We contributed to Lido DAO, P2P.org, =nil; Foundation, DRPC, Neutron and invested into 150+ projects
In 1975, Frederick Brooks published the book "The Mythical Man-Month" where he explained why "adding manpower to a late software project makes it later."
Twenty years later, the second edition included a new chapter titled "The Mythical Man-Month after 20 Years" which reaffirmed the book's original insights.
In 2025, it seems like the perfect time to imagine what "The Mythical Man-Month After 50 Years" chapter might look like in the era of AI agents operating at machine speed.
Brooks explains that as the number of team members increases, the number of communication channels grows quadratically according to the formula n(n−1)/2. This number grows very fast.
While Brooks does not explicitly define an "optimal" team size, his work emphasizes that smaller, well-coordinated teams are more efficient due to reduced communication overhead.
The "Two Pizza Team" rule, popularized by Jeff Bezos at Amazon, states that if a team needs more than two pizzas to be fed, it's too large - typically translating to 6-10 people.
It has been shown by other practitioners that the optimal team size is "seven plus or minus two".
Assume that we would like to build an AI-only team with no hierarchies. What would be the maximum number of AI agents in such a team (assuming that communication is the only bottleneck)?
A human being can process 4 "tokens per second" on average, while DeepSeek V3 produces 60 tokens per second.
DeepSeek has 60/4=15 times more bandwidth, and the size of the team can be 15≈4 times larger for AI agents, ranging from 20 to 32 for the "Two-Pizza Team" (or whatever AI agents eat).
QwQ-32B has emerged as a surprising competitor to DeepSeek R1, demonstrating comparable benchmark performance despite having a significantly smaller parameter count. This model produces 450 tokens per second.
Based on this speed, the optimal AI agent team size would range from 55 to 99 agents.
DeepSeek V3 is a large-scale Mixture of Experts (MoE) model that requires 16 NVIDIA H100 GPUs per instance.
Since each AI agent in this flat structure requires its own model instance, the total GPU requirements are 32×16=512 H100 GPUs for a team of 32 agents.
With a $2 per hour rental cost of an H100 GPU, this results in an estimated cost of $2×512=$1024 per hour per team.
Larger teams bring more cognitive diversity, allowing for a greater variety of perspectives and solutions.
They enable cross-disciplinary innovation by combining expertise from different domains, leading to novel insights and unexpected breakthroughs.
Additionally, larger teams can explore problem spaces more deeply by testing more hypothesis, thereby accelerating discovery and enhancing solution robustness.
Assume the total budget for GPUs is $250 billion (the other $250B are for pizza). Price of a single H100 GPU = $30,000.
Number of H100 GPUs Stargate can buy 30,000250,000,000,000=8,333,333 GPUs.
Each team requires 512 GPUs.
Total number of teams: 5128,333,333≈16,276
So, they can run approximately 16,276 teams with a $500 billion budget, assuming half is spent on GPUs.
A quick search reveals that the global AI research community comprises 300,000–500,000 individuals engaged in research activities, with 10,000–20,000 distinct research groups across academia, industry, and hybrid organizations.
With this number of research groups, we currently see about 2 fundamental AI breakthroughs annually.
Stargate could quadruple this capacity (since AI doesn't sleep), potentially leading to 8 or more fundamental AI breakthroughs per year.
The numbers speak for themselves: with over 16,000 teams working around the clock, we can systematically explore research directions and validate hypotheses at a scale that matches the complexity of modern scientific challenges. This represents a clear, quantifiable increase in our capacity to advance human knowledge.
Excited to see what's going to happen!