The numbers don't immediately grab you. Six million euros. A Series A extension, essentially. Barcelona, not Boston or the Bay Area. Pharmacelera isn't a household name, even in the insular world of computational drug discovery.
But dig into what the company is actually doing—and who's suddenly paying attention—and you start to see the outline of a different argument about how molecules should be predicted, and ultimately made.
Pharmacelera closed its €6 million equity round in February 2026, led by Madrid-based Heran Partners with participation from Clave, Inveready, and Bio&Tech Smart Capital. The money will bankroll a U.S. expansion and further development of its platform, which uses quantum mechanics calculations—not the headline-grabbing quantum computers, but classical physics applied rigorously—to generate three-dimensional molecular descriptors. Those descriptors then feed machine learning models. It's a hybrid approach, physics meets pattern recognition, and it sidesteps a problem that's become increasingly awkward for the AI-everything crowd: garbage in, garbage out.
"We're not trying to reinvent quantum computing," CEO Enric Gibert said in a recent interview. "We're using quantum mechanics because chemistry is quantum mechanics. Everything else is an approximation."
That's a subtle dig. The generative AI wave sweeping through drug discovery—NVIDIA's BioNeMo platform, partnerships with Eli Lilly announced just last month, Schrödinger's $2.5 billion deal with Novartis—has captured attention and capital. But a quieter cohort of platforms is placing a different bet: that physics-grounded methods will win the accuracy war, even if they're less flashy. Even if they don't generate breathless press releases about "foundation models" every quarter.
The Data Problem Nobody Talks About
Here's what the AI evangelists don't always say out loud. Machine learning models are only as good as their training data. And in biology, data is sparse. Noisy. Expensive to generate. A dataset of 10,000 protein-ligand binding affinities sounds impressive until you realize that chemical space contains something like 10^60 possible drug-like molecules. You're working with a rounding error.
Physics-based methods offer a way around this. They solve equations rooted in first principles—quantum mechanics, molecular dynamics, thermodynamics. They don't need millions of labeled examples because they're not learning patterns. They're calculating reality, or at least a close-enough approximation of it.
Pharmacelera's approach, refined over more than 25 years of research at the University of Barcelona, calculates electrostatic, steric, and hydrophobic molecular fields in three dimensions. The company then compresses those fields into descriptors that ML models can process. The result, according to Gibert, is better generalization. Models trained this way can extrapolate to novel chemical scaffolds—structures they've never seen before—with more reliability than pure data-driven systems.
It's a claim that's hard to verify from the outside. Drug discovery timelines stretch across years, and most work happens behind the walls of pharma R&D departments under heavy confidentiality. But the company's track record suggests someone believes it. Pharmacelera has pulled in roughly €2.2 million in non-dilutive grants: a €960,000 EIC Accelerator award under Horizon 2020 between 2020 and 2022, a €626,000 CDTI/FEDER-backed project in 2023–2024, and €600,000 from the Catalonia government in 2023. It raised about €1 million in 2021 via Capital Cell.
Add this latest round, and the company has built something durable without needing to chase Silicon Valley mega-rounds.
The Synthesizability Bottleneck
One of Pharmacelera's less obvious strengths is its integration with Enamine, the Ukrainian chemical vendor that maintains the REAL Space library—48 billion virtual compounds, roughly 80% of which can be synthesized and delivered within three to four weeks. That marriage of virtual screening and make-on-demand chemistry matters more than it might seem.
Because computational predictions, no matter how elegant, are worthless if they point to molecules that can't actually be made. Or that take six months and $200,000 to synthesize. The industry has learned this the hard way. AI models have suggested plenty of "hits" that turned out to be synthetically intractable or prohibitively expensive. Pharmacelera's exaScreen engine queries all 48 billion Enamine compounds, and the partnership—expanded in December 2022—tightens the loop from screen to synthesis. In October 2025, the company added eMolecules' eXplore and Synple libraries to the mix.
This is part of a broader shift. Virtual libraries have exploded in size. Enamine's REAL Space contains 48 billion compounds. WuXi's portfolio tops 8.6 billion. These aren't physical molecules sitting on shelves; they're recipes, reaction pathways that can be executed on demand. The economics of drug discovery are changing as a result, though not always in ways the headlines capture.
The Crowded, Messy Middle
Pharmacelera isn't alone in the physics-first camp. QSimulate, a Boston-area startup, announced fresh financing in early 2026 (exact details remain unpublished) and released QUELO v2.3, which the company says delivers 1,000× speedups for quantum mechanical calculations. QSimulate counts five of the world's top 20 pharmaceutical companies as customers, plus collaborations with Google, Mitsui, and JT Pharma. SandboxAQ, the Alphabet spinout, is commercializing what it calls "Large Quantitative Models"—AQBioSim and AQChemSim—through a partnership with Deloitte. Qubit Pharmaceuticals in France is working with NVIDIA on hybrid quantum-classical pipelines.
Last year, IonQ, AstraZeneca, AWS, and NVIDIA demonstrated what they claimed was a quantum-accelerated synthesis workflow, showing a 20× speed improvement for Suzuki–Miyaura coupling reactions. Entos, meanwhile, has built a family of ML surrogates (the OrbNet suite) that approximate quantum mechanical calculations at dramatically reduced computational cost.
These aren't identical approaches. Some are classical quantum mechanics made fast via machine learning shortcuts. Others involve actual quantum hardware, though those remain mostly experimental. But they share a thesis: molecular interactions are fundamentally quantum phenomena, and modeling them accurately at scale requires physics, not just statistical pattern matching.
Whether that thesis will prove correct is still an open question.
The Money Flows, But Clinical Proof Lags
The AI drug discovery sector is projected to grow from $3.25 billion in 2026 to $10.29 billion by 2031, a 25.9% compound annual growth rate, according to Mordor Intelligence. Other forecasts run higher—Kings Research pegs 2030 at $36.06 billion. Quantum computing applications in pharma, still nascent, are expected to climb from $450 million in 2025 to $811 million by 2030.
Big Pharma is spending real money on this stuff. Eli Lilly is constructing an internal AI supercomputer with NVIDIA. Recursion Pharmaceuticals deployed the BioHive-2 superPOD. Schrödinger locked down $150 million upfront from Novartis in late 2024 for its free energy perturbation (FEP+) platform, a physics-heavy approach focused on late-stage lead optimization. The company posted $180.4 million in software revenue for 2024.
But here's the uncomfortable truth. Clinical proof points remain scarce. Insilico Medicine reported positive Phase IIa topline results in November 2024 for ISM001-055, an AI-designed candidate for idiopathic pulmonary fibrosis. That was significant—one of the first widely cited examples of a fully computationally generated molecule clearing a meaningful clinical hurdle. Generate:Biomedicines is advancing GB-0895, an AI-designed antibody, into Phase 3 trials for severe asthma.
Those are exceptions. The rest of the field is still working through preclinical validation. And not every story ends well. BenevolentAI, once a poster child for AI-driven discovery, underwent a painful restructuring in 2024–2025. The company is delisting from the stock exchange and returning to its original research mission after commercial setbacks. The longer validation cycle for drug discovery—measured in years, not quarters—makes it brutal for venture-backed business models.
McKinsey estimates that generative AI could unlock $60 billion to $110 billion in annual value across pharma and medical products, mostly by accelerating early-stage discovery. BCG modeling suggests 30–50% time reductions and 25–50% cost savings in preclinical work. But adoption remains uneven, and those projections assume a level of integration that most companies haven't achieved yet.
Eroom's Law Still Holds
All of this is unfolding against a stubborn reality. Drug R&D productivity has been declining for decades, a phenomenon known as Eroom's Law—Moore's Law spelled backward. Since 1950, the number of new drugs approved per billion dollars spent has roughly halved every nine years. Deloitte's 2024 analysis found that Big Pharma now spends an average of $2.23 billion per approved asset, with an R&D return of just 5.9%. (That figure got a boost from GLP-1 blockbusters; without those, it's worse.)
The hope is that AI—whether generative, physics-grounded, or some hybrid—can reverse that curve. Maybe. The pressure is certainly there. But the industry has seen plenty of technological revolutions that promised transformation and delivered incremental improvement. High-throughput screening in the 1990s. Combinatorial chemistry. Genomics. Each changed the game somewhat. None broke Eroom's Law.
Regulatory Terrain Is Shifting
The regulatory landscape is at least starting to catch up. The FDA released draft guidance in January 2025 outlining an AI credibility framework for drug and biologics submissions, though discovery-only tools remain outside its scope for now. The European Medicines Agency finalized its Reflection Paper on AI in the medicines lifecycle in October 2024, emphasizing risk-based governance and traceability. The EU AI Act entered into force last August.
Open science efforts are also expanding. The OpenFold consortium, focused on protein structure prediction, welcomed Bristol Myers Squibb, Novo Nordisk, and NVIDIA in 2024 and 2025. NVIDIA's BioNeMo platform now supports foundation models for biology, with adoption accelerating among both biotechs and major pharmaceutical companies.
Whether these frameworks will accelerate or slow innovation is unclear. Pharma is notoriously conservative about adopting new methods, especially if they complicate regulatory submissions. But the conversation is happening.
The Long Bet
Pharmacelera's €6 million round isn't going to move markets. It's not a unicorn valuation or a splashy acquisition. But it does reflect a calculation: that accuracy at scale—especially when synthesizable chemistry is only weeks away—matters more than sheer generative volume. That physics-first AI may not dominate the headlines today, but could prove more durable over the long slog of drug development.
McKinsey's quantum computing analysis suggests that hybrid approaches—quantum-informed descriptors, machine learning surrogates for expensive quantum mechanics calculations—will deliver near-term value. True fault-tolerant quantum hardware, the kind that might actually revolutionize molecular simulation, remains a post-2030 story. Maybe longer.
For now, Pharmacelera and its cohort are making a different argument. Not louder. Not flashier. Just grounded in the physics of how molecules actually behave. Whether that's enough to crack Eroom's Law—or just chip away at its edges—is a question that will take years to answer.
But it's a bet worth watching.
