Welcome, humans.

Where is America’s answer to China’s DeepSeek? Well, it might be this company.

ICYMI: Reflection AI just hit the news in a big way. The company is now in talks to raise $2.5B at a $25B valuation, backed by NVIDIA and potentially JPMorgan Chase.

That's up from an $8B valuation from back in October 2025, and a nearly 46x increase from their previous $545M valuation less than a year ago.

They also just signed a $6.8B deal to build South Korea's largest AI data center, with the U.S. Commerce Secretary at the signing ceremony.

Why all this money? To counter China's dominance in open-source AI. Because while you might have heard of DeepSeek, there's also…

• Moonshot AI (who makes Kimi).

• Z.ai (formerly Zhipu AI) who makes the GLM models.

• MiniMax (founded by ex-SenseTime engineers), who just IPO'd in Hong Kong.

• Alibaba's Qwen, the most-downloaded model family on Hugging Face w/ over a billion downloads.

• Plus ByteDance's Doubao and SeeDance models, and Tencent's Hunyuan.

…all releasing frontier-level open models. The US used to have Meta (who has been absent for almost two years as they rebuild their key AI lab), and Google’s Gemma 4, which has now been downloaded over 2M times, plus smaller upstarts like Arcee and their Trinity models.

However, right now, an Andreessen Horowitz partner estimates that 80% of US startups are building on Chinese base models. And zero of those frontier (meaning state of the art) open models are coming from the US (although Gemma 4 might have just changed that, at least for models in its weight class).

In our latest podcast episode, we talk to Ioannis Antonoglou, co-founder and CTO of Reflection AI, who explains why every single frontier open-source AI model right now comes from China, how mixture-of-experts architecture lets a trillion-parameter model run like a small one, and why he believes open-source will surpass closed labs.

Here's our favorite parts:

• (22:13) "Every frontier open model is Chinese": Ioannis drops the stat that stopped us in our tracks—right now, zero frontier open-weight models come from the US. Reflection wants to change that.

• (30:28) "All else being equal, open wins every day": His thesis for why open-source models will eventually be the default choice over closed ones—and it's not just idealism.

• (8:52) Mixture of Experts, explained for humans: Grant tries three different analogies to understand MoE architecture. Ioannis patiently corrects every single one. The real explanation is way more interesting.

• (34:22) "Conductors of an agent orchestra": Ioannis's vision for what coding looks like in a year—you won't be writing code, you'll be directing a team of AI agents who write it for you.

• (33:14) "It is insane how far we've come": After 14 years in AI, Ioannis says today's coding agents are "day and night" compared to where the field was. His current favorite? Claude Code.

• (12:31) The trick that makes trillion-parameter models fast: How MoE architecture lets a model with a trillion parameters run inference using only 32B active parameters at a time. That's how DeepSeek does it too.

• (28:02) Safety through transparency, not secrecy: Why Ioannis thinks open-source AI is actually safer than closed models—and the Linux analogy that makes it click.

• (43:18) "We build it in America, but we build it for the world": The most quotable line of the interview—and the philosophy behind why Reflection releases models that any country can fine-tune to its own language and culture.

Why listen to Ioannis? Because he co-created AlphaGo at Google DeepMind—the AI that famously beat the world champion at Go in 2016. He spent over a decade building DQN, AlphaZero, MuZero, and the reinforcement learning behind Gemini. Now he's raised $2B (and possibly $2.5B more…) to give frontier AI models away for free.

Why watch this? Because the (22:13) stat about China's open-model dominance and the (8:52) MoE explainer will give you a better understanding of where AI is actually headed than most headlines—whether you're building with open models, evaluating them for your company, or just trying to understand why everyone keeps talking about DeepSeek.

Watch and/or Listen now: YouTube | Spotify | Apple Podcasts

P.S. Ioannis says Reflection is building a full family of models at different sizes—including some you might be able to run locally. When Grant asked if he could run one on his laptop, Ioannis gave the most tantalizing "maybe" we've ever heard. Jump to (38:59) for that.

Real quick: Want to see your AI-adjacent product or service show up right here, below these podcast promos? Click the button below to advertise to our 675K+ readers!

More From The Neuron…

Three recent interviews you’ll definitely want to check out (pick whatever looks interesting to you and dive in!):

80% of US Factories Have Zero Robots. Google Wants to Fix That.

Google's Intrinsic CTO Brian Gerkey co-created ROS, the open-source software running 1M+ robots including NASA's. He explains why most US factories still have zero automation—and how generative AI is about to change that. YouTube | Spotify | Apple Podcasts

The Secret Labs Where AI Learns to Work

Surge AI hit $1.2B in revenue without a dime of VC. VP of Product Nick Heiner reveals why even the best models still fail ~40% of workplace tasks, and predicts a $1B company with one human employee by 2030. YouTube | Spotify | Apple Podcasts

Solo Founders Are Taking Over (Carta's Data Proves It)

Carta's CMO Nicole Baer reveals what's really happening to startups right now. If you're thinking about starting something—or wondering how AI is reshaping the startup landscape—this is the one. YouTube | Spotify | Apple Podcasts

Dive deeper with these resources:

• Reflection AI — Ioannis's company; explore what they're building and open roles

• WSJ: Reflection AI Eyes $25B Valuation — the funding story

• WSJ: Korea Data Center Deal — the $6.8B sovereign AI play

• NVIDIA Blog: The Future of AI Is Open and Proprietary — featuring Reflection's CEO alongside Mistral, Perplexity, and Cursor

• Reuters: $25B Valuation Report — more details on the raise

Last week, we went live with the "AI for Total Beginners" 5-Level Starter Stack—a framework for going from "I've never used AI" to "AI saves me 10 hours a week." No coding required. Read the companion blog here. Here are some highlights:

• (34:54) Level 1 — Projects: They set up a "Job Search Manager" project from scratch, showing how organizing your AI into project folders means you never have to re-explain yourself.

• (43:14) The laziest hack that actually works: They asked Claude to write its own custom instructions by looking up Anthropic's best practices documentation. The AI literally taught itself how to help them better.

• (1:23:51) Level 3 — Skills: The unlock that changes everything. They turned a single conversation into a reusable skill that auto-generates custom cover letters for every job it finds—callable with one sentence.

• (1:44:10) Level 4 — Automations: They scheduled the whole thing to run every 15 minutes, creating a fully automated job search agent that finds postings, filters by salary and location, and writes cover letters while you sleep.

It's at 5,700+ views and counting, so you know this resonate with people!

And if you haven’t subscribed yet, please do! Click the image below to go to our channel and hit “subscribe” to get notified right when new videos go live.

We have a goal to hit 50K subscribers by the end of the year (if not 100K), and we’re almost to 20K! If you like learning about AI, and already watch some of our videos, do us a favor and click here to subscribe today.

Dive deeper with these resources:

• Reflection AI — Ioannis's company; explore what they're building and open roles

• WSJ: Reflection AI Eyes $25B Valuation — the funding story

• WSJ: Korea Data Center Deal — the $6.8B sovereign AI play

• NVIDIA Blog: The Future of AI Is Open and Proprietary — featuring Reflection's CEO alongside Mistral, Perplexity, and Cursor

• Reuters: $25B Valuation Report — more details on the raise

Stay curious,

The Neuron Team

P.P.S: Love the newsletter, but don’t want to receive these podcast announcement emails? Don’t unsubscribe — adjust your preferences to opt out of them here instead.