The Real AI Race Isn’t Model vs. Model Anymore. Here’s Where Poolside Actually Fits.

The closed frontier is down to two names, and one of them just changed

Start with the giants, because Poolside’s whole pitch only makes sense against them. Anthropic raised $65 billion in May 2026 at a $965 billion valuation, overtaking OpenAI, and now reports a $47 billion annualized revenue run rate. It wins roughly 70% of head-to-head enterprise deals against OpenAI. OpenAI still holds the consumer lead through ChatGPT. Google sits in the same tier as both. xAI does not: nine of its eleven original co-founders have left, and the company has effectively been absorbed into SpaceX at a combined valuation near $1.25 trillion. Meta has fallen further back still.

Three companies, in other words, control almost all of the closed frontier’s value. That concentration is exactly the thing Poolside co-founder Eiso Kant says he’s building against: “I want a world with 50 to 100 AI labs, not four.”

China isn’t catching up anymore. On open weights, it’s already ahead

The second half of the picture is the one Poolside is responding to. By this spring, eight of the top ten Chinese AI models were open-weight, and Alibaba’s Qwen family alone had already passed Meta’s Llama in cumulative Hugging Face downloads. Chinese-developed models now account for roughly 30% of all open-model downloads globally. DeepSeek and Qwen combined went from 1% to 15% of the global AI market in twelve months.

Moonshot’s Kimi K3, released July 17 at 2.8 trillion parameters, is being billed as the largest open-source model in the world. Zhipu AI’s GLM-5.2 launched a month earlier under an MIT license with a full 1-million-token context window. The gap between the single best closed model anywhere and the best Chinese open model is down to roughly 36 Elo points, and it’s narrowing. China is now setting the terms of what “open” means in this industry, by volume, by download count, by default choice.

That backdrop is why the same week Poolside launched, the White House accused Moonshot of distilling Kimi K3 from Claude Fable 5 in violation of Anthropic’s terms of service, Treasury threatened sanctions, and a coalition of roughly 200 companies, including Y Combinator, urged Trump not to ban Chinese open-weight models outright. Nvidia’s Jensen Huang publicly pushed back on what he called Wall Street “misunderstanding Kimi again.” Open weights stopped being a technical detail. It’s a live policy fight now.

Where Poolside sits

Poolside isn’t trying to out-scale anyone. Its new model, Laguna S 2.1, is 118 billion total parameters against Kimi K3’s 2.8 trillion, roughly 24 times smaller. Co-founder and co-CEO Jason Warner announced it plainly: “Laguna S 2.1 sits at the top of its weight class and competes with open models many times larger, while remaining small enough to run on a single NVIDIA DGX Spark.” One developer, @analogalok, ran the full model on a single consumer RTX 4090 and posted the numbers: 11.5 tokens per second decode, 250,000 tokens of context, no data center required.

Be honest about the ceiling, too. Per VentureBeat’s own comparison, GPT-5.6 Sol still scores 88.8 on Terminal-Bench 2.1, Claude Fable 5 scores 88.0, and Kimi K3 itself scores 88.3. Laguna doesn’t touch that tier. What it does is prove that an American, openly-licensed model can be efficient enough to run on hardware a single developer owns, which is a genuinely different value proposition than “biggest model wins.”

Kant’s framing for why that matters: “you should be able to own intelligence no one can turn off.” That’s a specific pitch, aimed at exactly the companies worried about depending on a closed API that a vendor could reprice, restrict, or shut off tomorrow.

The model was never the whole product

A raw Laguna-vs-Kimi comparison undersells what’s actually happening in this market. In March 2026, the LangChain engineering team took a coding agent from the Top 30 to the Top 5 on Terminal Bench 2.0, moving its score from 52.8% to 66.5%, without changing the underlying model at all. Every point of that gain came from the surrounding system: self-verification loops, loop detection, better context management. Multiple independent write-ups of that case study converged on the same number: with the model held constant, the harness accounts for roughly twenty times the performance difference that swapping models does.

A “harness” is just the software wrapped around a model that decides what tools it can use, what it’s allowed to touch, how it remembers what it already tried, and when to stop and check its own work. The model supplies judgment. The harness supplies everything that turns judgment into a finished, reliable task. By mid-2026, every serious AI lab ships both a model and a harness, because the models have converged enough that the harness is where the real competition happens.

Kant has a strong, specific opinion here, and it cuts against a popular industry assumption: “I think MCP and tools are stupid.” His argument is that stuffing fifty tool definitions into a system prompt is the wrong shape for long-horizon work, not that agents shouldn’t have capabilities at all. His preference is giving an agent direct access to a real environment, a virtual machine with the binaries installed, rather than a menu of tool calls to choose from. Whether or not you agree, it’s a real architectural bet, not a footnote.

Poolside’s own launch made the same point structurally, even if nobody framed it that way. The announcement didn’t just say “here are the weights.” It named the harnesses the model already works inside: kilocode, Cline, pidotdev, OpenClaw, and Poolside’s own agent, pool. A model with no harness behind it is a research artifact. A model shipped alongside five compatible harnesses is a product.

The cloud layer is the third leg, and Poolside treated it as a first-class decision, not an afterthought

The same launch tweet listed exactly how and where the model runs: weights on Hugging Face in BF16, FP8, INT4, and NVFP4, official GGUF and MLX conversions, served through OpenRouter or Vercel’s AI Gateway, deployable via Baseten, runnable locally through Ollama, vLLM, or SGLang, and optimized for NVIDIA TRT-LLM on a single DGX Spark. That’s Poolside explicitly acknowledging that “here’s an impressive model” and “here’s something a team can put into production” are two different claims, and that the second one depends entirely on infrastructure choices most benchmark coverage never mentions.

Eiso Kant has said the same thing about his own company’s internal build process. Poolside’s “Model Factory” runs 10,000 to 20,000 experiments a month with fewer than 105 people, and it’s what let this specific model go from the start of training to a public launch in under nine weeks. Kant’s own summary of why that speed matters more than raw scale: competing at the frontier is about iteration speed, not scale, and iteration speed requires infrastructure that matches the ambition of the research. That’s an infrastructure thesis. Model quality was the output. The Factory was the investment.

The level set

Put the three pieces together and Poolside’s real position gets clearer than any single benchmark comparison shows. The closed frontier has consolidated to two or three companies with enormous capital and enormous leverage over anyone who builds on their APIs. The open frontier is currently being defined by Chinese labs shipping bigger models faster than anyone in the West. Poolside isn’t trying to win either of those games outright. It’s betting that the unit of competition is now the full stack, model plus harness plus the infrastructure that runs it, and that owning all three openly is worth more than owning the single largest model for a news cycle.

Whether that bet pays off depends on something no benchmark measures: whether teams building real products would rather assemble their own stack from an open model, an open harness, and infrastructure they control, or just call the best closed API and not think about any of it. Right now, both bets are being placed at once, by companies with real capital behind each one. The real score is whether the open, assemble-it-yourself stack can compound as fast as the closed one can spend.

By Anthony Batt – 20+ years building software and digital media products at scale. Podcasting host at Future-Proof Podcast by CO/AI.