Fal has doubled its valuation since April as customers pile onto its platform. Meanwhile, data center builders are bolting retired jet engines onto their sites because the grid can't keep up.
Photo by Taylor Vick on Unsplash
Fal has raised $1.3 billion at a $16 billion valuation. In April, the San Francisco startup was worth $8 billion.
The round was led by Bessemer Venture Partners, with Nvidia, Sound Ventures and T. Rowe Price taking part. Fal also announced its Series E and Series F together, which brings the total raised across the two to $1.65 billion. Sequoia led the earlier round.
Fal hosts and runs AI models that generate video, images and audio. That puts it in the business of inference: running a model that has already been trained so it can produce outputs for users. Training gets the headlines. Inference is where customers actually pay.
Bloomberg reported that Fal passed $800 million in annualized revenue earlier this year. The company's valuation has doubled in the time since April. Investors are pricing in demand that is still climbing.
Fal is not alone. Baseten, Fireworks AI and Modal are also growing quickly and raising large rounds at higher valuations. The pull is strongest from customers building on open source models, who need somewhere to run them without buying their own hardware.
There is a catch. Several people familiar with these companies' financials say gross margins are thin, because buying or leasing compute is so expensive. Revenue is rising fast, but a large share of it goes straight to whoever owns the chips and the buildings around them.
The money is chasing inference. The constraint is electricity, and nobody can raise a round big enough to fix that quickly.
If inference demand is running ahead of supply, as this round suggests, the next question is what limits supply. Increasingly, the answer is electricity.
The International Energy Agency says connecting 100 megawatts to the grid in Northern Virginia takes about seven years. A Mitsubishi Power executive told Bloomberg that most new power capacity is expected to arrive by 2031. The delays are already costing business. Last month Microsoft pulled out as anchor customer for Nscale's data center in Loughton, England, after the local utility said it could not supply enough power until the early to mid-2030s.
Some builders are no longer waiting. The New York Times reported that certain AI data centers are generating their own power with engines pulled from retired aircraft and rebuilt to burn natural gas. “It’s all about getting power as fast as possible,” Rystad Energy analyst Audun Martinsen told the Times. He said jet-powered units can have a data center running in 24 months.
The adopters are big names. The Times, citing state permits and satellite imagery, reported that xAI installed eight converted turbines in Mississippi to help power its Colossus 2 data center in Memphis. They come from ProEnergy, which rebuilds a GE engine commonly found on Boeing jets.
OpenAI followed in 2025 with gas turbines at its first Stargate site in Abilene, Texas. Crusoe, the Denver-based startup building that campus, ordered 13 ProEnergy units in April. In September, Crusoe walked away from a $1.25 billion deal to buy 29 turbines from Boom Supersonic, which is developing a supersonic passenger jet and has started selling a version of its engine as a data center power plant. Crusoe told TechCrunch it still plans to use turbines.
The Trump administration is adding pressure by encouraging data centers to produce their own electricity.
The trade-offs are plain. According to the Times, converted jet engines are less efficient and pollute more than turbines designed for conventional gas plants, and it takes more of them to produce the same power. Some experts say efficiency and environmental concerns may matter more in a year or two.
Put the two stories side by side and a pattern shows up. Capital is flowing freely to the companies that sell inference. The physical capacity underneath them is slow to build, and operators are accepting dirtier, less efficient power to shorten the wait.
That suggests the scarce resource in AI is shifting from chips toward electricity and the time it takes to secure it. If that holds, the winners may be the companies that lock down power early, not just those with the best models or the fastest growth.
It also raises a question about margins. If inference providers already pay heavily for compute, and the buildings housing that compute depend on expensive, makeshift power, the squeeze on gross margins could last. That is analysis, not something the sources state. But a $16 billion valuation implies investors expect demand to outlast the shortage.
Sector watch: cloud
For the first time, Microsoft reported Azure's quarterly revenue in dollars. The number lets investors compare the big clouds directly, and shows how much of the growth rests on a few AI labs.
Sector watch: pricing
OpenAI and Anthropic have cut prices by as much as 80% in months. Cheaper tokens are real, but the evidence points to better models, not lower bills, as what drives usage.