Korva is a scroll of the internet’s best ideas.
Reselling GPUs caps what you can charge
Margin and volume
Profit is margin times volume, so thin margins on huge volume can still work, and investing in those businesses is not crazy. The harder problem is that many of today's AI infrastructure companies are GMV businesses, where the main activity is reselling tokens, GPUs, and similar goods.
The electricity ceiling
What matters is the terminal value of the good delivered to the customer. If the product is hosting GPUs so someone else can run their models, the customer has a fair question: why pay more than the cost of the electricity?
Margins at scale
That is why margin matters even for young companies. A seed or Series A company can run thin margins, but as it scales toward a public company it needs a margin structure that supports a public business. Inference companies like Fireworks sit in the mid 30s percent, against 65 to 80% for traditional software.
Why do marketplaces report GMV at all?
Gross merchandise volume is the total value of sales passing through a platform, before the platform's share is taken out. eBay, for example, reports gross merchandise volume alongside its much smaller revenue. For a business that resells goods, GMV can be enormous while the revenue it keeps is a thin slice. Investors compare GMV with gross profit to see how much of each sale the business holds on to.
Why do software margins reach 65 to 80 percent?
Software is copied at almost no cost, so serving one more customer adds little expense. Most of the spending sits in engineering and sales, which do not grow with each additional user. AI inference differs because each request runs on GPUs that consume electricity and hardware time. Each query to a large model requires computation on specialized chips, a cost that traditional software mostly does not carry.
How does electricity shape GPU hosting costs?
Data centers pay for the power that GPUs draw and for the cooling that removes the resulting heat. Nvidia's published specification for its H100 data center GPU lists a power draw of up to 700 watts per chip. A rack of such chips runs continuously, so the electric bill accumulates around the clock. Data center operators track efficiency with a ratio called power usage effectiveness, which compares total facility power with the power reaching the computing equipment.
More from 20VC with Harry Stebbings
