Do this

Stay safu

"A few things jump out to me from this list, and they’re more interesting than any individual ranking. 1. AI is starting to fragment into multiple frontiers. For the last few years, “AI progress” mostly meant better language models. This list suggests that’s changing fast: World Labs is betting on spatial intelligence, Physical Intelligence/Skild on general robotic intelligence, humans& on collaborative intelligence, ElevenLabs on voice, and Etched/Lightmatter on entirely different compute assumptions. World Labs explicitly frames spatial intelligence as a frontier beyond words, while Physical Intelligence is already showing improved generalization in robotic foundation models. My view: the next major AI breakthrough may not look like GPT-6. It could be a model that understands space, acts reliably in the physical world, or coordinates with humans over days rather than answering prompts. 2. The center of gravity is shifting from training → inference. Groq, Fireworks, Baseten, Together, Etched, Fractile, Modal — that is a lot of companies essentially betting that running intelligence becomes the bigger problem than creating it. I think that thesis is probably right. Agents change the economics dramatically: one chatbot response might require a handful of model calls; an autonomous workflow can require hundreds or thousands. Suddenly latency, throughput, memory and cost per token become strategic variables rather than infrastructure trivia. Etched now explicitly describes itself as building “frontier inference clusters,” not merely chips. 3. Compute itself is becoming a distributed-systems problem. Lightmatter and Credo are particularly interesting to me for this reason. We’ve spent years obsessing over the GPU, but the next limitation is increasingly how quickly and efficiently thousands of accelerators communicate. Lightmatter is explicitly building photonic interconnects for scaling AI training and inference, while Credo’s products target high-speed, energy-efficient connectivity for AI infrastructure. This is why I think Lightmatter may be one of the most underappreciated companies on this list. If the bottleneck moves from computation to communication, value migrates with it. 4. The NVIDIA moat is attracting attacks from every direction—not just “another GPU.” That may be the most important hardware pattern here: Cerebras: make the chip enormous. Etched: specialize the chip around AI workloads. Groq: optimize around inference execution. Tenstorrent: rethink the broader compute/software architecture. Lightmatter: attack interconnect rather than compute. The interesting question isn’t necessarily “Who replaces NVIDIA?” I doubt anyone cleanly does. It’s whether NVIDIA’s current vertically integrated stack gets unbundled at specific layers. 5. Open models may push value away from the model itself. DeepSeek, Qwen, Kimi, Z.ai, MiniMax, Mistral and Reflection collectively suggest something uncomfortable for frontier-model valuations: high-quality intelligence may become much less scarce. If the capability gap between closed and open models keeps narrowing, the profit pools move toward things that remain scarce: compute → inference → proprietary data → distribution → workflow integration → physical infrastructure. That would make companies such as Fireworks, Together, Surge, CoreWeave and Lightmatter more important precisely because models themselves become cheaper. 6. China is no longer one “DeepSeek wildcard.” It looks like an ecosystem. This stood out strongly. DeepSeek Kimi Qwen Z.ai MiniMax ByteDance Seed That’s not one exceptional company. It’s several independent teams exploring different model, pricing, open-weight and distribution strategies. My expectation is that Chinese labs put sustained downward pressure on model prices globally. Their biggest disruption may therefore be economic rather than simply winning benchmark #1. 7. Robotics is converging toward the foundation-model playbook. Physical Intelligence and Skild are much more interesting to me than “who has the coolest humanoid demo.” The big conceptual transition is: one robot → one task → bespoke software becoming: many robots → many tasks → shared foundation model. Physical Intelligence explicitly describes its work as general-purpose physical intelligence, and its recent research focuses on robotic foundation models generalizing to new tasks. If that abstraction works, the robot body eventually risks becoming the commodity. The intelligence layer becomes the platform. That is why I’d watch PI/Skild more closely than Figure. 8. humans& is interesting because it challenges a surprisingly unquestioned assumption. Almost every current AI product assumes: 1 human ↔ 1 AI But real work happens through teams, institutions and relationships. The genuinely interesting humans& question isn’t “can they make another assistant?” It’s whether multi-person persistent context, shared memory, delegation and AI-mediated coordination become their own model architecture/product primitive. That is much more speculative than Claude getting better at coding—but that’s exactly why I like it as a wildcard. 9. World models might connect several currently separate categories. World Labs is perhaps the company I’d watch most carefully. Spatial/world models initially sound like a 3D-generation category. But if successful, they potentially connect: video generation → simulation → robotics training → planning → embodied reasoning → interactive environments. World Labs itself is explicitly connecting spatial intelligence, world models, simulation and robotics, including through its July acquisition of SceniX. If this direction works, “world model” could eventually become as important a phrase as “language model.” 10. Data isn’t dying; the valuable kind of data is changing. I think Surge is more interesting than it initially appears. The internet gave frontier labs enormous quantities of cheap text. But once models know most of what’s easily scraped, improving them increasingly involves expert judgment, difficult reasoning traces, environments, evaluations and human preference data. So the data industry may transition from: “How many labels can you produce?” to “Can you reliably provide intelligence the model doesn’t already have?” That creates a potentially much higher-value business. What I’d personally watch over the next 12–24 months Not fundraising announcements. These signals: World Labs: Do world models start being used for robotics/simulation rather than just impressive 3D generation? Physical Intelligence / Skild: Can the same model reliably transfer across robots and genuinely unseen tasks? humans&: Is there actually a new collaboration primitive here, or does it collapse back into a conventional agent product? Lightmatter: Does photonic scale-up move from promising engineering into significant hyperscaler deployments? Etched: Can specialized architecture demonstrate an order-of-magnitude economic advantage, rather than 20–30%? Groq / Fireworks / Baseten / Together: Which layer of inference actually captures margin? There probably won’t be room for all of them. DeepSeek / Qwen / Kimi: Does the open/Chinese frontier remain ~months behind closed US models, or effectively converge? Surge: Does expert post-training data become materially more valuable as generic data saturates? Thinking Machines / SSI / Reflection: Do any of these supposedly exceptional research teams reveal an actual technical idea that the incumbents aren’t already pursuing? The broader pattern I see The first phase of modern AI was basically: More compute → bigger model → better chatbot. The next phase looks much messier: better architectures + cheaper inference + interconnect + memory + world models + physical intelligence + specialized data + persistent agents. That’s actually bullish for smaller, unusual companies. When progress depends almost entirely on spending $100B training a frontier model, incumbents dominate. When progress starts depending on solving dozens of different bottlenecks, a company like Lightmatter, World Labs, humans&, Etched or Physical Intelligence can discover something disproportionately important. That’s the main reason I find this particular list interesting: AI development looks like it’s broadening from a model race into a systems race." - ChatGPT Sol 5.6

@theworldlabs — Probably my favorite wildcard. Spatial intelligence/world models are a fundamentally different frontier from better chatbots, with potential implications for robotics, simulation and 3D creation; reports in early 2026 put its fundraising valuation around the ~$5B area, which is still relatively modest for that upside. @humansand — One of the strangest and therefore most interesting bets: persistent, collaborative AI designed around groups of humans rather than one-user/one-assistant interactions; at $4.48B after a $480M seed, you’re paying far less than frontier-lab valuations for a genuinely differentiated thesis. @LightmatterCo — AI’s next bottleneck may be moving information rather than calculating it, and Lightmatter’s photonic interconnect thesis attacks that at the physics level; its last clearly reported valuation was about $4.4B, which looks interesting relative to the size of the problem. @Etched — A wonderfully binary bet: throw away general-purpose GPU flexibility and build silicon specifically for transformers. Its valuation was around $4.5–5B in early 2026 before its latest round, so the upside is enormous if specialized inference actually wins. @physical_int — “One foundation model for many robots” is one of the cleanest disruptive theses in AI; the catch is valuation, with funding talks already above $11B, so a lot more optimism is priced in. @tenstorrent — Rather than merely making another accelerator, it’s attempting an alternative AI compute ecosystem spanning accelerators, RISC-V and software; extremely technically ambitious and one of the few bets that could attack CUDA lock-in from below. @SkildAI — The idea of selling a reusable robot brain across embodiments is huge, but at >$14B after its January 2026 round, I find the concept more attractive than the valuation. @thinkymachines — Mira Murati plus an unusually dense frontier team makes it impossible to ignore, particularly if its work produces new training/customization paradigms; however, the completed seed valuation was roughly $10–12B, while later $50B discussions make it much less of a valuation play. @deepseek_ai — Less wildcard now because it has already proven itself, but its core thesis remains highly disruptive: frontier capability at radically different economics. The downside is that its current fundraising target reportedly implies roughly a $74B valuation, so much of the discovery has already happened. @ReflectionAI_ — Frontier-quality open intelligence would be an enormous structural event, but Reflection has already jumped to a reported $27.5B valuation, making this a high-upside but increasingly expensive speculative bet. @cerebras — Wafer-scale compute is genuinely different rather than “NVIDIA but smaller”; if the architecture proves superior for enough training/inference workloads, it could create an entirely new compute branch. @GroqInc — The thesis becomes more compelling in an agent world: if agents make hundreds of sequential model calls, ultra-fast deterministic inference suddenly matters far more than it did in chatbot economics. @Kimi_Moonshot — Kimi is increasingly credible at the frontier and could be China’s next DeepSeek-style surprise, although the latest completed valuation is already around $30B, with discussions potentially going substantially higher. @AsteraLabs — Not sexy conceptually, which is exactly why it’s interesting: AI clusters increasingly live or die on connectivity and memory plumbing, giving Astera leverage over a bottleneck most people don’t think about. @HelloSurgeAI — The contrarian data bet: frontier progress may increasingly depend on expensive expert human judgment rather than more scraped tokens, potentially giving high-quality post-training data suppliers disproportionate power. @ssi — No product and enormous uncertainty, but Ilya Sutskever claiming a different research path makes the upside exceptionally asymmetric; SSI was valued around $32B before NVIDIA’s recent partnership. @FireworksAI_HQ — If open models keep closing the capability gap, the company that can make them dramatically cheaper and faster to operate becomes powerful infrastructure. @baseten — Its bet that enterprises will increasingly mix open and closed models creates a potentially huge neutral inference layer; its latest financing reportedly values it around $11–13B. @credoai — Making high-speed copper work deeper into AI clusters can materially alter networking cost and power economics. @togethercompute — A neutral training/inference layer becomes more important in a world with dozens of competitive open models instead of three dominant APIs. @ZhipuAI / Z.ai — China’s independent frontier-model ecosystem is developing extremely quickly, and Z.ai has emerged as one of its strongest players, albeit now at a very rich public valuation. @MiniMax_AI — Multimodal breadth plus China’s intense model competition makes it a credible source of unexpected capability gains. @ByteDanceSeed_ — Enormous compute, strong research, and real-world deployment scale make it one of the most underappreciated frontier organizations. @Alibaba_Qwen — Less wildcard than some names above, but the downstream influence of a leading open-weight model family makes it exceptionally important. VAST Data — AI increasingly bottlenecks on feeding compute rather than merely owning compute, making data architecture surprisingly strategic. @cohere — Sovereign/private models could become much more important if governments and large enterprises reject dependence on a handful of US frontier APIs. @MistralAI — Its combination of European sovereignty and relatively open deployment makes it strategically more interesting than another closed frontier challenger. @CoreWeave — Less conceptually exotic, but the emergence of specialized AI cloud as a category has already altered how frontier labs access compute. @nebiusai — Similar neocloud thesis with additional geographic and ecosystem optionality, especially in Europe. @CrusoeAI — The increasingly important thesis here is that power, not chips, may become AI’s ultimate scaling constraint. @ElevenLabs — If agents become conversational and persistent, voice becomes infrastructure rather than a content-generation niche. @modal — Serverless GPU infrastructure could become a natural execution layer for highly dynamic agent workloads. @Figure_robot — Humanoids offer enormous upside, but the company is farther toward vertically integrated product deployment than the purer robotics-intelligence bets. @fractile_ai — Attempt to rethink inference hardware around the specific structure of modern models rather than legacy compute assumptions.