403 Access Denied
Brace for the Platform Wars
This vibes have shifted in the AI world in the last few weeks. Two examples:
First: Slack cut off Glean (one of the most well-funded “put AI to work” startups) from their API. Glean integrated with Slack both as a convenient, chat-based interface to their product but also to automate many of the workflows that often live inside Slack. It was competing with Salesforce’s AI vision.
Second: Windsurf got bought by OpenAI and Anthropic cut them off Claude 4 models. Windsurf users could no longer use Claude 4 to generate code (so far, there is a workaround but users need to BYOK — bring your own keys — which is high friction and higher cost, to say the least).
The Slack change cuts deeper: not only did they cutoff a competitor, they have made it very hard for customers to access their own data via the API. There are restrictions to copying, storing, indexing your own Slack data, and API methods like conversations.history or conversations.replies have been rate-limited to 1-call-per-minute and 15-messages-per-call. Slack is ‘where work happens’ and it’s clear that Salesforce intends to be the only one capitalizing on the AI-fication of any work inside that platform.
Of course, the interweb was all abuzz — warfare is not something tech has seen in a while, and our memories are short. Platforms start with openness to invite as many builders as possible, and when the use cases get real or threatening, their tune changes overnight. Was it Sun Tzu who said “Let your enemies build on your API so you can get cut them off later”?
A walk down history: we’ve been here before
I was at Twitter when they shut down access to their graph, killing off House Party (in favor of Periscope which they bought), and when Instagram shut them off its images. I was at Zynga when Facebook brought the kibosh on OpenGraph and more generally exposure in the feed. I was at Amazon when Amazon shut down many ecommerce APIs and in turn got stonewalled by Google.
In the early days of Twitter, the vision was for it to be a protocol. The API was very open and 3rd party developers built far more compelling products and use cases on the API than the native product itself — everything from full-on clients with many unique features to marketing and analytics tools. The API proved to be an excellent laboratory to see what worked. But the API program was unceremoniously wound down when the company realized they needed to own the end-to-end experience to grow DAU and ad revenue as Wall Street demanded, and also to have dibs on future features. All 3rd party apps, some real businesses, died overnight. It left people up in arms, but without a choice.
Login-with-Twitter (oauth) was one of the few APIs left open, as it served to grow the platform. Some startups figured out how to use it to get users to bring over their Twitter graphs and bootstrap growth. House Party and Periscope were two such competitors. Twitter acquired Periscope, gave it all the fuel it need to grow and House Party saw 403s the next day, their growth stalling without the fuel of a pre-existing social graph. They shut down soon after.
Instagram, similarly, blocked Twitter from rendering posts so whenever users encountered a shared Instagram post, they would be forced to click the URL and go to Instagram.
Facebook was an open platform in the early days — remember FBML? You could build mini-apps inside of Facebook with all their data and the graph. Your app could publish activities to their feed. That’s how games like Zynga’s got their rise — on the sweet oxygen of a rich social graph and the many hundreds of millions eyeballs on the newsfeed. Until one day Facebook decided it wasn’t in their best interests.
Something that we would find it hard to believe today: All of Amazon’s catalog used to available freely via an API. It was one of the first API products Amazon offered, named ECS (E-Commerce Services). Google Search was available via an API — you build a ‘wrapper’ search engine using it, and they let you. Not only Search, their Ads system was an API call away. I was a young engineer at Amazon working on ads, and we would send a few keywords and targeting information like IP address and user agent via a SOAP/XML packet to a Google endpoint and get back three ads to show to our users. Google got no data in exchange (except for click tracking), and they trusted us to report back metrics like impressions. We were able to build so much smarts into how we queried their API and optimize for every cent we could. One fine day, that ended.
The coming war: harder and faster
The internet, after the dot-com crash was a giant open platform, looking for more use cases, inviting developers to build. That was the original “web 2.0”: the vision was open APIs, data portability, widgets and mashups. When the value became obvious, the platforms locked it all down and built gigantic businesses. The AI era is not going to be different.
Amazon didn’t start competing with their AWS customers (higher-level features like video calling, email service, etc) until a decade into it. Facebook didn’t shut down their open graph API for years. The AI platforms wars are speedrunning the previous wars.
Platforms today are competing with their customers sooner and harder. OpenAI bought Windsurf, as AI-assisted coding has risen to the top 2 use cases. They’ve hired the Instacart CEO to be ‘CEO of Applications’, which sounds like more competition with their API customers. They have also blocked OpenRouter from directly routing to their more capable models (users have to sign up on OpenAI and can BYOK). Retell AI (a voice agent platform for others to build on) is now directly building call center agents and medical clinic agents. Horizontal software like Slack are locking down. Content owners are hiring more lawyers. There are three fronts in this war: LLMs, distribution, data.
The stakes with AI are the highest they’ve ever been in tech, and no holds are barred.
MCP sounds amazing. The ability to access everything through my Claude app? As a user, I would like nothing better. But once shit gets real, no service will want Anthropic to be disintermediating them. I bet MCP will be as forgotten as OpenAI’s GPT Store before next year.
The uncomfortable truth about AI is that the prize is just too large, and any access or data can be directly and immediately monetized. If you’re an AI company, you need to have a data acquisition strategy as much as you need a customer acquisition strategy.
Prepare your grand strategy
If you’re a founder in AI, here are some things you can do:
Expect the hardest competition from the large labs or platforms, especially if you catch lightning in a bottle. Build some advantage they can’t copy easily (”do things that don’t scale” is evergreen wisdom - all large companies have similar disadvantages).
Expect everyone to play dirty. All’s fair in war.
Build optionality: have no single dependency or point of failure. Use multiple models, ideally an open weights model in the mix. Don’t rely on frameworks or unnecessary abstractions, especially when writing these is so easy now.
Distribution strategy is Startup 101. Startup 102 in the AI era is Data strategy. Get rights to any data you can early on, get as close to the source of data as you can, without relying on intermediaries.
Build partnerships early. While every large company is walling off their garden to own the AI applications on that data, it's easier said than done — just how compelling do you find Apple's or Google's AI applications? Open can beat closed, if done right. But you're up against old incentive structures that don't work anymore. Find the right incentive structures for this new world, and it may give you an edge long enough to survive.
“Never tell me the odds!” — Hans Solo, Star Wars: The Empire Strikes Back