Dispatch from SynBioBeta 2025
From trillion-token models to stalled scale-ups: why synbio needs a systems stack.
I spent three days at SynBioBeta 2025, and like everyone else, I barely scratched the surface. I didn’t attend every session, I don’t build with wetware, and I write from the edge of the field — not from its centre. These reflections are shaped by that partial vantage point, and any inaccuracies are my own. Still, even from the margins, one thing was clear: synthetic biology is not short on invention. What it lacks is integration. The field is brimming with capability — trillion-token models, enzyme cascades, strain design platforms, engineering enzymes with 10^50 permutations — but still struggles to translate that capability into deployable, durable systems. The excitement is real. So is the bottleneck.
Commodities and capability compression
The panel on commodity-scale synbio opened with a blunt provocation: are we fools for using biotech tools originally developed for ~$x0,000/kg pharmaceuticals to make goods that need to sell for $1/kg? The answer: not if we use those tools to do what chemistry can’t.
Panelists made the case that in agriculture, we’ve hit the wall of chemistry. Biology can solve drought stress, nutrient uptake, and resistance to tightening pesticide regulation — things small molecules can’t touch. Panellists pointed out that pheromone-producing microbes offer mode-of-action novelty without toxicity and how biology can now unlock new phosphorus pathways and reduce the nitrogen load that chemistry has failed to manage for decades.
In commoditized sectors, synbio’s job isn’t to match on price — it’s also to exceed on function. But the economic viability of that model hinges on biological products doing something radically better than incumbent tools. This demands a reframe from "biotech as cheaper chemistry" to "biotech as different economics entirely." Value-based pricing is possible in ag, but only if synbio solutions are drop-in, field-validated, and superior.
AI for Biology: table stakes, but no stack
If capability is outrunning economic logic in the field, it’s also outpacing infrastructure at the algorithmic level. AI was everywhere at SynBioBeta, obviously. The field is now saturated with trillion-token protein models yet there is still no canonical pipeline for what comes next. We have foundation models. What we don’t have is a foundation stack, so to speak.
What’s missing, imo, is the “model-to-value” arc — a structured, shared workflow that moves from sequence prediction to validated, manufacturable, GMP-grade product. Each company is building its own end-to-end scaffolding. Everyone trains on their own data. Everyone codes their own scoring functions, constructs their own expression assays, and defines their own manufacturability thresholds. There are no shared reference assays. No CASP-for-function. And so each team becomes its own infrastructure provider — slowing the field’s ability to generalize, compare, or plug into shared supply chains.
Several founders made the case — for a more open, composable ecosystem: sequence–function datasets, standard benchmarks for activity and expression, and API-friendly layers that decouple modelling from downstream validation.
Importantly, what matters now isn’t whether you use AI — it’s whether your AI survives contact with the lab. Across companies, the emphasis is shifting from model novelty to experimental validation. The real differentiators are:
Can your models generalise beyond your training data?
Are you solving a specific biophysical constraint (expression, yield, thermostability) — or just hallucinating structure?
Do you have a tight experimental loop with fast-turnaround assays and clear criteria for manufacturability?
In this sense, AI has become table stakes — but functional validation is the new currency. Clinical partners, regulators, and strategic buyers aren’t asking for novel architectures. They’re asking: does it work? What’s your assay? What’s your benchmark? Where’s the data? Until we have robust, interoperable infrastructure that links model output to field-deployable, regulator-grade inputs, the promise of AI in biology will remain structurally under-realised.
Capital journeys that don’t work
But even when the models work and the function is real, another constraint looms: capital fit.
There’s a widening gap between the companies being built (infrastructure-heavy, slow-margin) and the investment structures available to them (venture-speed, exit-centric). The hard truth: industrial bio lacks a repeatable M&A or IPO pathway. In therapeutics, exits through licensing or acquisition are enabled by standard milestones and predictable inflection points. In synbio, we don’t have that. As a result, companies are being pushed into premature verticalisation—owning manufacturing, regulatory filings, and go-to-market functions before they have product-market proof.
One proposed solution was to follow early biotech’s playbook: lean into outsourcing. Platform biotech companies survived by narrowing their core competency (e.g. discovery) and partnering for the rest (e.g. trials, distribution). This logic might yet work in industrial bio: design the pathway and outsource the plant. Pharma will only engage if the platform is both enabling and essential. You can’t pitch a partnership — you have to demonstrate that you’ve already built something irreplaceable. Novartis BD team at a panel reinforced this: deals are bought, not sold.
This is also where philanthropic and public innovation agencies — from DARPA to ARPA-H— may play a structural role, de-risking early infrastructure for translation, and allowing startups to stay lean.
The infrastructure bottleneck
At every panel, infrastructure came up as the silent constraint. Michael Tai (Boston Bioprocess) noted that companies can’t even get on the table with CMOs if they’re below certain titer and volume thresholds. There’s a pre-pilot valley of death: your 10L strain might work, but no one can help you scale it until it's "mature" — a maturity no one will fund you to reach.
Even when there is steel, the downstream is what breaks. Several speakers highlighted that while we’ve optimized strain engineering and fermentation, downstream purification (DSP) remains neglected. Panellists framed the entire question of scale as not just bioproduction, but the modelling of 5L to 500,000L correlations. If we want synbio to industrialise, we need predictive models that connect early strain performance to full-process economics. And that means data. High-quality, real-world, open data.
IFF’s protein engineering tournament with Harvard was cited as a rare pre-competitive example. Arzeda called for more such shared datasets — in part to make model building easier, but also to collectively debug what the field actually needs.
Regulation, reputation, and public trust
In agbio, the regulatory system isn’t just slow — it’s incoherent. In a panel, a founder explained how: USDA reviewers rejected their SBIR application due to anti-GMO sentiment, while federal research agencies were simultaneously funding transgenic microbe field trials! The frameworks haven’t caught up with the tools.
Meanwhile, the industry’s public reputation remains caught in an older era. Labelling remains opaque, and the distinction between “bioidentical” and “bioengineered” ingredients confuses both consumers and regulators. If synbio wants to scale, it needs to communicate clearly.
David Willetts, chair of the UK’s Regulatory Innovation Office, described this well: we need not just regulation, but regulatory navigation. In complex fields like Synbio, there is no single regulator — RIO’s model is to provide a concierge service, mapping routes through food, health, and environmental standards and aligning regulation with investment and procurement. It’s a model that could be adopted internationally when it proves its role.
Biosecurity is underbuilt — and under-watched
The dual-use risk was a background hum at SynBioBeta — and, in my view, still remains dangerously unaddressed. We now have generative platforms capable of designing enzymes, peptides, and capsids with high specificity — yet we lack basic oversight tools. There are no audit trails for model outputs. No watermarking of biological designs. No red-teaming protocols. No traceable provenance for how a viral vector, pathway, or payload was engineered.
This is a brittle place to build from. If synthetic biology is to be as powerful and accessible as the field hopes, we need the biosecurity equivalent of an LLM safety stack: model fingerprinting, structured red-teaming, traceability of design steps, real-time anomaly detection, and intentional-use gating. The alternative — waiting for a misuse event to build oversight infrastructure — is a failure mode.
This was echoed by the National Security Commission on Emerging Biotechnology (NSCEB). Their 2024 report made the point unambiguously: the U.S. biotechnology sector is structurally dependent on overseas supply chains — particularly Chinese firms like WuXi AppTec — for essential inputs. Over 79% of U.S. biopharma companies rely on China at some stage of production. That’s not just an economic risk — it’s a national security exposure. The playbook is familiar: copy IP, scale with state subsidies, dominate supply and use that dominance to exert geopolitical leverage — as seen with China’s export controls on gallium and germanium.
Without rebuilding secure, domestic biomanufacturing capacity, the U.S. risks losing sovereign control over its own medicines, food inputs, and foundational infrastructure. The NSCEB’s recommendation to designate biotech infrastructure and data as critical infrastructure — with the protections and investments that status entails — should not be optional. The same applies to creating a federated network of domestic manufacturing hubs, mandating the disclosure of single points of failure in supply chains, and backing the scale-up of secure precommercial capacity. We will be seeing a new era of bioeconomic statecraft, and I’m also hoping Europe will also take a leadership role.
What I’m excited about from SynBioBeta 2025
A few architectural ideas stood out to me, as identified across sesssions — connecting scientific ambition to deployment logistics, regulatory realism, and planetary scale.
Universal immunity, with a distribution stack to match. Jacob Glanville’s universal flu vaccine at Centivax is a credible moonshot: conserved epitope targeting, broad strain coverage, and a path to clinical trials within 9 months. But what elevated it was the delivery layer. Dean Kamen’s self-administered microneedle patch sidesteps the cold chain, replaces syringes, and works at a fraction of the dose. The bigger idea: immunity at a planetary scale isn’t just a big problem. It’s also a delivery problem. And that means co-designing the molecule, the device, the regulatory route, and the distribution channel — from day one.
A genome writing project for cell reprogramming. Jason Gammack’s call for a “Genome writing project” hit a nerve. We now have the ability to write mammalian genomes. But we don’t have the biological metrology to predict how cells will behave when perturbed. This proposal — akin to the Human Genome Project — would create a national-scale, open infrastructure for decoding cell state, response, and rewiring potential.
Test-time compute as a strategy for biological search. Rather than relying on ever-larger pre-trained models, some teams are shifting the computational load to inference — using test-time computing to evaluate billions of sequences in silico before wet-lab work ever begins. This flips the architecture: from model-centric to search-centric. It’s especially powerful in constraint-rich or safety-critical domains (like gene therapy), where brute-force training offers diminishing returns. What matters here isn’t scale but targeted adaptability.
Enzyme platforms for circular manufacturing. Enzyme platforms that aim not at pharmaceuticals, but at waste. Teams are building AI-guided cascades that valorise industrial byproducts — from fatty acid streams to lignin derivatives. This reframes synbio also as infra, not an ingredient. Instead of drop-in replacements, these enzymes are metabolically native to the problem space: industrial conditions, variable inputs, and marginal cost environments. It’s regenerative manufacturing, powered by biological specificity — and it hints at a new design space for industrial metabolism.
What needs building next
The field has outgrown its sandbox. Synbio no longer struggles with proof-of-concept science — it struggles with missing systems. From discussions/observations, a new blueprint is emerging in my view: not just for tools but for the translation architecture, industrial infrastructure, and institutional coordination that must now define the field’s second act. A few things that we could build:
I. A translation stack for AI-generated biology
Foundation models are no longer novel — they’re default. The problem isn’t prediction. It’s the product. Every team is building their own stack leading to a lack of shared translation infrastructure and so every AI-native startup becomes its own infrastructure company… What’s missing imo is a shared, composable “model-to-value” stack:
Reference validation assays with cross-platform benchmarks
APIs that link design outputs to real-world wet-lab execution
Interoperable data formats from in silico design to CMC filing
Clear translation routes from predictive models to regulatory-grade products
II. Predictive downstream logic
The strain works. The fermentation scales. And then, downstream processing kills the project. Multiple panels flagged the same issue: downstream purification is the new bottleneck — still artisanal, still under-modelled. We need to bring downstream-aware engineering into the design loop:
Predictive downstream processing models that guide strain selection, not just fermentation — addressing blind spots in scale-up.
High-fidelity simulation engines that can map from 5L to 500,000L with confidence, integrating feedstock, metabolite, and DSP parameters into unified techno-economic projections.
Modular, reprogrammable CDMO networks that allow early-stage companies to plug in at sub-100L scale without years of sunk investment or bespoke contracts.
Shared benchmarking assays, design rules, and open datasets — so “translation” doesn’t mean rebuilding a black box for every new protein or pathway.
III. Industrial infrastructure for pre-commercial scale
Even when the strain performs and the downstream is modelled — you still need a tank. And for most startups, that’s where things break in the pre-pilot infrastructure gap: companies can’t access CMOs until they’re already producing at scale, but no one wants to fund them to get there. The result is a Catch-22: prove your titer before you can build your tank. The field needs real infrastructure for pre-commercial scale — the translation layer between 5L prototypes and commercial manufacturing. What’s needed:
A federal biobond mechanism to finance pilot-to-demo scale capacity — treating bio like semiconductors or chips (by Andrew Endy).
Open-access fermentation and bio-foundry infrastructure in the 1L–100L range — especially in underserved regions.
Domain-specific testbeds (e.g. microbial corals, enzyme-enabled carbon upcycling, genome-edited crops) that provide regulators and funders with real-world data, not just PDFs.
IV. A workable regulatory interface layer
Many at SynBioBeta had a regulation story. Most of them were bad. Whether it was USDA rejecting a field trial over anti-GMO sentiment or startups navigating 5 different agencies with no map, the message was the same: regulation seems to be structurally misaligned with how Synbio works now. What’s needed is a navigation layer, not more rules:
Regulatory concierge services (like the UK’s RIO) to help startups find viable paths through multi-agency domains
Pre-consensus labelling frameworks for bioengineered vs bioidentical — to reduce consumer confusion and litigation risk
Unified regulatory decision registries — anonymized but public — so the field can learn from precedent
Cross-border regulatory harmonisation, especially in agbio and food — to make approvals portable and scale global supply chains
V. Cultural & scientific infrastructure for a programmable biology era
Today, we can encode non-canonical amino acids, build autonomous fermenters, and simulate entire molecular systems—yet we can't reliably manufacture insulin at a global scale; we don't have a shared vocabulary for cell states, and we still write plasmids like we’re in 2005…This is what Endy calls a prebiotic state: where programming is possible, but we haven’t built the compilers, standards, or even the workforce to do it predictably.
This is crucially a comprehension gap—a systems-level lag between what’s possible and what we know how to integrate safely, justly, and at a population scale. Endy’s provocation: fund the layers beneath the application. Push upstream—toward biometrology, cell-level design standards, and national-scale pilot manufacturing. This is what allows synthetic biology to become routine engineering.
What we also lack is the talent, coordination, and culture to make it usable — and safe. Talent and education efforts should be further funded, e.g. a fellowship of bioprogrammers to train a new generation in composable, interpretable biological engineering.
As Drew Endy reminded the room: biology is democracy’s natural ally. It lets people provision atoms, joules, and bits where they are — giving agency at the edge. But that promise only becomes real if we build the substrate: the software stack, the biometrology, the coordination scaffolds, and the financing architecture that makes it all move.
Why philanthropy matters here
Synbio now lives at the intersection of science, manufacturing, software, regulation, and risk. Several opportunities are precommercial and pre-consensus — too early for VCs, too weird for the government, and too shared for corporations. Philanthropy is uniquely positioned to de-risk and de-bottleneck the system layers that commercial and government actors struggle to touch.
It can underwrite the “plumbing” of the field — shared validation, regulatory navigation interfaces, public data standards, and translation testbeds — that make the rest of synthetic biology runnable. These are structural assets, and while they offer no immediate exit or procurement pathway, they are precisely what’s needed to accelerate all downstream activity. Funders can make this real by backing efforts like open manufacturing hubs in underserved regions, coordination scaffolds for cross-agency regulation, and robust benchmarking initiatives that enable plug-and-play infrastructure across the field.
Philanthropy is also well placed to ignite the next wave of moonshots — not by bankrolling the whole stack, but by seeding the first 10%. That means commissioning scoping papers, convening institutional coalitions (e.g. NSF + DOE + ARPA-H), and prototyping use cases that give vision shape. Especially given the significant funding cuts in the US. Moonshot coordinated research programmes, national biosecurity red-teaming infrastructure, or supporting underserved geographies and neglected problems (e.g. waste valorization, orphan biosecurity protocols) begin with system architecture and legitimacy — not capital. Philanthropy can fund the work that lets these ideas exit the whiteboard and enter public investment pipelines.
outro
These reflections only capture one slice of SynBioBeta — and are shaped by my own vantage point as a translator. The question now is how we build the systems to let synbio matter — not just at the scale of pilot plants or product launches, but at the scale of planetary infrastructure. Synthetic biology holds the potential to redraw the boundaries of what’s possible: how we grow food, manufacture goods, restore ecosystems, and rewire carbon, nitrogen, and energy flows across the biosphere. But that future won’t be delivered by capability alone. It will depend on the regulatory, financial, and cultural systems we build around the tools. On coordination. On infrastructure. On the stack.