The Number That Changed Everything
In January 2025, DeepSeek published a technical report that did something unusual in the AI world: it told the truth about what things actually cost. According to the DeepSeek R1 Technical Report, the company trained a frontier-level reasoning model for under $6 million. This wasn’t vaporware or a lab toy. This was a model that competed credibly with systems estimated to have cost over $100 million to train. The gap between those two numbers isn’t just a data point. It’s a recalibration of what’s actually possible in this industry.
When I was at McKinsey, we spent months on efficiency analyses. We’d find 15-20% cost reductions and call them wins. This wasn’t 15%. This was a 94% reduction in training costs for equivalent capability. That kind of gap doesn’t stay closed for long. Within days of the announcement, the market understood what this meant: the fundamental unit economics of AI infrastructure were broken.
The reaction in equity markets wasn’t subtle. Nvidia lost roughly $600 billion in market capitalization in a single trading session, marking the largest single-day wipeout in U.S. stock market history. People talk about market overreactions, but this one was clarifying. Investors weren’t wrong. They were pricing in what should have been obvious: if training costs collapse toward single-digit millions instead of three-digit millions, the entire pyramid of VC funding, infrastructure spending, and AI startup burn rates needs restructuring.
The Startup Portfolio Suddenly Looks Different
YCombinator’s winter 2025 batch accepted roughly 450 companies. Within the first quarter, more than 40% of the AI-focused founders in that cohort pivoted their infrastructure strategy. That’s not a gradual shift. That’s a category-wide reset happening in real time. The founders didn’t pivot because they lost faith in their core ideas. They pivoted because the financial assumptions underlying their go-to-market plans had evaporated.
Here’s what changed: six months earlier, an AI startup’s infrastructure costs were predictable and punishing. You built your model, you rented expensive GPU clusters, you burned $2-5 million a month on compute, and you hoped to show enough progress to raise a Series B before that math broke you. This wasn’t inefficiency. This was the market structure. Every credible AI company operated under similar constraints. The barrier to entry was capital, not talent or idea quality.
DeepSeek R1 made those constraints optional. Suddenly, a team with $10 million in seed funding could train a competitive model instead of just fine-tuning someone else’s. The distribution of what’s possible shifted. That’s the kind of shift that breaks business models built on scarcity.
When Commoditization Moves From Theory to Practice
Andreessen Horowitz’s 2025 State of AI report documented something that venture capitalists usually avoid saying plainly: infrastructure costs that defined AI startups dropped 34% year-over-year as open-weight models flooded the market. This matters because cost reductions at that scale change which kinds of companies can exist. When your infrastructure costs drop by a third and everyone else’s drop by a third, the relative playing field stays level. But it means the cash burn rate that made sense at the old cost structure is now profligate.
I watched a Series A AI startup in my portfolio walk through the numbers with their CFO two weeks after DeepSeek’s announcement. Their monthly infrastructure spend was $3.2 million. Under the old assumptions, that was acceptable runway math. Three months of burn, sustainable at their growth rate. After the DeepSeek announcement and the rapid emergence of alternative open-weight models, their CFO ran the same workload through different infrastructure combinations. The monthly bill dropped to $2.1 million. That doesn’t sound dramatic until you realize what it means: they now have four months of runway instead of three at the same cash position. That’s a genuine extension of corporate life.
But here’s where it gets complicated. Every startup that did this math also understood something deeper. If they can reduce costs by 30%, so can their competitors. The cost advantage is temporary. The real question becomes: what do you do with that runway extension? Do you hire faster? Do you extend your timeline to profitability? Do you price more aggressively knowing your cost structure is better than the market assumes?
The Price War Nobody Wanted but Everyone Expected
OpenAI moved first in the obvious direction. In April 2025, they released o3 at $10 per million output tokens, a 75% price reduction from o1 pricing. This wasn’t a competitive move designed to expand the market. This was a response to margin compression. When open-weight alternatives become genuinely competitive, proprietary models have to choose between pricing power and relevance. OpenAI chose relevance.
What’s interesting about this decision is what it signals about OpenAI’s cost structure. You don’t cut prices by 75% unless you can absorb it. That suggests their infrastructure costs are either lower than the market assumed or their margin profile was comfortable enough to take the hit. Most likely both. But here’s what it means for every other AI startup: the price floor just moved down. Significantly.
This creates a specific kind of pressure. If you’re a Series B AI startup with a model that costs $X to serve and the incumbents are now pricing at $10 per million tokens, you can’t charge $15 and compete on price. You have to find something else. Better accuracy on specific tasks. Faster inference. Better developer experience. Something that justifies a different positioning.
The Burn Rate Question That Matters Now
Every AI startup’s CFO is running the same scenario analysis right now. The question isn’t whether costs will stay low. They will. Open-weight models are here, training efficiency will keep improving, and the old cost structure was an anomaly born of scarcity. The question is what burn rate makes sense when you have three-to-five times as much runway per dollar of capital.
Some startups will use this to hire more aggressively. Others will extend their timeline to profitability. The smart ones will do both strategically, based on where they think the market is heading. But every single one is making a different decision than they would have made six months ago.
The startup world always adjusts to market structure shifts. Usually this takes eighteen months, a few spectacular failures, and a lot of expensive mistakes. This time, the adjustment is happening faster because the change was dramatic enough that it forced everyone to do the math at once. That’s actually healthy. It means the misallocated capital gets reallocated sooner.
What’s your startup’s burn rate looking like after the market moved? I’d genuinely like to hear what founders are actually doing with this new runway.