Will OpenAI's $5b burn wipe out startups? (no, and here's why)

Birds fly. Cows eat grass. Startups lose money. While most startups burn a few hundred thousand to a few million a year, OpenAI is reportedly hurtling towards bankruptcy because they’re burning $5 billion a year (of their total of ~$14b in investment). 

Your hands might be a bit sweaty if you’re building an AI product on OpenAI’s APIs. The reason they’re valud at $80 billion is because thousands of startups use GPT models as a core part of their product and even more use them to power their most advanced features. 

What happens if OpenAI has to shut down? Would it wipe startups out wholesale? Or could open-source models save us?

The easy answer is: You’ll probably be fine, even if OpenAI isn’t. The more complicated answer is: Building foundational AI models is expensive. You have to buy GPUs, run inference and pay elite engineers.

If you want to feel better about your AWS bill, the cost of running ChatGPT is about $700,000 per day, which meant OpenAI lost $540 million in 2023. 

While losing $500M loss per year, they could ride Microsoft's $10B investment for another 20 years. At $5B per year, there are three options: 

First, they could dramatically increase prices. So many startups rely on OpenAI that they’d probably retain a good amount of customers. Second, they could slow down their research, but that would mean ceding their advantage to competitors. Third, they could raise more cash. 

If you’re building a product on AI, you probably don’t want to pay more or build on outdated tech. Luckily, option three is most likely:

Getting more money would be easy. Few investors wouldn’t invest in OpenAI given its growth. OpenAI isn't reinventing the wheel here. It launched ChatGPT for free and now has 200 million users on its platform each day. 

That's not to say that it doesn't make any money. In fact, Open AI's annual revenue doubled to $3.4 Billion since late 2023, thanks to revenue generated from its ChatGPT plus subscriptions & its API usage. 

But OpenAI isn’t on a quest to be a B2B SaaS vendor, but to build AGI. If they’re successful in building AGI, burning $5 billion will seem like the price of admission.

As the incumbent in the LLM space, OpenAI has a lot of power. If history is any indicator, incumbents steamroll competitors - like Microsoft did when it steamrolled Netscape Navigator by bundling Internet Explorer with PCs.

This time, Microsoft made a $10b bet on OpenAI. For a while, the GPT models were the only ones at the frontier. That’s no longer the case - which is great if you’re a product builder: Even if OpenAI burns the $13 billion invested in them and has to shut down, you can keep your AI product up and running: 

• Anthropic’s (backed by Amazon) Claude models are now state of the art. 
• Google’s Gemini Ultra outperforms GPT-4 on many benchmarks.
• Meta’s LLaMa 3.1 outperforms GPT-4 on some benchmarks.

Meta’s model is worth highlighting because it’s an open-source model. You can host it for free and adapt it to your circumstances. This means that whether you’ve build a feature or product on GPT models, you can keep them up, even if OpenAI is in hot water.

With OpenAI’s financial situation, investors in Open AI might ask the hard questions. What is their moat? How unique are the models really? How do they become profitable? 

OpenAI will have to answer these questions instead of relying on their dominant position. It might be on a quest to build AGI, but they need to make the financiers of their quest happy in the meantime.

As for us soon-to-be-replaced-with-AGI-folks, it’s worth exploring different models for our use cases instead of depending on one vendor. Platform risk is real: Just ask the developers who lost their businesses when the X API got more expensive. 

If we really are in an AI bubble, we need to be cautious. The signs are there: The economics dont work (yet), the current approach has reached a plateau, there is no clear category winner and hallucinations remain.

One thing is clear: We’re not going back to the pre-AI days. The LLM genie is out of the bottle. 

But if you’re building a product, you may want to diversify and try out different models depending on your use case: It’s worth getting familiar with different models and/or explore self-hosting so your product doesn’t rely on one (unprofitable) company’s financial decisions.

The battle between LLM providers continues. Will OpenAI win despite burning $5 billion? We’ll have to wait and see.