Liquid AI
Private venture-funded AI lab; monetizes via enterprise licensing of on-prem/edge models, cloud API access, and co-development/consulting (e.g., automotive, commerce). Pre-scale, not yet profitable.
Two real priced rounds; ~7.8x step-up from the ~$303M seed valuation (Dec 2023) to ~$2.35B post at Series A (Dec 2024). No Series B or secondary disclosed as of mid-2026; ~$2.35B remains the latest priced mark. Seed valuation is the disclosed post-money; the ~$37.6M vs ~$46.6M spread reflects an initial announcement later revised to total disclosed.
Earnings, margins, COGS & capex
Private, pre-revenue-scale. ~$297M raised across a Dec-2023 seed (~$37.6M announced, ~$46.6M total disclosed, at ~$303M valuation) and a Dec-2024 $250M Series A (AMD Ventures lead) at ~$2.35B post. No audited financials, no company-disclosed revenue; a third-party estimate pegs ARR near $13M but this is unverified. Value is architecture + team (liquid neural networks) + strategic partnerships (AMD, Mercedes-Benz, Shopify), not current cash flows. Company has publicly targeted profitability via API + licensing + services, an aspiration, not a reported result.
Revenue trend
Margins
n/a
n/a
n/a
COGS structure
Not disclosed. Structurally, inference COGS should be below transformer peers because LFM/LFM2 target far lower memory and higher tokens/sec on commodity/edge hardware (e.g., LFM2.5-1.2B cited at ~719MB memory and ~70 decode tok/s on a Samsung Galaxy S25 Ultra via llama.cpp, Jan 2026); training COGS (compute) remains the dominant cost line.
Capex
Not disclosed as a line item. Frontier model training compute is the principal capital sink; AMD partnership provides hardware optimization/access that may offset some GPU spend. No owned-datacenter disclosures.
Latest earnings
n/a
No formal guidance. Company has publicly stated an aim of profitability and general availability of LFMs (cloud API + on-prem), with model releases rolling out through 2025-2026 (LFM2 Jul 2025, LFM2.5 Jan 2026).
- Total raised
- ~$297M (seed + Series A)
- Last valuation
- ~$2.35B post-money (Dec 2024)
- Lead investor
- AMD Ventures (Series A)
- Flagship edge model
- LFM2.5-1.2B (~719MB RAM on Galaxy S25 Ultra, on-device)
Growth drivers
- Enterprise on-prem/edge licensing where data sovereignty + low latency rule out cloud LLM APIs
- Automotive embedded intelligence (Mercedes-Benz in-car voice/assistant, first production targeted H2 2026)
- Commerce (multi-year Shopify partnership; models live in Shopify search/recommendation)
- AMD co-optimization opening AMD-silicon device/enterprise channel
- Efficiency edge — sub-billion-to-1.2B models running under ~1GB RAM on phones -> viable on phones/embedded/robotics
Bull & bear
If small, efficient, on-device models are where the next unit-volume of AI inference actually lives, Liquid AI owns a genuinely differentiated architecture, a Tier-1 founding team, and marquee proof points (AMD, Mercedes-Benz, Shopify) -- a credible path to a category-defining edge-AI franchise well above the current $2.35B mark.
- Real architectural edge: LFM2.5-1.2B running on a phone in ~719MB of RAM at usable decode speed is not a transformer parlor trick -- it opens device/embedded TAM the big clouds monetize poorly
- AMD as investor + hardware co-optimizer gives a distribution and silicon tailwind as AMD contests NVIDIA
- Anchor enterprise deployments (Mercedes-Benz embedded voice, Shopify commerce search) validate on-prem/edge willingness-to-pay in regulated, latency-sensitive verticals
- Founding team from MIT-CSAIL (Daniela Rus lab) is among the strongest technical benches outside the top-3 labs -- attracts talent and durable IP
- Edge/sovereign inference is structurally underserved by OpenAI/Anthropic cloud-first models -- Liquid isn't fighting them where they're strongest
A ~$2.35B valuation on an unverified ~$13M ARR estimate, in a segment where hyperscalers give away capable small models (Phi, Gemini Nano, Llama) and silicon vendors bake AI on-device, is a bet on architecture novelty surviving both commoditization and a brutal capital race -- the base rate for sub-scale labs is acqui-hire, not IPO.
- Free, well-distributed small models from Microsoft (Phi), Google (Gemini Nano) and Meta (Llama) cap pricing power for a paid edge model
- Valuation-to-revenue gap is extreme; monetization (licensing + API + services) is early and unproven at scale
- Transformer efficiency gains (quantization/distillation/MoE) keep eroding the liquid-architecture advantage
- On-device AI is increasingly a silicon feature (Apple, Qualcomm, Arm) -- vendors may not need a third-party model
- Concentration on a handful of anchor partners; the Mercedes-Benz first production deployment is still a target (H2 2026), not yet shipped -- delay or non-renewal would materially dent the story
- Staying at the frontier is capital-hungry; competing against firms with 10-100x the war chest risks a down-round or absorption
What it is worth
Last-priced private round (Dec 2024 Series A) + qualitative venture read; no public comps multiple is meaningful given unverified revenue.
Small-model commoditization by hyperscalers + silicon-native on-device AI cap pricing; capital race forces a down-round or acqui-hire below $2.35B despite strong technology.
Holds around the ~$2.35B Series A mark; grows into the valuation via enterprise licensing + partner-led deployments, with a future up-round contingent on demonstrated ARR traction and the Mercedes production launch landing.
Category-defining edge-AI franchise; on-device/sovereign inference scales into automotive, robotics, and enterprise licensing -> next round well above $2.35B, potential multi-$B+ franchise if ARR compounds off anchor logos.
Anchor is the ~$2.35B post-money from the AMD-led $250M Series A (Dec 2024), still the latest priced round (no Series B disclosed as of mid-2026). No public price, no audited revenue; a ~$13M ARR third-party estimate (unverified) implies a ~180x ARR multiple if taken at face value -- a bet on architecture + team + TAM, not current economics. Treat any implied multiple as indicative only.
SWOT
Strengths
- Differentiated non-transformer architecture (liquid neural networks) from MIT-CSAIL founders (Ramin Hasani, Mathias Lechner, Alexander Amini, Daniela Rus) -- deep technical moat and credibility
- Best-in-class efficiency — small models matching/beating larger transformers on targeted benchmarks at a fraction of memory/compute
- Strategic backer + hardware partner in AMD (investor and co-optimization), plus marquee deployments (Mercedes-Benz, Shopify)
- Edge/on-device positioning sidesteps head-to-head cloud-scale spend war with OpenAI/Google/Anthropic
Weaknesses
- Tiny revenue vs valuation -- ~$2.35B on an unverified ~$13M ARR estimate implies steep execution risk
- No public financials; monetization model still maturing (licensing + API + services unproven at scale)
- Brand/distribution far behind Llama/Gemini/Phi in developer mindshare
- Dependent on a small set of anchor partners; concentration risk
- Architecture novelty cuts both ways -- smaller tooling/ecosystem than the transformer stack
Opportunities
- On-device/edge AI wave (phones, cars, robotics, IoT) where efficiency and offline operation are decisive
- Data-sovereignty / regulated verticals (auto, healthcare, defense, finance) needing on-prem models
- Robotics + embodied AI where low-latency, low-power inference matters
- Ride AMD's push to contest NVIDIA in AI silicon as the software showcase
- Licensing efficient models to hardware OEMs and enterprises as a per-device/per-seat annuity
Threats
- Free open-weight small models from far better-capitalised labs attack Liquid AI's exact price point: its own LFM2.5 launch post benchmarks against Alibaba's Qwen3-1.7B, Google's Gemma 3 1B IT, Meta's Llama 3.2 1B Instruct and IBM's Granite-4.0-h-1b — a competitive set that is Apache-2.0 or open-weight and costs licensees nothing, so Liquid must monetise a benchmark delta over a zero-price alternative rather than a capability gap.
- Platform owners are absorbing the on-device model layer Liquid AI sells into: Apple's Foundation Models framework ships a system on-device model to developers and, at WWDC 2026, added a public LanguageModel protocol (iOS/macOS 27) letting developers swap between local and cloud inference through one API — which Google's 8 Jun 2026 announcement implements for cloud Gemini via the Firebase Apple SDK — giving device OEMs and app developers a default on-device path that requires no third-party LFM licence.
- Severe funding and scale disadvantage versus the frontier labs Liquid AI competes with for talent, compute and enterprise deals: its last disclosed round was the $250M Series A announced 13 Dec 2024 with no Series B announced as of Jul 2026, against Anthropic's $65B Series H at a $965B valuation (May 2026) and OpenAI's $122B raise at an $852B valuation (Mar 2026) — roughly two to three orders of magnitude of capital.
- Revenue appears concentrated in a small number of named, non-exclusive partnerships whose economics are undisclosed and whose deployments are unproven: Shopify (multi-year search/recommender licence announced 13 Nov 2025, 'financial terms are not disclosed') and Mercedes-Benz MBUX (multi-year deal announced 23 Apr 2026, with first production deployment only targeted for H2 2026) — so a single partner's renewal or in-housing decision is material, and the Mercedes revenue has not yet shipped to production.
- Liquid AI's non-transformer architectural differentiation is narrowing by its own hand and by ecosystem gravity: its 2026 releases are increasingly hybrid mixture-of-experts designs (LFM2-24B-A2B, Feb 2026; LFM2.5-8B-A1B, May 2026) that converge on mainstream sparse-transformer practice, and it ships LFM2.5-230M through transformer-optimised runtimes (llama.cpp, MLX, vLLM, SGLang, ONNX) — commoditising efficiency into a tooling-and-quantisation race rather than a defensible architecture.
- Strategic dependence on AMD is a reach constraint in the on-device market: AMD Ventures anchored the Series A and Liquid AI's optimisation story runs across AMD's CPU/GPU/NPU stack, but the high-volume edge silicon that its smartphone, vehicle and IoT targets actually run on is dominated by Qualcomm, Apple, Arm and MediaTek — so its closest silicon ally does not control the sockets its models most need to win.
Moats, dependencies & bottlenecks
Moats
Genuine IP + research lead, but architecture advantages in AI have historically been copied/eroded fast.
Elite bench; but talent is mobile and acqui-hire-able.
Measurable today; commoditization pressure from free small models is intense.
Distribution + validation, but partnership-based, not lock-in.
Dependencies
Investor + hardware/compute + co-optimization partner Lead Series A investor; strategic alignment is an asset but ties roadmap/optics to AMD's silicon fortunes.
Frontier training still leans on scarce accelerators despite the efficiency thesis.
Training + API hosting Standard hyperscaler dependency for the cloud-served side.
Revenue + validation Early revenue concentration; the Mercedes production launch (targeted H2 2026) and renewals are load-bearing for the narrative.
Pre-profit; future rounds needed -- down-round risk if AI-lab sentiment cools.
Advantages
- Architectural differentiation (liquid neural nets) -- not a me-too transformer lab
- Demonstrated extreme efficiency enabling true on-device/offline inference
- Tier-1 research pedigree and IP from MIT-CSAIL
- AMD strategic backing (capital + silicon + go-to-market)
- Production and near-production deployments in demanding verticals (automotive, commerce)
Weaknesses
- Valuation far ahead of demonstrated, verifiable revenue
- Opaque financials; monetization unproven at scale
- Small share of developer mindshare vs Llama/Gemini/Phi
- Customer/partner concentration
- Exposure to small-model commoditization by hyperscalers and silicon vendors
Bottlenecks
- Converting technical/benchmark superiority into recurring enterprise revenue at scale
- Developer ecosystem + tooling maturity relative to the transformer stack
- Capital required to hold the frontier against far-better-funded rivals
- Distribution reach vs free hyperscaler small models
- Talent retention amid intense AI-lab poaching
Top signals & trends
Top signals
Strategic, not purely financial -- signals hardware conviction in the architecture.
Marquee, safety-critical vertical; first production deployment targeted H2 2026 (not yet shipped).
Second anchor logo; Liquid-powered search live on storefronts.
Concrete edge-efficiency proof point (LFM2.5-1.2B under ~720MB on a phone).
Rich multiple; heavy execution risk priced in.
Direct pressure on paid-edge-model economics.
Trends
Core tailwind for Liquid's thesis; edge/robotics/IoT unit volumes.
Favors deployable, non-cloud-locked models.
Compresses pricing power for a standalone paid model.
Expands edge-AI market but risks disintermediating third-party model vendors.
Sub-scale labs face tougher rounds and consolidation pressure.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Investor + accelerators + co-optimization; strategic silicon supplier.
GPUs for frontier training.
Amazon Web Services / Microsoft Azure / Google Cloud Cloud training + API hosting infrastructure.
Private; model distribution and open-source tooling channel.
Multi-year embedded in-car voice/AI partnership (Apr 2026); first production deployment targeted H2 2026, North America.
Multi-year commerce AI partnership (Nov 2025); Liquid-powered search live in production.
Data-sovereignty, latency-sensitive verticals via licensing + on-prem deployment.
Open-weight Llama small models are the free default for on-device/edge developers.
Phi line directly targets efficient small-model / edge use cases.
On-device Gemini Nano shipped into Android/Pixel; huge distribution.
On-device AI in mobile/auto silicon; can bundle models with hardware.
Native on-device models on iPhone reduce need for third-party edge vendors.
Private (France); efficient open-weight models, similar efficiency positioning.
Private; smaller/cheaper API tiers pull cloud demand, less direct on edge.
Private; frontier cloud models (context only, not an edge competitor per se).
Private; open-model hub/distribution and small-model ecosystem gravity.