Original Research · Arc Shift Ventures

Startup programs measure everything except the sale.

We went looking for one number from every major corporate startup-partnership program we could find — cloud-credit programs, corporate accelerators, incubator partnerships, big-tech startup teams: the rate at which startups in the program become commercial customers or partners of the corporate running it. The programs publish credits issued, cohorts graduated, alumni funding raised, jobs created. Exactly one program model publishes conversion. It isn’t a coincidence which one.

Talk about a program you run — or one you’re in See the numbers

Four numbers that frame the problem

60–80%

of US accelerator programs leave startups worse off than never joining, in the largest study yet run — ~750,000 startups across 329 programs. The gains concentrate in a small elite; selection, not acceleration, explains most of the rest.

Baek & Hegde, NBER Working Paper 35063, April 2026 · Working paper, not yet peer-reviewed; disclosed sample · Grade A · Conflict disclosed: one author directs an accelerator the paper places in its top-performing tier

280,000

startups have received credits from the largest cloud provider’s startup program — US$6 billion issued over the program’s first ten years. Number of published figures on how many became paying customers of that cloud: zero.

The provider’s own 10-year program retrospective, 2023 · vendor-published · Grade B

$31.2B

raised by the alumni of another tech giant’s accelerator, with 109,000 jobs created across 1,700+ alumni. Number of published figures on alumni who became cloud or enterprise-partnership customers: zero.

The company’s 2025 accelerator impact report · vendor-published · Grade B

>50%

of startups that pilot with one carmaker’s venture-client unit convert to paid follow-up projects — from 280+ pilots, 6,000+ startups assessed. The one program type that publishes a conversion rate is the one whose unit of activity is a purchase order.

The unit’s own figures, via an industry interview · self-reported · Grade B

The core finding

The funnel has no bottom — by design, not by accident

Every startup program reports a funnel: startups reached, startups engaged, credits consumed, demo days held, alumni dollars raised. Read enough of these reports side by side and you notice the same absence in all of them. The metrics stop precisely at the point where a business unit would have to say yes.

This is not a measurement gap. Measuring conversion is easy — the venture-client model does it with a supplier number and a PO. It is a design choice. Activity metrics (cohort size, credits issued, applications received) can all be hit by the program team alone. A conversion metric — startup becomes paying customer, pilot becomes production contract, referral becomes account — requires someone outside the program to commit budget. Programs are staffed, funded and judged on the numbers they can control, so those are the numbers that exist.

The proof is the counter-model. The venture-client approach — pioneered in German industry in 2015, adopted across several industrial groups, and productised by a specialist firm that a major consultancy acquired in 2023 — defines the engagement itself as a purchase: the startup gets a supplier number and a paid order on day one. Because the activity is a transaction, the conversion rate is measurable, and so it gets measured and published: the pioneering carmaker’s unit reports over 50% of pilot alumni winning paid follow-up projects; a 2024 industry survey of venture-client practice reports companies with dedicated venture-client units adopting 25% of piloted solutions versus 10% without one. Grade those numbers B — they come from the model’s own vendors — but note what they are: the only startup-program conversion rates published anywhere in this field.

The falsifiable claim: among the major cloud-credit programs, corporate accelerators and startup-partnership teams we surveyed, none publishes a startup-to-commercial-customer conversion rate; the venture-client model does. Show us one credits program or corporate accelerator that publishes its conversion rate with a disclosed denominator, and this page gets rewritten.

Where it actually breaks

Not at engagement. At the handover to a revenue owner.

Stage 1 — Engagement: measured obsessively. Applications, cohort selections, credits activated, event attendance, delegation visits. This stage is fully owned by the program team, so it is instrumented to the decimal. It is also where nearly all published numbers live.

Stage 2 — Pilot / PoC: counted, sometimes. Pilots started, PoCs run. One prominent bank-backed fintech incubator publishes 250+ startups through programs, US$370M+ raised by alumni, 65+ corporate partners “enabling pilots and PoCs” — activity and startup-side outcomes. What is not published: how many of those PoCs became commercial contracts with the partner banks. (Incubator/press figures, Grade B.)

Stage 3 — Commercial conversion: unowned, unmeasured, unpublished. The startup-to-customer handover belongs to a sales team or business unit that never signed up to the program’s goals — the same unowned handover that kills corporate pilots (see Beyond the Pilot: 81% of corporates convert fewer than 1 in 4). The program’s incentive is to graduate the cohort and report the raise. The business unit’s incentive is its own locked plan. Nobody’s compensation moves when a program startup becomes an account, so nobody counts it.

Two corroborating anchors. The European Innovation Council measured what happens when engagements are tracked end-to-end: 6% of 1,500+ corporate-startup engagements reached a signed deal (EIC 2025, Grade A — the low anchor). And the NBER accelerator study found that founder quality and prior financing predict outcomes far more than program participation — meaning a program reporting its alumni’s raise totals is largely reporting its own admissions filter, not its effect. (Baek & Hegde, Grade A.)

Honest caveat: absence of a published number is not proof the number isn’t tracked internally. Cloud programs almost certainly track credits-to-consumption conversion privately — these are disciplined businesses. But a decade of impact reports without a single conversion figure, from companies that publish everything favourable, is evidence of what the favourable numbers are not.

What actually works

Give the funnel a bottom: one terminal event a budget owner must sign

If you run a program

Define the terminal event before the cohort. A purchase order, a production contract, a referral-attributed account — one named event that requires a budget owner outside your team. If you can’t name it, you’re running marketing, which is fine, but budget and report it as marketing.
Publish your conversion rate, with the denominator. The venture-client pioneer’s >50% follow-up rate does more for its startup dealflow than any credits headline. The number doesn’t need to be high to be credible — it needs to exist.
Source cohorts from business-unit problems, not open applications. The venture-client insight: start from a named internal customer with a named gap, then find the startup. Adoption at 25% vs 10% in the 2024 venture-client industry survey is a vendor number (Grade B), but the mechanism — demand-first sourcing — is the consistent thread in every model that converts.
Pay the startup from day one. A paid pilot at PO-triggering price forces procurement, legal and a cost centre through the pipes before scale-up, when stakes are low. Free pilots defer every hard conversation to the moment enthusiasm is weakest.
Report one activity metric per outcome metric. For every “startups engaged,” one “startups converted.” The discipline changes what the team optimises within a quarter.
Measure time-to-first-PO, not time-to-demo-day. Demo day is a date you control; a PO is a date the business controls. The gap between them is your program’s real friction number.

If you’re the startup in one

Treat credits as cost reduction, never as traction. US$350k in cloud credits changes your burn, not your business. No investor underwriting you, and no enterprise buying from you, counts it as revenue — neither should you.
Ask the program its own conversion number. “How many alumni have a commercial contract with you or your corporate partners?” A crisp answer means you’re in a funnel with a bottom. A pivot to alumni-raise totals answers the question too — just differently.
Get introduced to the budget owner, not the innovation team. The program’s real asset is warm access to business units. Insist the intro lands on someone who owns a P&L line your product touches; multi-thread past your program champion before the cohort ends.
Prefer programs whose parent buys — if the corporate is your customer. A venture client pays you and gives you a supplier number; an accelerator gives you a stage. If you sell to enterprises like the sponsor, the PO is worth more than the demo day.
Time-box the program against CAC. Three months of program time is three months of founder-hours. If the expected value of program-brokered contracts (not meetings) doesn’t beat your normal pipeline, attend the opening, skip the middle, use the logo.
Use demo day for the audience it actually converts: investors. The evidence that demo days create enterprise customers is nonexistent; the evidence they create investor meetings is decent. Aim the asset at the target it can hit.

The honest caveat: the venture-client conversion numbers on this page come substantially from the model’s inventor and his firm, since acquired by a major consultancy — the people with the strongest incentive to publish them. Publication bias runs in both directions: programs that convert publish, programs that don’t stay silent. What survives that filter is the structural point, not any single rate: models that define engagement as a transaction can be measured; models that define engagement as activity resist measurement, structurally.

The citation audit

Four numbers this field repeats that don’t survive a source check

“60% of corporate accelerators fail within two years”

Attributed to a widely-cited industry research firm

Unverifiable as quoted

The number exists only in the headline of a 2019 research page from a widely-cited industry research firm. The publicly visible article contains no sample, no methodology, no definition of “fail,” and no study behind the figure — the body sits behind a trial-signup wall, and no underlying dataset has ever surfaced in anything citing it. It is quoted across consulting decks as an established measurement. As far as public evidence goes, it is a headline.

“Corporate accelerators cut a startup’s success rate to 8%”

A venture builder’s internal analysis, via startup press

Quoted far beyond its weight class

Real analysis, quoted far beyond its weight class. The source is a 2020 startup-press newsletter item describing a venture builder’s internal analysis of one founding cohort (startups founded 2013): baseline “success” 11%, accelerator 12%, corporate accelerator 8%. Sample size undisclosed; “success” undefined publicly; never peer-reviewed. The outlet’s own readers spotted the flaw immediately — by the same yardstick the world’s best-known accelerator is a ~93% “failure.” Directionally interesting, and consistent with the NBER finding; a measurement it is not. Grade B.

“96% of our accelerator alumni survive, while over 80% of startups fail”

A global tech company’s accelerator impact report

Vendor stat on a zombie baseline

Two problems compound. First, selection: the program admits a tiny, heavily screened, often already-funded fraction of applicants — the NBER study shows startup quality sorting into programs is first-order, so high alumni survival is what selection alone predicts, program effect or none. Second, the baseline: “over 80% of startups fail” has no consistent primary source — it circulates in variants (80%, 90%, “9 out of 10”) traceable to no single disclosed study, and it describes lifetime failure of all startups, compared here against a few years’ survival of funded ones. A true-sounding comparison built from a selected numerator and an unsourced denominator.

“80% of the world’s unicorns run on [our cloud]”

A major cloud provider, citing a market-data firm’s unicorn list

True-ish, but a category error where it’s deployed

This is a market-share observation about cloud infrastructure among late-stage companies — and it appears inside the provider’s flagship credits-program retrospective, where it functions as evidence the program produces winners. It measures nothing about the program: not credits-to-customer conversion, not program causality, not even whether those unicorns ever took the program’s credits. The startup-program version of “our alumni include a household name.”

A pattern worth naming: all four are numbers about activity or association standing in for numbers about conversion — which is the whole mirage in miniature.

Working on a startup program — or stuck inside one? I’d like to compare funnels.

I connect startups to corporate innovation teams across eleven APAC markets, so I watch daily where program engagement does and doesn’t become revenue — from both sides of the table. I’ve run the corporate build, and I’ve sat with a hundred-plus founders on the Building Real podcast talking about what actually converted.

If you run a partnership program and want to know what your funnel’s bottom looks like — or you’re a founder deciding whether a program is worth three months — reach out. Happy to compare notes even if there’s nothing to sell.

Book a 30-min conversation Connect on LinkedIn

Method & honesty note

How this was put together

We collected the published metrics of the major cloud-credit programs, corporate accelerator research (academic and commercial), fintech incubator outcomes, and the venture-client literature (Gimmy et al. and industry surveys), and traced every conversion-adjacent statistic toward its primary source. Grades: A = peer-reviewed study or disclosed-sample primary research · B = industry-published or self-reported data, sample not independently verified · EST = reasoned estimate. Body text describes source categories rather than naming organisations; the full named citations sit in the appendix below.

The central claim — that conversion is unpublished everywhere except the venture-client model — is an absence claim, and absence claims are falsifiable but not provable: we can only report that we looked and did not find. Where a number is soft, it is labelled soft on the page rather than rounded into confidence.

The full limits file — eight notes, unabridged

The core claim is an absence claim. “No program publishes conversion” can be falsified by one counterexample but never fully proven; we surveyed the major programs’ public materials, not the entire field. A program publishing a conversion rate in an obscure annual report would partially defeat the claim — and we’d welcome it.
Unpublished ≠ unmeasured. The cloud providers almost certainly track credits-to-consumption internally; banks may track PoC-to-contract privately. The finding is about what’s published — which shapes how the field talks, benchmarks and budgets — not about internal dashboards we cannot see.
The counter-model’s numbers come from its salesmen. The >50% follow-up rate and the venture-client survey figures are published by the model’s inventor, his firm, and its consultancy acquirer. No independent replication of venture-client conversion rates exists. We use them as proof that conversion can be measured and published — not as proof of the model’s superiority at those specific rates.
Publication bias cuts both ways. Programs that convert have every incentive to publish; silence correlates with weak numbers but doesn’t prove them. Conversely, the venture-client literature’s strong numbers survive precisely because weak venture-client units don’t write reports.
The academic base is thin and mismatched. The best study (Baek & Hegde) is an April 2026 NBER working paper — large sample, not yet peer-reviewed, and one author directs an accelerator the paper places in its top-performing tier — and it measures startup outcomes, not corporate-side conversion. Kohler 2016 is qualitative (n=40 interviews). Seitz et al. 2024 covers 15 German programs. Almost no academic work measures the corporate side of the funnel at all — which is itself part of the finding, but limits triangulation.
SEA-specific conversion data is nearly nonexistent. The page’s audiences sit in Singapore and ASEAN; nearly all graded evidence is US/European. Arjun’s practitioner vantage point (BLOCK71, NUS GRIP, corridor work) fills this gap anecdotally, and the page should not imply otherwise.
Figures may drift. The cloud provider’s credit total ($6B in the 2023 retrospective, ~$7B on current pages), the accelerator’s alumni count (1,100+ in earlier coverage, 1,700+ in the 2025 report) and the incubator’s totals are moving vendor numbers; the page anchors each to a dated source. A pre-publication re-check of appendix items 1, 4, 5 and 11 is recommended.
The 2026 SEA programs are too young to judge. The newest SEA cohorts (an agentic-AI accelerator, an AI credits program) launched 2025–26; their absence of conversion data is a function of age, not necessarily design. The page’s structural argument applies to their announced metrics (credits, cohort size), not to outcomes they haven’t had time to produce.

Appendix

Every number on this page, with source, grade and provenance

Sources appendix — 20 entries

Named sources appear here so every number can be checked; the body text deliberately describes categories rather than singling out organisations.

1 · 280,000 startups; US$6B credits issued (2013–2023) Grade B
Amazon, “How AWS Activate has helped more than 280,000 startups,” Oct 2023

Amazon’s own 10-year retrospective; no external audit. Later AWS pages cite ~US$7B — the 2023 page’s $6B is used as the anchored figure.

2 · No published AWS Activate credits→paying-customer conversion Grade A (absence)
Absence finding across aws.amazon.com/startups pages and the retrospective above

Verified absence in Amazon’s own program publications; internal tracking unknowable.

3 · “~80% of unicorns run on AWS” Grade B
Same Amazon retrospective, citing PitchBook’s unicorn list

Vendor claim; market-share observation, not a program outcome. Audited on-page.

4 · $31.2B alumni funding; 109,000 jobs; 1,700+ alumni; 87 countries Grade B
Google 2025 Accelerator Impact Report · PDF

Google-published; startup-side outcomes only; no customer-conversion figure anywhere in report.

5 · “96% alumni survival vs >80% of startups fail” Grade B
Google inaugural Accelerator Impact Report

Vendor comparison; baseline unsourced. Audited on-page.

6 · 60–80% of accelerator programs leave startups worse off; ~750,000 startups, 329 programs; selection is first-order Grade A
Baek & Hegde, “Beyond Demo Day: Sorting and Value Added in Startup Accelerators,” NBER WP 35063 (April 2026) · NYU Stern summary

Disclosed-sample working paper (not yet journal-published — noted in limits).

7 · BMW Startup Garage: 6,000+ startups assessed; 280+ pilots, 26 countries; >50% of alumni win paid follow-up projects Grade B
Munich Startup interview · bmwstartupgarage.com

Self-reported by the program; consistent across BMW’s own materials over multiple years.

8 · Venture-client model definition and BMW “10x more startup adoptions than CVC” Grade B
Gimmy, Kanbach, Stubner, König & Enders, “What BMW’s Corporate VC Offers That Regular Investors Can’t,” Harvard Business Review, July 2017

Practitioner-academic article by the model’s inventor; case description, not an outcome study.

9 · Venture-client units vs none: pilot rate 54% vs 20%; solution adoption 25% vs 10%; 56% vs 10% adopting within 12 weeks Grade B
27pilots, 2024 State of Venture Client Report · summary (Gimmy, Medium)

Vendor survey by the model’s commercial champion; respondent n not disclosed in public summary. Conflict of interest flagged on-page.

10 · 27pilots acquired by Deloitte (Jan 2023); model at 20+ corporates; “10–15 pilots/yr, ROI of 3” Grade B
27pilots announcement · consultancy.eu · tech.eu

Acquisition factual (multiple outlets); ROI and pilot-count claims vendor-published.

11 · Tenity (ex-F10): 250+ startups since 2015; ~70% still active; US$370M+ raised by alumni; 65+ corporate partners; SGD 300k incubation tickets Grade B
fintechnews.ch · tenity.com · startupticker.ch

Company/press figures (plus Batch-1 research file, Julius Baer section); no published cohort→bank-contract conversion rate (absence finding).

12 · EIC: 6% of 1,500+ corporate–startup engagements reach a signed deal Grade A
European Innovation Council corporate-partnership data, 2025 — as sourced and graded on arcshift.ventures/beyond-the-pilot

Carried over from the Beyond the Pilot sources; disclosed engagement count (1,500+ at 120+ corporates).

13 · 81% of corporates convert fewer than 1 in 4 pilots Grade A
500 Startups, 2017, n=100 corporate innovation programmes — as graded on arcshift.ventures/beyond-the-pilot

Cross-reference to existing page; disclosed sample.

14 · “60% of corporate accelerators fail after 2 years” Unverifiable
CB Insights, May 2019

Headline only; visible body contains no methodology, sample, or definition; remainder paywalled. Audited on-page.

15 · Stryber analysis: 11% baseline / 12% accelerator / 8% corporate-accelerator “success” Grade B
Sifted (Maija Palmer), May 2020

Secondhand report of internal analysis; single 2013 founding cohort; n and success-definition undisclosed. Audited on-page.

16 · Kohler corporate-accelerator framework, n=40 interviews Grade A
Kohler, T., “Corporate accelerators: Building bridges between corporations and startups,” Business Horizons 59(3), 2016

Peer-reviewed; qualitative (interviews at Intel, Samsung, Orange, Cisco programs); design framework, not conversion rates.

17 · 223 alumni startups from 15 German corporate accelerators; specialization trade-off in post-program performance Grade A
Seitz, Lehmann & Haslanger, “Corporate accelerators: design and startup performance,” Small Business Economics 62 (2024)

Peer-reviewed, disclosed sample; measures startup performance, not corporate-side conversion — itself evidence of where the literature’s lens sits.

18 · Steve Blank “innovation theater” critique Opinion
Blank, “Why Companies and Government Do ‘Innovation Theater’ Instead of Actual Innovation,” Oct 2019

Practitioner essay; used for framing only, never as data.

19 · Google SEA: 25-startup agentic-AI accelerator, US$350k credits; AI Innovation Corridor w/ EnterpriseSG Grade B
Google Cloud press corner via Batch-1 research file, ~May 2026

Program-launch announcement; inputs only, no outcome data yet (program too young).

20 · AWS ASEAN: APJ Innovation Hub Singapore (Jun 2025); AI Springboard — up to 300 SG businesses, up to S$350k credits Grade B
press.aboutamazon.com · mti.gov.sg (via Batch-2 research file)

Launch announcements; activity commitments, no conversion data.