The ai startup stack
AI companies overbuy infrastructure and underbuy distribution. At pre-revenue the model is rarely the bottleneck — explaining it and finding the first ten users is. This is the order most AI teams get value in.
The layers that matter
01
Compute
Rented GPU capacity you can turn off. Owning hardware at this stage locks up cash you need for hiring.
02
Research automation
Pulling structured data from the web on a schedule, so evaluation sets and lead lists build themselves.
03
Narrative
Decks and demos. Technical products lose more deals to unclear explanation than to weak models.
04
Outbound
One repeatable channel to reach design partners. Not a CRM yet — a list and a sequence.
What to skip early
Three categories cost AI startups money before they earn any:
- Dedicated or reserved GPU capacity before you know your steady-state load.
- An enterprise CRM for a pipeline you can still hold in your head.
- Observability and analytics suites before you have daily users to observe.
How the stack changes at seed
Once you have paying design partners, two things get added: a real pipeline record so handoffs survive people leaving, and support or telephony if your buyers are non-technical. Compute spend usually becomes your largest line and deserves its own review, not a subscription audit.
A reasonable first month
Week one, get inference running on rented capacity and measure cost per request. Week two, automate the research loop that feeds your evaluations. Week three, build the deck and demo. Week four, run outbound to twenty named accounts and count replies, not opens.
What it costs
Most pre-seed AI teams run this on roughly $100–$400 a month plus variable compute.