DEEPSEEK OPENAI

CFOTO/CFOTO/Future Publishing via Getty Images

AI

DeepSeek could upend the funding math for AI apps

Some AI startup CEOs say they’re already using DeepSeek’s cheaper reasoning model, R1, over OpenAI’s o1.

Startup founders are already testing DeepSeek‘s R1 reasoning model as an alternative to OpenAI‘s o1—and are finding it’s much cheaper.

“The application layer just got a whole lot more interesting,” Keval Desai, managing partner at Shakti, tells PitchBook News. Almost overnight, DeepSeek has thrown into question the assumed cost of building AI apps. This has huge implications for how their makers are financed.

DeepSeek’s R1, a Chinese large language model, has thrown the VC industry for a loop.

AI engineers are finding that both running application programming interface calls on DeepSeek and building on top of its open-source model are several-fold cheaper than using OpenAI’s models. DeepSeek is currently selling API access at a discount, but even when DeepSeek’s pricing per million tokens rises in February, engineers still expect it to cost a fraction of the cost of running on o1 or Claude 3.5, its market competition.

The surge of startups already testing DeepSeek has not gone unnoticed by OpenAI. “We will pull up some releases,” CEO Sam Altman said Monday.

And despite concerns over embedded content biases and security risks, engineers are still testing the model.

If AI engineers start widely using R1, it could change the startup funding math for those developing AI apps. So far, fundraising budgets have relied on more costly assumptions.

How much cheaper is DeepSeek, anyway?

DeepSeek’s R1 “costs 93% less to use than o1 per each API [call],” Gavin Baker, managing partner at crossover investor Atreides Management, posted on X. That doesn’t necessarily mean it’s 93% cheaper to get a result. It also depends on the quality of R1’s responses, which affects the number of prompts needed.

But DeepSeek is also open source, so developers can host their own version, fine-tune their own model and run it on their own infrastructure. This is “a first for a state of the art reasoning model,” said Joe Floyd, GP at Emergence Capital.

DeepSeek has also optimized its architecture to use less memory. “It can be run locally on a high end work station and does not seem to have hit any rate limits which is wild,” Baker said.

The lack of a rate limit is critical: DeepSeek has done optimizations on token output, which means the architecture of the open-source model itself requires less memory than running on o1. That, combined with no rate limits, means that running DeepSeek becomes exponentially cheaper over time, said Daniel Chesley, principal at Work-Bench.

“OpenAI charges a tremendous margin and they’re a real business, and DeepSeek is something out of a hedge fund and it was a project, so they don’t really have margin just yet,” Chesley said.

Experts have long predicted that foundational models would become commodities, and DeepSeek is one more step in that direction.

“For startups looking to leverage third-party models, this is a game-changer because inference is cheaper than ever before,” Chesley said.

CFOTO/CFOTO/Future Publishing via Getty Images

Learn about our editorial standards.