Apoha Raises $36M to Teach AI How Matter Behaves in 2026

Every frontier AI model on earth has learned from the same exhausted well: text scraped off the internet and images lifted from the web. Apoha just raised $36 million on the conviction that the next leap will not come from a bigger model trained on that same data, but from an entirely new kind of data that does not exist online at all. The London startup teaches machines to read the wave forms that materials produce when shaken in liquid, capturing how a substance smells, tastes, and reacts in ways no language model has ever encountered.

What Actually Happened

On June 3, 2026, Apoha emerged from stealth with $36 million in total funding, combining a 2024 seed round with a fresh raise this spring led by European venture firm Singular. Participating investors include Draper Associates alongside existing backers Redalpine, Seedcamp, Wilbe, and Nucleus, plus a grant from Innovate UK. The company was founded in 2021 by CEO Shamit Shrivastava, a mechanical engineer with a PhD from Boston University and postdoctoral research at Oxford who holds patents on liquid wave-form analysis, and COO Anshika Srivastava, a former executive director at Goldman Sachs.

Apoha builds AI models for designing new materials, spanning proteins, food products, and even paints, using a data type the company calls "liquid intelligence." Its method measures the wave forms a material generates when it is suspended in a liquid and then subjected to controlled physical stresses. Those patterns, the company argues, correlate to properties like smell, taste, and reactivity that conventional AI systems have no way to perceive. The first commercial product, called VIBE, short for Variations in Inter-facial Behaviour Under Excitation, produces more than 1,000 distinct numerical descriptors of a material's behavior in a single run that takes minutes rather than the days or weeks traditional lab tests require.

The early results are concrete, not theoretical. In a partnership with pharmaceutical giant Boehringer Ingelheim, Apoha's system showed greater than 90% precision in identifying high-risk antibody candidates and outperformed 12 industry-standard tests across 236 clinical-trial antibodies. In food, it helped a company find a plant-based protein substitute within two weeks. A separate collaboration with Ethris applies the approach to predicting the behavior of lipid nanoparticles. The company has completed roughly 40 customer projects with a team of just 25 employees, a sign that the technology is generating commercial pull rather than sitting in a lab.

Why This Matters More Than People Think

The AI industry has spent three years quietly panicking about running out of data. The frontier labs have already trained on essentially the entire public internet, and the marginal text or image left to scrape adds little. Synthetic data helps at the margins but risks models learning from their own echoes. Apoha sidesteps the entire problem by manufacturing a category of data that was never on the internet to begin with: direct physical measurements of how matter behaves. That reframes the data ceiling from a wall into a door, because the physical world is effectively infinite as a data source if you build the instrument to read it.

This matters because the highest-value problems left for AI are increasingly physical, not linguistic. Designing a drug, formulating a battery electrolyte, engineering a protein, or inventing a better food texture are not problems you solve by predicting the next word. They require understanding how molecules and materials actually behave under real-world conditions, and that understanding has historically been locked inside slow, expensive wet-lab experiments. Apoha's claim is that it can compress weeks of lab work into minutes while generating richer data than the lab produced, which would attack the single biggest bottleneck in materials science and drug discovery at once.

There is a structural reason investors are interested in proprietary physical data rather than another model. Models are increasingly commoditized, with open weights from Meta, Alibaba, and DeepSeek closing the gap on closed labs every quarter. Data that no one else has is the opposite of commoditized. If Apoha's wave-form measurements genuinely capture properties that conventional characterization misses, the company is building a moat made of a dataset that competitors cannot simply download or distill. In an industry where model advantages evaporate in months, a proprietary data-generation engine is the rarer and more defensible asset.

The economics reinforce the point. A single wet-lab assay to characterize an antibody or a novel polymer can cost thousands of dollars and tie up specialized equipment and staff for days, and large pharmaceutical and materials firms run such tests by the tens of thousands every year. If Apoha can deliver richer characterization in minutes for a fraction of the cost, the savings compound across an entire research pipeline, not a single experiment. That is the kind of order-of-magnitude shift in cost and speed that turns a niche tool into infrastructure, because it does not just make existing research cheaper, it makes whole classes of experiments economical that were previously too slow or expensive to attempt at all. The companies that adopt it first could iterate through candidate molecules and formulations at a pace their competitors physically cannot match.

The Competitive Landscape

Apoha is entering a crowded and well-funded race to point AI at the physical world. Isomorphic Labs, the DeepMind spinout, is using AlphaFold-derived models to design drugs. Microsoft released MatterGen, a generative model for inventing new materials. Orbital Materials and a wave of other deep-tech startups are training models to discover compounds computationally. Each of these, however, largely works in silico, predicting structure and properties from existing databases and simulation. Apoha's bet is different in kind: it generates fresh empirical data from real physical samples rather than predicting from what is already recorded.

That distinction is the crux of its differentiation. A model like MatterGen is only as good as the experimental data it was trained on, and high-quality physical measurements are scarce, inconsistent, and expensive to produce. Apoha is selling the layer underneath, the measurement engine that produces the empirical ground truth those models are starved for. In theory, that makes Apoha complementary to the generative players rather than purely competitive with them, because a generative materials model trained on a thousand VIBE descriptors per sample would be far better grounded than one trained on sparse legacy lab data.

The historical parallel is the way new instruments, not new theories, have repeatedly unlocked entire sciences. The microscope created microbiology, the telescope created modern astronomy, and X-ray crystallography made structural biology possible by revealing the shape of DNA. In each case the breakthrough was a device that let humans perceive something previously invisible, and a flood of discovery followed. Apoha is positioning VIBE as an instrument of that lineage for material behavior in liquids. If the analogy holds even partially, the company is selling shovels in a gold rush where everyone else is selling maps of where the gold might be.

Hidden Insight: The Next AI Moat Is a Sensor, Not a Model

The conventional wisdom is that AI advantage comes from algorithms and compute. Apoha's thesis quietly inverts that. Its core intellectual property is not really a neural network; it is a measurement apparatus and the physical insight that wave forms in excited liquids encode material properties. The AI is downstream of the data. This is a profound reframing because it suggests the durable frontier of artificial intelligence over the next decade may belong to whoever can instrument the physical world, not whoever has the cleverest transformer architecture. The bottleneck is perception, not cognition.

That has been the quiet pattern in every domain where AI actually changed an industry. Self-driving progress was gated less by the driving policy than by the lidar, radar, and camera rigs that let cars perceive the road. Protein folding leapt forward because decades of crystallography and cryo-electron microscopy had built a dataset rich enough to train on. In each case the sensor and the data came first, and the model followed. Apoha is making an explicit wager that materials and chemistry are the next domains to follow that sequence, and that the company holding the novel sensor captures the value the model-builders cannot.

The deeper implication touches how we think about intelligence itself. Today's frontier models are, in a literal sense, deprived of senses. They have read everything humans wrote about taste and smell and texture, but they have never tasted, smelled, or felt anything. Apoha is attempting to give machines a crude form of those missing senses through physical wave data, which is a fundamentally different project from making a language model larger. If it works, it points toward a future where the most valuable AI systems are the ones wired to novel sensors that let them perceive slices of reality humans cannot, rather than the ones that merely recombine human text more fluently.

The bear case, however, is serious and deserves to be stated without softening. Skeptics point out that "liquid intelligence" is an unproven concept dressed in evocative language, and the core scientific claim, that wave forms reliably encode properties like smell and reactivity across wildly different material classes, is far from established. Critics argue the impressive Boehringer Ingelheim result on antibodies may not generalize to paints or foods, and that a method tuned for one material family could collapse on another. The risk is that Apoha has found a clever correlation in a narrow domain and is over-extrapolating it into a universal platform. With only 40 projects and 25 employees, the evidence base is thin, and the history of materials-discovery startups is littered with elegant measurement methods that never scaled into reliable products.

What to Watch Next

Over the next 30 days, watch whether Apoha converts its Boehringer Ingelheim and Ethris collaborations into named, multi-year commercial contracts rather than pilots. The difference between a pilot and a paid platform deal is the difference between a promising demo and a real business. Watch also for any independent validation of the VIBE results, because a method that produces 1,000 descriptors is only as valuable as the proof that those descriptors predict outcomes better than the 12 industry-standard tests it claims to beat. Third-party replication, not company press releases, is the signal that matters.

Over 90 days, the question is breadth: can Apoha show the same precision across genuinely different material classes? A result that holds for antibodies, plant proteins, and an industrial coating would suggest a general platform. Results that work only inside biopharma would suggest a valuable but narrower niche tool. Watch which industries the company prioritizes with its fresh capital, because that choice reveals where management actually believes the technology is strongest. Expansion of the 25-person team into chemistry and materials-science hires would signal a serious push beyond the pharma beachhead.

Over 180 days and into 2027, the real test is whether a frontier AI lab or a major materials company decides it needs Apoha's data badly enough to partner deeply or acquire it. If proprietary physical data becomes the contested resource the model-data-ceiling thesis predicts, then a company generating 1,000 novel descriptors per sample in minutes becomes strategically valuable far beyond its current size. Watch for partnership announcements with generative-materials players, and watch whether the broader narrative in AI shifts from "bigger models" toward "better physical data," which would put Apoha squarely in the center of the next funding cycle.

There is one more thread worth pulling, and it concerns who ultimately controls this data. If physical-measurement data becomes the scarce input that frontier AI depends on, then the labs and instruments that generate it acquire quiet leverage over the entire stack. A company like Apoha could find itself in the position of a critical supplier to model-builders far larger than itself, the way a specialized sensor maker can become indispensable to an entire industry despite its modest size. That is a powerful place to sit if the thesis holds, but it also invites the giants to build or buy their own measurement capability rather than depend on a startup. The coming year will reveal whether Apoha can entrench itself as the standard source of physical-behavior data before a deeper-pocketed rival decides to replicate the instrument in-house.

The frontier of AI may not be a larger model that has read more of what humans wrote, but a machine given a sense humans never had.

---