YC after ChatGPT: research appendix

What founders are building · What company values tell us · Research appendix

This appendix preserves the complete original analysis, including its methodology, detailed tables, charts, sensitivity tests, data audit, technical workflow guidance and proposed future research. It uses the 5 November 2026 snapshot and the separate September valuation collection. The two shorter articles draw from this material; no fresh valuation collection has been performed.

---

What changed in YC’s company mix, and what would make those products economically useful? The directory shows where founders are building. Assessing their value requires separate evidence about customer outcomes and operating costs.

The sector analysis covers 2,532 publicly listed YC companies from the 2023–2026 batches, captured on 5 November 2026. The 2026 figures are a snapshot through that date, not a completed-year total. All cross-cohort industry percentages use current directory classifications. They measure how YC’s directory labels past batches today, not necessarily the sectors those companies were originally pitched or funded under. This is a comparison of listed cohorts, not a verified history of YC’s investment decisions; establishing that history would require archived descriptions and classifications from each batch. The valuation research has different coverage and methods; its figures are presented separately.

How to read the categories: “Industrials” includes manufacturing and robotics, aviation and space, energy, defense, drones, climate, agriculture and automotive businesses. It is not a pure robotics or hardware category. An industrial company can also sell to businesses. Here, B2B and Industrials are directory labels: each company is counted once under its first listed primary category. That accounting choice prevents double-counting; it does not make their customers or business models mutually exclusive.

Three findings frame this analysis:

1. Industrial activity expanded in both count and share. Industrials rose from 12 of 494 listed companies in 2023 (2.4%) to 162 of 827 in 2026 (19.6%). Manufacturing and Robotics is the largest named industrial subcategory across the four years.
2. B2B infrastructure gained ground within a larger cohort. It rose from 39 of 621 companies in 2025 (6.3%) to 86 of 827 in 2026 (10.4%). Within Infrastructure, execution and reliability grew strongly in count, but their combined share changed more modestly.
3. Higher average valuations do not establish broad improvement. A separate September 2026 re-run reports means of $48.2m for pre-2023 batches and $85.8m for recent batches, but medians of $5.1m and $5.2m. Removing 10 companies from each group leaves both means at $22.4m, but removes unequal sample fractions. Our new equal-percentage trimming tests and cohort-level results below show why that distinction matters.

ChatGPT launched publicly on 30 November 2022. That date provides context, not evidence that every later YC company uses AI or that AI caused these changes.

Read the industry changes, infrastructure drilldown, or separate valuation evidence.

The portfolio broadened into industrial markets

Using one current directory classification across all cohorts, Industrials rose from 12 of 494 companies in 2023 (2.4%) to 162 of 827 in 2026 (19.6%). Its share grew roughly eightfold. The category is not synonymous with AI.

A falling share does not necessarily mean retreat. B2B’s share declined from 67.6% to 55.5%, but its count increased from 334 to 459. Fintech also recovered: its count doubled from 35 in 2025 to 70 in 2026, while its share rose from 5.6% to 8.5%. The data supports a broader company mix, rather than a simple abandonment of software or fintech.

Inside B2B: what businesses are actually being built?

Infrastructure increased from 39 companies in 2025 to 86 in the 2026 snapshot. Counts rose 120.5%, while the entire cohort grew 33.2%, from 621 to 827. Infrastructure’s share of the entire cohort rose from 6.3% to 10.4%, a 4.1 percentage-point increase. This is evidence of a composition change as well as a larger cohort.

Coverage gap: 152 of 459 B2B companies in 2026 (33.1%) have no listed subcategory, compared with 85 of 334 (25.4%) in 2023. The chart retains these companies. Named subcategories therefore describe only part of B2B, and changing classification coverage can affect the comparison.

B2B accounts for 1,569 companies across the four cohort years. The table uses YC’s listed subcategories. Every percentage uses the full B2B total, including the 480 companies with no listed subcategory. The named rows therefore provide an incomplete breakdown, not percentages recalculated over only classified companies. The descriptions explain their scope; they are not additional classifications assigned to individual companies.

Subcategory	Companies	Share of B2B	What it covers
Engineering, Product and Design	335	21.4%	Tools for building software and designing products
Infrastructure	182	11.6%	Systems and services supporting applications, data and AI workloads
Productivity	76	4.8%	Tools for everyday knowledge work and collaboration
Operations	74	4.7%	Tools for running business workflows and service processes
Finance and Accounting	63	4.0%	Accounting, tax and financial workflows
Security	60	3.8%	Protection of applications, systems and data
Sales	60	3.8%	Tools for sales teams and customer acquisition
Supply Chain and Logistics	57	3.6%	Shipping, warehousing and supply-chain coordination
Analytics	50	3.2%	Data analysis, shared metrics and business intelligence
Marketing	44	2.8%	Advertising, campaigns and audience development
Legal	42	2.7%	Legal research, document and case workflows
Retail	19	1.2%	Tools for retailers and commerce operators
Recruiting and Talent	14	0.9%	Hiring and talent sourcing
Human Resources	11	0.7%	Workforce administration and employee processes
Office Management	2	0.1%	Workplace and office administration
No subcategory listed	480	30.6%	The directory supplies only the broad category

Engineering, Product and Design is the largest named B2B subcategory, followed by Infrastructure. These are different from selling an application to an accounting, legal or sales team: they often provide the tools and systems used to build or operate other products.

For a concrete example of Analytics, Honeydew describes its product as a home for shared data logic. That is a specific data-management problem within the much broader B2B label. Named examples throughout this article illustrate product directions; they are not a representative sample of a batch and do not establish typical commercial performance.

Nearly a third of B2B companies, 480 of 1,569, have no listed subcategory. They remain in the total and are shown separately. Assigning them to sectors from names or a few tags would introduce a different, subjective classification method.

Inside Industrials: manufacturing, energy, defense and more

Industrials accounts for 279 companies. Every percentage below uses that full total, including the 43 companies with no listed subcategory. Manufacturing and Robotics makes up 41.2% of this group, but the category also includes aerospace, energy, defense, agriculture and other applications.

Subcategory	Companies	Share of Industrials	What it covers
Manufacturing and Robotics	115	41.2%	Production, quality inspection, robots and automation
Aviation and Space	32	11.5%	Aircraft, aerospace and space systems
Energy	30	10.8%	Power, batteries and energy-system tools
Defense	30	10.8%	Products serving defense applications
Drones	9	3.2%	Uncrewed aircraft and supporting systems
Climate	9	3.2%	Climate-related technologies and infrastructure
Agriculture	8	2.9%	Farming technologies and agricultural inputs
Automotive	3	1.1%	Vehicle and automotive-business technologies
No subcategory listed	43	15.4%	The directory supplies only the broad category

Bucket Robotics works on manufacturing defect detection using CAD and synthetic data. Ionworks provides simulation software used by battery and automotive companies. These examples show why “Industrials” should not be read as “companies that manufacture hardware.” Software can serve an industrial customer or process.

B2B and Industrials are directory categories, not mutually exclusive business models. An industrial company can sell to other businesses. For these calculations, each company is counted once under its first listed primary category.

Which subcategories changed most?

To compare portfolio composition, the table below shows both company counts and the percentage of the entire cohort. The denominator is 494 listed companies in 2023 and 827 in 2026. These selected categories illustrate the shift; the downloadable table includes every subcategory and all four years.

Subcategory	2023 count	Share of 2023 cohort	2026 count	Share of 2026 cohort
B2B: Infrastructure	26	5.3%	86	10.4%
B2B: Engineering, Product and Design	77	15.6%	66	8.0%
Industrials: Manufacturing and Robotics	0	0.0%	66	8.0%
Industrials: Energy	2	0.4%	21	2.5%
Industrials: Defense	0	0.0%	23	2.8%

Infrastructure expanded in both count and cohort share. Engineering, Product and Design declined on both measures between these endpoints. Within Industrials, Manufacturing and Robotics is the largest contributor to the increase in company count, with Energy and Defense also contributing.

A zero means no company in that cohort carries that subcategory in this snapshot. It does not establish that YC had no exposure to the underlying technology. Classifications can change, and older companies may be reclassified or absent from the directory. These are portfolio-composition comparisons, not evidence that one sector generates better returns.

Download the complete subcategory breakdown for 2023–2026 (CSV).

What Infrastructure companies actually build

Infrastructure is a broad directory label. To inspect it more closely, we reviewed the snapshot descriptions of all 182 companies in B2B → Infrastructure and assigned each to one primary function. This is an author-created taxonomy, not an additional YC classification. Products can span several functions; each company appears once so the counts add up. Descriptions reflect the 5 November snapshot, including for older cohorts.

The distinction matters: this category includes agent runtimes, databases, model-serving systems, payment tools and specialized applications. It cannot be treated as a count of AI-agent infrastructure alone.

The functional mix, year by year

Each cell below is a company count. The denominator is the Infrastructure group for that year, not all B2B companies or the entire cohort.

Primary function	2023	2024	2025	2026
Execution and deployment	3	7	6	15
Models and compute	2	3	2	8
Data and context	3	4	5	8
Connections and transactions	7	1	5	12
Reliability and controls	2	1	5	12
Training and evaluation	1	1	5	8
Developer and hardware tools	6	5	0	2
Specialized models and applications	2	2	3	9
Other or unclear	0	7	8	12
All Infrastructure companies	26	31	39	86

Download every company assignment and its source description (CSV). These counts describe what companies say they build. They do not measure customer spending or prove product performance.

What the functional breakdown actually tells us

Infrastructure grows from 39 companies in 2025 to 86 in 2026, a 120.5% increase. Within our author-created grouping, Execution and Deployment grows from six to 15, and Reliability and Controls from five to 12. Together they rise from 11 to 27 companies, but their share within Infrastructure moves only from 28.2% to 31.4%. The larger count is striking; the change in composition is more modest. Much of the increase accompanies growth of the whole category.

This distinction limits the technical narrative. The snapshot supports growing activity around running and controlling systems. It does not establish that these functions displaced the rest of Infrastructure. The classifications are also based on current short descriptions, not historical product specifications or independent product tests. Moving a multifunction product between groups can change these small counts.

Drill down another level: the components

Function	More specific components	Examples from the snapshot
Execution and deployment	Sandboxes and agent computers; Orchestration and agent cloud; Hosting and customer deployment	Freestyle, Rivet, Cakework
Models and compute	Inference serving and optimization; Context and prompt compression	Texel.ai, The Token Company
Data and context	Search and retrieval; Memory and durable state; Databases and data movement; Extraction and web data	Trieve, Mem0, Arroyo, Indexical
Connections and transactions	Tool and API gateways; Agent communications and interfaces; Payments wallets and metering	SuperAPI, Inkbox, ZeroDev
Reliability and controls	Observability and incident response; Permissions identity and action checks	Odigos Technologies Inc., Decawork
Training and evaluation	Evaluations and test environments; Training data; Training optimization	Magicflow 𐂂, Markov, RunRL
Developer and hardware tools	Software build and development; Hardware control and configuration	Portalform, Synnax
Specialized models and applications	Voice models and routing; Specialized or frontier models	Speko, Guide Labs

Who needs these components, and what would improve?

Consider a supplier-invoice workflow: retrieve the purchase order, compare it with the invoice, resolve exceptions and update an authorized record. Each infrastructure function addresses a different constraint. These are engineering interpretations and proposed measurements, not demonstrated benefits from the companies counted above.

Function	Buyer and problem	Where efficiency could improve	What to measure against the existing process
Execution and deployment	Application teams running multi-step work that can stop midway	Resume completed work instead of restarting the entire task	Accepted completion rate, recovery time and cost per completed workflow
Models and compute	Platform teams facing model latency, hardware limits and inference cost	Serve equivalent-quality outputs with less compute or delay	Cost per accepted outcome and 95th-percentile latency at matched quality and workload
Data and context	Data and application teams working with scattered or changing records	Retrieve relevant, current evidence with less manual searching	Correct-source retrieval, stale-record errors and human correction time
Connections and transactions	Integration teams maintaining APIs, communications and payments	Reduce integration upkeep and repeated manual transfers	Integration effort, successful authorized operations and duplicate transaction rate
Reliability and controls	Operations and security teams investigating incidents or restricting actions	Detect problems sooner and prevent invalid business changes	Detection and recovery time, unauthorized changes and false blocks
Training and evaluation	Model and application teams deciding whether a change is safe to release	Find task failures before deployment and reduce regressions	Held-out task performance, regressions detected and failures missed in production
Developer and hardware tools	Developers and hardware teams configuring build or device systems	Reduce setup, build and configuration rework	Time to a working configuration, build failures and operator intervention
Specialized models and applications	Product teams with domain or voice requirements	Improve a specific task relative to a general-purpose alternative	Domain acceptance rate, latency, escalation rate and total operating cost

For concrete product descriptions, Superserve describes virtual machines for long-running agents; RunAnywhere describes inference across hosted, on-premises and device environments; Salus describes validation before execution. These three examples explain proposed product roles. They do not represent the whole batch, demonstrate strong commercial traction or establish sector-wide performance.

What would establish that the infrastructure works?

The directory descriptions support the existence of these product offerings. They do not establish adoption, efficiency or commercial success. A stronger assessment would pair a named deployment with comparable before-and-after measurements: cost per successfully completed task, time to completion, error rate, recovery after interruption and unauthorized actions prevented. Those measures need a stated workload and baseline.

The practical finding is that Infrastructure contains multiple layers needed to operate software and agents, alongside specialized products that do not fit a simple agent stack. Its growing company count warrants investigation of these layers individually. It does not establish that every layer has equal demand or that the newest companies have solved the underlying problems.

CTO deployment checklist: workflow state, true operating cost and failure tests

The architecture question: who owns the workflow state?

The crucial boundary in an agent system is between a proposed action and a committed business change. A model can propose updating an invoice. The surrounding application must determine whether the user is authorized, whether the invoice version is still current, and whether an earlier attempt already performed the update.

Boundary	Failure to examine	Engineering requirement
Retrieval → model	A stale record is treated as current, or a restricted record enters context	Carry source version and access context; enforce retrieval permissions
Model → tool	A valid-looking request targets the wrong record or exceeds the user's authority	Validate arguments and authorize the concrete resource and operation outside the model
Tool → retry	A timeout hides a successful write; retrying creates a duplicate	Persist an operation identifier and reconcile the result before repeating the side effect
Approval → execution	The record or proposed action changes after approval	Bind approval to the reviewed action and record version; recheck before commit
Execution → recovery	The process restarts without knowing which steps completed	Persist workflow state and distinguish proposed, attempted, confirmed and failed actions
Execution → investigation	Logs show a model answer but cannot explain the business change	Correlate the request, source versions, tool calls, approvals and committed result

These are engineering requirements inferred from the workflow, not findings that any named YC company satisfies them. AWS’s discussion of safe retries and idempotent APIs explains why retrying a request needs a contract that prevents duplicate effects. OpenTelemetry’s signal model provides context for tracing activity across components. A trace supports investigation; it does not enforce authorization or guarantee a correct transaction.

This is also where product overlap becomes important. A sandbox provides an execution environment, but does not automatically supply business permissions. Memory retains information, but does not establish that it is current. An evaluation measures selected behavior, but does not replace runtime checks. Buyers should identify the component that owns each requirement and test the boundaries between products.

Measure the cost of an accepted outcome

A lower token price or faster model response can coexist with a more expensive workflow. A useful operational measure is:

Cost per accepted outcome = total workflow operating cost ÷ outcomes meeting the defined acceptance criteria.

Include inference, tool calls, compute, storage, retries and human review in the numerator. Report setup and integration costs separately over a stated period. Define acceptance before testing: the correct record changed, with authorization, required evidence and no duplicate side effect. Report completion rate alongside the unit cost so the system cannot appear efficient simply by declining difficult cases.

For illustration only, a workflow costing $1,000 and producing 900 accepted outcomes costs $1.11 per outcome. A replacement costing $700 but producing only 500 accepted outcomes costs $1.40. Total spend falls 30%, yet cost per accepted outcome rises 26%. These are hypothetical figures, not results from the companies in this dataset.

Multi-step reliability needs similar care. Ten independent steps that each succeed 98% of the time yield roughly 81.7% probability that every step succeeds. Real failures may be correlated, so this is an illustration rather than a prediction. Measure complete tasks, including recovery, instead of multiplying vendor benchmark claims.

A deployment test that can challenge the claims

Test the same representative workload against the current process, a simpler automated baseline and the proposed agent system. Keep workload difficulty and acceptance criteria consistent. Include missing information, conflicting records, permission changes, a timeout after a successful write, a restart mid-task and a repeated request. Report performance on routine and difficult cases separately.

Measure accepted completion rate, cost per accepted outcome, human review time, duplicate or unauthorized changes, recovery success and the 95th-percentile completion time. A system that pauses correctly for missing authority should be distinguished from one that silently fails or proceeds without permission. Set thresholds according to the workflow's consequences, rather than importing a universal success percentage.

Anthropic’s agent-evaluation guidance discusses evaluating tasks, reading execution transcripts and combining assessment methods. This supports a practical buying principle: require evidence at the level of the workflow being purchased. A component benchmark and a successful demonstration are useful inputs, but neither establishes reliable operation across the customer's workload.

Higher valuations at the top, with a similar median

Ben West’s original study, published on 29 August 2025, used publicly collected valuations and pre/post-ChatGPT batch membership as a proxy for AI use. It found little evidence of faster valuation growth at that point, while acknowledging missing data, extraction errors, economic conditions and the possibility that effects needed more time to appear. Some figures in the original article differ from the repository’s reproduced baseline, so the two versions should not be treated as numerically interchangeable.

Portfolio composition and financial outcomes are different questions. The valuation comparison below shows a higher mean for recent batches, but a similar median. It uses a separate sample of companies with available valuations, rather than the 2,532-company sector snapshot above. The underlying valuations have not been independently verified.

Reported two-year valuation measure	Pre-2023 batches	2023 onward batches
Mean valuation	$48.2m	$85.8m
Median valuation	$5.1m	$5.2m
Mean after removing the top 10 companies from each group	$22.4m	$22.4m

The reported mean is 1.78 times higher for recent batches, but the median differs by only $0.1 million. Removing the top 10 companies from each group leaves both means at $22.4 million. That sensitivity matters, but the groups differ in size: removing 10 companies excludes 0.9% of the older group and 3.9% of the newer group. Our equal-percentage tests below provide a different comparison.

The re-run also reports that 10.81% of recent companies with a usable two-year valuation reached at least $100 million, compared with 4.74% of older companies in that sample. This threshold measure adds information beyond the mean: the reported difference extends beyond a single exceptional company. It remains a result for companies with available valuations, rather than a population-wide probability of success. Reported robustness checks.

The mean difference is highly sensitive to the highest-valued companies, while the median changes little. The higher reported share reaching $100 million also matters: the evidence does not justify saying that every difference is confined to the top 10. Neither result establishes a broad improvement for the typical YC startup. These are reported valuations, expressed in June 2026 dollars, rather than revenue, profit or realized investor returns. The comparison alone cannot identify an effect caused by generative AI.

What our independent reanalysis actually found

This is a computational reproduction of the September collection, not a new collection of November valuations or independent verification of every financing round. All dollar figures below use June 2026 dollars.

The fixed top-10 comparison removes unequal fractions

The reproduced sample has 1,076 pre-2023 companies and 259 companies from 2023–2024. Removing 10 from each excludes 0.9% of the older group but 3.9% of the newer group. The resulting $22.4m means are therefore a useful sensitivity result, not evidence that equal portions of the two distributions are identical.

Treatment of the upper tail	Pre-2023 mean	2023–2024 mean	Recent / older
Keep all selected marks	$48.18m	$85.77m	1.78×
Remove highest 1% within each group	$21.64m	$40.30m	1.86×
Remove highest 5% within each group	$10.67m	$18.15m	1.70×
Remove highest 10% within each group	$7.89m	$9.76m	1.24×

Removal counts are rounded upward: 11 versus 3, 54 versus 13, and 108 versus 26. These are one-sided upper-tail trims, not symmetric trimmed means. The recent-group mean remains higher under these specifications, while the size of the difference changes. That does not resolve selection, measurement or macroeconomic confounding.

The positive-value geometric means are $4.95m and $7.27m, using 1,045 and 259 positive records respectively. The older sample contains 31 parsed zero marks, excluded only from the geometric mean. Those zero strings have not all been independently validated as economic write-offs. Threshold counts are 51/1,076 versus 28/259 at $100m, and 8/1,076 versus 4/259 at $1bn. These denominators are selected valuation records, not all eligible YC companies.

Combining 2023 and 2024 conceals a large difference

YC cohort year	Selected marks	Mean	Median	At least $100m	At least $1bn
2023	128	$33.38m	$5.19m	10/128	0/128
2024	131	$136.97m	$4.70m	18/131	4/131

The higher mean is much more pronounced in the 2024 records; its median is lower than the 2023 median. Both remain calendar-year-offset comparisons. Some 2024 entrants have not completed an exact 24 months at the September collection date. A two-year outcome for 2025 entrants cannot yet be observed under this design. These figures do not rank cohort quality.

A cleaned numeric field can still contain the wrong economic measure

After upstream cleaning, 100 of the 1,335 selected strings contain “raised,” “funding” or “investment” without an explicit valuation label; 66 contain multiple dollar amounts and 77 contain ranges. These flags overlap. The upstream parser takes the first dollar amount, which can represent capital raised or one endpoint rather than a comparable valuation. Flags identify records requiring review, not a verified error count.

Excluding the union of these flags leaves 946 older and 211 recent marks. Means become $50.69m and $100.85m; medians become $4.57m and $4.66m. The mean gap increases rather than disappearing. This is a diagnostic exclusion test, not a newly verified dataset: unflagged rows can still be wrong, and removing ambiguous strings can introduce further selection.

Coverage and collection vintage prevent a clean comparison

The raw database contains 6,438 company-name rows and 21,846 annual valuation rows. For 2023 it has 523 company-name rows, 360 marked collection-completed and 128 selected deduplicated two-year marks. For 2024 the corresponding counts are 663, 467 and 131. Name rows can include aliases, so these are processing counts, not identity-reconciled coverage rates. “Completed” describes collection status, not business survival.

Restricting to records collected on or after 6 September 2026 leaves nine older two-year marks versus all 259 recent marks. A common collection-vintage rule therefore leaves too little and too selected an older comparison group to rescue the headline with a consistent-vintage estimate.

The population problem is also visible outside the valuation database. YC's original Winter 2023 announcement reports 282 funded companies and its Summer 2023 announcement reports 229: 511 combined. Our current directory snapshot has 274 and 220: 494 combined. The net gap of 17 is not a verified delisting count. It requires identity-level reconciliation of aliases, additions, omissions and batch changes.

Download the reanalysis package: code, input fingerprint, selected marks, coverage and sensitivity results. The package documents execution and limitations; reproduction of a calculation does not validate its underlying records.

Collection errors limit the valuation conclusions

The re-run documents substantial contamination in automated collection: invented funding rounds, confusion between companies with similar names, duplicate company entries and algorithmic valuation estimates treated as reported valuations. It also identifies an asymmetric year-selection rule that allowed a later valuation into a purported one-year result. The revised analysis filters identified problems, applies manual corrections and uses stricter target-year selection.

That is evidence of corrective work, not proof that every remaining record is accurate. The report describes manual verification of post-2023 companies appearing in the top-20 lists, rather than verification of the entire database. Most pre-2023 figures retain the older collection, while recent cohorts were re-collected with a different model. The authors’ check finds substantial disagreement between collections even for historical years. In addition, the calculation includes companies with usable numeric valuations, so its sample is not the entire YC population. Report and limitations · Cleaning and selection code.

The study’s one-year and two-year marks use calendar-year offsets from the YC batch year. This improves comparability over using today’s valuation for every company, but it is not an exact 12-month or 24-month follow-up from each company’s entry date. Companies enter batches and raise rounds at different points within a year.

The qualified finding is that the cleaned re-run reports stronger valuation results among the highest-valued recent companies, with similar medians and material data limitations. It neither establishes better survival nor resolves the separate problem of comparing closure rates across companies of different ages.

Valuation, commercial performance and survival need different evidence

A financing valuation records the price attached to a transaction. Revenue shows customer spending; profit depends on the costs of delivering that revenue. An acquisition can produce very different outcomes for founders and investors depending on its terms. None of these measures can stand in for all the others.

The same care applies to company status. Our directory snapshot contains 2,318 active, 93 acquired and 121 inactive companies. Older cohorts have had more time to develop, close or be acquired. Comparing their inactive shares directly with 2026 entrants would give newer companies an artificial advantage.

A fair closure comparison needs consistent definitions, verified event dates and equal follow-up periods, or a survival model that accounts for incomplete observation. The directory snapshot cannot provide that analysis. The valuation study’s calendar-year offsets improve age comparability for its own question, but do not supply the missing closure histories.

A November update changes one company, not the whole study

Legora announced a $550 million Series E at an $8.8 billion post-money valuation on 21 October 2026, above the $5.6 billion mark discussed in the September report. This is a concrete example of how quickly individual financing marks can change.

Our 5 November check reviewed 10 selected companies discussed in the re-run. It did not recollect the full valuation dataset or recalculate cohort statistics. The September results above retain their original collection date and inflation adjustment. Legora’s October announcement is shown as a separate nominal-dollar event.

The research question and the tests we could not run

The primary follow-up question is now explicit: within the Winter and Summer 2023 YC batches, did companies whose core product depended on AI at entry have a different probability of obtaining non-YC follow-on equity financing within 24 months than comparable non-AI-core companies? This is an association question. Both batches follow ChatGPT's launch, so this comparison alone cannot identify the launch's causal effect.

The intended outcome is a verified qualifying financing event, with amount and valuation recorded separately. A final protocol must define treatment of SAFEs and insider rounds before outcome estimation; debt and secondary sales should be recorded separately. AI status must come from dated product descriptions, with two independent coders, an explicit core-product rule and an unknown category. Current tags, current headcount and batch year cannot substitute for historical exposure, baseline size and incorporation date.

Requirement	Actual audit result	Consequence
Reconciled entry population	Announcement, directory and valuation-database counts differ	No common population denominator yet
Historical AI/non-AI classification	Dated founder descriptions are recoverable for some companies; full coverage and independent double coding are not complete	No validated treatment/control assignment
Exact baseline and financing dates	Database stores one valuation string per company-year; collection timestamps are not financing dates	No exact 12- or 24-month outcome panel
Baseline covariates and comparison groups	No consistent historical founder, size, initial financing or non-YC control panel assembled	Matching and adjusted causal models not run
Operating performance and investor returns	No verified cohort panel of revenue, retention, margins, dilution or cash distributions	No operating-performance or investment-return estimate

This audit found useful primary records: Twenty's July 2023 founder launch describes an open-source CRM, while Sweep's August 2023 launch describes an LLM coding bot. Runa Capital records Twenty's November 2024 financing. These establish that some historical evidence is recoverable; they are not a matched pair, a representative sample or completed independent coding.

The date problem is substantive: Foundation Capital's February 2024 Tennr post describes an investment made in May 2023. Treating publication date as the transaction date would shift the observation by nine months. A confirmed positive financing event also cannot turn another company's missing record into a confirmed negative.

No propensity-score matching, difference-in-differences effect, failure-rate comparison or risk-adjusted return is reported. The necessary inputs have not passed the checks above. A future pre/post study would additionally need a credible counterfactual trend, checks for concurrent sector shocks and an account of AI spillovers into controls. A non-significant pre-trend test would not prove identification. World Bank methodological discussion.

The completed work here is narrower: directory recounting, source/schema inspection, computational reproduction, percentile-tail sensitivity, cohort separation, ambiguous-string exclusions and collection-vintage restriction. These improve transparency and reveal fragility; they do not convert the study into causal investment evidence.

An investment decision needs several kinds of evidence

Evidence is better assessed by the question it answers than by a single score. A verified financing announcement can establish a transaction price while providing little evidence about paid demand or profitability. An audited financial statement improves verification; it does not alone establish causal advantage or future returns.

Decision	Evidence to request
Is demand durable?	Paid invoices reconciled to revenue, customer cohorts, retention, concentration and sales-cycle history
Does the product improve economics?	Comparable workflow results, gross margin after compute and support, customer acquisition cost and payback, burn and runway
Can the advantage persist?	Evidence of switching costs, distribution or data rights; tests under falling model prices and loss of a key model supplier
Is the investment attractive at this price?	Entry valuation, ownership, dilution, liquidation preferences, downside scenarios and exit assumptions
Have investors actually earned returns?	Dated contributions and distributions, realized proceeds separated from remaining marks, fees and a stated benchmark

Venture-return research explicitly addresses the selection problem created by which investment outcomes become observable. Cochrane's research on venture capital risk and return is one example. A financing valuation, an acquisition label and a profitable exit are different observations.

The current article supplies directory evidence and attributed financing research. It does not supply verified operating cohorts or investor cash flows. Improving that evidence requires collecting and validating those records, not assigning the article a higher score for a more elaborate method section.

What founders and business buyers can take from this

The B2B label covers several distinct markets. Infrastructure companies sell systems that other products depend on. Accounting, security, legal and operations companies address particular business functions. Industrial software may serve manufacturers, energy operators or defense customers without producing hardware itself.

For founders, the sector breakdown helps identify the actual market being entered: the customer, workflow and buying decision matter more than the broad AI label. A larger cohort share indicates more activity in that category; it does not establish customer demand, lower competition or better returns.

For business buyers, rising valuations are evidence of financing interest. Product decisions still require evidence of reliability, integration, operating cost and measurable results in the buyer’s own workflow.

In the 5 November 2026 directory snapshot, the 2026 cohort has a larger Industrials share and more B2B companies than the 2023 cohort. Infrastructure also increased in both count and cohort share. These comparisons describe current classifications of publicly listed companies. They do not establish historical selection intent, AI causation or superior sector performance.

Our reproduction confirms a higher mean in the selected recent-company records and a similar median. Equal-percentage tail trimming preserves a smaller or larger mean gap depending on the cutoff, while separating 2023 from 2024 exposes substantial cohort differences. These remain preliminary descriptive signals because ambiguous amounts, incomplete follow-up, missing records and mixed collection vintages prevent a clean outcome comparison. For a business buyer, the practical question remains whether a product delivers reliable, accepted outcomes at a lower total operating cost. Company counts and financing valuations cannot answer that question.

Historical announcements suggest an enterprise emphasis

YC’s historical batch announcements use different category definitions. They provide context for an enterprise emphasis, but they do not form a standardized time series.

Batch announcement	Reported B2B/Enterprise share	Reported Fintech share
Winter 2022	34%	24%
Summer 2022	39%	21%
Summer 2023	70%	10%

Summer 2022 separately classified 13% as DevTools. That prevents a straightforward interpretation of the enterprise figures as like-for-like growth from 34% to 70%. Simply adding DevTools to one year would not demonstrate that all the remaining definitions align. These announcements suggest a stronger enterprise emphasis; they do not measure its precise size on a consistent basis or establish that ChatGPT caused it. Sources: Winter 2022, Summer 2022, Summer 2023.

The directory analysis above instead applies one snapshot’s classifications across the 2023–2026 cohorts. That makes category definitions more consistent within this analysis, while retaining the limitations of current classifications and directory coverage.

Sources and methodology

The conclusions are supported by the cited evidence within its coverage and verification limits. Directory counts describe listed companies and their current labels; company descriptions are not independent product tests. The valuation collection has not been independently verified record by record for this article.

The source for the sector analysis is YC’s public company directory, retrieved through its public directory search service on 5 November 2026. The dataset contains 2,532 unique company IDs in Winter 2023 through Fall 2026. It is a dated snapshot, not a forecast or a complete record of every company originally admitted.

Each company is counted under its first listed primary industry and, where available, its next listed subcategory. Missing subcategories remain explicitly unclassified. The B2B and Industrials breakdowns total 1,569 and 279 companies respectively. Percentages in those two tables use their respective industry totals; the cohort comparison uses all companies listed in each cohort year. Rounding can prevent percentages from adding to exactly 100%.

Current classifications are applied to every cohort. They are not archived classifications at admission. Directory additions, removals, reclassification and reporting delays can affect comparisons. YC’s historical batch announcements use their original category definitions and are presented separately from this snapshot analysis.

The original valuation results come from the September 2026 yc-growth re-run. The input fingerprint, selected records, calculation code and outputs are downloadable above. Collection dates end on 6 September 2026; this is not a November recollection. The source’s calendar-year windows, current directory snapshot and unresolved record validity remain separate limitations.

Selected company descriptions illustrate product functions, while financing announcements provide dated evidence about individual transactions. Neither establishes typical commercial outcomes, causes of success or investor returns. This article is research and general information, not investment advice or an endorsement.

Download the 2,532-company research dataset (CSV) · Download the detailed sector counts (CSV).