For most of the last decade, the AI frontier had a simple shape: a handful of general-purpose labs—OpenAI, Google DeepMind, Anthropic—racing along one axis, model scale, toward one destination, artificial general intelligence. That shape is now breaking apart. A generation of leaders who helped build those incumbents has walked out to run smaller, sharper organizations, and they are not all running in the same direction.
Four of them define the new topology of the frontier. Each has taken a different bet on what matters most: safety, human collaboration, spatial understanding, or raw scale wired to policy. None of these labs is trying to be everything. That focus is the story.
Ilya Sutskever's SSI: the safety-first bet
Ilya Sutskever, OpenAI's former chief scientist, has built the most single-minded of the new labs. Safe Superintelligence describes itself, per its own materials on ssi.inc, as a company with “one goal and one product: a safe superintelligence”—and states plainly that “superintelligence is within reach.” The organizing principle is sequencing: SSI says safety and capabilities should advance in tandem, with safety always staying ahead.
What changed SSI from a thesis into a scaling operation is compute. In a joint announcement carried by nvidianews.nvidia.com and globenewswire.com, SSI and NVIDIA disclosed a long-term strategic partnership in which NVIDIA also made an investment in the company, with access to Vera Rubin systems expected to increase SSI's compute by an order of magnitude. SSI's own updates page (ssi.inc) frames that as a 10x jump. Sutskever's read, in the NVIDIA announcement, is that SSI now has “research that is worthy of scaling up.” No revenue or margin figures have been disclosed on SSI's public pages—this is a lab optimizing for a single long-horizon output, not a product line.
Mira Murati's Thinking Machines: the human-centered bet
Mira Murati, formerly OpenAI's chief technology officer, has anchored Thinking Machines Lab on a different premise. The lab's stated position, on thinkingmachines.ai, is that “the future worth building is human,” and that AI trained for autonomy alone crowds people out. That is a direct rebuttal to the fully-autonomous-agent framing dominant elsewhere.
The engineering follows the thesis. Thinking Machines has described (thinkingmachines.ai) “interaction models” that continuously take in audio, video, and text, then think, respond, and act in real time, arguing that “interactivity should scale alongside intelligence” rather than being an afterthought. On distribution, the lab has gone open: it released Inkling, its first open-weights model, a multimodal mixture-of-experts transformer with 975B total parameters, 41B active, and a 1M-token context window, with full weights available. Like SSI, it has secured frontier-scale compute—a multi-year NVIDIA partnership to deploy at least one gigawatt of Vera Rubin systems, with NVIDIA taking a significant investment (thinkingmachines.ai). No 2026 revenue figures appear in its public materials.
Fei-Fei Li's World Labs: the spatial bet
Fei-Fei Li, the ImageNet pioneer, is betting that the language-model paradigm has a ceiling that only geometry can break through. Per worldlabs.ai, Li's position is that spatial intelligence is AI's next frontier, and that world models are needed for machines to understand and interact with the real world. World Labs frames its work as generative AI that can “understand and interact with the world,” spanning uses from storytelling to simulation.
The robotics angle sharpens the distinction. World Labs argues that simulation is what unlocks the next generation of robotics, and that training robots is not like training large language models—a pointed claim in a field where most bets route back to the LLM. To push it, the company announced the acquisition of SceniX to advance spatial intelligence for robotics (worldlabs.ai). It is the clearest architectural departure among the four: a lab whose core primitive is space, not text.
Alexandr Wang's Meta Superintelligence Labs: the scale-and-policy bet
Alexandr Wang, the Scale AI founder now leading Meta Superintelligence Labs, is the one insider building inside an incumbent rather than outside it—which makes his bet the reverse of the others. His framing is about inputs at national scale. At the AI Impact Summit, per a transcript on singjupost.com, he named four building blocks for AI policy and scale: talent, energy, data, and compute.
The lab is shipping and provisioning against that view. Meta Superintelligence Labs announced Muse Spark, the first model in a new Muse series and, per about.fb.com, the first large language model from the lab. On compute, Meta signed a multi-year, multi-generation agreement to deploy up to 6 gigawatts of AMD Instinct GPUs, described as a portfolio approach to a more diversified stack (ir.amd.com, about.fb.com). Wang has paired the scale story with governance language—that Meta's models need model cards, evaluation benchmarks, red teaming, and risk assessments before release—and a long-term vision he calls “personal superintelligence.”
What to watch
The through-line is fragmentation into focused bets, and the near-term signals are concrete:
- Does safety-first survive scaling? SSI's 10x compute jump is the first real test of whether its “safety stays ahead” sequencing holds once the system is large enough to matter.
- Does open-weights become a wedge? Thinking Machines shipping Inkling with full weights sets it against the closed incumbents; watch whether adoption follows the philosophy.
- Does spatial intelligence produce a shipped robotics result? The SceniX acquisition is an input; a working simulation-to-robot pipeline would validate Li's departure from the LLM paradigm.
- Does scale-plus-policy cohere? Wang has to make 6 gigawatts of GPUs, a new model line, and a governance posture add up inside Meta—the hardest integration of the four.
The incumbents still hold the center. But the frontier is no longer a single race down one axis. It is four labs, four theses, and four founders who decided the bet worth making was the one nobody else was.
