Liquid AI’s Hasani says 80% of enterprise AI never reaches production

Ramin Hasani, co-founder and chief executive of Liquid AI, has a number for the corporate AI graveyard: 80% of production-grade efforts never reach production quality. He draws it from roughly 200 Fortune 500 engagements over two and a half years, recounted in a Latent Space interview. His diagnosis cuts deeper than the statistic. Failed pilots push companies back towards cloud services, he argues, while developers chase raw token counts when, by his estimate, about 90% of generated tokens are useless.

Watch Desk analysis

What happened

Hasani's core argument is that the next decade of AI belongs to whoever solves efficiency and customisation: the difference between a demo and a model that genuinely works in a car, on a laptop or inside an enterprise's existing systems.

The company he co-founded exists to prove it, and its roots are a two-millimetre worm. Modelling human neurons stalled at TU Wien in 2013 because their spiking behaviour resists the mathematics used to train networks, so his team turned to C. elegans, a creature that shares about 78% of its genome with people and whose 302 neurons behave in smooth, continuous, fully differentiable ways. The resulting liquid time-constant networks, he says, steered a robot better than any system of the era, and Daniela Rus brought the group to MIT in 2017.

The technical claim is that transformers won what he calls the "hardware lottery": they are essentially matrix multiplications, exactly what GPUs accelerate, but attention's cost grows quadratically as context grows. Liquid builds hybrids, chosen by an automated search system it calls meta-AI and tuned to the device they run on, with LFM 2.5 on the market and LFM3 in development.

The commercial claims are unusually specific. A Mercedes-Benz partnership announced in April 2025 puts a roughly 600 MB multimodal model inside the car, running on a chip that costs about $100, with first deployments this year on North American Generation 3 vehicles. Liquid models serve Shopify's Shop app, over a billion requests a month by his account, and open-weight downloads have passed 40 million, running at roughly 1.5 million a week.

The business model skips token APIs entirely. Liquid assigns an applied ML team to each design partner, charges for the engagement and then yearly per-device licences, and is beta-testing a platform it says will let customers, in his words, own their own intelligence.

The numbers

• 80% never reach production
  Hasani's failure rate for production-grade enterprise AI efforts, drawn from work with roughly 200 Fortune 500 companies over two and a half years.
• 90% of tokens wasted
  His estimate of the share of generated tokens that create no value, as developers inside firms chase maximum output.
• A 600 MB model in the car
  The Mercedes-Benz system handles voice, camera input and text reasoning on a chip costing about $100, deployed over the air.
• A billion requests a month
  The volume Hasani claims for Liquid models serving Shopify's Shop app stack.
• 40 million downloads
  Total open-weight downloads so far, at roughly 1.5 million a week, on the company's account.

Why it matters

Enterprise AI pilots stalling is the industry's open secret. Here is a practitioner who has sat across the table from some 200 large companies putting a figure on it. If the estimate is even half right, much of the money going into corporate AI is buying demos, and the useful questions change: what does production quality mean here, and how much of what we generate does anyone actually use?

It also sketches a different economic shape for the field. Hasani argues value is shifting from inference tokens to customisation, spanning pre-training through post-training, and he is blunt that on-device AI is not yet ready to carry the load: in Apple Intelligence and Galaxy AI, he says, roughly 90% of calls still go to the cloud because device models are not powerful or reliable enough. That is either the market gap or the wall, depending on whose models you believe.

Our read

Treat the 80% as an informed practitioner's estimate, not audited data, and remember that the man quoting it sells the cure. But the pattern behind the number matches what enterprises say about their own programmes, and the sharpest thinking in this interview is economic rather than architectural: tokens are cheap to make and easy to waste.

The worm-to-production story is a genuinely interesting counter-bet in a field that has largely converged on one design, and Liquid is one of the few companies testing it against real commercial traffic rather than benchmarks alone. If you run an AI programme, the takeaway is not to rush out and buy liquid models. It is to define production quality before the pilot starts, and measure your token waste honestly.

What to watch

• The customisation platform
  Whether the beta promising customers their own trained intelligence actually ships, and which design partner signs first.
• Mercedes, on the road
  Whether in-car deployments genuinely reach North American Generation 3 vehicles this year at scale.
• LFM3
  Whether the promised gains in quality, stability and latency control show up in benchmarks others can run.
• The 80% figure
  Whether other enterprise practitioners publicly endorse or dispute the failure rate.

Discussion spark: If 80% of enterprise AI really never reaches production, who is failing: the models, the buyers, or an industry that sells pilots faster than it can run them?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.