Inc. Magazine
DeepSeek and other startups prove that scaling with a small team is possible by focusing on efficiency and innovation. By using AI, applying a “mixture of experts” strategy, and learning from competitors, companies can grow quickly without relying on heavy funding or large teams, achieving high productivity with fewer resources.
How to Do More With Less by Following DeepSeek’s Playbook
Author: Peter Cohan
Scaling with a tiny team is nothing new. For example, Instagram—which Facebook acquired for $1 billion in 2012, according to the New York Times, was able to grow rapidly to 30 million users with a mere 13 employees before it joined what is now known as Meta Platforms. And then there’s DeepSeek, a Chinese AI startup that built the most popular Apple app store AI chatbot at a fraction of the cost and time of rivals such as ChatGPT. More specifically, DeepSeek spent about $5.6 million in three months to build a complex reasoning model rivaling OpenAI’s ChatGPT o1.
This spells good news for all those Zombie Unicorns—venture-capital backed startups worth more than $1 billion that have been unable to go public for the last three years. Rather than raising more capital, which is nearly impossible right now anyway, cash-strapped and stalled startups should stop fundraising and start innovating.
Why? Capital scarcity puts pressure on companies to build a great product and lead their organization in a way that encourages each team member to be highly productive. Delaying venture capital also enables founders to maintain control of their equity and keep outside directors from possibly forcing them out of their companies.
But back to DeepSeek’s innovation playbook, which hinges on three key elements:
Divide and Conquer
DeepSeek was able to scale quickly through a “mixture of experts” approach to development. This approach trains a mix of specialized neural networks in say 100 fields—from accounting through zoology—that share information when needed. A generalist network coordinates which “expert” network shares information with other expert networks, when there is overlap in their fields, rather than one giant neural network learning everything about everything, notes the New York Times. Put another way, this is sort of like having multiple cores in a processor or even a highway to reduce traffic jams.
Whether you’re developing AI or workflows for your teams, there’s power in removing bottlenecks and sharing knowledge strategically across teams.
Beg, Borrow, and Steal
DeepSeek also used distillation—extracting knowledge from a larger AI model to create a smaller one. In January, Berkeley researchers used distillation to recreate OpenAI’s reasoning model for $450 in 19 hours. Soon after, “researchers at Stanford and the University of Washington created their own reasoning model in just 26 minutes, using less than $50 in compute credits,” according to CNBC. OpenAI has labeled DeepSeek’s distillation approach as theft, but CNN reports that, according to tech venture capitalist Bill Gurley “the core algorithm everyone uses was developed at DeepMind,” Google’s AI lab. “No one disputes that. The vast majority of LLM insights and breakthroughs are ‘borrowed.’”
Where in your organization could redundant workflows be eliminated? Where are you wasting time by repeatedly reinventing the wheel? And are you taking steps to learn from your competitors or are you burying your head in the sand under the guise of “differentiation.”
Use AI
Even though DeepSeek’s development approach is worth emulating for its agility, it would be unfair not to acknowledge that other AI came before DeepSeek, which played an undeniably big role in speeding up DeepSeek’s product development. And AI can help all sorts of companies scale—not just builders of LLMs. Here are two other examples of companies using AI to scale rapidly with less capital:
- Gamma has 50 million users and “tens of millions” in revenue with 28 people. Gamma, a maker of software for building websites and creating presentations, uses 10 AI tools for activities such as customer service, image generation, data analysis, and coding – to make its 28 employees so productive “they can do the work of 200,” CEO Grant Lee told the New York Times. “We get a chance to rethink that, basically rewrite the playbook.”
- Anysphere hit $100 million in annual recurring revenue in less than two years. The maker of Cursor – an AI-first coding platform – grew fast with 20 employees. Rather than selling to companies, Cursor quickly won more than 360,000 individual developers who paid an average of $276 a year, noted Spearhead. Anysphere has now raised $175 million in funding, with plans to add staff and conduct research, the company’s president Oskar Schulz told the Times.
The takeaway for business leaders is simple: Be innovative, only spend what you need, and you’ll soon find yourself well on your way to scaling fast with a small team.
Credits: TCA, LLC.