When every model can think, the one that remembers wins.
@OpenAI should buy @Dropbox**.
[@AnthropicAI](https://x.com/@AnthropicAI) should buy [@Box](https://x.com/@Box).**These are the two most obvious acquisitions in technology, and the window to do them just opened.
I was President of Dropbox from 2019-2023. I spent those years helping build the FCF story the market still prices the company on. That story is true. It is also beside the point.
Let's start with the thing many people believe. The file is dead.
SaaS killed it as everything became links, customers became rows in Salesforce, designs became objects in Figma, documents dissolved into Google's cloud. Then mobile buried it. The iPhone shipped in 2007 with no visible filesystem, and Apple did not add a Files app until 2017. For fifteen years, betting on files was betting against the direction of computing.
But the file never died. It simply disappeared from the interface.
SaaS did not delete your files. It captured them: the schema, the workflow, the distribution. You got a cleaner screen. The vendor got your data model. Mobile did not delete the filesystem. It hid it behind icons. The file went underground, and the market wrote it off.
Then AI arrived and reversed the abstraction.
Look at what the labs ship, not what they say. Anthropic's Skills are folders of markdown. Agent behavior is configured in CLAUDE.md and AGENTS.md sitting in a repo. MCP launched with a filesystem server as a reference implementation. Claude Code and Codex do their work inside a directory. Memory, when it persists, persists as markdown on disk. Context engineering, the highest-leverage discipline in applied AI right now, is file curation with a better title.
Skills are folders. Memory is markdown. Context is a directory tree.
Every example on that list is a developer tool. Developers are on the frontier of knowledge work. Where they go, everyone else follows a few years later. And developer work orbits a repository: the durable, versioned, permissioned home for everything the team produces. Knowledge work has the same center of gravity. I watch it at Jasper every day. A marketing department's real repository is its website and everything that feeds it, the pages, the assets, the campaigns that attract and convert. The website is marketing's production branch.
General knowledge work has had a git repo for twenty years. The industry gave it the dullest name imaginable: file sync and share.
And the playbook for buying it has already run once. In 2018, Microsoft paid $7.5 billion for GitHub, kept it neutral, and watched it become the distribution spine for Copilot. That was the acquisition of the developers' file layer. The open question is who acquires everyone else's.
Now watch what developers are doing right now. They have stopped writing the code and started architecting the systems that write it. The rest of knowledge work will follow the same arc: the marketer stops writing the campaign and starts architecting the system that writes it. An agent needs durable objects it can read, edit, version, permission, cite, and hand to another agent. That is a file. The model provides the intelligence. The filesystem provides the memory, the state, and the continuity.
The file is not coming back as an interface. It is coming back as infrastructure.
Call this the great file reversal.
And here is why it is worth real money. The most universal unit of context is still the file. Files are portable. Files are permissionable. Files are inspectable, versionable, local or remote. They are read by people, by models, and by software. And they move between applications without asking the application owner for permission. Try that with a Salesforce record.
A file is also more than a container. The name carries meaning. The folder carries context. The versions carry the history of the thinking. The permissions carry authority. The collaborators reveal relationships. Over years, every person and every company quietly builds an ontology of what matters, and that ontology lives in its files. String fifteen years of it together and you get something no vector database can reconstruct: a context graph. Identity plus content plus hierarchy plus permissions plus history.
Context is the scarce complement to intelligence. Models are converging. Context compounds. A frontier model without your files is a genius with amnesia.
In the land of brilliant amnesiacs, the keeper of memory is king.
That is why two of the most strategically important assets in AI may be companies the market still thinks of as cloud storage.
Microsoft has OneDrive. Google has Drive. Apple has iCloud. Each welded to its parent's stack. Outside the giants, the entire independent file layer is the two companies that defined file sync and share: Dropbox and Box. Roughly $7 billion and $4 billion of market cap, trading around 7 and 12 times free cash flow, priced by a market still fighting the last war. Storage is a commodity. Sync is a feature. Growth is over. All true in 2019. All beside the point now.
You are not buying storage. You are buying the context graph, with a home address.
The OpenAI trade: $DBX.
OpenAI has the intelligence. It has the distribution. It has the dominant consumer interface for AI. What it does not yet own is a durable, independent, user-controlled memory layer that lives outside the conversation.
Dropbox does.
Dropbox is 700 million registered accounts across 180 countries, 18 million paying, $2.5 billion of ARR, and close to a billion dollars a year of free cash flow. Registered is not active, sure, but Dropbox’s desktop app is already installed on hundreds of millions of devices. It is still the largest independent identity-and-content base outside the platform giants, and the funnel is the point.
What OpenAI actually gets:
The context. Dropbox is where people keep the most important information in their lives. Photos. Videos. Contracts. Manuscripts. But the real value is not just the files. It is how users have organized them: folder structures, filenames, versions, and sharing relationships that map how they think. Fifteen years of user-built ontology, pre-labeled and nearly impossible to recreate. And because Dropbox often spans both personal and professional life, it gives OpenAI a natural bridge from consumer memory into work.
Now for the part that determines whether this deal is genius or suicide: OpenAI cannot treat Dropbox as a data acquisition. It gets no blanket right to read, train on, or repurpose users’ files. The moment people believe otherwise, users leave, regulators arrive, and the asset burns. The product must be a trusted bridge: users explicitly point their AI at the parts of their lives they choose, can revoke access at any time, and know that training is off by default. Not as a policy promise, but as an architectural guarantee. The acquirer owns the infrastructure. The user owns and controls the content. That distinction is the entire deal.
Get it right and the payoff writes itself. Help me understand my financial life. Find the warranty for the refrigerator. Organize thirty years of family photographs. The files exist. The organization exists. The AI needed a bridge people already trust. The folders you approve become the context window. A type of long-term memory to the chat conversation's short-term memory.
The footprint. ChatGPT has the sessions. Dropbox has the residency: nearly two decades of living inside people's desktops, consumer and enterprise both, plus the sync engine, the metadata, the indexing, the recovery, and Magic Pocket, the exabyte-scale storage system Dropbox built when it pulled its data off AWS. Boring problems. Also the exact problems an agent must solve before it can safely act on a computer. ChatGPT already connects to Dropbox as a data source today. The acquisition is vertical integration of an existing trajectory, not a lark.
The value of undo has not been recognized by most yet. Agents will make mistakes. They will overwrite the wrong document and apply an instruction too broadly. Dropbox already ships version history, recovery, and rewind. Point that at agents and every action gets a diff, a log, and a rollback. Intelligence makes autonomous action possible. Versioning makes it survivable.
The financials. Nearly a billion dollars of annual free cash flow, against a company reportedly burning $27 billion this year on compute. Not decisive but not nothing.
The trust. Users used to tell me, "Dropbox is the longest subscription I have ever continuously paid for. Longer than my mortgage." Two decades of holding people's tax returns. You cannot A/B test your way into that.
And the price. At OpenAI's $852 billion valuation, all of Dropbox at a healthy premium, debt included, let’s call it $11 billion. A bit over one percent of the company for the personal context layer of AI. The cheapest strategic asset anyone will buy this decade.
The Anthropic trade: $BOX.
Box is a permission system disguised as a file product. Underneath the documents: identity, access controls, classification, retention, legal holds, residency, audit. It serves more than 100,000 organizations, 69 percent of the Fortune 500 by its own count, did $1.2 billion of revenue last fiscal year, accelerated to 11 percent growth last quarter, and already books a tenth of revenue from its new AI tier.
What Anthropic actually gets:
The permission graph. An enterprise agent cannot just answer, can I understand this document. It has to answer: is this employee allowed to see it, can the output leave the company, is the source under legal hold, which version is approved, does this action need a human. The intern asks the agent about the layoffs document and the agent, running on Box permissions, says nothing. An agent without those controls is a demo. An agent inside them is enterprise software.
The distribution. Every Box contract is a pre-cleared vendor path into a large enterprise: the CIO relationships, the procurement scar tissue, two decades of passed security reviews. Anthropic sells the most trusted models in the enterprise. Box sells the most trusted place enterprise content lives. Same buyer. Same pitch. Half the sales cycle.
The head start. This is not a speculative adjacency. Box already ships an MCP server, already runs Claude inside Box AI, already adopted Skills, and is building agent-facing automation across its platform. The two companies are building adjacent halves of the same product. The acquisition compresses a multiyear partnership roadmap into one architecture.
The control plane. In the agent era, every enterprise will maintain a library of approved skills, templates, policies, and evaluation sets. Those are files. Box becomes the governed repository for what agents are allowed to do: which skills, which departments, which models, who approved the change. The employee sees Claude. The enterprise sees Box underneath it.
At Anthropic's $965 billion valuation, all of Box at a premium is well under one percent dilution.
And the symmetry is the tell. Dropbox is personal memory. Box is organizational memory. ChatGPT is the consumer default. Claude is enterprise to its core. The pairing is not clever. It is obvious.
The objections take one line each.
They are ex-growth. You are not buying growth. You are buying position, trust, and cash flow in the scarcest layer of the stack.
Connectors are enough. Both labs already connect to both products, which proves the direction and disproves the sufficiency. A connector rents context. An acquisition owns the architecture.
Enterprises chose them because they are neutral. So keep them neutral. Let every model in, win on depth of integration, and never degrade a competitor's access. Exclusivity would torch the asset on day one.
Antitrust. These are the rare deals regulators should want: they arm the last independents against the Microsoft and Google bundles. The giants could never buy these companies; each already owns the competing product. The labs can, because they own none.
Why would they sell. Both trade at single-digit and low-teens multiples of free cash flow, in stock that has gone sideways for years, while the acquirers print the fastest-appreciating paper on earth. @drewhouston and @levie did not spend twenty years building these companies to run terminal-value stories.
And here is why the window is now. Anthropic confidentially filed for an IPO on June 1. OpenAI filed on June 8. Both now have a path to public currency and public scrutiny, and every strategic gap in their stories will get priced daily. The file layer is one of them: cheap, scarce, and impossible to recreate.
Whoever holds the files holds the context. Whoever holds the context becomes the interface. Intelligence without context is a demo. Intelligence with memory, permissions, and state is an operating system.
Steve Jobs spent a decade hiding the filesystem. The labs will spend the next decade buying one back.
**The file was never dead. It was waiting for a reader.
**
(Disclosure: I was President of Dropbox from 2019 to 2023. At the time of this writing I have no position in either $DBX or $BOX
