The Math That Changed Everything
Let me cut straight to it. On January 27, 2025, Nvidia lost $593 billion in market value in a single day. That’s not a correction. That’s not a normal market adjustment. That’s the stock market collectively processing the fact that the infrastructure moat everyone built their investment thesis around had just developed a massive leak.
The leak came courtesy of DeepSeek, a Chinese AI lab that released R1 in January and essentially did what everyone said was impossible: built a frontier-class reasoning model for $5.6 million in training costs. OpenAI’s GPT-4 cost hundreds of millions. DeepSeek’s R1 cost $5.6 million. That’s not 10 percent cheaper. That’s not 50 percent cheaper. That’s roughly 1/50th the price for a model that performs in the same competitive league.
I’ve spent a decade reading balance sheets and earnings calls. I know how to smell when a company’s competitive advantage is structural versus when it’s temporary. This felt structural.
The Inference Cost Collapse Nobody Planned For
Here’s the thing nobody talks about enough: training cost matters less than you think for most AI startups. What matters is what happens after the model is trained. That’s inference. That’s the moment a customer asks your system a question and your infrastructure has to run it.
The a16z State of AI 2025 report documented something wild: inference costs dropped roughly 90 percent year-over-year between 2024 and 2025. Ninety. Percent. That’s not a gradual efficiency gain. That’s a regime change.
So now let’s talk about the $500K monthly bill. A mid-market AI startup building a customer-facing application using OpenAI’s o3 model at launch was paying $10 per million input tokens. A small team could easily burn through $500K a month depending on usage volume. But if that same startup switched to a self-hosted open-weight alternative running on standard infrastructure, they’d pay under $0.50 per million tokens. That’s a 20x cost reduction, the difference between being a growth story and being bankrupt.
The choice suddenly stopped being theoretical and became urgent.
The Startup Exodus in Real Time
By Q4 2025, over 60 percent of AI startups surveyed by Lightspeed Venture Partners had either switched their primary model provider or were actively evaluating alternatives. Sixty percent. That’s not “companies reconsidering their vendor strategy.” That’s a wholesale market repricing.
I want to be careful here because the tech press loves a good “AI winter” narrative and a good “everyone’s pivoting” narrative in equal measure. Neither captures the actual moment. Startups aren’t panicking. They’re being rational. If your cost structure just got hit with a 20x multiplier in the other direction, you don’t stay loyal to your original vendor out of gratitude. You run the numbers again.
The real story is that the economics of AI inference became contestable. Before January 2025, most startups assumed they had to use proprietary models from companies with massive training budgets and closed ecosystems. DeepSeek didn’t prove those models were bad. It proved they were optional. And once something becomes optional, unit economics become destiny.
What This Means For The Current Playbook
Let me break down what happened to the venture-backed AI startup playbook. The old version looked like this: Build something interesting on top of OpenAI or Anthropic. Raise money from VCs who believe in the application layer. Scale fast. Worry about margins later.
That playbook assumed inference costs would stay flat or decline slowly. It assumed you could build a venture-scale company while paying premium prices for proprietary model access. The math worked if you believed the cost curve was asymptotic. It didn’t work if you believed it was collapsing.
The DeepSeek R1 Technical Report made the cost collapse credible. Suddenly the question wasn’t whether open-weight models could compete. It was whether proprietary pricing could survive the comparison. Those are very different questions.
For founders with $500K monthly inference bills, the re-evaluation isn’t about being disloyal. It’s about unit economics. A 20x cost reduction on your largest operating expense either compounds into years of extra runway or it gets passed to competitors who move faster. There’s no middle ground.
The Uncomfortable Truth For Everyone
Here’s what I think is actually happening beneath the noise. The AI infrastructure market just became efficient. When inference costs drop 90 percent in a year and open-source alternatives reach parity with proprietary models, you can’t sustain a pricing model based on scarcity anymore. You have to compete on quality or features or service, not on being the only game in town.
That’s good for startups. It’s bad for the companies that built their entire business model on proprietary model access and lock-in. It’s particularly awkward for anyone who raised capital assuming they’d have years to figure out sustainable pricing before the market forced a reckoning.
The founders I know who handled this well didn’t panic or swing hard into religious devotion to open-source. They ran the numbers. They understood their unit economics. They made informed decisions about when to switch, when to negotiate, and when to build proprietary layers on top of cheaper infrastructure. That’s not a pivot story. That’s just competence.
The question for your own startup is simple: Have you recalculated your inference costs since January? If not, you probably should. The numbers might surprise you.