Why We Invested: Liquid AI
Global investment capital is pouring overwhelmingly into one trade: bigger data centers, bigger models, bigger power bills. We believe the market has not priced where intelligence actually gets deployed. Here is the number that matters: 99% of the world's processors live outside of data centers. More than two billion chips ship every year across 1.2 billion smartphones, roughly 260 million PCs, and roughly 90 million vehicles, before counting robots, wearables, and the 20 billion connected IoT devices already in the field. Every one of those surfaces is about to become intelligent. The question is not whether AI reaches the edge. The question is which company builds the software layer that gets it there. We believe that company is Liquid AI, and here is why we invested:
I. The Thesis: Cloud AI Has a Physics Problem
The dominant AI paradigm was architected for data centers, where memory is effectively infinite, compute is billed per token, and requests arrive in batches. On a device, none of that holds. Memory bandwidth is the constraint, not FLOPs. Latency is measured in milliseconds. The power budget is a battery. And the cloud itself is hitting a wall. The International Energy Agency projects data center electricity demand will more than double to roughly 945 terawatt-hours by 2030, more than Japan consumes today. There is not enough energy, semiconductor supply, or memory to serve the world's AI demand through cloud infrastructure alone. Most of the industry is answering this by compressing cloud models after the fact. Shrink the transformer, quantize it, and hope it fits. The results speak for themselves: stripped-down models that sacrifice quality, or 200 to 500 millisecond cloud round-trips that sacrifice latency, privacy, and reliability.
Liquid took the opposite approach. Instead of retrofitting data center architectures onto devices, they derive the model architecture from the device itself.
II. The Technology: First Principles, Not Fine-Tuning
Liquid AI was founded in 2023 by researchers from MIT CSAIL, the team that invented Liquid Neural Networks, a class of brain-inspired systems published across Nature Machine Intelligence, NeurIPS, ICLR, and Science Robotics. The core asset is a hardware-in-the-loop design system that takes device constraints as inputs and generates foundation model architectures optimized for that exact deployment target: CPU, NPU, or GPU. The output is the Liquid Foundation Model (LFM) family, open-weight models that collapse the size-to-quality curve. One public data point tells the story: Liquid's 2.6 billion parameter model scores 88.1 on IFEval, the standard instruction-following benchmark, against 79.0 for the original GPT-4. Frontier-class capability, running locally on a phone.
Since January 2026, LFMs have been downloaded at roughly one million per week, placing Liquid among the fastest-growing open-weight AI labs in the United States, in company with Google, Meta, Microsoft, and Nvidia.
III. The Commercial Validation: Global Investors and Partners
Liquid has already secured a strong roster of strategic investors and commercial partners to accelerate growth across core operating verticals:
AMD (NASDAQ: AMD). AMD led their $250 million Series A at a reported $2.2 billion valuation in December 2024, one of the largest Series A rounds in AI history. The commercial proof followed: in January, AMD published a joint demonstration of a fine-tuned LFM summarizing a full 60-minute meeting entirely on a Ryzen AI processor in 2.7GB of RAM. The comparable 30B-parameter transformer required 15.2GB, a 476% larger footprint, which is the difference between running on a mass-market 16GB laptop and not running at all. AMD is now the first AI PC platform with full tri-engine LFM support across CPU, GPU, and NPU.
Mercedes-Benz (XETRA: MBG). In April, Mercedes-Benz announced a multi-year partnership with Liquid AI to scale embedded, on-device intelligence across models running third- and fourth-generation MBUX in North America, built directly into the in-house Mercedes-Benz Operating System (MB.OS). Liquid's models will power the MBUX Virtual Assistant, processing speech, language understanding, and reasoning locally on the hardware already inside the vehicle, no cloud round-trip required, with first production deployment targeted for the second half of 2026. Read that timeline again. Automotive hardware cycles run up to five years; this went from partnership announcement to production path in months. Jörg Burzer, Mercedes-Benz's CTO of Development and Procurement, framed it as laying the foundation for the next generation of intuitive, multimodal in-car experiences. In an industry where time-to-market is measured in model years, Liquid ships at software speed, and one of the world's most demanding luxury OEMs just made that bet in public.
Shopify (NASDAQ: SHOP). Shopify's CTO, Mikhail Parakhin, has publicly stated that no one else is delivering sub-20ms inference on real workloads, and that Liquid models with roughly 50% fewer parameters beat Alibaba's Qwen and Google's Gemma while running 2 to 10x faster. Parakhin's own framing: that is what it takes to power interactive commerce at scale. Understand what sub-20ms unlocks. A platform processing trillions of interactions cannot put a 100-plus millisecond model in the purchase path; the latency-conversion relationship is one of the oldest laws in e-commerce, dating back to Amazon's famous finding that every 100 milliseconds of added latency cost roughly 1% of sales. At sub-20ms, generative recommendation and search stop being a batch process and become part of the live transaction, at inference costs that make trillion-request economics work. And note the strategic tell: this is a cloud deployment. The same architecture derived from edge constraints turns out to be structurally cheaper and faster inside the data center too, which means Liquid's addressable market is not edge instead of cloud. It is both. In a market generating roughly $7 trillion in annual GMV, latency is revenue, and Liquid operates where transformers cannot.
Insilico Medicine (HKG:3696). In drug discovery, Liquid's 2.6B model outperformed Google's TxGemma-27B, a model more than 10x its size, on 13 of 22 pharmacology benchmarks, running entirely on private infrastructure so proprietary wet-lab data never leaves the building. The model also posted a 98.8% success rate on industry-standard multi-parameter molecular optimization benchmarks and produced better correlation scores across 2.5 million measurements on 689 protein targets than frontier cloud models. This is the proof point that matters for every regulated industry: you no longer have to choose between state-of-the-art performance and data sovereignty, and in a projected $25 billion AI drug discovery market, that tradeoff collapsing is the entire investment case.
G42. In June 2025, Liquid entered a multifaceted commercial partnership with G42, the Abu Dhabi technology group that took a $1.5 billion strategic investment from Microsoft and operates some of the largest AI infrastructure in the world. The mandate: jointly develop and deploy private generative AI internationally, from the Middle East and North Africa to the Global South, across investment and banking, consumer electronics, telecommunications, biotech, and energy, with G42 companies Core42 and Inception providing the infrastructure and co-design layers. Announced in the wake of the U.S.-UAE AI acceleration partnership, this is the sovereign AI trade in physical form: nations and enterprises that want frontier intelligence without shipping their data to someone else's cloud.
IV. The Moat: Upstream of Everyone
Liquid sits at the interface between foundation model architecture and silicon, in deep co-design partnerships with chipmakers including AMD, Arm, and Qualcomm. Every deployment sharpens their understanding of real hardware constraints, which improves the design system, which produces better architectures, which deepens the silicon partnerships. It is a compounding loop. Here is why we believe the incumbents cannot simply copy it: the hyperscalers' entire economic model depends on the cloud round-trip. Every query routed to a data center is billed compute. They are structurally disincentivized from making on-device intelligence too good. Liquid has no such conflict.
Think of it as the ARM of AI, the cross-platform intelligence layer embedded in the silicon itself, monetizing per device rather than per token.
V. The YXS Perspective: Our Investment Playbook, Executed
Readers of our published "America 2026 Playbook" will recognize the fit. Liquid AI sits at the intersection of three of the four stacks we identified.
In the Cognitive Stack, it is Applied Intelligence in its purest form: models that do not just chat, but execute inside real systems, in cars, phones, factories, and labs. In the Energy Stack, it attacks the constraint directly. If compute is bounded by electrons, then the company delivering frontier-quality intelligence at a fraction of the energy cost is not just an AI play. It is an energy play, and one of the few in the market that reduces demand on a strained grid rather than adding to it. And through G42, it reaches the third dimension we wrote about in January: sovereignty. As capital and compute realign along national lines, the demand for AI that runs privately, locally, and under the owner's control is not a niche. It is the procurement standard for governments, banks, and critical industry across half the world.
Liquid AI is not a bet on our framework. It is the framework, priced.
Important Disclosures. This article is provided by YXS Capital for informational purposes only. It references publicly available information, including company announcements, partner publications, press coverage, and third-party industry data from sources including Bloomberg, AMD, Mercedes-Benz Group, G42, the International Energy Agency, and IDC. YXS Capital has not independently verified third-party information and makes no representation or warranty as to its accuracy or completeness. Nothing herein constitutes an offer to sell, or a solicitation of an offer to buy, any security or investment product, and nothing herein constitutes investment, legal, tax, or accounting advice or a recommendation with respect to any investment or strategy. Any offer or solicitation with respect to any YXS Capital fund will be made only pursuant to definitive offering documents, which will contain material information not included herein and which supersede this article in its entirety. This article contains statements of opinion, market views, and forward-looking statements, including projections and estimates regarding companies, markets, and technologies; such statements involve known and unknown risks and uncertainties, actual results may differ materially, and YXS Capital undertakes no obligation to update them. References to specific companies, including portfolio companies, are for illustrative purposes only, do not represent all investments made by YXS Capital, and should not be construed as an indication that any investment has been or will be profitable. Past performance is not indicative of future results. Investments in venture capital involve substantial risk, including the possible loss of the entire amount invested. Recipients should not rely on this article in making any investment decision and should consult their own advisors. Views expressed are those of YXS Capital as of the date of publication and are subject to change without notice.
