The Silicon Valley Startup Quietly Building the Future

Misha Laskin doesn’t currently fit the stereotype of a Silicon Valley founder. He doesn’t discuss civilizational destiny in his posts. He doesn’t wear a vest to Sushi Tech keynote addresses. He worked for DeepMind for years, developing the math that subtly drives half of the products that investors are currently debating. At some point last year, he decided that leaving DeepMind and starting his own business would be the most beneficial course of action. Reflection AI, the business he and Ioannis Antonoglou founded, raised $8 billion in a market that is either overheating or finally becoming honest, depending on who you ask.

The number isn’t what makes Reflection intriguing. Such numbers are no longer as shocking. The pitch beneath the number is what it is. While the majority of the Valley built closed, proprietary frontier models in 2024 and 2025, Laskin observed that the Chinese open-source competition was catching up, something that very few people in San Francisco wanted to publicly acknowledge. Alibaba’s Qwen and DeepSeek were no longer merely low-cost substitutes. American startups were discreetly moving over to save money as they approached the frontier. Brian Chesky of Airbnb stated in public that Qwen is crucial to his business. It’s no small admission.

Reflection is wagering that the US needs its own open-source frontier lab, not out of patriotism but out of necessity. There is no serious dispute that the closed models from Anthropic and OpenAI are exceptional. Additionally, they are costly, take a long time to customize, and are overkill for many developers. Engineers I’ve spoken to over the past year seem to believe that the person with the largest model won’t win the next stage of artificial intelligence. Whoever makes the most practical one available will win.

You can hear the same conversation taking place at various tables in any SoMa co-working space. Inference costs are compared by the founders. Engineers are arguing over whether to pay closed-model API fees that compound at scale or refine a Qwen variant. Last year, Jerry Liu, the creator of the productivity app Dayflow, told NBC that closed-model expenses were costing each user about $1,000. For a consumer product, that is not a sustainable base. It’s an excuse to search elsewhere.

That is precisely what Reflection’s pitch tackles. The company is positioning itself as the American solution to a question that most of its peers haven’t yet asked by building models in the open and training them in the United States. It’s actually unclear if the strategy is effective. For decades, open-source companies have had difficulty making money, and the industry is harsh. The European attempt at something similar, Mistral, has struggled. Reflection may or may not avoid the same gravity.

However, the contrast is difficult to ignore. While consumer AI devices like the Rabbit r1 became jokes and Meta reported a $13.7 billion quarterly loss on Reality Labs, Laskin’s team has been performing the unglamorous task of training models that businesses might actually want to use without blowing up their runway. No headset is present. No demonstration of spatial computing. Not a keynote on transforming the human experience.

It is feasible. Similar to how Cisco subtly outlasted Pets.com in the previous cycle, Reflection emerges as one of the key businesses of this decade. It might also end up as a footnote. In any case, it’s intriguing that the most ambitious wager in Silicon Valley at the moment may be the one no one is discussing, made by a physicist who appears remarkably uninterested in being discussed.

You get the impression that the loudest period in the Valley may be coming to an end as you watch it play out. There’s a quieter movement underneath.