AI Unfiltered: Real Shifts Behind the Hype—From NVIDIA’s Cosmos 3 to Market Moves

NVIDIA just launched Cosmos 3, an AI model capable of directly generating robot action sequences. In our last episode, we talked about the real shifts behind the hype. Today, we're seeing two of them hit at once: a fundamental change in AI's physical capabilities, and a wave of private giants preparing to reshape the public markets. The headlines today are dominated by a few major players. NVIDIA also unveiled Alpamayo 2 Super, a 32-billion-parameter model designed to give level 4 robotaxis 360-degree perception and reasoning.

This isn't just about driving; it's about explaining decisions. Meanwhile, the private markets are getting ready for their public debut. SpaceX, OpenAI, and Anthropic are all preparing for offerings in mid to late 2026. The valuations are staggering: SpaceX is targeting one-point-seven-five trillion dollars. OpenAI is valued at 852 billion. We'll dig into the numbers behind those numbers later. They are… revealing. In funding news, Geordie AI secured thirty million dollars in a Series A. The company provides security and governance for enterprises deploying AI agents, and it’s reporting one-thousand-three-hundred percent ARR growth.

That’s not a typo. It signals that the agentic AI era is creating a new, urgent need for control. Separately, Mecka AI raised sixty million dollars to train robots using human motion data collected from body sensors and iPhones. And in research, a new framework from PromptSE AI showed it could predict drug side effects by having the AI reason through biological mechanisms, a significant step beyond simple pattern matching. Let's go deeper on NVIDIA. The launch of Cosmos 3 is not an incremental update. This is a shift in kind.

Models like GPT-4o and Gemini can describe what’s in a video. Cosmos 3 watches a video… and then generates the specific action sequences for a robot to replicate the task. We're talking joint angles, gripper positions, force feedback. It outputs movement. It does this using a Mixture-of-Transformers architecture. One part is a vision-language tower for reasoning. The other is a diffusion-based generation tower that creates the physical action sequences. It was trained on twenty trillion multimodal tokens, including four hundred million videos.

NVIDIA is offering two versions. A 16-billion parameter model, Nano, for real-time inference on a workstation. And a 64-billion parameter model, Super, for datacenter-scale synthetic data generation. Critically, it’s all available under an open license. The catch? The Edge variant, for deployment on actual hardware-constrained robots, is still forthcoming. So the brain is here. The body is still waiting. Now for the money. The planned public offerings for SpaceX, OpenAI, and Anthropic are not just big—they are a stress test for the entire market's belief in AI.

Let's look at the filings. SpaceX is targeting a one-point-seven-five trillion dollar valuation. But its S-1 filing shows the AI division lost six billion dollars in 2025 and another two-point-five billion in the first quarter of this year. That loss is being offset by Starlink's four-point-four billion dollar operating profit. Public investors will be asked to value a company where the most profitable division is underwriting the least profitable one. It’s a similar story at OpenAI. The company is valued at 852 billion dollars and has nine hundred million weekly active users.

But it expects to lose fourteen billion dollars this year. Profitability is not projected until around 2030. Anthropic is the outlier. It’s projecting its first operating profit in this quarter—559 million dollars on ten-point-nine billion in revenue. But even they caution this is supported by discounted compute contracts, and sustained profit isn't guaranteed. These aren't tech companies in the traditional sense. They are capital-intensive research projects selling access to their experiments, and the market is about to decide what that's worth.

One part of the industry is teaching machines to interact with the physical world, one joint and one action at a time. The other is preparing to sell that future on the public market, valued in trillions before it fully exists. Today, both stories accelerated. The gap between what is physically possible and what is financially promised is where the next decade will be defined.