OpenAI is Dropbox, Anthropic is Box.com
Attention conservation notice: loosely grounded pontification about “AI”, a prose genre of which there is no shortage of these days. On the plus side, I wrote this entirely by hand, which is more than you can say about most of this stuff.
Briefly: Dropbox’s business is a hardware-constrained commodity that competes directly against the world’s largest and best-capitalized technology companies, all of which possess huge structural advantages that limit Dropbox’s upside. OpenAI is just the same.
The rest of this writeup will just expound on this a bit more. If you feel that you can work out the parallels and implications yourself (and I don’t think it’s that hard), feel free to close the tab now.
Dropbox’s core product competes against Google Drive, Microsoft OneDrive, and Apple iCloud, which are bundled for free with the operating systems that come preinstalled on every device that consumers and businesses buy. The heart of Dropbox’s technology — its backend storage systems and client sync software — might be something like 10% better than those competitors’ offerings. However, all that engineering prowess ultimately produces a thin, barely-differentiated layer over an underlying hardware commodity (datacenter capacity), which makes their product into a commodity in turn.
Now, Dropbox is doing fine: as of writing, it has annual revenues over $650M and a market cap of around $7.8B. Dropbox won’t go out of business anytime soon. However, it is extraordinarily unlikely to break out of its niche and become much more influential or much more valuable.
Dropbox’s competitors can cross-subsidize storage from their other lines of business, which are far more valuable and differentiated, and this creates huge headwinds for Dropbox as it tries to navigate to a different part of the parameter space. Dropbox experiences sharp price discipline in their core business, and any expansion into adjacent product lines suffers from two facts. First, their competitors have already successfully colonized much of the potential adjacent ecosystem: consider the fates of Dropbox Paper (Dropbox’s version of Google Docs and Microsoft Word) and Dropbox Carousel (which tried to take on Google Photos and iPhoto). Second, any novel product discovery by Dropbox basically amounts to free market research for its competitors, who can use their huge capital and labor advantages to clone anything that succeeds.
Today, the big independent US companies developing frontier-class large neural networks are OpenAI and Anthropic. Their businesses are structurally extremely similar to Dropbox.
Frontier models are expensive to build, but the frontier moves continually, and for many use cases, near-frontier models serve as close-to-drop-in replacements for each other. That makes models akin to data storage systems: a thin, barely-differentiated layer over an underlying hardware commodity (in this case, GPUs rather than disk drives). And like Dropbox, OpenAI’s competitors are among the largest, best-capitalized technology companies on Earth: Google, Microsoft, Facebook, etc.
What about Anthropic? Let’s talk about Box.com.
Box.com is Dropbox’s twin with a business suit and briefcase: it sells storage services, but primarily to businesses, which pay a premium for various “enterprise” features, accessory to storage, that normal people never think about. Indeed, I suspect that if you have never worked at a white-collar job in a large corporation, you might not even know what Box.com does.
In much the same way, Anthropic is OpenAI’s twin with a business suit: OpenAI’s flagship product is a chat app that a billion people have installed on their mobile phone, whereas Anthropic’s flagship is a coding harness and inference API that has spread like wildfire in the business realm but that normal people never think about. (If you live in the Bay Area, then I ask you to believe me that most normies are not, in fact, vibe coding with Claude).
As the saying goes, history never repeats itself exactly, though sometimes it rhymes. OpenAI and Anthropic differ from Dropbox and Box.com in various ways: the depth and caliber of talent they have attracted, the rapidity of adoption, the broader utility of the techology, and more. However, structurally, I can currently see no reason to believe that they will not ultimately suffer a similar fate: it is hard to believe that they will succeed in staying ahead enough, for long enough, that Google and Microsoft and perhaps other competitors won’t eventually marginalize them through distribution and development of more successfully differentiated complementary products.
People today observe, correctly, that Claude Code, Codex, and Cursor are meaningfully better than Google Antigravity and Microsoft Copilot. Dropbox’s sync engine was, and is, likely meaningfully superior in quality to Google Drive or Microsoft OneDrive. That did not save Dropbox from its fate.
One difference that may cut particularly deep, and not in favor of OpenAI and Anthropic, is that both of the independent model labs have financially engineered themselves into rather extreme positions, the likes of which Dropbox and Box.com never did. It seems possible that rather than settling down into “normal” profitable businesses, the big independent model labs simply implode under the weight of various forms of debt and equity revaluation.
Lastly, if you want someone else’s take on all this, I suggest reading Ben Evans’s essay Ways to think about token pricing, which I find canny, nuanced, and insight-dense in a way that most other writing on this subject is not.
Well, that’s it. That’s the writeup.
Ok, one last navel-gazing thing. If you’re wondering why I wrote this: I’ve been making this observation privately to colleagues for a couple of years now, and I said it in passing on fedi in October 2025; I wanted to write this down more completely, partly for posterity (so that I would have the entertainment of either looking amazingly prescient or utterly foolish in a few years), and partly so that I could point colleagues at this instead of recapitulating my entire rant each time.
And that really is it.
…
…
What, you’re still here? It’s over. Go home.