45x in 12 Months: The $25B AI Bet With No Shipped Product

Reflection AI is trying to raise $2.5 billion at a $25 billion valuation. There is no product.

The numbers have moved fast, even by AI industry standards. Reflection closed a $130 million round at a $545 million valuation in March 2025. Seven months later, a TechCrunch report covered a $2 billion raise that pushed the valuation to $8 billion. Now, according to Reuters, the Nvidia- and JPMorgan Chase-backed startup is in talks for another $2.5 billion that would value it at $25 billion pre-money. The jump from $545 million to $25 billion in under 12 months is a 45-fold run on a company that has not shipped its core product.

The pitch is not a product demo. It is a geopolitical argument.

Reflection AI was founded in March 2024 by Misha Laskin, who led reward modeling for Google DeepMind's Gemini project, and Ioannis Antonoglou, who co-created AlphaGo — genuine DeepMind reinforcement learning pedigree. The company launched a research agent called Asimov in July 2025. As of early March 2026, that agent remains on a waitlist and runs on third-party models, not Reflection's own. The open-weight frontier model at the center of the company's pitch has not been publicly released. The team numbers roughly 60 people, mostly AI researchers and engineers. No customers. No shipped model. No public benchmark results from the flagship system.

What Reflection has is a story that Washington wants to hear.

The Trump administration has made clear it wants a U.S.-based competitor to DeepSeek — the Chinese lab whose January 2025 model release sent shockwaves through Washington and contributed to the export control review that followed. The logic runs through national security channels: if American researchers and sovereign governments default to open-weight models for inference and fine-tuning, the question is which lab supplies them. Reflection is positioning itself as the answer. JPMorgan's participation is reportedly through its Security and Resiliency Initiative — a name that says plainly where the logic originates.

The investors who backed the October 2025 round — Nvidia, Disruptive, DST, 1789, B Capital, Lightspeed, GIC, Sequoia's profile notes CRV, Citi, and individuals including Eric Yuan and Eric Schmidt — were buying the same thesis. Nvidia's check is the most legible: it wants open-source AI running on its hardware, and it wants that ecosystem to be Western. The others followed a strategic lead.

The $25 billion valuation is where the geopolitical story runs into basic skepticism. No AI lab has closed a $25 billion funding round on the basis of an unreleased model and a waitlist. The valuation chain reflects anticipated national importance, not demonstrated capability. "We realized this was a gap we were the only ones positioned to fill," a company representative told Turing Post in March. The gap is real: there is no American open-weight frontier model. But filling a gap and being worth $25 billion are different things, particularly when the timeline to product is unknown and the team is 60 people.

The counterargument is that the option itself has value. If the U.S. government is going to direct capital toward Western open-source AI, the lab with DeepMind pedigree, Nvidia backing, and a waiting list of researchers wanting to work on a non-Chinese frontier model is exactly the asset that scales in a policy-driven funding environment. The investors are not paying for what Reflection has built. They are paying for what it would be worth if Washington keeps spending.

Whether that bet is correct depends on whether the administration's stated preference for a U.S. DeepSeek translates into sustained procurement, favorable export rulings, and preferential treatment from cloud providers — and whether Reflection's team of 60 can actually build the model before a better-capitalized competitor arrives. Meta, with Llama and a research team an order of magnitude larger, is the obvious alternative. Google DeepMind has the weights. Mistral has been publishing. The window for a standalone Western open-source lab to establish itself as the default is not infinite.

The Reuters reporting on the current round is here. TechCrunch covered the October 2025 $2 billion raise and the founding team here. The Turing Post profile of the company's product status is here. Sequoia's profile of the Asimov agent, which notes it runs on third-party models, is here.