Institutional Diligence in Frontier Tech: A Buy-Side Playbook
Updated: Sep 5
We advised a client against a growth round in enterprise AI.
The product worked and te demo was genuinely impressive. What stopped us was a slide, a wall of forty-odd enterprise logos under the heading "Customers." So we asked a boring question: how many of these are paid production contracts rather than pilots? The number came back in single digits. Of those, most sat on annual renewals with shallow integration and nothing that would make switching painful.
Eighteen months on, that company is still converting pilots. The logo wall has grown. The revenue line hasn't moved the way the model said it would.
That wasn't a clever call. It came from asking a question the deck wasn't built to answer.
The public data caught up later. MIT's NANDA study put the share of enterprise generative AI pilots delivering no measurable P&L return at 95%. The methodology has been fairly criticised since, so take the exact figure with salt. The direction is what matters, and it matches what sits in most data rooms we open.
Capital moved into frontier tech years ago but due diligence didn't move proportionally with it. That gap is what this is about.
We've read several hundred decks across these sectors in the last two years. This is what actually holds up. For allocators underwriting frontier positions, our Buy-Side M&A Advisory works on exactly this: execution risk, commercialization pathways, and whether the capital structure survives contact with reality.
Why Frontier Tech Demands a New Diligence Framework
The Risk Profile Is Asymmetric
Frontier bets are early, and a good number of them are genuinely unknowable. Pretending otherwise is where investment committees get into trouble. Market timing is unpredictable, infrastructure barely exists, and the regulations are still being drafted by people who don't fully understand the technology either.
Which means TAM projections and LTV:CAC ratios become theatre. You can build the model. It just won't tell you anything.
Diligence here shifts from validation to conviction. You're underwriting a paradigm, not a business plan, and that takes fluency in both the technology and the capital structure. Most ICs have one or the other. Very few have both sitting in the same room.
Tech Validation ≠ Product-Market Fit
A frontier startup will usually have IP or a prototype that works. Getting from there to a commercially viable product is a different problem with completely different failure modes. A validated model or a functioning carbon capture unit tells you almost nothing about cost curves, manufacturability, or whether a customer will actually sign.
The space between those two things is what we call the execution delta. Most diligence processes measure the technology carefully and the market carefully, then quietly assume the bridge between them exists.
An example. In 2025 we ran diligence on a direct ocean capture company. The science held up under scrutiny, which is more than you can say for a lot of what we see. What didn't hold up was the revenue model, which depended entirely on carbon market liquidity the company had no ability to influence. The engagement ended up changing three things: the technology roadmap, the revenue model, and the go-to-market approach. Not one of those changes came from questioning the CO2 science or underlying chemistry.
A Playbook for Institutional Diligence in Emerging Sectors
1. Engineering Truth Filters
Everyone claims a breakthrough and every deck has advisors with impressive affiliations. What matters is where the IP came from, what's genuinely proprietary, and how close it sits to being replicated by someone with a bigger budget.
Two questions do most of the work for us.
In agentic AI, the pilot-to-production question from the opening is the whole game. Pilot logos cost a company nothing and tell you nothing. Ask how many became paid contracts, at what value, on what renewal terms. Then ask a harder one: how much of this stack did the company actually build, and how much is a wrapper on a frontier model whose pricing and capability it doesn't control? Companies that can't answer the second question cleanly are one API price change away from a different business.
In quantum, ask what logical error rate the claimed application actually requires. This is the one that ends conversations. Microsoft Quantum's resource estimates (Beverland et al., 2022) put a commercially relevant chemistry problem, simulating a ruthenium catalyst for carbon fixation, at 2,740 logical qubits and a logical error rate of 3.0 x 10^-17. Factoring RSA-2048 needs 25,481 logical qubits. Both imply hundreds of thousands to millions of physical qubits. Those estimates are a few years old and the hardware has moved, but the order of magnitude hasn't. So when a company leads with physical qubit counts and never states the logical error rate its own use case demands, it's answering a question that has nothing to do with its revenue line. Ask directly and watch what happens.
In synthetic biology and precision fermentation, ask what the largest vessel is that the organism has actually run in, and what happened to titer, rate and yield at that scale. Bench numbers from a two-litre run rarely survive the move to 10,000 litres, because oxygen transfer, contamination risk and heat management all behave differently once you leave the flask. Then ask the question most decks skip entirely: what does downstream processing cost? Separation and purification frequently run over half of COGS in bio-based molecules, and a model that carefully forecasts fermentation while waving at recovery is telling you where the founders' attention has been.
Underneath those, three things we always trace:
Code lineage and research citations. Published work with the founders as authors is a different signal from published work the founders cite.
The team's academic versus industrial history. Scientific depth without shipping experience is the most common structural gap in this asset class, and it's the hardest to fix post-investment.
Which layers of the stack are protected and which are configuration of something freely available.
Our Private Equity & Venture Capital Advisory work applies the same lens to early-stage deployment.
2. Capital Stack Sensitivity
Deep tech doesn't scale on software burn models. Quantum hardware, AI, life sciences, green hydrogen, advanced materials: all of them need capital formation that tolerates nonlinear timelines and heavy asset intensity.
So most frontier companies end up running blended stacks. Grants, equity, venture debt, strategic money, and eventually project finance if they get that far.
Founder-side advice treats this as a financing question, and there's plenty of good writing on it from that angle. For an allocator it's something else. It's a diagnostic. Each layer prices a different risk, which means the mix tells you who has already underwritten what, and what nobody has been willing to touch. I've written more on where these structures break down in Why Most Capital Stack Designs Fail, and How Operators Fix Them for Connectively.
Capital source | What it actually prices | What it tells an allocator |
Non-dilutive grant | Technical merit, judged by a technical panel | Scientists validated it. Buyers didn't. |
Equity | Terminal outcome | Check whether tranches attach to technical milestones or to calendar dates |
Venture debt | Near-term cash predictability and collateral | Pre-revenue, this borrows against the next equity round, not cash flow |
Strategic or corporate | Option value on future supply | Look for ROFR and exclusivity that quietly caps your exit |
Project finance or offtake-backed | Contracted cash flow | Only available post-FID. Its absence tells you the company isn't bankable yet |
The EIC Accelerator makes the pattern easy to see: up to €2.5 million as a lump-sum grant, plus €1 million to €10 million in direct equity or quasi-equity. A European deep tech company can therefore carry real third-party validation that says nothing whatsoever about whether anyone wants to buy the product. In the US the same logic runs through different doors. An ARPA-E award is a technical endorsement from a technical panel. A DOE Loan Programs Office commitment is closer to a bankability judgment, and it arrives much later. Knowing which one a company holds tells you which risk somebody else has already priced.
Three things we test on any blended stack.
First, what the grant calendar does to the runway. Most grants pay in arrears against audited milestones, which makes them receivables with a lag, not cash. Companies present them as cash constantly. Sometimes they genuinely don't know the difference.
Second, whether a venture debt covenant triggers before the next technical milestone. If the covenant test lands before the milestone that releases the next equity round, the lender is running the company, not the board. We've seen this twice in eighteen months and it never once surfaced in a management presentation. You have to go find it in the loan documents.
Third, what sits senior to the equity you're buying. Grants can claw back. Venture debt takes priority. Strategic investors sometimes hold preferences stacked above yours. Model the downside waterfall before you spend any time on the upside case, because in this asset class the downside case is not a remote scenario.
We go deeper on the method in capital stack diagnostics. The short version: capital structuring is diligence.
3. GTM Strategy in Markets That Don’t Exist Yet
When the category has no incumbents, commercialization gets vague fast. An e-fuels producer might be waiting on offtake contracts, subsidies, and OEM partnerships that don't exist in enforceable form yet, and the deck will describe all three as though they're signed.
Synthetic aviation fuel is the clearest public example of what that looks like at sector scale. When Argus Media launched its eSAF price index in November 2025, electrolytic SAF was trading near $3,000 per tonne in the Amsterdam-Rotterdam-Antwerp region. Roughly thirteen times conventional jet fuel, three and a half times bio-based HEFA. More telling: of about seventy tracked UK and EU eSAF projects, not one had reached a final investment decision as of that assessment. EU mandates start in 2030 and plants take three to four years to build.
Seventy companies with working chemistry. None of them with demand a bank would lend against. That is a commercialization problem wearing a technology costume, and it is exactly the failure mode diligence should catch before an IC vote rather than after.
What we look at:
Pilot customers and where the money came from. A pilot funded out of an innovation budget is not a purchase signal. A pilot funded out of an operating budget is a completely different conversation.
Regulatory readiness, meaning whether the company qualifies today under IRA, RED III, FDA or EPA, what certification costs, and how long it takes. Not whether a supportive scheme exists somewhere.
Offtake quality. An LOI, an MOU, and a binding take-or-pay contract are three different instruments and only one of them finances a plant. People conflate them deliberately.
Red Flags LPs and ICs Often Miss
Each of these has a specific tell if you know where to look. The sectors differ but pattern mostly doesn't.
Narrative overshoot, where the pitch outruns what's technically achievable. In biomanufacturing this usually looks like a titer figure from a two-litre bench run sitting a few slides away from a cost model built on 50,000-litre economics. The general tell: the technical appendix and the financial model quietly use different assumptions. Read them side by side.
Academic advisors used as logos. Common in quantum and photonics, where a famous name from a national lab does a great deal of work on a slide and none in the company. The tell is that no advisor has a documented time commitment and none of them appear in the company's own patents or publications.
Fuzzy IP ownership in university spinouts. Advanced materials companies are the worst offenders, partly because the underlying chemistry often predates the company by a decade. The tell is a licence agreement that's missing from the data room, or one that's there but contains field-of-use restrictions and reversion clauses nobody on the deal team has actually read.
Science depth with operational blind spots. You see it most in robotics and hardware, where the prototype is genuinely excellent and nobody has priced a contract manufacturer. The tell is straightforward: nobody in the leadership team has taken a physical product through a manufacturing scale-up.
Valuations built off hype-cycle comps. AI is where this is currently most acute, though grid storage went through the same thing. The tell is a comparable set at a different technology readiness level, priced in a different rate environment, usually two years ago.
Several hundred decks in, the pattern is consistent. The best ones are almost never the flashiest. They're the ones where the technical appendix and the financial model agree with each other.
We've been wrong in the other direction too, and it's worth saying. Being too conservative on timelines in this asset class costs you entry price on companies that do get there. Call it discipline rather than pessimism. What it comes down to is insisting that the timeline in the deck has something behind it.
How Institutional Partners Can Win Here
Frontier tech is diligenceable. It needs a different lens, applied earlier than most processes apply it. The edge sits with allocators who can hold conviction and skepticism at the same time, which in practice means reading a technical claim and a capital structure with equal fluency.
Arete Ventures works with GPs, LPs, and asset allocators on diligence built for frontier sectors, mapping execution deltas across capital and operations, and supporting ICs with operator judgment rather than another framework. When scaling problems show up after close, our Performance Improvement Advisory turns that judgment into operating change.
If you're underwriting a frontier position and want a second read before the IC vote, we're happy to share how we'd approach it.
Frequently Asked Questions
How do you diligence a frontier tech company with no revenue?
Work the execution delta instead of the financials. Establish what the technology does today, what commercial threshold it has to clear, and what sits between the two. Then test IP provenance, scale-up capability, and whether any customer has committed operating budget rather than innovation budget.
How should grants, equity, and venture debt be combined in a frontier tech capital stack?
Each layer prices a different risk. Grants validate technical merit, equity prices the terminal outcome, venture debt prices near-term cash predictability. Sequence them so covenant tests and grant milestones fall after technical inflection points, never before. And model grants as lagged receivables, not cash.
What is the execution delta in deep tech due diligence?
It's the operational distance between a technology that works and a business that returns capital. Cost curves, manufacturability, regulatory qualification, customer readiness. Technology and market size both get diligenced carefully. The bridge between them usually doesn't get diligenced at all.
What are the biggest red flags in frontier tech due diligence?
Five recur: narrative that outruns technical feasibility, academic advisors used as credibility logos, unresolved IP ownership in university spinouts, leadership with deep science but no scale-up experience, and valuations anchored to hype-cycle comparables instead of achieved milestones.
How is buy-side diligence different from a technical audit?
A technical audit asks whether the technology works. Buy-side diligence asks what it would take for that technology to return capital inside a fund's holding period, and whether the capital structure and regulatory pathway can support that timeline.
