BUY-SIDE ADVISORY | PRIVATE AND PUBLIC MARKETS
Home / Private Equity & Venture Capital / Deep Tech Due Diligence
Frontier Technology
Deep Tech Due Diligence: Finding the Metric That Gates Revenue
Every frontier company reports a number that moves faster than its revenue. In AI, quantum, tokenization, biomanufacturing and robotics, the investor’s job is to find the slower number that decides revenue and to price the distance between the two.
By Gaurav Shah
Managing Partner, Arete Ventures · Investment committee member
Published November 2026
22 min read
Illustration: Arete Ventures.
In Brief
In every frontier sector, the deck reports a proxy that is easier to move than the number that decides revenue. We call that second number the gating metric.
The gating metric sits where a customer pays: paid production on the customer’s own data in AI, logical error rate for a named workload in quantum, titer in the largest vessel in biomanufacturing, interventions per operating hour in robotics.
Each sector has a reading that separates a fundable company from a promising one, and a deal term that fits the gap. Two exhibits below set out both.
A gap belongs in the price or the conditions of the deal. A memo that only notes it leaves the risk with the fund.
Delay to the gating metric costs more than it did in 2021, because required returns are higher and follow-on capital outside AI has thinned.
LPs can test a GP’s frontier underwriting with six questions, and boards can track the gating metric every quarter.
I sit on venture capital and private equity investment committees that vote on robotics, deep tech and life sciences deals, after sixteen years as an operating partner. As a chemical engineer, I have built and run industrial plants. In deep tech due diligence, the hardest debate at committee is usually about which number the company is asking us to believe. Whether the science works comes up less often than people outside the room expect.
Frontier companies report what they can move quickly: benchmark scores, pilot counts, physical qubits, wallets, titer in a shake flask, robot demo videos. These numbers are real, and they sit upstream of the number that decides whether customers pay, and at what margin. That number moves slowly, often depends on people outside the company, and is the one the financial model quietly assumes.
This piece names that number for each major frontier sector, sets out what a good reading looks like, and shows the deal term that fits a gap in each. It also explains why the time it takes to reach the number costs more now, and it ends with questions for LPs and boards.
Most frontier deals are priced on the number a company can move this quarter.
What Is a Gating Metric?
A gating metric is the measure that has to be true, at a paying customer, for the revenue in the financial model to arrive.
Two tests identify it. It is measured at a paying customer, outside the lab or the demo. And a specific line in the model, usually revenue timing or gross margin, depends on it directly. Gating metrics tend to move slowly, because something outside the company has to change: a customer’s data, a larger vessel, a regulator’s view, a supplier’s process.
The proxy is the number that reaches the deck. Proxies are rarely false; a benchmark score is a real score. A proxy is worth as much as it predicts the gating metric, and diligence is the work of finding out how well it does.
Figure 1
The proxy reaches the model’s assumption long before the gating metric does
Illustrative. Source: Arete Ventures.
The map below sets out the criteria we use to evaluate a deep tech startup in each sector. The last column is our view of what a good reading looks like, with outside benchmarks where they exist.
Exhibit 1
The gating metric map
Sector
What the deck reports
What gates revenue
What good looks like (Arete view)
What the deck reports: Pilots, benchmark scores, ARR run rate
What gates revenue: Paid production on the customer’s own data, at a gross margin that survives inference cost
What good looks like: Most ARR from production contracts; margin after inference on a dated path toward the high 50s, where surveyed AI companies expect to be by 2027
What the deck reports: Users, integrations, workflows built
What gates revenue: What a customer keeps only by staying, once the model vendor ships the same feature
What good looks like: Leaving would cost a customer months of work; the product runs on more than one model vendor
What the deck reports: Volume generated, benchmark lifts
What gates revenue: Performance of models trained on it, on real data they have never seen
What good looks like: Customer-run tests on held-out real data show gains on the rare cases that matter; clear rights to the seed data
What the deck reports: Physical qubits, fidelity records
What gates revenue: Logical error rate and gate count for a named workload
What good looks like: A published resource estimate for the target workload, met by a dated roadmap; revenue before fault tolerance modeled on its own margins
What the deck reports: Best samples, lab performance
What gates revenue: Yield and repeatability across production batches
What good looks like: Batch-to-batch yield data and at least one completed customer qualification
What the deck reports: Wallets, transactions, value locked
What gates revenue: Fees from users who would transact without incentives; claims that hold up legally
What good looks like: Fee revenue above incentive spend; a legal opinion on the holder’s claim; custody outside the issuer’s bankruptcy
What the deck reports: Titer in a flask or bench reactor
What gates revenue: Titer, rate and yield in the largest vessel, plus recovery cost
What good looks like: Titer held at each step up to the largest vessel run; downstream cost from actual runs; no jump to commercial scale without data from an intermediate vessel
What the deck reports: Demo videos, pilot sites
What gates revenue: Interventions per operating hour across customer sites; service cost per unit
What good looks like: Intervention rate falling month on month at every site, including the worst; the tenth installation far faster than the first
What the deck reports: Memoranda of understanding, pipeline, reactor designs
What gates revenue: Binding offtake and incentive eligibility under current rules; for nuclear, licensing milestones and first-unit cost
What good looks like: Most planned output under executed offtake; an eligibility opinion under current rules; for nuclear, a licensing schedule with dated milestones
Select a sector to go to its section. Source: Arete Ventures. AI margin reference: ICONIQ, State of AI 2026.
How Should a Gap Change the Price of a Deal?
Through the terms.
A gap the committee can measure today goes into the price. A gap that resolves on a known date goes into a tranche or a closing condition. A gap that depends on someone outside the company goes into contracts and structure, and a gap no one can size calls for a smaller first check or a pass. Here is how that plays out by sector.
Figure 2
Where a gap belongs in the deal
Source: Arete Ventures.
Exhibit 2
The term that fits the gap
Sector
The common gap
The term that fits
AI applications
The common gap: Margin after inference below the model
The term that fits: Price on post-inference margin at next year’s usage; tie part of the round to the share of revenue from production contracts
AI-native platforms
The common gap: Dependence on one model vendor
The term that fits: A condition that the product runs on a second vendor by a set date; board reporting on vendor pricing and terms
Synthetic data
The common gap: Gains not yet shown on real data
The term that fits: Release the second tranche on a customer-run test on held-out real data; confirmed rights to seed data at closing
Quantum computing
The common gap: Revenue assumed before a resource estimate supports it
The term that fits: Take that revenue out of the base case; smaller first check, with follow-on reserved for a logical error rate milestone
Tokenization and blockchain
The common gap: Unclear holder claim; activity bought with incentives
The term that fits: Legal opinion and custody structure as closing conditions; value revenue net of incentives
Biomanufacturing
The common gap: Largest-vessel titer unproven
The term that fits: Second tranche on titer at the largest scale; size the round to fund the months from mechanical completion to steady output
Robotics and physical AI
The common gap: Intervention rate unknown across sites
The term that fits: Tranche on the fleet-wide intervention rate; service cost per unit in every board pack
Energy transition and nuclear
The common gap: Offtake unsigned or eligibility uncertain
The term that fits: Executed offtake and an eligibility opinion as conditions; for nuclear, capital staged by licensing milestone
Source: Arete Ventures.
From our work
$360M
Deep tech Series C
Technical diligence for the GP
We worked this way on a $360 million deep tech Series C. The GP needed to know whether the technology could reach commercial scale on the plan’s timeline, and whether the revenue, growth and cash flow in the model matched what the technology could deliver. We ran the technical diligence and rebuilt the model’s operating assumptions from what we found. The GP invested at a lower valuation, and closing was made conditional on a technical milestone: shipping new features and service lines that customers depended on and that kept the company ahead of competitors. The gap went into the price and into the closing terms.
Into the price
A lower valuation
Into the terms
Closing conditional on a technical milestone
Why Does Time to the Gating Metric Cost More Now?
Frontier companies are long-duration assets. The return investors require has risen since 2021, and follow-on capital outside AI has thinned, so a slip costs more than it used to.
Illustrative
What a One-Year Slip Costs
A company expects to reach its gating metric in 18 months. Once it does, the next investor or a buyer would value it at $600 million. Investors require a 35% annual return for this risk. Today’s value of that outcome is about $383 million. If the milestone slips by twelve months, today’s value falls to about $283 million, before counting the bridge round the extra year needs.
Figure 3
Today’s value of a $600 million outcome, by months of delay
$99 million
Value lost to a one-year slip
26%
Share of today’s value lost at a 35% required return
20%
Share lost at a 25% required return
The higher the required return, the more each month of delay costs. The date of the gating metric deserves as much diligence as the metric itself.
Long-term rates have moved well above the levels of 2021, and venture dollars have concentrated in AI; current figures are in the Where Markets Stand box below. Outside AI, follow-on rounds are thinner and more selective. A quantum, biomanufacturing or tokenization company that misses its gating date may find its next round led by insiders, priced lower, or not available at all.
For a GP, the practical response starts with reserves. Size them to the expected date of the gating metric plus six months, and check that the company reaches the metric before its cash runs out. When the date slips, decide early whether to fund the bridge, because late decisions tend to be made at the worst price.
When the date slips, decide early whether to fund the bridge, because late decisions tend to be made at the worst price.
Exits point the same way. Evercore’s review of the secondary market in the first half of 2026 found that top-quartile venture interests clear at modest discounts while weaker ones fare far worse. Strategic buyers pay for the gating metric because that is what they can use, and secondary buyers discount for whatever gap remains, so the proxy earns little at exit.
The sections that follow apply the map sector by sector. Each gives the short answer, the evidence, what we test and the question we press at committee.
Sector 1 of 8 · AI applications
Why Do So Many AI Pilots Stall Before Revenue?
A pilot proves the model works on data the company chose. Revenue depends on the customer’s data, inside the customer’s workflow, at a margin that survives inference cost.
The deck reports
Pilots, benchmark scores, ARR run rate
What gates revenue
Paid production on the customer’s own data, at a gross margin that survives inference cost
MIT’s NANDA initiative reported in 2025 that about 95% of the enterprise generative AI pilots it studied delivered little or no measurable impact on profit and loss. Tools bought from specialist vendors succeeded about 67% of the time, roughly three times the rate of internal builds, which is good news for vendors whose product reaches production.
Margin is the second gate. Bessemer’s State of AI 2025 found that its fastest-growing AI cohort averaged about 25% gross margin, often negative. ICONIQ’s 2026 survey of about 300 AI builders put average gross margin on AI products at 45% in 2025, with companies expecting 59% by 2027. Inference sits in cost of goods and grows with usage. Falling model prices help less than they seem: Epoch AI found that the price of a given level of model performance fell between 9x and 900x a year, depending on the task, and that saving reaches every competitor and every customer as well.
Figure 4
AI gross margins are rising, from a low base
Sources: Bessemer Venture Partners, State of AI 2025; ICONIQ, 2026 State of AI. The 2027 figure is what surveyed companies expect.
What We Test: AI Applications
Revenue from production contracts, separated from paid pilots and design partners.
Accuracy on the customer’s own data, measured by the customer.
Gross margin after inference at next year’s expected usage, with model prices both falling and rising.
What a customer would have to rebuild to leave.
The Committee Question: AI Applications
Which revenue would remain if every pilot ended today?
Sector 2 of 8 · AI-native platforms
What Stays With an AI-Native Platform When the Model Vendor Ships the Same Feature?
Only what a customer would lose by leaving: the data and workflow history that live in the platform and nowhere else.
The deck reports
Users, integrations, workflows built
What gates revenue
What a customer keeps only by staying, once the model vendor ships the same feature
Bessemer notes that large labs are moving into applications, and that some fast-growing applications sit close enough to core model functions to risk being called thin wrappers. Each quarter, more of a platform’s feature list overlaps with what model vendors ship. What is left after that overlap decides the exit multiple. For a buyer, that dependence shows up in the exit multiple, and our note on exit multiple diligence shows how to test it before signing.
What We Test: AI-Native Platforms
Switching cost per customer, in months of work and dollars.
The share of each customer’s workflow state held only in the platform.
Dependence on a single model vendor, and the cost of moving to another.
The model vendor’s terms on pricing and data use.
The Committee Question: AI-Native Platforms
If the leading model vendor shipped your top three features tomorrow, what would a customer lose by switching?
We have reviewed AI-native platform companies at investment committee. The answers that hold up name specific things: years of workflow history, approvals and permissions built into the platform, certified connections to systems of record that took months to set up. An answer about the interface or the team’s speed usually means the platform holds little a customer would miss.
Sector 3 of 8 · Synthetic data
When Is Synthetic Data a Moat?
When models trained on it perform well on real data they have never seen, and the company holds clear rights to the real data it starts from.
The deck reports
Volume generated, benchmark lifts
What gates revenue
Performance of models trained on it, on real data they have never seen
Synthetic data fills gaps where real data is scarce or private: rare events for autonomous systems, edge cases in robotics, patient records that cannot be shared. The risk is circular training. A 2024 study in Nature found that models trained on recursively generated data degrade irreversibly, and that rare, low-probability cases disappear first. Those rare cases are what customers pay a synthetic data company for. The authors also concluded that data from genuine human interactions becomes more valuable as generated content spreads.
What We Test: Synthetic Data
The share of the customer’s production training data that is synthetic, and what happened to performance as that share rose.
How the customer validates on real data, and who runs the test.
Who owns the real seed data, and what the license allows when the contract ends.
Whether the generation method holds up when the customer’s real-world data shifts.
The Committee Question: Synthetic Data
What happened to your customers’ real-world results as the synthetic share of their training data went up?
We have reviewed synthetic data companies at investment committee. A company that answers with measured results from its customers has a data advantage it can price. When the answer is a demo, the customer’s own team may soon do the work.
Sector 4 of 8 · Quantum
How Should Investors Read Quantum Computing Milestones?
Start from a named workload, work out the logical error rate and gate count it needs, and compare that with what the company has shown and promised, with dates.
The deck reports
Physical qubits, fidelity records
What gates revenue
Logical error rate and gate count for a named workload
Error correction has started to work. In Google’s Willow results, published in December 2024, each step up in surface code size, from 3x3 to 5x5 to 7x7, cut the encoded error rate by a factor of 2.14. Google’s team was plain about the distance left: today’s error rates sit around one in a thousand, and useful applications need around one in a trillion.
Resource estimates are moving too. In May 2025, a Google researcher estimated that 2048-bit RSA could be factored in under a week with fewer than one million noisy qubits; the 2019 estimate needed 20 million. IBM targets its Starling system in 2029, running 100 million gates on 200 logical qubits.
Figure 5
Quantum: the distance left, and how fast the estimates are moving
Sources: Google Quantum AI, December 2024; C. Gidney, arXiv:2505.15917, May 2025.
What We Test: Quantum
A published resource estimate for the workload the company plans to sell.
Logical qubits and error rates demonstrated, against the roadmap and its dates.
Revenue before fault tolerance, from cloud access, research contracts, software, services or sensing, and its margins.
For quantum materials, quantum dots and sensing: yield and repeatability across production batches, and customer qualification results.
The Committee Question: Quantum
Which published resource estimate does your roadmap meet, and in which year?
Sector 5 of 8 · Tokenization and blockchain
What Separates Real Usage From Paid Activity in Tokenization?
Fees from users who would transact without incentives, and claims on real assets that would hold up if tested.
The deck reports
Wallets, transactions, value locked
What gates revenue
Fees from users who would transact without incentives; claims that hold up legally
Tokenized real-world assets are growing from a small base. On 7 November 2026, rwa.xyz showed $39.4 billion of distributed tokenized real-world assets, against $281.6 billion of stablecoins, with treasury funds and private credit leading. On-chain activity is easy to count and easy to buy, since token incentives and wash trading inflate wallets and transaction counts.
For real-world assets, the second gate is legal. A token is worth the holder’s claim on the asset, and that claim depends on who holds the asset, which entity issued the token, which law governs it and what happens if the issuer or custodian fails.
Figure 6
Tokenized real-world assets are still small next to stablecoins
Source: rwa.xyz, accessed 7 November 2026.
What We Test: Tokenization
Fee revenue by user cohort, with incentives stripped out.
The legal opinion on the holder’s claim.
The custody arrangement, and how the assets are kept out of the issuer’s bankruptcy.
The regulated channels that reach the institutional buyers in the plan.
The Committee Question: Tokenization
What would fee revenue have been last quarter if no incentives had been paid?
Sector 6 of 8 · Biomanufacturing
Why Does Biomanufacturing Scale-Up Break the Model?
Because titer, rate and yield usually fall as vessels grow, and recovering and purifying the product often costs more than making it.
The deck reports
Titer in a flask or bench reactor
What gates revenue
Titer, rate and yield in the largest vessel, plus recovery cost
A strain that performs in a flask meets different mixing and oxygen transfer in a large fermenter, yet models often carry the bench titer forward. A 2016 review of bioprocesses put product concentration and purification at about 50% to 80% of total processing costs, while decks often show recovery as a single assumption.
History says to expect a slow ramp. RAND studied 44 first-of-a-kind process plants and found that more than half failed to reach their production goals in the second six months after start-up, with performance explained mostly by how much of the technology was new and whether the plant handled solids. A later RAND study traced solids-handling problems to physical and mechanical causes, rarely the chemistry. Biology adds contamination and strain stability.
Figure 7
Recovery and purification often cost more than making the product
Source: G. M. Forde et al., Physical Sciences Reviews, 2016.
What We Test: Biomanufacturing
The largest vessel run, and titer, rate and yield at each step up in scale.
Downstream cost per kilogram from actual runs.
Who on the team has started up a commercial plant.
Months of cash funded between mechanical completion and steady output.
The Committee Question: Biomanufacturing
Who on this team has started up a commercial plant and carried it to nameplate?
I ask it first because I have built and run plants. In therapeutics, the gate is clinical and regulatory, and specialist life sciences funds price it with their own methods. The manufacturing side is where an operator’s eye adds most.
Sector 7 of 8 · Robotics and physical AI
What Should Investors Measure in Robotics and Physical AI?
Interventions per operating hour across customer sites, and the service cost of keeping each unit running.
The deck reports
Demo videos, pilot sites
What gates revenue
Interventions per operating hour across customer sites; service cost per unit
We have not found a published intervention benchmark for robot fleets, so we apply the standard autonomous vehicles already meet. California’s DMV requires test permit holders to report disengagements, each time a safety driver takes control because of a technology failure or other concern, along with miles driven. Permit holders logged more than 9 million test miles in the year to November 2025. A robot fleet deserves the same discipline: how often a person has to step in, per operating hour, at every site including the worst one.
Service cost is the second gate. A robot with an attractive bill of materials still loses money if every site needs custom engineering.
Figure 8
What a fleet report should show: every site, including the worst
Illustrative data. Source: Arete Ventures.
What We Test: Robotics
Intervention logs by site and by month, across the whole fleet.
The company’s definition of uptime, and what it leaves out.
Service cost per unit over time.
How long the tenth installation took compared with the first.
The Committee Question: Robotics
What was the intervention rate at your worst site last month?
Sector 8 of 8 · Energy transition and nuclear
Where Do Energy Transition and Nuclear Fit?
As one frontier sector among several, gated less by the technology than by binding offtake, incentive eligibility under the rules in force today and, for nuclear, licensing and the cost of the first unit.
The deck reports
Memoranda of understanding, pipeline, reactor designs
What gates revenue
Binding offtake and incentive eligibility under current rules; for nuclear, licensing milestones and first-unit cost
Policy has moved fast. The Department of Energy’s former Loan Programs Office, now the Office of Energy Dominance Financing, said in January 2026 that more than $83 billion of earlier loans and conditional commitments were being restructured, revised or eliminated, and named lower electricity prices and winning the AI race among its aims.
A project modeled on earlier rules needs a fresh eligibility read, while one serving data center demand or firm power may find support others no longer get. Advanced nuclear and small modular reactors add licensing milestones and first-unit cost, where the first-of-a-kind record from process plants applies in full.
What We Test: Energy and Nuclear
Executed offtake contracts with creditworthy buyers.
An eligibility opinion under current rules.
For nuclear, the licensing schedule and where the application stands.
The basis of the first-unit cost estimate, and who carries an overrun.
The Committee Question: Energy and Nuclear
Which revenue is contracted today, and under which rules does it qualify?
Funding the path
Match the Capital to the Milestone
Each kind of capital tests something different, and a frontier company usually needs several, in order, to reach and then repeat its gating metric.
Figure 9
Each kind of capital fits one milestone, in order
Source: Arete Ventures.
Exhibit 3
Which capital fits which milestone
Capital
The milestone it fits
What the provider tests
The common mistake
Grants and government programs
The milestone it fits: The science and the first demonstration
What the provider tests: Technical merit and public purpose
The common mistake: Planning around a program that can change faster than the build
Venture equity
The milestone it fits: Reaching the gating metric for the first time
What the provider tests: The team and its path to the metric
The common mistake: Paying for a first commercial plant with equity priced like a software round
Venture debt
The milestone it fits: Runway between two equity milestones
What the provider tests: Confidence that the next equity round will happen
The common mistake: Using it to bridge to a milestone further away than the loan term
Project finance and asset debt
The milestone it fits: Repeating a proven unit
What the provider tests: Operating history of identical units; contracted revenue
The common mistake: Expecting it for the first unit, which history says will underperform
Strategic capital and customer prepayment
The milestone it fits: A customer that needs the product
What the provider tests: The customer’s need, and the rights it gets in return
The common mistake: Granting rights that narrow the pool of exit buyers
Source: Arete Ventures.
The order matters. Grants and equity carry a company to its gating metric. Venture debt is safe only between equity milestones, with a lender that expects the next round. Project finance arrives once the metric has been shown on a unit the company can repeat. Strategic money helps when a customer needs the product, but a right of first refusal or an exclusive license can quietly remove the buyers who would have paid most at exit. Fund the plan on government programs as they stand at signing, with a case in which they fall away.
For limited partners
For LPs: How Can You Tell Whether a GP Can Underwrite Frontier Risk?
Ask the GP to name the gating metric for each of its largest frontier holdings, and to show one deal where diligence changed the terms.
Six questions, and what a good answer sounds like:
What is the gating metric for each of your five largest frontier holdings, and where does it stand?
Good answer One number per company, dated and measured at a customer.
Show us one deal where diligence changed the terms.
Good answer The tranche, closing condition or price change, and the finding behind it.
Who on the team has been accountable for scaling a technology?
Good answer A named person, and the plant, fleet or production system they ran.
How do you set reserves for holdings outside AI, now that follow-on rounds are thinner?
Good answer Reserves tied to each company’s gating date.
How do your marks move when a gating metric slips?
Good answer Marks linked to the metric. A weak answer points to the last round price.
Which exit route does each holding assume, and what does that buyer pay for?
Good answer Strategic buyers, public markets and secondaries kept apart, each with the metric it will test.
Our note on capital stack diagnostics covers the fund-level questions on IRR, DPI and continuation vehicles.
For boards
Five Questions for Every Quarter
What is our gating metric, and what did it do this quarter?
Which number in our last board deck is a proxy?
If the milestone slips six months, how much cash is left, and who funds the gap?
Which outside party could move our gating metric, such as a model vendor, a regulator or an offtaker, and what have we done about it?
What will the next lead investor test first, and do we already have the evidence?
How we work
How a Pre-IC Read Tests the Gating Metric
In about four weeks, we name the milestone the company claims, find the metric that gates revenue, trace it into the model, test it with evidence from outside the data room, and turn the gap into terms.
Figure 10
The pre-IC read in five steps
About four weeks · faster when the vote is close
1
Name the claimed milestone.
The number the deck leads with, and the date it promises.
2
Find the gating metric.
The measure at a paying customer that the revenue depends on.
3
Trace it into the model.
Which revenue, margin or capex line moves when the metric moves.
4
Test it outside the data room.
Customer calls, published resource estimates, plant and fleet data, on-chain records.
5
Turn the gap into terms, or into a pass.
The term from Exhibit 2 that fits, with the reasons in writing.
This is the core of our technical due diligence on frontier technology deals for GPs and LPs. When the vote is close, we run a shorter version on the two or three assumptions that decide it. We work on fixed fees, so advising against a deal costs us nothing.
Where Markets Stand · As of 7 November 2026
The backdrop for frontier deals
10-year Treasury yield
4.84%
Against below 2% for most of 2021. The Federal Reserve held rates in late October, and traders have raised bets on another hike.
Global venture funding, Q3 2026
$159B
AI: $102B (64%)
A record 27 rounds of $1 billion or more.
Secondary market, first half of 2026
$121B
Up 19%, with GP-led deals at $65 billion, the larger share.
Frequently Asked Questions
Frontier technologies are fields where the core science or engineering is still being proven at commercial scale. They include AI and AI-native software, synthetic data, quantum computing and quantum materials, blockchain and tokenization, biomanufacturing and life sciences, robotics and physical AI, and advanced energy, including nuclear. Investors often use “deep tech” for the same set. For an investment committee, what they share matters more than the label: technical risk has to be retired before market risk can be tested, and revenue waits on a milestone the company cannot yet show. That milestone, and its date, decide most of the value.
A gating metric is the measure that must be true at a paying customer for the revenue in a company’s financial model to arrive. It differs from the proxies a deck reports, such as pilots, benchmark scores or physical qubits, and it usually moves more slowly because something outside the company has to change.
Test revenue from production contracts separately from pilots, accuracy on the customer’s own data, gross margin after inference at next year’s usage, and what a customer would lose by switching if the model vendor shipped the same features. Bessemer found its fastest-growing AI cohort averaged about 25% gross margin, so margin deserves the same weight as growth.
Start from the workload the company plans to sell, find a published estimate of the logical qubits and error rate it needs, and compare that with what the company has demonstrated and promised, with dates. Then ask what it sells before fault tolerance, because that revenue funds the years in between.
Use grants and equity to reach the gating metric, venture debt only to bridge between equity milestones when a lender expects the next round, and project finance once the metric is proven on a repeatable unit. Strategic capital helps when a customer needs the product, but watch for rights that narrow the pool of exit buyers.
Someone who can name the gating metric for the sector, test it with evidence from outside the data room, and turn the gap into price and terms. Look for a technical due diligence consultant who has sat on investment committees and run the kind of operation the company is building, and who is paid the same whether the deal closes or not.
Start with the gating metric for the sector and what a good reading looks like, then test the date the company expects to reach it and the cash it has until then. Check whether anyone on the team has scaled the technology before, and decide which deal term fits any gap you find.
A buyer at exit pays for the gating metric, whatever the deck led with.
Arete Ventures provides operator-led private equity and venture capital advisory to GPs, LPs and family offices. Gaurav Shah sits on venture capital and private equity investment committees, voting on deals in robotics, deep tech and life sciences, and spent sixteen years as an operating partner at a San Francisco private equity firm. He is a chemical engineer with six US patents.
Sources
MIT NANDA, “The GenAI Divide: State of AI in Business 2025,” as reported by Fortune, 18 August 2025.
Bessemer Venture Partners, “The State of AI 2025,” August 2025.
ICONIQ, “2026 State of AI: The Builder’s Economy,” July 2026.
Epoch AI, “LLM inference prices have fallen rapidly but unequally across tasks,” 12 March 2025.
Google Quantum AI, “Making quantum error correction work,” Google Research blog, December 2024.
IBM Quantum, “IBM lays out clear path to fault-tolerant quantum computing,” June 2025.
rwa.xyz, tokenized real-world asset dashboard, accessed 7 November 2026.
E. W. Merrow, K. E. Phillips and C. W. Myers, “Understanding Cost Growth and Performance Shortfalls in Pioneer Process Plants,” RAND R-2569-DOE, 1981; E. W. Merrow, “A Quantitative Assessment of R&D Requirements for Solids Processing Technology,” RAND R-3216, 1986.
California DMV, autonomous vehicle test miles, December 2024 to November 2025.
Crunchbase News, Q3 2026 global startup funding, October 2026.
US Treasury yields as reported 7 November 2026.
Examples marked illustrative are hypothetical. Figures from outside sources are as of the dates shown. Arete Ventures is not a registered investment adviser, and nothing here is investment advice.