NextFin News - A Cambridge, Massachusetts startup spun out of MIT is arguing that the most valuable layer of artificial intelligence is about to move out of the data center and onto the devices people already own. Liquid AI, a foundation-model company built on "liquid neural networks" rather than the transformer architecture that powers ChatGPT and Gemini, has raised $250 million and signed production deals with Shopify, Mercedes-Benz, and MacPaw on the premise that latency, privacy, and cost will push inference to the edge. The bet is not that cloud AI disappears; it is that the cloud loses its monopoly on running intelligence.
For three years, the dominant assumption in generative AI has been straightforward: models get bigger, training runs get longer, and inference happens in hyperscale data centers owned by a handful of cloud providers. Liquid AI is built on the opposite conviction. Its models are designed around the hardware constraint first — the fact, as the company puts it, that 99% of the world's processors live outside data centers and come with hard limits on latency, memory, power, and connectivity. That framing turns the industry's scaling race into a distribution question: if capable models can run on a phone, a car, or a laptop without a cloud round-trip, the economics of AI shift from who owns the most GPUs to who owns the device layer.
The company's commercial footprint is still small but widening. In November 2025, Shopify struck a multi-year, multi-million-dollar agreement to license Liquid's models for search and a co-developed generative recommender system; the first production deployment is a text model that completes a search in under 20 milliseconds. In April 2026, Mercedes-Benz entered a multi-year partnership to run Liquid's models on-device in its MBUX vehicles, with a North American rollout expected in the second half of 2026. And in August 2026, MacPaw announced a partnership to bring on-device AI to millions of Mac users through its Eney assistant. Around those deals sits $250 million of Series A funding raised in December 2024, led by AMD, which values the company at $1 billion.
The market backdrop explains why investors are listening. One research firm projects the global edge AI market will grow from roughly $30.9 billion in 2026 to $225.5 billion by 2035, a 24.7% compound annual rate; another puts the market at $118.7 billion by 2033, growing 21.7%. An industry forecast cited in market research holds that 40% of enterprise AI inference will run at the edge by 2030, up from 10% in 2024. Those are forecasts, not results — but they mark the direction of travel that Liquid AI is trying to monetize.
Why the Cloud's Grip on Inference Is Under Pressure
The first pressure point is latency. In interactive applications — product search, in-car voice assistants, desktop agents — a round-trip to a cloud endpoint adds delay that compounds with every step of an agentic workflow. Shopify's chief technology officer, Mikhail Parakhin, put the benchmark bluntly in the partnership announcement:
"I've seen a lot of models. No one else is delivering sub-20ms inference on real workloads like this."
A sub-20-millisecond response is not a marginal improvement; it is the difference between an AI that feels instantaneous and one that feels like it is thinking.
The second pressure point is privacy and data sovereignty. When an assistant runs on a Mac or inside a car, personal data never leaves the device. That matters for regulated industries and for consumers who would rather not have every query streamed to a third party. Oleksandr Kosovan, MacPaw's chief executive, framed the design principle directly: "We believe intelligence should live where people work: private by design, fast by default, and be able to reach the cloud when that's the better tool." The hybrid phrasing is important — this is not a cloud-versus-edge argument, but a question of which workloads belong where.
The third pressure point is cost. Liquid AI's open-weight license is free to download, run, and fine-tune, including in commercial products, until a company passes $10 million in annual revenue. When the marginal cost of running an agent drops toward zero at the device layer, the competitive advantage in inference shifts from who has the cheapest API to who has the most efficient local implementation. That dynamic puts downward pressure on cloud inference pricing precisely as hyperscalers are racing to cut token costs.
The fourth pressure point is hardware. AMD's strategic partnership with Liquid AI optimizes the models for AMD Instinct GPUs, CPUs, and neural processing units, and Apple's silicon already carries substantial on-device AI capacity. The hardware is no longer the bottleneck it was five years ago; the question is whether model architectures can be compressed enough to fit it. Liquid AI's answer is that they can: its LFM2.5-2.6B model, released in July 2026, decodes at 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395 CPU, while staying under 2.5 gigabytes of memory.
What Liquid's Architecture Actually Changes
Liquid AI's distinguishing claim is architectural, not just engineering. Its models are based on "liquid neural networks," a class of systems that uses differential equations to model dynamic, continuous-time decision-making, rather than the static feed-forward layers of a transformer. The practical consequence is adaptive computation: the network can allocate more processing to ambiguous inputs and less to straightforward ones, instead of running the same fixed number of operations on every token.
The company's model family spans roughly 230 million to more than 40 billion parameters, designed to slot into different hardware envelopes. Its mixture-of-experts design, LFM2.5-8B-A1B, carries 8 billion total parameters but activates only 1 billion at inference time. The 2.6-billion-parameter model — 2.69 billion, to be precise — was pre-trained on about 34 trillion tokens and supports a 128,000-token context window, dimensions that fit on consumer hardware while retaining enough context for agentic planning and tool use.
The efficiency claims are specific and testable. On production-like tasks, Liquid says models with about 50% fewer parameters have outperformed Qwen3, Gemma3, and Llama 3 while delivering inference speeds two to ten times faster. Shopify's CTO repeated the comparison in his own words:
"in some use cases, a model with ~50% fewer parameters beats Alibaba Qwen, and Google Gemma, and still runs 2-10x faster."
Those are controlled comparisons on specific workloads, not a general capability ranking — but they are the kind of benchmark that matters for latency-sensitive production systems.
The architecture also targets long-context efficiency. Liquid's LFM2.5-DSpark release in August 2026 claimed up to 3.2 times faster inference across hardware from Nvidia's H100 data-center GPU to Apple's MacBook, using speculative decoding to accelerate output. If that holds at scale, it suggests the efficiency advantage is not confined to tiny models — it travels up the size spectrum.
Still, the burden of proof is on the startup. Transformers benefit from a decade of ecosystem lock-in: tooling, talent, optimized kernels, and a massive body of fine-tuning recipes. Liquid AI is asking developers to adopt a different programming model for reasons that are economic today but unproven at the scale of a frontier lab. The company's own CEO, Ramin Hasani, frames the ambition cautiously: "At Liquid, our mission is to build the most capable and efficient AI systems at every scale."
The Second-Order Question: Who Captures the Value
Here is the question the market is not asking loudly enough. If inference migrates to the edge and its marginal cost falls toward zero, who actually captures the value — and who loses the franchise?
The most likely answer is counter-intuitive. The migration of inference to devices does not destroy the cloud; it clarifies it. Training frontier models remains capital-intensive and compute-constrained. In August 2026, a months-old infrastructure startup backed by Nvidia, Volta Infrastructure, was reported to have signed a $10 billion, six-year compute supply contract with an AI lab — one of the largest infrastructure commitments disclosed in the sector. Deals of that size are a bet that training compute is the scarce resource, not inference. If Liquid AI is right, labs may lose pricing power on inference while remaining tethered to the physical infrastructure required to build the next generation of intelligence. The "compute landlord" thesis survives — it just relocates its leverage upstream.
The real winners of an edge-native world are the owners of the device layer. Apple, Mercedes-Benz, and MacPaw each control a surface where AI meets the user, and each has an incentive to keep intelligence local. For Apple, on-device models defend the privacy narrative that underpins its brand. For Mercedes, in-car inference eliminates connectivity dependency and reduces the energy cost of cloud data centers. For MacPaw, the combination of Liquid's models with its own Elix inference engine and Mnemos memory layer could become shared infrastructure distributed to thousands of Mac developers through its Setapp marketplace. The locus of value in the AI stack is migrating: training compute stays scarce, but inference becomes a commodity that lives on the device.
That migration has a third-order implication for model providers like Liquid AI itself. In a world where inference is free at the edge, the monetizable asset is not the token — it is the model's fit to a specific hardware-software stack and the developer tools that make deployment trivial. Liquid's LEAP platform, which lets developers fine-tune a model and bake it to a runtime, is the product that matters if the company wants recurring revenue rather than one-time licenses. The open-weight license that makes adoption frictionless also caps the upside: free below $10 million in revenue is a growth strategy, not a margin strategy.
The Counter-Thesis, and What Would Prove It Right
The strongest case against Liquid AI's thesis does not come from skepticism about efficiency — it comes from the hyperscalers themselves. Microsoft, Google, Amazon, and Meta are all pushing inference toward the edge: custom silicon, distributed inference, and on-device models that still report back to the cloud. If the cloud providers own the edge runtime, the "escape" from the cloud monopoly becomes a migration within the same ecosystem. The device layer would be local, but the value capture would not.
A second objection is technical. Agentic workflows that plan, call tools, and reason over long horizons may hit a quality ceiling on small models. A 2.6-billion-parameter model can handle a well-defined task; it is less clear that it can replace cloud-scale reasoning for open-ended problems. If the most valuable AI workloads remain too complex for on-device models, the edge becomes a complement to the cloud rather than a substitute — and Liquid AI's total addressable market shrinks accordingly.
Both objections are serious, and neither is answered by benchmark charts alone. The falsifying signal is concrete: if by the end of 2027 the share of enterprise AI inference running at the edge remains below roughly 15% — a trajectory inconsistent with the 40%-by-2030 projection — or if Liquid AI's commercial footprint stays limited to pilot-scale deployments rather than expanding beyond its current anchor partners, the structural-displacement thesis weakens materially. Watch the Mercedes rollout in the second half of 2026: a production vehicle with on-device generative AI is a far stronger signal than another partnership announcement.
What Comes Next
The near-term outlook is sentiment-driven. Edge AI is a crowded theme, and valuations in the space can move on narrative before revenue catches up. Liquid AI's $1 billion valuation from its December 2024 round is a high bar for a company whose disclosed commercial relationships are still in early production.
The medium-term outlook depends on fundamentals: whether the Shopify deployment expands beyond search, whether the Mercedes integration reaches production vehicles on schedule, and whether the MacPaw assistant ships to millions of users later this year. Each of those is a binary, observable milestone.
The long-term outlook is structural. If the economics of latency, privacy, and cost continue to favor on-device inference, the industry's center of gravity shifts toward efficient models and the hardware that runs them. AMD, Apple, and other edge-silicon vendors benefit from the same trend that benefits Liquid AI. Pure-play cloud inference providers, by contrast, face a slow compression of token pricing as capable models move closer to the user.
Three scenarios frame the path. In the base case, edge inference grows steadily, Liquid AI becomes a credible specialist supplier to device-layer companies, and the cloud retains dominance over training and complex reasoning. In the upside case, on-device models cross a quality threshold that makes them viable for agentic work, and the value capture shifts decisively to the device layer. In the downside case, hyperscalers absorb the edge into their own stacks, small models hit a capability ceiling, and Liquid AI remains a niche architecture for latency-sensitive tasks.
The cloud did not lose the computing war to the PC, and it will not lose AI to the edge overnight. But Liquid AI's argument is not that the cloud dies — it is that the cloud's monopoly on intelligence does. If inference becomes a commodity that lives on the device, the companies that own the surface where users meet AI will write the next terms of the market, and the model providers that survive will be the ones that fit the hardware, not the ones that merely scale the biggest.
Explore more exclusive insights at nextfin.ai.
