Nobody Checked the Physics: How Venture Due Diligence Went Missing Where the Checks Are Largest

Table of Contents

Fifteen years ago I spent a large share of my working weeks in rooms that JP Morgan and HSBC had arranged. The banks picked a small number of people they trusted on a subject, in my case cybersecurity, and put us in front of the hedge funds, corporate investors and sovereign wealth funds they most wanted to keep. The investors had read the filings. They knew which vendor claims had failed at a customer site and they wanted to know which of the companies they were pricing had built something that worked. Those sessions were the most demanding conversations of my career.

The invitations stopped. In July an investor called about an orbital data center company and asked whether the physics worked. Nobody on the pitch side had said they asked, and I wrote up what the physics said. This article is what I found when I went looking for whether the missing question was my nostalgia or a change I could measure.

Part of it is measurable and part of it is inference, and I will mark which is which. Venture financing split into two markets. For the median startup, raising got slower and stricter. For the frontier deals, where the checks are largest and the physics is hardest, closing compressed to days, and the visible substitute for the domain expert a bank used to put in the room became the names on the cap table. The inference is that in those deals the physics check was skipped. The alternative reading, that it was done privately, is one the investors could confirm and have not.

No public dataset measures technical review hours per deal, so the inference is drawn from deal speed, from who leads the rounds, and from the ratio of capital to revenue in the sectors that raise fastest. Quantum, where private venture investment reached $4.9 billion in 2025 against $1.4 billion of market revenue, is the sector where that ratio is widest.

Two Venture Markets

Among startups on Carta’s platform that raised a Series A in the fourth quarter of 2024, the median time since the seed round was 774 days. In the fourth quarter of 2021 the same interval was 420 days. On Carta’s platform, the median wait between a seed and an A is 84 percent longer than it was three years earlier. The A-to-B gap grew 97 percent over the same period.

Investors became more selective after 2022. Bridge rounds rose to 40 percent of all seed-stage venture rounds in 2024. Down rounds ran at 19 to 20 percent of all investments through every quarter of 2023. For most companies, capital got harder to raise and the process got longer.

The frontier deals moved the other way, at least in what could be seen from outside. Thinking Machines Lab, founded in February 2025 by former OpenAI CTO Mira Murati, closed a $2 billion seed round in July 2025 at a $12 billion valuation, led by Andreessen Horowitz with Nvidia, AMD, Cisco, Accel, ServiceNow and Jane Street participating. TechCrunch noted that earlier coverage in June had put the round at $10 billion; the valuation rose by a fifth in the weeks before signing. The company had shipped no public product and had disclosed little about what it was building.

Ilya Sutskever’s Safe Superintelligence raised $2 billion in April 2025, valued at $32 billion, with no product; the Financial Times, which broke the story, reported that even the investors had received only limited insight into the research.

In April 2024, TechCrunch reported that xAI had proposed raising $3 billion at a pre-money valuation of $15 billion. Investors asked for larger allocations, and when the round closed on 26 May, xAI had raised $6 billion at a pre-money valuation of $18 billion. Mistral raised a $113 million seed four weeks after it was founded, before it had a product and days after its first employees started.

TechCrunch described the pattern in April 2021, in a piece asking how firms were handling diligence when deals that once took months closed in days: the timeline for reaching conviction and executing due diligence had compressed, and some firms, one Sapphire Ventures investor said, were completing their diligence before a financing opened, a practice he called pre-diligence.

Both sets of numbers are true at once. The review period lengthened for the startup that needs eighteen months of revenue to justify an A, and closing compressed for the company whose founder’s last employer is the diligence. An observer of deep tech sees only the second market, because that is where deep tech is funded, and concludes that diligence has gone. What the public record shows is narrower: it has become unobservable from exactly the deals where it was needed most. Whether the lead investors performed a private technical review before those closes, or a pre-diligence before the round opened, is something they have not said, and nothing requires them to.

Dakin Sloss, who runs Prime Movers Lab, a venture firm that invests only in what it calls breakthrough science, made the point from the inside in a September 2022 piece for Fast Company. Venture firms, he wrote, cannot tell whether a company has developed a true breakthrough, a small iteration of existing technology, or pseudoscience that will never work, and he attributed that to two decades of capital flowing to mobile apps, social media, SaaS and fintech. His proposed remedy was to hire people who understand the science across several industries alongside the MBAs. I have not found a survey that counts how many firms did.

Four Reasons the Check Left the Room

Deployment pressure

Venture funds raised record amounts in 2020, 2021 and 2022 and then found fewer companies they wanted to back at the prices founders expected. US venture funds ended 2023 with $311.6 billion of committed but uninvested capital, a record, by the count of the NVCA, which uses PitchBook data. The two organizations’ Venture Monitor for the first quarter of 2024 put $227.9 billion of it, 73.1 percent, in funds raised between 2020 and 2022. Those funds have investment periods, typically three to five years, after which uncalled capital goes back to the limited partners and the management fee on it stops.

Carta’s data on fund deployment shows 2022-vintage funds had called just 43 percent of their capital two years in, the slowest pace of any recent vintage. PitchBook’s dry powder dashboard put half of all private-market dry powder in funds two to five years old by the second quarter of 2025, close to the record 54.4 percent. PitchBook’s Hilary Wiek described the consequence: managers under pressure to deploy write larger checks to win the deals they are willing to back.

A partner with $400 million to place before a 2027 deadline and a limited partner base that wants AI exposure has one incentive, to be in the round, and none to spend three months on thermodynamics.

Herding

David Scharfstein and Jeremy Stein published the model for this in 1990, in Herd Behavior and Investment in the American Economic Review. A manager whose reputation depends on being judged competent will follow the decisions of other managers and ignore private information that points the other way, because being wrong alongside the crowd is read as bad luck and being wrong alone is read as incompetence. The model was written for corporate investment and it describes venture syndication with uncomfortable precision.

In the first quarter of 2025, AI and machine learning companies took 57.9 percent of global venture dollars, $73.1 billion, by PitchBook’s count, with OpenAI’s $40 billion round accounting for most of it, and the headline PitchBook chose for its own report was that fear of missing out was driving the dealmaking.

Followers relied on the lead investor’s name, not its review

William Megginson and Kathleen Weiss showed in a 1991 Journal of Finance paper that the presence of a reputable venture investor certifies a company to later investors, lowering underpricing at IPO and the cost of going public. The measurement dates from the 1980s, when lead investors did the review before lending their names. When the lead investor’s own process has compressed to weeks, followers inherit a signal that may not be backed by verification. Every follower who reads a16z or Sequoia on the cap table as a substitute for its own review is relying on a check it cannot confirm was performed.

A lead investor puts in the most money and does the most work; the followers in a syndicate get the certification without the cost of producing it.

Sovereign wealth funds said as much to the IFSWF and PwC in their 2021 survey, Partnering for Success, which drew responses from 21 institutions: a majority, 58 percent, did not yet have the resources or expertise to lead a consortium or go it alone and preferred to be part of one, and the report’s conclusion was that sovereign funds therefore depend on general partners for deal flow and for expertise in due diligence. That is free-riding on diligence, stated in a survey, by the kind of institution that co-led PsiQuantum’s $1 billion round in 2025.

The independent expert layer was unbundled

The rooms I described at the start were a product of how investment banks paid for research. Analyst-hosted expert calls, field trips and curated one-to-one sessions were bundled into the trading commissions that institutional clients paid, and the bank’s selection of the expert was the guarantee of quality.

MiFID II, effective 3 January 2018, required investment firms in Europe to separate research payments from execution commissions. In ESMA’s interpretation, published in April 2017 ahead of the rules taking effect, arranging a meeting with a company’s management was not research but a discrete service that had to be paid for on its own, and a bank-arranged session with an outside expert was the same kind of service.

Greenwich Associates, now part of Coalition Greenwich, measured the result: the largest European institutions cut external equity research budgets by 19 percent in 2018 alone, and by the firm’s count the European equity research business shrank by 20 percent under MiFID II. A Capital Access Group survey found only 5 percent of fund managers would pay a broker to arrange a meeting.

Coalition, then a separate research firm and since 2020 part of the same company, counted analyst headcount across twelve major banks falling 8 percent in the first half of 2019, to 3,500, on pace for the sharpest annual drop since it began counting in 2012. In May 2025 HSBC dismissed more than two dozen analysts including its global head of fixed income research. The UK and the EU relaxed the unbundling rule in 2024; the FCA’s PS24/9 reintroduced the option to bundle from 1 August 2024. The departments that had curated the experts were already gone.

The function migrated. Commercial expert networks such as GLG, AlphaSights and Third Bridge grew into an industry with about $2.5 billion of revenue in 2024, by the estimate of Inex One, an expert-network aggregator. They changed the selection model: the client chooses from a marketplace of available experts and pays per call. The responsibility for knowing which specialist to ask, which used to belong to the bank, now belongs to the client. For public-market work that is a workable substitute.

To find out whether a proposed orbital data center can radiate its heat, the client has to know which expert to ask before any call is useful. Nobody in the marketplace tells them. The other destination was the cap table itself. What replaced the expert’s assessment of whether the thing works, in what an outside observer can see, was the observation that Benchmark had taken a board seat.

The $2 Trillion Factory

Aswath Damodaran, who teaches valuation at NYU Stern, has been trying to get the AI conversation onto numbers that both finance people and technologists can argue about. His framing on the BiggerPockets Money podcast was to separate the companies building the factory from the companies that hope to sell what comes out of it. The figures that follow are his estimates from that conversation, not audited industry accounts.

By his count the industry has invested more than $2 trillion in data centers and their inputs since 30 November 2022, the day ChatGPT was released, which he calls the largest build-out of any new business in history. The revenue from AI products and services, on the most generous count he could find, was about $250 billion over the trailing twelve months. To justify $2 trillion of capital at ordinary margins, he estimates the steady-state revenue of the AI product business would need to reach $8 to $10 trillion a year, and he puts the capex growth at $700 to $800 billion a year.

He applied three tests on air. Is it possible? Yes: total compensation paid to employees worldwide is about $26 trillion, so an AI industry that replaced labor at scale could in principle collect a share of it. Is it plausible? At $10 trillion, he says, the number is at the edge of plausible, because reaching it requires AI to displace a large fraction of the workforce. Workers who have been displaced are no longer the customer base that buys the products. Is it probable? He assigns it a low probability, and he is careful to say that the data do not support the word bubble either.

His written analysis of the sector, separate from the podcast, applies the same arithmetic to OpenAI, priced at $500 billion by its most recent investors on $13 billion of trailing revenue.

Damodaran and Bradford Cornell named the mechanism the Big Market Delusion in a 2020 paper. Each company is priced as if it will be one of the two or three winners in a market large enough to justify its own capex. The prices are individually defensible and, in the paper’s term, a collective overpricing, because the sum of the winners’ shares exceeds the market. He calls it a feature of every big new market rather than a bug in this one, and it is the same behavior the herding model predicts from a different starting point.

The other mechanism is financing arranged inside the industry. On 17 August 2026 SB Energy’s registration statement disclosed that Nvidia had provided residual-value guarantees covering up to $105 billion of OpenAI’s lease and power obligations at a data center campus in Pike County, Ohio, that SB Energy will build and operate under a 20-year lease to OpenAI and that will host Nvidia compute exclusively. The guarantee covers an initial 4.25 gigawatts, with an option on a further 3.75, and the first capacity is due in 2028. Nvidia also invested $1.5 billion in SB Energy, whose existing shareholders include OpenAI.

The guarantee was first discussed at $250 billion. It was cut to $105 billion after Nvidia’s shares fell on the larger figure.

Jensen Huang rejected the word circular on X, on the grounds that OpenAI will pay the lease. He is right that the structure differs from a supplier lending its customer the purchase price. Nvidia has also said the site could take about 1.5 million of its GPUs per generation, and the reading I take from the arrangement is that the supplier is now underwriting the customer’s ability to occupy the building. Fortune’s summary of the test is the correct one: whether AI companies can generate enough outside revenue to justify spending financed from within the industry.

The same investors, in the same syndicate structures and in several cases with the same strategic backers, are now pricing quantum companies.

Quantum Investors Are the Most Exposed

Private venture investment in quantum reached $4.9 billion in 2025, a 192 percent increase over 2024, by the count of the Quantum Economic Development Consortium (QED-C). PitchBook, using a narrower definition, has $3.9 billion across 127 deals. The same QED-C report puts the entire quantum computing market at $1.4 billion of revenue for the year. I covered the report when it came out; one year of new venture capital was 3.5 times one year of the whole market’s revenue.

Two things changed in the quantum capital base in 2025. First, PitchBook’s data show venture-growth capital jumped from roughly 1 percent of quantum deal value in 2024 to 27.5 percent in 2025. Second, the investors who wrote the largest checks changed. PitchBook’s list of the top quantum investors by capital deployed in 2025, as Fortune reported it, has BlackRock at $1.7 billion and Nvidia at $1.6 billion, followed by Baillie Gifford, Ripple Impact Investments and Temasek; no specialist quantum fund is on it.

Quantinuum’s $600 million round at a $10 billion pre-money valuation brought in Quanta Computer, NVentures and QED Investors among its new investors, alongside Honeywell, JPMorganChase, Mitsui and Amgen, who reinvested; Honeywell’s announcement named no lead. PsiQuantum’s $1 billion Series E, which the company said valued it at $7 billion, was led by funds managed by BlackRock with Temasek and Baillie Gifford. Neither round was led by a specialist quantum fund. Temasek, on PsiQuantum’s lead line, and the Qatar Investment Authority, among its new investors, are sovereign wealth funds of the kind the IFSWF survey described.

The 2021 SPAC class went public carrying multi-year revenue projections of a kind a traditional IPO prospectus does not carry, because underwriters will not take the liability for them. By my arithmetic in that piece, the three companies whose decks included 2025 forecasts projected about $1.2 billion of combined revenue for the year and delivered about $138 million. Rigetti reported $1.8 million of revenue for the three months to 30 June 2025, and its market capitalization on the day it reported was about $5.2 billion. D-Wave reported $3.1 million in the same quarter.

IonQ is the one member of the cohort whose revenue line resembles what the decks projected. It reported $130 million for 2025, the first quantum hardware company past $100 million. In August 2026, after second-quarter revenue of $80.1 million, it guided to $280 to $290 million of organic revenue for the year.

On 8 September it raised that to $450 to $460 million, but the increase folds in five months of SkyWater, the semiconductor foundry it acquired on 31 July, whose own first-half revenue was $317 million; the raise is mostly foundry revenue. Even IonQ’s $130 million for 2025 fell short of the $237 million its 2021 deck had projected for the year.

Quantum Computing Inc. is the case that shows who does the checking when nobody in the round says they have. Its shares rose about 1,200 percent over the year to November 2024 on announcements of a thin-film lithium niobate chip foundry, purchase orders from a research institute in Asia and from the University of Texas at Austin, and a NASA contract. On 27 November 2024 the short seller Iceberg Research published Quantum Computing Inc.: the Phantom Chip Foundry.

Iceberg reported that the company had removed all references to the University of Texas from its press release the day before. A professor at the university had told Iceberg the order was not reviewed before the announcement. The photographs the company had published of its foundry, Iceberg wrote, showed a laboratory where the company had described a mass-production facility on five acres.

Iceberg also noted that the company’s convertible financing had come from Streeterville Capital, an entity controlled by John Fife, whom the SEC had charged in September 2020, together with five other companies he controls, with acting as an unregistered securities dealer.

Capybara Research followed in January 2025 with separate allegations about the NASA relationship. Securities class actions followed both. I am not adjudicating the allegations; the SEC and the courts will. They came from a short seller with a position to profit from, who did the work of checking whether the foundry existed, and no investor on the register has said publicly that it asked to see the fab before the announcements.

Two rounds I assessed on this site show that the difference between an engineering-backed round and a narrative-backed one is visible to anyone who looks. QuantWare’s $178 million Series B was raised on shipped hardware and a manufacturing process third parties had used. OQC’s £260 million Series C came with a cost breakdown I could follow. Nothing in the pipeline requires anyone to look before the wire goes out.

The snake oil and quantum washing I write about are the retail end of the same failure. A company that stacks quantum, AI and blockchain in a pitch is betting that nobody in the room can evaluate any one of the three, and in the current pipeline that is a good bet. The vendors who sell premature cryptography products on Q-FUD and the founders who raise on a buzzword premium are working the same weakness in the audience from opposite directions.

What Biotech Does That Quantum Does Not

Biotech venture investing has a technical diligence norm that deep tech lacks, and the reason is the FDA. A drug cannot reach a market without staged, data-driven approval, so an investor who skips the preclinical data review is ignoring the only thing that determines whether the asset exists. The industry built its diligence around that fact.

External scientific advisory boards are expected of a credible company and their absence is read as a red flag. Specialist scientific diligence consultancies review mechanism of action, preclinical packages, IND readiness and manufacturing plans as a standard engagement, and investors and their advisers describe bringing in outside scientists to confirm that data were collected under the right conditions.

Quantum computing has no validation gate comparable to the FDA’s staged approval. It has competing measures of progress and no accepted one; I wrote in August that every vendor measures it differently. There is no norm of an independent scientific board, and no engagement that a growth-stage investor is expected to commission before leading a $1 billion round.

The one rigorous technical review of a quantum company I have seen documented was a government’s, and it reached the public only through a freedom-of-information release in March 2026. Before Australia committed to PsiQuantum’s Brisbane facility, two assessors from a Technology Due Diligence working group, one from the Defence Science and Technology Group and one from the National Measurement Institute, visited PsiQuantum’s Palo Alto headquarters in October 2023 to evaluate its staff levels, facilities and infrastructure.

Australia commissioned the review before committing public money. BlackRock and Temasek may have done the same before the Series E. If they did, they have not said so, and nothing required them to.

Cleantech 1.0 Already Ran This Experiment

Between 2006 and 2011 venture investors put more than $25 billion into cleantech and lost more than half of it. The post-mortem by Benjamin Gaddy, Varun Sivaram and Francis O’Sullivan at the MIT Energy Initiative, Venture Capital and Cleantech: The Wrong Model for Clean Energy Innovation, found that more than 90 percent of the cleantech companies funded after 2007 failed to return even the initial capital invested, and that the companies developing new materials, chemistries or processes returned a sixth of their capital. Cleantech, they concluded, clearly did not fit the risk, return or time profiles of traditional venture capital.

Solyndra, which had received half a billion dollars in federal loan guarantees, went bankrupt. KiOR went public in 2011 and filed for bankruptcy in 2014. Bessemer’s own retrospective reaches the same diagnosis: generalist investors applied a software playbook to hardware whose unit economics were unproven, and the science risk they had not priced was the risk that killed the companies.

The MIT authors also named who they expected to fund the next round of hardware: pension funds, family offices and sovereign wealth funds, which they described as set up to wait decades for returns but often inexperienced technology investors. That was 2016. Temasek and the Qatar Investment Authority, both sovereign wealth funds, were in PsiQuantum’s 2025 round.

The diagnosis reads as a description of quantum with the nouns changed. Long development cycles, capital-intensive hardware, physics that has not been demonstrated at the scale the pitch requires, and investors whose experience is in a different asset class. Cleantech had one advantage quantum lacks: the products it promised, solar panels and biofuels, could be tested against a known market price. A quantum computer’s product is an advantage on a problem class that has not been demonstrated at commercial scale by anyone, which is why I keep asking why customers are buying machines that cannot do anything yet.

One Recommendation

I invest in quantum startups through Quantum.Partners and Applied Quantum does technical due diligence for investors, so I am describing a market I have an interest in. With that on the table, the recommendation is aimed at limited partners and growth-stage investors. A venture firm racing a competitor to close will not volunteer to slow down.

Limited partners should require, as a term of their commitments, that any deep-tech check above a threshold they set is preceded by an independent technical feasibility review that the fund did not write and the founder did not choose. The model is biotech’s, where the review is a condition of the asset existing and the industry organized itself around that.

The specialist firms exist. Aspen Quantum Consulting describes its team as including a former co-director of the Joint Quantum Institute. h-bar, Simon Devitt’s consultancy, says it has reviewed more than $100 million of quantum investments across two dozen startups. My own practice is one more. All of them are small, because nobody in the pipeline is required to hire them. Demand for them will come from the people whose money is at risk, not from the intermediaries who are paid to deploy it.

The review has to answer a narrow question: does the physics permit what the pitch requires, at the scale the valuation assumes, on the timeline the capex implies. For the orbital data center in July the answer took an afternoon with the Stefan-Boltzmann law. For a quantum company the questions are the ones I use in the capability framework. Has the error correction that the roadmap assumes been shown below threshold on the company’s hardware, and what error rate has been demonstrated on how many physical qubits? What does the logical qubit count the business plan requires imply for the physical qubit count and the cryogenics?

Those are checkable. A reviewer with the right background can answer them in weeks, which is faster than the median founder’s wait for a Series A and slower than the frontier round closes now. That difference in speed is the whole cost of the fix, and the $4.9 billion of 2025 is the scale of capital now exposed to the question.

What Replaced the Curated Expert

The banks still run conferences, and Goldman and Morgan Stanley still introduce capital to managers. What they no longer do is choose the expert and put her in the room as a condition of the relationship. The regulators who unbundled research in 2018 were addressing a real conflict, and the partial reversal in 2024 came too late to rebuild what had been cut. When the departments disappeared, the independent technical judgment they had provided was replaced by a name on the cap table, a signal that can be copied without being earned.

Whether the trade was worth it is now an empirical question. The 2021 quantum SPAC class was the first data point: by my count in the SPAC piece, it returned about 12 cents of 2025 revenue for every dollar its decks had projected. The 2025 cohort attracted $4.9 billion in a single year with no more public evidence of an independent technical review than the last one had. The next set of results is due in about the time it takes a Series E to reach its next round.

The post appeared first on PostQuantum - Quantum Computing, Quantum Security, PQC.