← The Vault
The Big Story

Why running a model costs more than building it

The AI industry is hitting a predictable bottleneck: it is much cheaper to train a model once than to run it for millions of users. As companies like Baseten raise billions to streamline 'inference'—the process of answering a query in real-time—others like Snap are spinning off AI units to manage the massive compute bills. This shift reveals that in the current market, the hardest problem isn't intelligence; it's the cost of constant calculation.

Edition № 053Room: The Big Story18 June 20262 min readSources: 4
Article

Building an AI model is a one-time intellectual investment, but running it is a perpetual tax on your bank account. As venture capital follows the path of least resistance, we are seeing a disconnect between the companies building the intelligence and the companies struggling to pay the electricity bill for that intelligence to speak.

While high-profile firms like OpenAI recruit top-tier researchers to improve model performance, companies like Baseten are raising massive funding rounds to solve the 'inference' problem. Inference happens every time someone clicks 'generate' on a chatbot, requiring thousands of processors to perform billions of floating-point math operations in a fraction of a second. This is fundamentally different from the training phase; it is an incessant, high-stakes sprint where every microsecond adds a tangible cost to the bottom line.

The physics of the inference tax

To make a model answer a question, the server must pass your prompt through layers of neural networks, each one holding millions or billions of parameters that need to be read from memory and multiplied. If you run a large model on a standard chip, the latency makes it feel sluggish and the cost per query makes it unprofitable. Inference startups build specialized software infrastructure to batch these requests and optimize where the data sits in memory, essentially trying to squeeze more work out of expensive hardware like H100 GPUs before the electricity meter stops running.

This reality is forcing a reckoning for everyone else. Snap spinning off its AI video team is a tacit admission that constant, high-fidelity AI generation at scale is expensive enough to threaten a stable business model. We are moving out of the phase where 'cool demos' were enough and into the phase where engineering efficiency—simply making a model faster and cheaper to run—is the most valuable skill in the business. If your business depends on these models, the question is no longer 'how smart is your AI?' but 'can your margins survive the cost of every keystroke?'

Sources
← PreviousThe Fog of AI GovernanceNext →Elastic’s move into automated debugging
Tomorrow's edition · free

Liked this one? The next lands at breakfast.

Every story in tomorrow's AI news, rebuilt in plain English — five minutes, sources linked, free forever.

By joining you agree to receive Article's daily newsletter — unsubscribe in one click. Privacy

← Back to the Vault