Speedrun vs YC:
what the companies build

Physical and biological products: 4.3% at Speedrun, 14.0% at YC

We looked for a physical product, integrated robot, industrial production system, biological therapeutic or biological computing substrate as a principal deliverable. Software and data sold to those industries were counted separately. Both proposed and current products qualify.

The gap appears in both cohort comparisons: 5.1% versus 11.6% for Speedrun SR006 and YC Winter; 3.8% versus 16.1% for SR007 and YC Summer. These are shares of the publicly listed records, with different cohort sizes and disclosure dates.

Nine YC candidates remain unresolved. Records that did not trigger the keyword search may include missed physical products. There has been no independent review of these labels.

Download all 567 records, labels and sources · Download cohort totals

Applied AI: 79% of the Speedrun sample, 60% of YC’s

The pilot assigns one primary technology family based on what the company principally builds or delivers. Physical systems can still use AI. “Software; AI unspecified” means the reviewed evidence does not establish an AI implementation; it does not mean the company uses no AI.

• Applied AI: 38/48 Speedrun companies (79.2%) versus 29/48 YC companies (60.4%). These labels describe declared product approaches, not independently audited code or model ownership.
• Physical/biological families combined: 4/48 Speedrun companies (8.3%) versus 13/48 YC companies (27.1%). This combines hardware/electronics, robotics, industrial processes and biotech.
• Broad AI product language is much closer: 38/48 versus 36/48 after selected website enrichment. That signal can include companies building hardware for AI and must not substitute for a core-technology classification.
• Speedrun’s physical-system examples include Quo Labs/Sam, Syncere, Galileo Space and KIRA.
• YC examples include Piris Labs, Hlabs, 6thSense, Ultrasonium, ProvenMetal, Gamgee and Parasma.

Among software products, both samples are mostly applied AI

Among companies classified as software products, 29/32 Speedrun companies and 22/24 YC companies fall into applied AI. Under this coarse technology rubric, software-only effective diversity is 1.45 for Speedrun and 1.33 for YC. The sample sizes differ, so this is descriptive rather than an equal-size ranking.

The sample does not establish differences in algorithms, model ownership, training methods or proprietary data. Those would require a separate technical review.

Software products, services and physical systems

The operating-model rubric distinguishes software vendors, delivered services, hybrid businesses, physical products, financial operators, marketplaces, research/IP, and content experiences. It asks who does the work and what the customer receives. It does not infer subscriptions from an ARR claim or a product login.

• Physical products/systems: four versus twelve. Parasma is instead research/IP, explaining why YC's physical/biological technology count is thirteen.
• Delivered services plus hybrids: nine versus four. This includes professional work and operational services; it is not a profitability or labour-efficiency measure.

The sample’s operating-model diversity estimate is higher for YC, but the ranking can reverse in resamples. Full calculations and category-change tests appear in the sensitivity checks.

Pricing evidence is available for only 8 of 96 companies

Reviewed sources establish a charging unit or mechanism for eight companies. The other 88/96 are unknown in this evidence set, not necessarily undisclosed everywhere. Fifteen company websites were checked specifically to resolve ambiguities; this was not a comprehensive pricing-page crawl.

• Sam: product purchase with no current subscription for the advertised buyer group. [8]
• Hammock: employer pricing per employee per month, with a quote. [9]
• Athena: fixed legal transaction fees; Concorda: legal-service charges plus software access. [10, 11]
• Maestro: subscriptions; Mecone: proposed licensing of financial indices. Their other potential revenue sources remain unverified.
• Didit: usage-based identity checks; TryNearby: monthly restaurant subscriptions.

The sources establish what companies deliver more often than how they charge. Revenue-model diversity, margins and pricing power remain unmeasured.

Customer markets: 16 in the Speedrun sample, 13 in YC’s

The Speedrun pilot spans sixteen primary market labels; YC spans thirteen. Effective diversity is 12.59 versus 10.11, accounting for concentration rather than simply counting labels. The largest three markets account for 43.8% of Speedrun sample companies and 52.1% of YC sample companies.

• Industrial and energy exposure is equal: eight companies each. The products differ: operational software alongside physical production, sensing and chemistry.
• Finance and insurance exposure is equal: seven each. Within that market, banking software, financial institutions, collections platforms and trading firms should not be collapsed.
• Media and entertainment: six Speedrun versus one YC.
• Developer and data infrastructure: four Speedrun versus ten YC. This includes physical networking, biological computation and software, the category includes several technologies.
• Businesses dominate both buyer mixes: 34/48 and 37/48 have an exclusively business buyer label. Nine versus five have an exclusively consumer label; mixed and government cases remain separate.

The market ranking can reverse in resamples. Several categories contain one company, and changing the category definitions could change the result. The sample is too small to establish that Speedrun serves more markets across its portfolio.

Company comparisons

Rosebud AI appears in both programmes

Rosebud AI appears in Speedrun SR001 and YC Summer 2019, with the same domain and founder on the two official profiles. That confirms at least one historical shared company; founding-year metadata differs between the profiles. Other exact-name matches remain identity leads until domains and founders are checked. None is automatically counted as shared from its name alone. [14, 15]

Text similarity does not establish competition

The downloadable company-pair search compares descriptions across the 139 Speedrun and 428 YC records. We also reran it after removing generic AI/software language. Shared words can produce high scores for different products, so these candidate pairs are excluded from the diversity calculations.

Implications and unanswered questions

1. Speedrun’s industry coverage may be broader than its technology mix. The sample covers 16 markets, but 38/48 companies fall into applied AI. Classifying the remaining 91 recent Speedrun companies would test whether this holds beyond the sample.
2. YC’s hardware and biology exposure accounts for much of the technology gap. The 567-record screen identifies physical or biological deliverables at 14.0% of YC records and 4.3% of Speedrun records. A second reviewer should check the classifications and audit records that did not trigger the keyword search.
3. Software companies in the samples look more alike. Applied AI accounts for 29/32 Speedrun software products and 22/24 YC software products. Comparing model providers, training, data ownership and system design would test differences within that category.
4. Both programmes back businesses that perform work for customers. Nine Speedrun and four YC sample companies deliver services or combine them with software. Contracts, human hours per job and margins would show whether these companies operate differently from conventional service firms.
5. Speedrun’s directory now contains few Gaming tags. The tags fall from 19/25 in SR001 to 0/80 in SR007. Archived applications and launch descriptions could separate a change in selection from later pivots and retagging.
6. The observed mix cannot explain selection. We lack applicant and rejection data. Different portfolios could reflect different applicant pools, acceptance decisions or willingness to publish company profiles.
7. The data cannot tell investors which programme produces better companies. We have no comparable revenue, survival, exit or return data. Cohort age, starting stage, investment size and follow-on funding would also need to be controlled.
8. More categories do not establish lower investment risk. Companies in different markets may depend on the same model providers or customer budgets. Investment weights and shared dependencies would be needed to assess portfolio diversification.

Search the 96 company classifications

Each row includes the classification rationale and source. These are single analyst/AI-assisted labels, not independent human validation. “High confidence” means clearer supporting public evidence, not a validated company claim or a measured probability.

96 companies

No companies match these filters.

Cohorts, sources and sample selection

We collected all seven public Speedrun cohort lists and a YC-OSS directory snapshot on 16 October 2026. The comparison uses 139 records from Speedrun SR006/SR007 and 428 from YC Winter/Summer 2026. Stealth and other unlisted companies are excluded. [1, 2]

The full collection contains 322 Speedrun and 6,307 YC directory records. These counts include older cohorts and are not the comparison denominators. Speedrun’s FAQ describes recent residencies of 70–80 companies, while SR006 lists 59. The actual historical cohort size is not established, so its disclosure rate is unknown. Duplicate names and slugs also need identity checks before the historical records can be counted as distinct companies. [3]

Sample selection

• Speedrun: 24 companies from SR006 and 24 from SR007.
• YC: 24 from Winter 2026 and 24 from Summer 2026.
• Within each cohort, records were sorted by a fixed SHA-256 hash of programme plus directory ID using seed string 20261016. The first 24 were selected before reading company models.
• Equal cohort weights make this a balanced two-cohort comparison. It is not the investment-count-weighted composition of either entire 2026 portfolio.
• Spring 2026 YC companies are collected but excluded from this pilot. Fall 2026 is ongoing and excluded. The census records their current public counts separately.

SR006 ran late January–April; SR007 ran July 27–October 11. YC's Winter and Summer Demo Days were March 24 and September 10. These are nearby comparison periods, not perfectly synchronized admission or disclosure windows. SR007 had only recently finished at the snapshot date. [4, 5, 6]

Speedrun’s current Gaming tags, by cohort

Speedrun’s current directory tags mark 19/25 SR001 records as Gaming, compared with 2/59 in SR006 and 0/80 in SR007. a16z documented an expansion beyond gaming in March 2025. Archived profiles would be needed to establish what every company did when it joined. [7]

“AI” appears in 35/59 SR006 headlines and 41/80 SR007 headlines, compared with 89/198 YC Winter and 105/230 YC Summer headlines. This counts wording in the same directory field; a company can use AI without putting it in its headline.

Sensitivity tests and sampling uncertainty

Effective categories account for both the number of labels and how evenly companies are spread across them. A value of four means the same diversity as four equally common categories. The calculation is exp(Shannon entropy); it depends on the labels and the sample.

4,000 bootstrap draws, sampling with replacement within each of the four pilot cohorts. The intervals describe variation conditional on the pilot and its labels; they are not causal confidence intervals, do not incorporate non-disclosure or coding error, and use a finite-sample-biased plug-in diversity estimator. They should not be treated as a formal discovery test after examining multiple measures.

• Exclude medium-confidence labels: technology diversity remains higher for YC, at 3.63 across 40 cases versus 2.29 across 38 Speedrun cases. Unequal retained sample sizes limit direct comparison.
• Merge physical/biological technology families: effective technology diversity remains higher for YC, at 2.71 versus 2.03.
• Use inverse concentration rather than Shannon diversity: technology inverse-Simpson values are 2.47 versus 1.56, retaining the direction.
• Merge the game/content category into software: YC's operating-model diversity falls to 3.97 versus 3.17; the pilot direction survives the label change.
• Equal-weight the thirteen shared markets: software-product shares become 62.4% Speedrun versus 49.4% YC; applied-AI shares become 82.4% versus 70.4%. Several market cells have only one observation, so this is a fragile sensitivity check, not a matched-market estimate.
• Current snapshots, not admissions: a pivot can change a company's classification. Original application-time diversity has not been measured.

The later cohort has more technology variety in each programme’s sample. Each cohort contributes only 24 companies; Spring YC is excluded. These four samples cannot establish a trend over time.

Missing data and limits

• Full classification: 96 companies are source-coded. The accessible recent comparison frame has 567 records, all included in the narrower physical-deliverable screen. The full multidimensional taxonomy is applied only to the 96-company pilot.
• Independent coding: a second reviewer should label a blinded sample and resolve disagreements. Automated or analyst confidence is not inter-rater reliability.
• Disclosure: undisclosed, stealth, removed and deceased companies may differ systematically from visible companies. The historical total is not complete by assumption.
• Time: the pilot compares current descriptions from nearby but not identical periods. It does not measure what founders proposed at selection.
• Technical depth: no code, model weights, training economics, patents, benchmarks or independent engineering review was used to rank novelty or defensibility.
• Commercial detail: billing evidence is incomplete. Private contracts, implementation obligations, human labour and revenue splits could alter operating-model labels.
• Sampling and category design: only 24 companies per cohort. Sector cells can be tiny. Alternative granularity and ambiguous-case choices can change rankings.
• Selection: applicant/rejection denominators are absent. Portfolio composition cannot identify whether programme preferences or different applicant pools caused the difference.
• Outcomes: no claim about survival, revenue quality, returns, founder quality or whether diverse portfolios perform better.

Next research steps

Classify the remaining 91 recent Speedrun and 380 selected YC companies across all dimensions. Add a Spring YC comparison. Ask an independent reviewer to label a blinded sample and audit the physical screen. Collect pricing, delivery responsibilities and original launch descriptions where companies have pivoted.