://www.sec.gov/Archives/edgar/data/0002130357/000162828026068676/anthropicpbc-sx1ax1.htm) implies about an [85% gross margin](https://fourweekmba.com/ai-anthropic-s-1-85-5-gross-margin-up-from-29-4-a-year-ago/) for the first 9 months of 2026, up from about 29% a year earlier, which means that for every dollar a customer pays, roughly 85 cents is left after the cost of serving it.

However, Anthropic doesn't officially report a gross margin itself, and the compute spent training the next model is booked under research and development instead of cost of revenue. The filing also doesn't split training cost from serving cost, so there's no clean way to back into a margin that includes training.

But we can definitely infer that margins after training and serving are a lot lower than 85%, and the cherry on top is that training isn't a one-time cost. It's what a lab has to keep spending to stay ahead of everyone else.

As we can imagine, problems start to compound for the labs when their lead is short. Epoch AI [estimates](https://epoch.ai/data-insights/open-closed-eci-gap) that open-weight models have trailed the frontier by about 4 months on its capabilities index since January, and they're available at a fraction of the price.

Curious to see how public investors will react when the company ends up spending billions for only a short-term 4-month oligopoly. This quickly becomes a hard business to defend on tokens alone.

The next greenfield for these labs is in the application layer. The value moves up into defaults, the workspace, memory, connectors, and the agent that decides which tool gets used for a task. Put all of that on top of the model, sell it as one subscription like the $20 Claude plan, and you get what I'll call the bundle for the rest of this post.

This is a lucrative layer to own as a hedge because the value built does not wear off when the model underneath becomes a commodity. If anything, they get stickier the longer someone uses them, since a workspace full of your docs and a memory full of your preferences provides deeper lockin and doesn't move to another lab with a config change.

In *The Innovator's Solution*, Clayton Christensen and Michael Raynor call this the law of conservation of attractive profits. When one layer of a product commoditizes, the profit doesn't disappear, it moves to the layer next to it.

One of Christensen's go-to examples was the IBM PC. In the early 1980s, IBM built its PC from off-the-shelf parts and outsourced the processor to Intel and the operating system to Microsoft.

As manufacturing capabilities caught up, assembling PCs quickly became something anyone could do. IBM's profits got quickly competed away. However, the profit margins of the PC business moved one layer over, into Intel's chips and Microsoft's Windows, the two proprietary pieces inside every box. Read more about it in [Skate to Where the Money Will Be](https://hbr.org/2001/11/skate-to-where-the-money-will-be).

## Open Weights Presents Sophie's Choice

Open-weight models are the closest thing AI has to open-source software. With open source, like Linux or Postgres, anyone can download the code, run it on their own servers, and change it to fit their needs.

With an open-weight model, the lab behind it publishes the weights, the billions of numbers the model learned during training, so anyone can download and run the model on their own hardware and fine-tune it for their own use. It isn't fully open source, since most labs still keep their training data and code to themselves, but the model is yours to run. Meta's Llama, Google's Gemma 4, DeepSeek, Alibaba's Qwen, and Nvidia's Nemotron are some of the better-known examples.

Open-weight models operate in a paradox. They give app builders a way out of the bundle, since a builder no longer has to rent its model from the same lab that's shipping apps to compete with it, while also pushing those very labs further into bundling as much as they can. Since the faster open models catch up, the less a frontier token is worth on its own, and the harder a lab has to lean on everything wrapped around the token.

For builders, the way out looks something like this:

1. Route by task, using a decision head like Jev to send queries to open models for high-volume work and saving frontier models for the hard tail.
2. Build and keep your own telemetry. Traffic that runs on your own models tells no one else what to copy.
3. Deploy where the bundle can't go, like on-prem, private cloud, and regulated or air-gapped settings.
4. Price on outcomes instead of seats, since a seat is exactly what the $20 plan undercuts.
5. Be the opposite of a mile wide, but an inch deep. Embed in vertical workflows, compliance, and observability.

I believe all 5 are worth doing, because tons of super smart founders are doing these exact things today. However, finding success and leaving the gravitational pull of Dario and Sam is hard.

The best open-weight models still trail the frontier models from labs like Anthropic, OpenAI, and Google. Epoch's 4-month gap is measured on public benchmarks, but on private benchmarks, which are much harder to train toward, [one analysis on LessWrong](https://www.lesswrong.com/posts/rJcCrXyEsJKmmDpWG/how-far-behind-are-open-models) finds open-weight models are 8–10 months behind. So for the hardest and most valuable tasks, most builders still end up paying a frontier lab.

The bigger problem for startups selling software to businesses is distribution. Take a startup selling an AI whiteboard to product teams. It could run every request on open weights, but the people it wants to sell to may already have Whiteboards inside the Claude plan their company pays for. Swapping the model under the product takes a config change, while convincing a team to buy a separate tool, and leave a workspace they already live in, is much harder.

## The Central Bank Gets a New Governor

There's a second force pushing in the same direction, and it showed up this week. Jensen Huang [resigned as Nvidia's CEO](https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000082/nvidiaannouncesceotransiti.htm), and the company that has been described as AI's "central bank" now has an interim CEO, Colette Kress, who was its CFO.

A central bank keeps money moving through an economy by lending, setting the terms of credit, and stepping in as a backstop when others won't. Nvidia has played a similar role in AI.

It used the cash from selling GPUs to invest in, lend to, and guarantee the companies (many neoclouds) that buy its GPUs, which kept the capital for the AI buildout flowing even when traditional lenders were nervous. That also means a big part of the industry's funding depends on how one company decides to spend its balance sheet.

Some context for anyone outside the infrastructure world. Neoclouds like CoreWeave and Nebius buy GPUs and rent them out by the hour. Nvidia was often their supplier, their investor, their anchor customer, and the guarantor on their leases, all at once. [Thomas Tunguz counts](https://x.com/ttunguz/status/2120931491115135287) that over the past 2 years Nvidia committed $42B to M&A, put $61B into equity stakes across 229 startups and 13 public companies, and extended $164.5B in lease guarantees and capacity backstops.

It's too early to know whether new leadership pulls back on any of that. Many think that a CFO is traditionally hawkish. But if new commitments tighten, the squeeze lands first on the neoclouds, on Nvidia-funded startups, and on open-weight models, where Nvidia has committed $26B over 5 years to building its own. The hyperscalers and the labs with their own supply or access to Google's TPUs mostly don't feel it.

You might expect traditional banks to fill the gap. [THE DECODER](https://the-decoder.com/building-ai-data-centers-is-becoming-a-stress-test-for-banks/) makes the case that they won't, at least not at the risky end. Banks lend senior and secured, meaning they get paid back first and hold the contracted leases as collateral, and several are already hitting their concentration limits on data center lending. They aren't going to take equity-like risk on a speculative GPU cluster.

For the purposes of this piece, the net effect is that the shift favors integrated players, the ones who own the chips, the cloud, the model, and the workspace. That strengthens the bundle.

## What Can't Be Bundled

So what's left for everyone else? I believe a lab will eventually bundle any product it can rebuild without your help, much like the predatory M&A sharks testing business moats.

In [Irreducible Context](https://gufamily.org/writing/irreducible-context.md), I argued that the only part of a document a person really adds is what a model couldn't have predicted from the prompt. I believe the same is true for a company.

If your product works the same for every customer and doesn't depend on something only your company has, like customer permissions, official records, or a regulatory license, there's nothing in it a lab couldn't predict, and a lab can copy it. Those things only you have are your company's irreducible context.

| Layer | Bundled or defensible | Why |
| --- | --- | --- |
| Generic productivity (decks, docs, notes, summaries) | Bundled | Predictable from a prompt and the same everywhere |
| Thin vertical wrappers | Bundled | The value comes mostly from the model |
| Systems of entry (CRM, notes, tickets) | Bundled | AI makes the data easy to move, so switching costs disappear |
| Entitlements and permissioned actions | Defensible | A lab doesn't know your org, so it can't decide who's trusted to do what |
| Systems of record where the record is the asset (ledgers, custody, claims) | Defensible | Only the original can move money or pay a claim |
| Licenses, liability, and capital | Defensible | A lab can't hold your banking license or absorb the loss when something goes wrong |
| Distribution and logged-in users | Defensible | The bundle has to win them one by one |
| Decision data you own | Defensible | Your labels and thresholds aren't in anyone's training set |

The top three rows are what a lab can learn by watching usage and then ship. The rest are things a lab can see but doesn't hold. A model can learn what a wire transfer looks like, but it can't decide who is allowed to send one, and it can't be the place where the record of that wire actually lives.

I split systems of record into two rows on purpose, because the CRM space is having this exact argument right now. Newer CRMs like Attio pitch that AI can pull your data out of Salesforce, clean it up, and load it somewhere else in a fraction of the time a migration used to take, and I think they're mostly right. A CRM records things that happened somewhere else, like a call, an email, or a deal stage someone typed in, so a good enough copy is worth about the same as the original.

A bank's ledger is different. It doesn't describe your balance, it *is* your balance, and everyone else reconciles to it. You can copy every row to another system, but only the original can actually move the money. Almost any data can be exported now. A sales team can run on a copy of its CRM, but nobody can spend a copy of their bank balance.

## Storefront or Supplier

If the agent becomes the storefront, incumbents risk becoming interchangeable suppliers underneath it, the same way Amazon sellers did. The customer asks the agent, the agent picks the supplier, and the supplier's brand shows up as a line in a tool call.

Banking already has two visible answers to this. (I work at JPMorgan, so to be clear, everything here is based on public reporting, and these views are my own.)

Schwab is [piloting its own MCP server](https://invezz.com/news/2026/11/13/schwab-mcp-server-claude-extended-hours-options/) so clients can connect the LLM they already use to their accounts, starting this week with a small group on Claude, who can even place trades as long as they approve each one. JPMorgan [blocks agents](https://www.beri.net/article/schwab-jpmorgan-block-outside-ai-agents-credentials-browser-automation-read-only-mcp-money-movement) that log in with a client's credentials and allows read-only access through connectors like Plaid, with no money movement.

The risk of being only an API is that your traffic becomes someone else's demand data, and the context and memory live in their agent. That's the AmazonBasics setup again, just one layer down.

So should an incumbent open its doors to outside agents like Schwab, or keep them out and steer clients to its own, like JPMorgan? I believe the answer is probably both. Own the first-party agent where the context lives, and open a narrow, governed door for outside agents, with permissions that widen as trust builds.

In [Trust Is the Rail](https://gufamily.org/writing/trust-is-the-rail.md), I wrote that agents move as fast as your trust does, and the same goes for agents you didn't build. Read-only today can become limited actions next year, once there's a track record to point to.

## Rent Generation, Own Decisions

If I were advising a large regulated company on where to spend, I'd boil it down to renting generation and owning decisions.

1. Let the bundle have generic productivity. Decks, docs, and meeting notes will be good enough in the $20 plan, so govern them rather than build them.
2. Buy and wrap frontier reasoning. Put it behind your own model gateway so that any one lab is replaceable.
3. Build and own the client-facing agents behind your login, the memory and context they carry, the entitlements they respect, and search over your systems of record.

That last item is close to what my team is doing. As I wrote in [A Bitter Lesson in Finding Needles](https://gufamily.org/writing/a-bitter-lesson-in-finding-needles.md), we now own both the agent that asks the questions and the search surface that answers them, and every query has to respect entitlements at the record level. None of that shows up in anyone's $20 plan, because none of it works the same anywhere else.

## Building With Open Weights

A lot of what I've written here about open weights, I didn't want to just take on faith. So I've been building with them on the side to understand what they can actually do and to form my own opinion. The most useful of those experiments is a [decision head on a frozen Gemma 4](https://ai.gufamily.org/gemma4-decision-head), a small classifier that sits on top of a local open-weight model and answers routing, scoring, and yes/no questions with a probability for every option.

It's calibrated, meaning its confidence matches how often it's actually right, so when it says it's 90% sure, it's right about 90% of the time. That lets it handle the requests it's sure about and abstain on the rest, so frontier tokens get spent only on what it can't decide. To me, this is what "own decisions" looks like in practice.

The mechanics turned out to be the easier part. With no labels, the head simply reproduces Gemma's own answer, and with 20 labeled examples per option it gained 2.5 to 18.9 points of accuracy across 14 intent benchmarks, with training cheap enough to run on a laptop.

The hardest engineering problem was making training stable, since Gemma's hidden states are large enough that my first version swung wildly from one step to the next until I normalized them. But the hardest problem for using this in the real world is the data. I didn't have a labeled set of real decisions of my own, only two small hand-written suites of 12 and 66 cases, so I had to lean on public benchmarks as a stand-in.

That gap is where I can see the competitive advantage sits. A lab can ship the same technique tomorrow, but it can't ship your labels, like the routing calls, escalations, and approvals your team has already made, or the thresholds you've set for when to trust the model. A company that has been carefully recording its own decisions already holds the ingredient that makes this work, and it's the one ingredient no lab has in a training set.

## Which Layer Is Somebody Else's AmazonBasics?

The LinkedIn post was about a product list, but the question underneath it is one I believe every company should ask itself about its own roadmap. Which of our layers is somebody else's AmazonBasics?

Generic productivity was always going to get cheap. I could be off on the timing, and open weights or a slower Nvidia could shift the math in ways I don't see yet. But I believe the layers that survive will be the ones a lab can see but can't ship.