Run the Numbers · Oct 5, 2026 · 54m · 39m saved
With Emily Man, Partner, Enterprise AI & Fintech, at Primary Venture Partners
A growth-stage startup's chief operating officer told Emily Man that the company's AI spending went from $200,000 last year to a run rate of $4 million.
Most AI money so far has gone into chips, data centers and models. Man invests in the layer above them, the applications, and she thinks the lasting value there comes from a company's own data and decisions rather than from the model it rents.
"My framework is the roadkill, the cockroaches and the cheetahs." — Emily Man
Emily Man, Partner for Enterprise AI and Fintech at Primary Venture Partners, on Run the Numbers, has spent most of her career investing in software for chief financial officers and has invested at every stage, from early rounds to growth, at Redpoint and Point72 Ventures before joining Primary.
The full interview is covered here so you can skip it. 54 minutes of audio, 15 minutes of reading.
Here are the 15 insights that matter.
Key Takeaways
Less than 10% of the world has tried a large language model, Man said, so applications have barely started to spread
Anthropic reportedly spends about 2.3x its payroll on compute, and she sees inference becoming a cost line as large as general and administrative expense
One company's AI spend went from $200,000 to a $4 million run rate in a year
She expects employee-level budget controls on AI models, like the controls that came with corporate cards
A model anyone can buy is "inherently beta"; the advantage comes from a company's own context and workflows
Payroll and money-movement software are "cockroaches" that survive AI; simple small-business tools are "roadkill"
Companies that vibe-coded their own software are finding it takes half to a whole engineer to maintain
AI works best where answers are objectively right or wrong, which is why code has led
Buyers tell an AI sales avatar more than a human rep, a Primary portfolio company found
1. The Five-Layer AI Stack
The host asked Man to set out the layers of the AI stack for finance people who pay the bills without always knowing what they buy.
"Maybe just to start like this is all stolen from Jensen, but I think he has a very distinct way of talking about the layer cake that is underpinning the AI era, right?" — Emily Man
She described a supply chain of five layers: energy, chips, infrastructure, models and applications
Energy is the power that feeds the hardware. Chips go into infrastructure, which is wired together for the models built on top. The applications connect what the models can do to a human result.
Most funding so far has gone into layers two through four
She compared the build-out of chips, infrastructure and models to the railroads and the broadband boom, where the pipes had to be laid first. Her partners at Primary have invested across those layers.
"But I myself focus a lot more on that last level, the applications." — Emily Man
The power grid is not ready
The host said he had not realized as a layman that there was not enough energy. Man said Primary has looked at small nuclear reactors placed next to data centers.
"But the way that the power grid works for us today is just not going to be set up anywhere near enough for what the demand is going to be from all the data centers that are coming online." — Emily Man
