Skip to content
Slow Variables
Ask

The value chain

Every part of building and using AI, from power and chips to the apps people use, earns money differently. This page is the site's judgement of where each part's economics are heading. A part is commoditising when it is becoming cheap and interchangeable, so the gain passes to buyers; holding when its profit is steady for now; and tightening when it is getting scarcer, so whoever owns it can charge more. Each judgement names the view that argues against it, where one was found, and what would prove it wrong. “Binds on” names the inputs each part depends on, with how tight the site's bottleneck scorecard reads each tonight. The readings behind every judgement are on its layer's page.

What the powers mean

Hamilton Helmer, a strategy consultant and investor who studied why some firms keep earning more than their investors could get elsewhere, found that an advantage lasts only when a benefit sits behind a barrier that stops rivals copying it, and named the kinds that recur:

a cornered resource
:
preferential access to something others need: a patent, a licence, a scarce input, a rare team
branding
:
buyers pay more for the same thing because of what they believe about who made it
counter-positioning
:
a newcomer's model the incumbent could copy only by damaging its own business
network economies
:
each user makes the product more valuable to the others, so the leader's lead feeds itself
process power
:
a way of working that improves cost or quality and takes a rival years to learn
scale economies
:
unit costs fall as volume grows, and a rival cannot match them without matching the volume
switching costs
:
a customer would lose money, time or work by moving to a rival

Where the map shows a shortage, some businesses could be built to meet it: businesses that could be built, with the profit the site's rent rule (its test of who keeps a profit that lasts) gives each and what would prove it wrong.

The chain in one picture: where the companies are, and which way this site judges each part's profit is headingChart for the counts, this site's judgement for the directions
  1. Compute & physical133
  2. Model121
    Holding, judgedFrontier labs
  3. Training input81
  4. Serving & orchestration132
    No judgement yetAgent security & identity
  5. Deployment & application379
    Commoditising, judgedCoding agents & dev tools
    No judgement yetTalent production / AI education & FDE trainingAutonomous services / AI-native service firms
  6. Adopters0
    Commoditising, judgedthe whole layer
  7. Labour & consumers0
    Holding, judgedthe whole layer
companies the tracker follows in the layer (a count)tightening, a judgement: getting scarcer, so whoever owns it can charge moreholding, a judgement: profit steady for nowcommoditising, a judgement: becoming cheap and interchangeable, so the gain passes to buyersa part with companies in it and no judgement on this page yet

The layers run in the order work flows, from power and chips to the people who use the products. The bars count each company on the market map below once, by the part of the chain the tracker files it under. The map sorts the same companies into its own finer layers and categories and lets a company sit in more than one, so its counts are larger and do not line up with these bars. The folded table lists the parts behind each bar. A bar is the tracker's reach, not the size of any market: a tall bar means many companies followed. The last layers hold none because the site reads them through surveys and statistics, not through companies.

Each direction is a judgement, not a measurement, drafted by Claude, a model made by Anthropic, whose own part of the chain (frontier labs) is among those judged: follow it to read the argument, the view against it and what would prove it wrong. 169 of the 849 companies sit in parts with no judgement yet, and 3 are on the map without a part of the chain. The parts with no judgement are ones the site has not yet written up, not ones it judged unimportant.

Short names are the tracker's own; each links to a plain statement of the judgement. RL is reinforcement learning, training by practice and grading; HBM is the fast memory stacked beside an AI chip; FDE is an engineer the seller places inside a customer; evals are tests of a model.

The numbers: companies followed and the judged direction, part by part
LayerPart of the chainCompanies followedDirection, judged
Compute & physicalSemiconductors & AI chip startups28Holding
Compute & physicalHyperscalers25Holding
Compute & physicalData-centre developers & power55Tightening
Compute & physicalMemory/HBM25Holding
Compute & physicalThe whole layer1331 tightening, 3 holding
ModelFrontier labs5Holding
ModelNeolabs (founded 2023+)108Commoditising
ModelOpen-weight & Chinese labs8Commoditising
ModelThe whole layer1211 holding, 2 commoditising
Training inputHuman/expert data & labelling22Holding
Training inputRL environments12Holding
Training inputRLaaS / post-training / continual-learning infra7Commoditising
Training inputSynthetic data & pipelines9Commoditising
Training inputPermissioned data markets & licensing12Holding
Training inputEvals19Commoditising
Training inputThe whole layer813 holding, 3 commoditising
Serving & orchestrationInference serving39Commoditising
Serving & orchestrationRouting12Commoditising
Serving & orchestrationAgent infrastructure42Commoditising
Serving & orchestrationAgent security & identity39no judgement yet
Serving & orchestrationThe whole layer1323 commoditising
Deployment & applicationDeployment / FDE / implementation services34Holding
Deployment & applicationTalent production / AI education & FDE training40no judgement yet
Deployment & applicationCoding agents & dev tools14Commoditising
Deployment & applicationVertical AI by domain40Holding
Deployment & applicationAutonomous services / AI-native service firms90no judgement yet
Deployment & applicationConsumer AI6Holding
Deployment & applicationAI-native rollups / buyout vehicles155Holding
Deployment & applicationThe whole layer3794 holding, 1 commoditising
AdoptersThe whole layer01 commoditising
Labour & consumersThe whole layer01 holding
On the map without a part of the chainApplied Materials, KLA, Lam Research3not filed under a part

Compute & physical

Model

Training input

Serving & orchestration

Deployment & application

Adopters

Labour & consumers

The judgements and profiles were drafted by a model from the companies' own filings and the site's readings, and each says whether a person has reviewed it. They are conditions, not forecasts: no company is ranked.

What makes a profit last

A shortage earns a profit for a while; it lasts only where something holds rivals off. The first figure shows where the judgements above, and the company profiles, say such a hold exists today, and of which kind. The second shows the test the site applies to each company it profiles: whether what it sells can be copied, and who owns what customers need in order to use it.

Where this site judges a lasting advantage exists today, and of which kindA chart of this site's judgements, not of measurements
The kinds of power, in plain words
scale economies: unit costs fall as volume grows, and a rival cannot match them without matching the volume
network economies: each user makes the product more valuable to the others, so the leader's lead feeds itself
counter-positioning: a newcomer's model the incumbent could copy only by damaging its own business
switching costs: a customer would lose money, time or work by moving to a rival
branding: buyers pay more for the same thing because of what they believe about who made it
a cornered resource: preferential access to something others need: a patent, a licence, a scarce input, a rare team
process power: a way of working that improves cost or quality and takes a rival years to learn
the part's own judgement names this power as present today (a judgement)na number: profiled companies in this part judged, each on its own, to hold the power (a judgement about companies, not about the part)neitherbeside a part: judged tighteningbeside a part: judged holdingbeside a part: judged commoditising

A power is a reason a firm keeps earning more than its rivals: a benefit that sits behind a barrier rivals cannot cross. The kinds are Hamilton Helmer's, each defined beside the grid. Every mark is this site's judgement, drafted by Claude, a model made by Anthropic, which is one of the frontier labs judged here: 5 of the 23 judged parts name any power as present today.

The two marks are separate judgements and they do not always agree. The tint judges the part as a whole; the number judges single companies, profiled one at a time. Most parts name no power while some companies inside them are credited with one: a company can hold an advantage its part of the chain does not. Samsung Electronics, Alphabet, Meta, Apple, Databricks are profiled in more than one part, so their powers are listed in the table and left out of the numbers. A blank row means no power was named, not that the part earns nothing.

Short names are the tracker's own; each links to a plain statement of the judgement. RL is reinforcement learning, training by practice and grading; HBM is the fast memory stacked beside an AI chip; FDE is an engineer the seller places inside a customer; evals are tests of a model.

The judgements: the powers each part names, and the profiled companies judged to hold each
Part of the chainDirection, judgedPowers the part's judgement namesCompanies profiled in this part alone, judged to hold a power
Semiconductors & AI chip startupsHoldingscale economies, switching costs, a cornered resource, process powerscale economies: ASML, NVIDIA, Broadcom, TSMC; switching costs: ASML, NVIDIA, TSMC; a cornered resource: ASML; process power: TSMC
Memory/HBMHoldingscale economies, a cornered resource, process powerscale economies: Micron, SK Hynix; process power: SK Hynix
Data-centre developers & powerTighteningscale economies, a cornered resourceno company profiled in this part alone
HyperscalersHoldingscale economies, switching costsscale economies: Microsoft, Amazon; switching costs: Microsoft, Amazon
Frontier labsHoldingscale economiesscale economies: OpenAI, Anthropic, xAI; switching costs: OpenAI, Anthropic; a cornered resource: xAI
Neolabs (founded 2023+)Commoditisingnone namedswitching costs: ElevenLabs; a cornered resource: Periodic Labs
Open-weight & Chinese labsCommoditisingnone namedno company profiled in this part alone
Human/expert data & labellingHoldingnone namednetwork economies: Mercor; switching costs: Scale AI; a cornered resource: Handshake AI; process power: Surge AI
RL environmentsHoldingnone namedno company profiled in this part alone
RLaaS / post-training / continual-learning infraCommoditisingnone namedno company profiled in this part alone
Synthetic data & pipelinesCommoditisingnone namedno company profiled in this part alone
Permissioned data markets & licensingHoldingnone namedno company profiled in this part alone
EvalsCommoditisingnone namedno company profiled in this part alone
Inference servingCommoditisingnone namedswitching costs: Baseten, Fireworks
RoutingCommoditisingnone namedno company profiled in this part alone
Agent infrastructureCommoditisingnone namedno company profiled in this part alone
Deployment / FDE / implementation servicesHoldingnone namedswitching costs: Palantir
Coding agents & dev toolsCommoditisingnone namedswitching costs: Cognition, Anysphere, Lovable
Vertical AI by domainHoldingnone namedswitching costs: Harvey; a cornered resource: OpenEvidence
Consumer AIHoldingnone namedno company profiled in this part alone
AI-native rollups / buyout vehiclesHoldingnone namedno company profiled in this part alone
AdoptersCommoditisingnone namedno company profiled in this part alone
Labour & consumersHoldingnone namedno company profiled in this part alone
Company profiled in more than one partThe partsPowers the company is judged to hold, not tied to either part
Samsung ElectronicsMemory/HBM; Semiconductors & AI chip startupsscale economies
AlphabetHyperscalers; Frontier labsscale economies, process power
MetaConsumer AI; Frontier labsnetwork economies, scale economies, counter-positioning
AppleInference serving; Consumer AInetwork economies, switching costs
DatabricksInference serving; Agent infrastructureswitching costs
PowerWhat it means
scale economiesunit costs fall as volume grows, and a rival cannot match them without matching the volume
network economieseach user makes the product more valuable to the others, so the leader's lead feeds itself
counter-positioninga newcomer's model the incumbent could copy only by damaging its own business
switching costsa customer would lose money, time or work by moving to a rival
brandingbuyers pay more for the same thing because of what they believe about who made it
a cornered resourcepreferential access to something others need: a patent, a licence, a scarce input, a rare team
process powera way of working that improves cost or quality and takes a rival years to learn
Who keeps the profit: the questions this site asks of each company, and where its answers send each oneA model of the rule; where each company lands is this site's judgement
  1. Can the company stop others copying what it sells?A patent, a secret or know-how that takes years to learn holds imitators off.
    yes →
    The company itself keeps the profit
    ASMLNVIDIABroadcomAMDTSMCMicrosoftAmazonAlphabet
  2. If it can be copied: does the company itself own something scarce that customers need in order to use it?Factories, a sales network, a licence or a base of users that a copier would also have to build.
    yes →
    The company itself keeps the profit
    xAIMetaPalantirAnysphereAppleDatabricks
  3. If it does not: does somebody else own that scarce thing?When copying is easy, whoever owns what customers need to use the product collects instead of its maker.
    yes →
    Whoever owns the scarce thing keeps the profit, which can be the company itself among others
    Judged to go to established firms that already own it; the company named can be one of them:MicronSK HynixSamsung ElectronicsPeriodic LabsEtchedHarveyOpenEvidenceChai Discovery
    Judged to go to the platforms that own what its customers need:OpenAIAnthropicElevenLabsCognitionLovable
  4. Otherwise: it is easy to copy and needs nothing scarce to reach customers.Rivals enter and cut prices until nothing above the cost of staying in business is left.
    so →
    Competition passes the gain to buyers
    OracleCoreWeaveNebiusFluidstackScale AIMercorHandshake AISurge AIBasetenFireworks
a question the rule asks, in order; the first yes settles ita profiled company, placed by this site's answers about it (a judgement)

A rent is profit beyond what it takes to keep a business going, earned because something holds rivals off. The steps follow David Teece, an economist who studied why the firm that invents something often fails to keep the returns. The drawing shows the rule, which has no numbers of its own. How large a lasting profit the rule gives each business follows the kind of rent and how long it is judged to last, and is in the folded table: it is a lasting profit beyond a normal return, not what the company earns today.

Where a company lands is this site's judgement. The answers about each were drafted by Claude, a model made by Anthropic, from the company's filings; Anthropic is itself placed here and competes with, sells to or buys from many of the others. Separate reviewing runs of the same model checked each profile against its sources; no person reviewed each one. The rule then decides. A company placed where whoever owns the scarce thing keeps the profit may itself be one of those owners; the rule files it there because others like it hold the same thing. Only companies the site has profiled appear, on each layer's page; a company on the map with no profile is absent, which is no judgement of it. Each placing judges the company's business inside the AI chain, not the whole company. None of it is advice to buy or sell, or a forecast of a share price.

Each profiled company and how the rule reads its business in the AI chain
CompanyThe step that settles itHow the rule reads its business in the AI chain
ASMLCan the company stop others copying what it sells?a monopoly-like profit, kept by the company itself
NVIDIACan the company stop others copying what it sells?a moderate profit, kept by the company itself
BroadcomCan the company stop others copying what it sells?a fat profit, kept by the company itself
AMDCan the company stop others copying what it sells?a thin profit, kept by the company itself
TSMCCan the company stop others copying what it sells?a monopoly-like profit, kept by the company itself
MicrosoftCan the company stop others copying what it sells?a fat profit, kept by the company itself
AmazonCan the company stop others copying what it sells?a fat profit, kept by the company itself
AlphabetCan the company stop others copying what it sells?a monopoly-like profit, kept by the company itself
xAIIf it can be copied: does the company itself own something scarce that customers need in order to use it?a thin profit, kept by the company itself
MetaIf it can be copied: does the company itself own something scarce that customers need in order to use it?a fat profit, kept by the company itself
PalantirIf it can be copied: does the company itself own something scarce that customers need in order to use it?a thin profit, kept by the company itself
AnysphereIf it can be copied: does the company itself own something scarce that customers need in order to use it?a thin profit, kept by the company itself
AppleIf it can be copied: does the company itself own something scarce that customers need in order to use it?a thin profit, kept by the company itself
DatabricksIf it can be copied: does the company itself own something scarce that customers need in order to use it?a thin profit, kept by the company itself
MicronIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to established firms that already own it; the company named can be one of them
SK HynixIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to established firms that already own it; the company named can be one of them
Samsung ElectronicsIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to established firms that already own it; the company named can be one of them
OpenAIIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to the platforms that own what its customers need
AnthropicIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to the platforms that own what its customers need
Periodic LabsIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to established firms that already own it; the company named can be one of them
ElevenLabsIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to the platforms that own what its customers need
EtchedIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to established firms that already own it; the company named can be one of them
CognitionIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to the platforms that own what its customers need
LovableIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to the platforms that own what its customers need
HarveyIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to established firms that already own it; the company named can be one of them
OpenEvidenceIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to established firms that already own it; the company named can be one of them
Chai DiscoveryIf it does not: does somebody else own that scarce thing?a thin profit, judged to go to established firms that already own it; the company named can be one of them
OracleOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
CoreWeaveOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
NebiusOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
FluidstackOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
Scale AIOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
MercorOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
Handshake AIOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
Surge AIOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
BasetenOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users
FireworksOtherwise: it is easy to copy and needs nothing scarce to reach customers.competed away to users

The market map

Every company the tracker follows, placed in a finer map of the chain. A company can sit in several places, and each place says which source put it there. The categories are Alex Brogan's taxonomy of the chain. Each placement is this site's, taken from the source named on it; a company with no named source sits in the usual category for its part of the stack.

8 layers · 42 categories · 270 parts · 849 companies · 917 placements · 0 categories not yet mapped · 33 kept off the map · 85 readings

How much of the map is verified: the share of each layer's companies that carry the tracker's verified flagChart
Physical inputs and compute supply23.4% · 34 of 145
Models and capability production13.9% · 15 of 108
Data, training and verification inputs9.6% · 9 of 94
Runtime, serving and orchestration10.2% · 10 of 98
Trust, control and transaction infrastructure1.9% · 1 of 52
Deployment and organizational diffusion3.7% · 3 of 81
Applications and operating models3.6% · 11 of 302
The whole map9.1% · 77 of 849
verified: the tracker's record ties the company to a filing or to a row in a dataset (bold on the map)not yet verified: no such check is recorded

Each bar is all the companies the map places in a layer, counted once. Verified is a flag on the tracker's own record of a company; the tracker's notes do not say who set each flag or when. It says the company itself was matched to a record; it does not say a person reviewed where the company sits on the map, and it says nothing about how good the company is. A company not yet verified has simply not been through that check; nothing here says its entry is wrong. The last bar is the whole map.

The numbers
LayerCompaniesVerifiedNot yet verifiedShare verified
Physical inputs and compute supply1453411123.4%
Models and capability production108159313.9%
Data, training and verification inputs949859.6%
Runtime, serving and orchestration98108810.2%
Trust, control and transaction infrastructure521511.9%
Deployment and organizational diffusion813783.7%
Applications and operating models302112913.6%
The whole map849777729.1%
How many companies the map holds in each category, with the few recorded as bought or closed marked, and the parts of each category where it has placed noneChart
no sale or closure recorded (counted 888 times across the categories)bought, or being bought (counted 18 times across the categories)closed (counted 4 times across the categories)a part of the category with a company placed in ita part with none yet

Each bar counts the companies the map places in a category, on a shared scale: the longest bar is the fullest category. A long bar means the tracker follows many companies there; it does not measure the market's size, its sales or its importance. A company placed in more than one category is counted in each. A company in more than one part of a category is counted once in it, so these counts add to slightly fewer than the placements above. Bought, being bought and closed are the states recorded on each company, with who and when in the card's list below; the split shows ownership only where the tracker has recorded it. A company with no recorded state is drawn pale, so the dark and outlined segments are a floor: some sales sit in the tracker's notes and have not yet been recorded as a state.

A category is split into parts, the finest grain of the map, drawn as the dots at the end of its row. An empty ring says where the tracker's coverage stops, not that nobody sells there: 47 of the 263 parts in scope hold no company yet. The category set aside as out of scope, which holds the remaining parts, is left out.

Categories holding few companies (5 or fewer): Model compression and adaptation (3), Settlement and billing (5), Liability and risk transfer (4), Managed AI operations (1). Thin here is a property of the tracker, too.

The numbers: companies in each category by ownership state, with the name of every part that holds no company
CategoryLayerCompaniesNo sale or closure recordedBought, or being boughtClosedThe parts with no company
1.1 Power and sitesPhysical inputs and compute supply414100none
1.2 Data-centre infrastructurePhysical inputs and compute supply191720none
1.3 Semiconductor supply chainPhysical inputs and compute supply302811none
1.4 Memory, storage and networkingPhysical inputs and compute supply181710none
1.5 Compute providersPhysical inputs and compute supply323200none
1.6 Compute utilization and controlPhysical inputs and compute supply10910Capacity management
2.1 General-purpose modelsModels and capability production373700none
2.2 Modality-native modelsModels and capability production434120Robotics and spatial intelligence
2.3 Specialized modelsModels and capability production262420none
2.4 Model compression and adaptationModels and capability production3300Quantization; Pruning; Model merging
3.1 Data originationData, training and verification inputs161600Enterprise workflow data; Negative-CAC data products
3.2 Data rights and exchangeData, training and verification inputs121200Bankruptcy and legacy-data acquisition
3.3 Data transformationData, training and verification inputs121200Cleaning and enrichment; Provenance
3.4 Synthetic dataData, training and verification inputs8800Synthetic-data auditability
3.5 Training environmentsData, training and verification inputs101000Domain-specific training gyms; Environment creation automation; Environment maintenance and version control
3.6 Reward and verification systemsData, training and verification inputs7601Reward models; Verifiers; Outcome telemetry
3.7 Post-training and learning infrastructureData, training and verification inputs201910Mid-training; Refresh and rollback infrastructure; Training-cost optimization
3.8 Evals and assuranceData, training and verification inputs171610none
4.1 Inference supplyRuntime, serving and orchestration292720none
4.2 Inference optimizationRuntime, serving and orchestration101000Speculative decoding
4.3 Routing and gatewaysRuntime, serving and orchestration131300Confidence-based cascades; Frontier-on-exception routing; Compute-provider routing; Outcome-based routing
4.4 Context and memoryRuntime, serving and orchestration212100Knowledge freshness
4.5 Agent runtimesRuntime, serving and orchestration141400none
4.6 Tools and actuationRuntime, serving and orchestration131300Physical actuation
5.1 SecurityTrust, control and transaction infrastructure161420none
5.2 Identity and authorityTrust, control and transaction infrastructure111010Authority and mandate registries
5.3 Provenance and auditTrust, control and transaction infrastructure9900Traceability; Decision logs
5.4 Outcome verificationTrust, control and transaction infrastructure8800Definition-of-done systems; Outcome telemetry; Attribution windows; Dispute resolution
5.5 Settlement and billingTrust, control and transaction infrastructure5500Escrow; Machine-to-machine procurement; Task-completion markets
5.6 Liability and risk transferTrust, control and transaction infrastructure4400Risk scoring; Loss-data networks
6.1 Technical implementationDeployment and organizational diffusion303000none
6.2 Context engineeringDeployment and organizational diffusion6600Evaluation-rubric creation; Human escalation design
6.3 Managed AI operationsDeployment and organizational diffusion1100Agent operations; Evaluation operations; Continuous improvement; Vendor and model management
6.4 Adoption and changeDeployment and organizational diffusion8800Workflow redesign; Political adoption
6.5 Talent productionDeployment and organizational diffusion404000none
7.1 Horizontal copilotsApplications and operating models262420none
7.2 Functional enterprise applicationsApplications and operating models202000none
7.3 Vertical workflow softwareApplications and operating models585800none
7.4 Autonomous servicesApplications and operating models898702none
7.5 Sensor–model–actuator networksApplications and operating modelsout of scope for now000out of scope for now
7.6 AI-native rollupsApplications and operating models10610600none
7.7 Consumer AIApplications and operating models121200none
LayerParts in scopeWith a companyWith noneCompanies placed in a category but in no part
Physical inputs and compute supply353410
Models and capability production211740
Data, training and verification inputs5439150
Runtime, serving and orchestration403370
Trust, control and transaction infrastructure3422120
Deployment and organizational diffusion3022810
Applications and operating models494904

each dot is one part of the category; filled means the tracker has a company thereBold: verified through a filing or a dataset(acquired), (acquisition pending), (closed): what happened to the companynot yet mapped: a gap in our coverage, not in the market

Layer 1

Physical inputs and compute supply

6 categories · 145 companies

1.1

Power and sites

41
Applied DigitalCore ScientificCrusoeIRENOkloTeraWulfAgentAmerican TerawattArborAyr EnergyBase PowerBoom SupersonicCalibrant EnergyCommonwealth Fusion+27
5/5 parts1 indicator
Companies, sources and readings
Generation
Placed by hand (review, 28 Sep 2026): Oklo assess, Commonwealth Fusion assess, Exowatt assess, X-energy assess
Roadmap: The AI data center stack, section “Power Generation”, read 28 Sep 2026: Arbor assess, Boom Supersonic assess, ERock assess, Inertia assess, Mainspring assess, Radiant assess
Storage and grid services
Roadmap: The AI data center stack, section “Transmission & Power Conversion”, read 28 Sep 2026: Agent assess, Moment Energy assess
Placed by hand (review, 28 Sep 2026): Base Power assess
Roadmap: The AI data center stack, section “Power Generation”, read 28 Sep 2026: Calibrant Energy assess, Redwood Energy assess
Transmission and interconnection
Roadmap: The AI data center stack, section “Transmission & Power Conversion”, read 28 Sep 2026: American Terawatt assess, Ayr Energy assess, GridAstra assess, Heron Power assess, Ionate assess, TS Conductor assess
Roadmap: The AI data center stack, section “Software & Orchestration”, read 28 Sep 2026: GridCARE assess, Squid assess, ThinkLabs assess
Roadmap: The AI data center stack, section “Construction & Maintenance”, read 28 Sep 2026: Gridware assess
Data-centre real estate and site development
Placed by hand (review, 28 Sep 2026): Applied Digital assess, Core Scientific assess, Crusoe assess, IREN assess, TeraWulf assess
Roadmap: The AI data center stack, section “Permitting & Site Selection”, read 28 Sep 2026: Geoforge.ai assess, Lorica assess, Nyxium assess, Paces assess, PermitPal assess, Scout assess, Spark assess
Power procurement and demand management
Roadmap: The AI data center stack, section “Software & Orchestration”, read 28 Sep 2026: Companion Energy assess, GridBeyond assess, Mercury assess, Verse assess

Readings: Training power draw growth (doubling time)

Venture money for dc power, as recorded by sub-layer.

1.2

Data-centre infrastructure

19
JetCool(acquired)AlcorAmperesandAravoltaCorintisDG MatrixFlexnodeFrore SystemsGeolockedHyperscale PowerLiquidStack(acquired)LucendMaintainXNovoLINC+5
5/5 partsno indicator
Companies, sources and readings
Construction and modular data centres
Placed by hand (from its own site, read 5 Oct 2026): Flexnode assess
Cooling
Placed by hand (review, 28 Sep 2026): JetCool (acquired: Flex, Nov 2024) assess, LiquidStack (acquired: Trane Technologies, closed Mar 2026) assess, Submer assess
Roadmap: The AI data center stack, section “Cooling Technologies”, read 28 Sep 2026: Corintis assess, Frore Systems assess, Geolocked assess, Lucend assess, NovoLINC assess, Phaidra assess, Reliability Engine assess
Power distribution
Roadmap: The AI data center stack, section “Transmission & Power Conversion”, read 28 Sep 2026: Amperesand assess, DG Matrix assess, Hyperscale Power assess
Racks and integrated systems
Placed by hand (from its own site, read 5 Oct 2026): Supermicro assess, Wiwynn assess
Data-centre operations
Roadmap: The AI data center stack, section “Construction & Maintenance”, read 28 Sep 2026: Alcor assess, MaintainX assess
Roadmap: The AI data center stack, section “Software & Orchestration”, read 28 Sep 2026: Aravolta assess
1.3

Semiconductor supply chain

30
AMDApplied MaterialsASMLBroadcomKLALam ResearchMatXNVIDIAPositronSamsung ElectronicsTenstorrentTSMCAxelera AICerebras+16
6/6 parts6 indicators
Companies, sources and readings
Accelerator design
Placed by hand (review, 28 Sep 2026): AMD (Advanced Micro Devices) assess, MatX assess, NVIDIA (NVIDIA Corp) assess, Positron assess, Tenstorrent assess, Cerebras assess, d-Matrix assess, Extropic assess, Fractile assess, Groq assess, SambaNova assess
The Inference Market Map, section “chips / edge inference silicon”, read 28 Sep 2026: Axelera AI assess, FuriosaAI assess, Hailo (acquired: Microchip Technology, closed 21 Sep 2026) assess, Mythic assess, Rebellions assess, Sima.ai assess
The Inference Market Map, section “chips / GPU incumbents”, read 28 Sep 2026: Intel assess
The Inference Market Map, section “chips / photonic & novel”, read 28 Sep 2026: Q.ANT assess
The Inference Market Map, section “chips / transformer ASICs”, read 28 Sep 2026: Rain AI (closed: ceased operating in 2025; patents bought by OpenAI (reported)) assess
neolabs.fyi, section “Compute & chips”, read 28 Sep 2026: Unconventional AI assess
Custom silicon and ASICs
Placed by hand (review, 28 Sep 2026): Broadcom assess, Etched assess
Semiconductor fabrication
Placed by hand (foundry, review, 28 Sep 2026): Samsung Electronics assess
Placed by hand (review, 28 Sep 2026): TSMC assess
Advanced packaging
Placed by hand (review, 28 Sep 2026): TSMC assess
Semiconductor manufacturing equipment
Placed by hand (review, 28 Sep 2026): Applied Materials assess, KLA (KLA Corp) assess, Lam Research assess
Placed by hand (lithography maker, review, 28 Sep 2026): ASML assess
Chiplets and interconnect
Placed by hand (review, 28 Sep 2026): Tenstorrent assess
Roadmap: The AI data center stack, section “IT & Networking”, read 28 Sep 2026: Kandou.ai assess

Readings: Rent concentration in semiconductors, Semiconductors' share of stack operating income, Chips' share of the stack's gross profit, Concentration among covered chip filers (HHI), Non-Nvidia share of AI compute shipped (H100-equivalents), Hardware price-performance growth (doubling time)

Venture money for semis, as recorded by sub-layer.

1.4

Memory, storage and networking

18
Astera LabsBroadcomCelestial AI(acquired)EnfabricaLightmatterMicronNVIDIASamsung ElectronicsAyar LabsDDNDriveNetsEriduNexthop AISK Hynix+4
6/6 partsno indicator
Companies, sources and readings
HBM and memory
Placed by hand (review, 28 Sep 2026): Micron assess, SK Hynix assess
Placed by hand (memory maker, review, 28 Sep 2026): Samsung Electronics assess
Storage
Placed by hand (from its own site, read 5 Oct 2026): DDN assess, VAST Data assess, WEKA assess
Networking silicon
Placed by hand (review, 28 Sep 2026): Broadcom assess, Enfabrica assess
Roadmap: The AI data center stack, section “IT & Networking”, read 28 Sep 2026: Eridu assess
Optical interconnect
Placed by hand (review, 28 Sep 2026): Celestial AI (acquired: Marvell, closed Feb 2026) assess, Lightmatter assess, Ayar Labs assess
Roadmap: The AI data center stack, section “IT & Networking”, read 28 Sep 2026: Xscape Photonics assess
Scale-up networking
Placed by hand (review, 28 Sep 2026): Astera Labs assess, NVIDIA (NVIDIA Corp) assess
Roadmap: The AI data center stack, section “IT & Networking”, read 28 Sep 2026: Upscale AI assess
Scale-out networking
Roadmap: The AI data center stack, section “IT & Networking”, read 28 Sep 2026: DriveNets assess, Nexthop AI assess

Venture money for memory networking, as recorded by sub-layer.

1.5

Compute providers

32
AlphabetAmazonCoreWeaveCrusoeFluidstackIRENMicrosoftNebiusOraclePrime IntellectRunPodTensorWaveAethirAkash Network+18
7/7 parts7 indicators
Companies, sources and readings
Hyperscale clouds
Placed by hand (review, 28 Sep 2026): Alphabet (Alphabet Inc.) assess, Amazon (Amazon.com Inc.) assess, Microsoft (Microsoft Corp) assess, Oracle (Oracle Corp) assess
Neoclouds
Placed by hand (review, 28 Sep 2026): CoreWeave (CoreWeave Inc.) assess, Crusoe assess, Fluidstack assess, IREN assess, Nebius (Nebius Group N.V.) assess, Lambda assess, Vultr assess
The Inference Market Map, section “clusters / GPU marketplaces”, read 28 Sep 2026: Hyperstack assess, Ori Cloud assess, Verda assess
GPU bare-metal providers
Placed by hand (review, 28 Sep 2026): RunPod assess, TensorWave assess, Voltage Park assess
Sovereign and regional AI clouds
Placed by hand (review, 28 Sep 2026): Firmus assess, Nscale assess
Edge and on-premise compute
Placed by hand (from its own site, read 5 Oct 2026): Armada assess, Oxide Computer assess
Compute marketplaces and brokers
Placed by hand (review, 28 Sep 2026): Prime Intellect assess, Hyperbolic assess
The Inference Market Map, section “clusters / decentralized-DePIN”, read 28 Sep 2026: Aethir assess, Akash Network assess, Gensyn assess, io.net assess, Render Network assess
The Inference Market Map, section “clusters / GPU marketplaces”, read 28 Sep 2026: Salad Cloud assess, Vast.ai assess
Compute financing and capacity contracts
Placed by hand (from its own site, read 5 Oct 2026): Compute Exchange assess, Ornn assess

Readings: Cloud backlog (remaining performance obligations), Cloud segment operating margins, Concentration among covered cloud filers (HHI), Hyperscaler capex to AI revenue (trailing year), Hyperscaler server useful lives (10-K estimates), Neocloud borrowing cost (CoreWeave facility rates), GPU rental price per hour

Venture money for hyperscalers neoclouds, as recorded by sub-layer.

1.6

Compute utilization and control

10
AnyscaleClaros(acquisition pending)Emerald AIHammerheadNiv-AIPado AIQUTWOSims Lifecycle ServicesSoma EnergyZymtrace
5/6 partsno indicator
Companies, sources and readings
Cluster scheduling
Placed by hand (review, 28 Sep 2026): Anyscale assess
GPU utilization
Placed by hand (from its own site, read 5 Oct 2026): Zymtrace assess
Workload placement
neolabs.fyi, section “Compute & chips — Sovereign AI”, read 28 Sep 2026: QUTWO assess
Energy-aware orchestration
Roadmap: The AI data center stack, section “Software & Orchestration”, read 28 Sep 2026: Claros (acquisition pending: Navitas Semiconductor, agreed Aug 2026) assess, Emerald AI assess, Hammerhead assess, Pado AI assess, Soma Energy assess
Roadmap: The AI data center stack, section “Cooling Technologies”, read 28 Sep 2026: Niv-AI assess
Hardware lifecycle and resale
Placed by hand (from its own site, read 5 Oct 2026): Sims Lifecycle Services assess
Layer 2

Models and capability production

4 categories · 108 companies

2.1

General-purpose models

37
AnthropicDeepSeekFlapping AirplanesMetaMiniMaxMistral AIMoonshot AIOpenAIPrime IntellectxAIZhipu AI01.AIAleph AlphaAlibaba Qwen+23
4/4 parts17 indicators
Companies, sources and readings
Closed frontier labs
Placed by hand (review, 28 Sep 2026): Anthropic assess, Meta (Meta Platforms) assess, OpenAI assess, xAI assess, Google DeepMind assess, Safe Superintelligence assess, Thinking Machines Lab assess
Placed by its sub-layer (neolabs): Flapping Airplanes assess, Core Automation assess, Discovery Loop assess, Humans& assess, Ineffable Intelligence assess, Mirendil assess, Ndea assess, Oak Lab assess, Recursive assess
neolabs.fyi, section “Applied — Agents, Multimodal, Open weights, Sovereign AI”, read 28 Sep 2026: Cohere assess
Open-weight labs
Placed by hand (review, 28 Sep 2026): Mistral AI assess, Prime Intellect assess, Alibaba Qwen assess, Nous Research assess, Reflection AI assess
neolabs.fyi, section “General-purpose models — Open weights”, read 28 Sep 2026: Deep Cogito assess, Essential AI assess
Placed by its sub-layer (neolabs): Dolphin AI assess
neolabs.fyi, section “General-purpose models — Open source, Open weights”, read 28 Sep 2026: Zyphra assess
Sovereign and national labs
Placed by hand (review, 28 Sep 2026): Mistral AI assess, Aleph Alpha assess, MBZUAI assess, Sakana AI assess, Sarvam assess
Low-cost frontier followers
Placed by hand (review, 28 Sep 2026): DeepSeek assess, MiniMax assess, Moonshot AI assess, Zhipu AI assess, 01.AI assess, Baichuan assess, StepFun assess

Readings: Frontier-lab revenue run-rates, Lab recoupment ratio, Lab concentration of routed tokens (OpenRouter, HHI), Price per unit of capability (cheapest METR-hour on OpenRouter), Price per unit of capability (cheapest near-frontier model, Artificial Analysis), Frontier labs' share of tracked venture dollars, Frontier lab valuation to run-rate multiple, Lab vertical integration (acquisitions and acquihires, trailing four quarters), Open-weights gap to the closed frontier (ECI points), Frontier lab gross margin (estimated), Model-layer token concentration on OpenRouter (HHI, top table), METR 50% time horizon, METR 80% time horizon, Frontier training compute growth (doubling time), Open-weights lag behind the closed frontier (months), ARC-AGI-2 best score, Internal-to-public deployment gap at the frontier

Venture money for frontier labs, neolabs, open weight labs, as recorded by sub-layer.

2.2

Modality-native models

43
Black Forest LabsLumaAdvanced Machine Intelligence / AMI LabsAssemblyAIAxiom MathCartesiaChai DiscoveryCuspAIDecartDeepgramEdison ScientificElevenLabsElorianEvolutionaryScale(acquired)+29
4/5 partsno indicator
Companies, sources and readings
Voice and audio
The Inference Market Map, section “modality / voice”, read 28 Sep 2026: AssemblyAI assess, Deepgram assess, Hume AI assess, PlayHT (acquired: Meta, Jul 2025) assess, Sesame AI assess, Ultravox assess
Placed by hand (review, 28 Sep 2026): Cartesia assess, ElevenLabs assess, Kyutai assess, Suno assess
Image
Placed by hand (review, 28 Sep 2026): Black Forest Labs assess
Placed by its sub-layer (neolabs): Elorian assess
The Inference Market Map, section “modality / image”, read 28 Sep 2026: Ideogram assess, Krea AI assess, Midjourney assess, Recraft assess
Video and world models
Placed by hand (review, 28 Sep 2026): Luma assess, General Intuition assess, Moonlake AI assess, Pika assess, Runway assess, World Labs assess
Placed by its sub-layer (neolabs): Advanced Machine Intelligence / AMI Labs assess
neolabs.fyi, section “World models & simulation — Video”, read 28 Sep 2026: Decart assess
The Inference Market Map, section “modality / video & real-time”, read 28 Sep 2026: Genmo assess, Hedra assess, Tavus assess
neolabs.fyi, section “Generative media & voice — Video”, read 28 Sep 2026: Moonvalley assess
Scientific and biological models
neolabs.fyi, section “AI for math & science — Math, Open source”, read 28 Sep 2026: Axiom Math assess
neolabs.fyi, section “AI for math & science — Drug discovery, Open weights”, read 28 Sep 2026: Chai Discovery assess, EvolutionaryScale (acquired: Chan Zuckerberg Biohub, Nov 2025) assess
neolabs.fyi, section “AI for math & science — Materials”, read 28 Sep 2026: CuspAI assess, Unreasonable Labs assess
neolabs.fyi, section “AI for math & science — Agents, Drug discovery”, read 28 Sep 2026: Edison Scientific assess, Latent Labs assess
neolabs.fyi, section “AI for math & science — Agents, Open weights”, read 28 Sep 2026: FutureHouse assess
Placed by its sub-layer (neolabs): Helical assess
Placed by hand (review, 28 Sep 2026): Isomorphic Labs assess, Lila Sciences assess, Periodic Labs assess
neolabs.fyi, section “AI for math & science — Agents, Materials, Open source, Open weights”, read 28 Sep 2026: Orbital assess
neolabs.fyi, section “AI for math & science — Agents, Open source”, read 28 Sep 2026: Physical Superintelligence assess
neolabs.fyi, section “AI for math & science — Drug discovery”, read 28 Sep 2026: Xaira Therapeutics assess
2.3

Specialized models

26
Liquid AIMagicArcee AICohereContextual AIdoubleAIFundamentalH CompanyHarmonicInflection AIJina AI(acquired)KindoMarqoMath, Inc.+12
6/6 partsno indicator
Companies, sources and readings
Vertical/domain models
Placed by hand (review, 28 Sep 2026): Magic assess, Harmonic assess, Math, Inc. assess, Poolside assess, Ricursive Intelligence assess, Thomson Reuters–Imperial Frontier AI Research Lab assess
neolabs.fyi, section “General-purpose models”, read 28 Sep 2026: doubleAI assess
neolabs.fyi, section “Applied”, read 28 Sep 2026: Inflection AI assess
Placed by hand (physics simulation models (review, 28 Sep 2026); neolabs.fyi lists it under Physical AI & robotics): UniversalAGI assess
Small language models
Placed by hand (review, 28 Sep 2026): Liquid AI assess, Arcee AI assess
Placed by its sub-layer (neolabs): PrismML assess
Decision and classification models
neolabs.fyi, section “Applied — Tabular / structured data”, read 28 Sep 2026: Fundamental assess
neolabs.fyi, section “Applied — Academic spinout, Open source, Open weights, Tabular / structured data”, read 28 Sep 2026: Prior Labs assess
neolabs.fyi, section “World models & simulation”, read 28 Sep 2026: Simile assess
Placed by its sub-layer (neolabs): Sooth Labs assess
Embedding and reranking models
The Inference Market Map, section “modality / embed & rerank”, read 28 Sep 2026: Cohere assess, Contextual AI assess, Jina AI (acquired: Elastic, Oct 2025) assess, Marqo assess, MongoDB assess, Nomic AI assess, Voyage AI (acquired: MongoDB, Feb 2025) assess
Computer-use models
neolabs.fyi, section “Applied — Agents, Open weights”, read 28 Sep 2026: H Company assess
Placed by hand (review, 28 Sep 2026): Standard Intelligence assess
Security models
Placed by hand (from its own site, read 5 Oct 2026): Kindo assess
2.4

Model compression and adaptation

3
Distil LabsInception LabsUnsloth
3/6 partsno indicator
Companies, sources and readings
Distillation
Placed by hand (from its own site, read 5 Oct 2026): Distil Labs assess
Fine-tuning
Placed by hand (review, 28 Sep 2026): Unsloth assess
Architecture and inference optimization
Placed by hand (review, 28 Sep 2026): Inception Labs assess
Layer 3

Data, training and verification inputs

8 categories · 94 companies

3.1

Data origination

16
MercorProlificBee MapsDatacurveHandshake AIHuman ArchiveInvisible TechnologiesKledLila SciencesMicro1NATIX NetworkParetoPeriodic LabsScale AI+2
5/7 parts2 indicators
Companies, sources and readings
Expert networks
Placed by hand (review, 28 Sep 2026): Mercor assess, Prolific assess, Datacurve assess, Handshake AI assess, Invisible Technologies assess, Micro1 assess, Pareto assess, Scale AI assess, Surge AI assess, Turing assess
Physical sensor data
Placed by hand (from its own site, read 5 Oct 2026): Bee Maps assess, NATIX Network assess
Robotics and motion data
Placed by hand (from its own site, read 5 Oct 2026): Human Archive assess
Scientific and experimental data
Placed by hand (review, 28 Sep 2026): Lila Sciences assess, Periodic Labs assess
Consumer interaction data
Placed by hand (from its own site, read 5 Oct 2026): Kled assess

Readings: Labs buying from the training-input layer (enter bell), Expert-data market run-rate (Mercor)

Venture money for human data, as recorded by sub-layer.

3.2

Data rights and exchange

12
RSL CollectiveCreated by HumansDataradeDecentriqDefined.aiDuality TechnologiesHuman NativeProRataProtegeTollBitTroveoVermillio
5/6 partsno indicator
Companies, sources and readings
Permissioned licensing
Placed by hand (review, 28 Sep 2026): Created by Humans assess, Protege assess
Data marketplaces
Placed by hand (review, 28 Sep 2026): Datarade assess, Defined.ai assess, Human Native assess
Data brokers
Placed by hand (from its own site, read 5 Oct 2026): Troveo assess
Privacy-preserving data exchange
Placed by hand (from its own site, read 5 Oct 2026): Decentriq assess, Duality Technologies assess
Consent and rights management
Placed by hand (review, 28 Sep 2026): RSL Collective assess, ProRata assess, TollBit assess, Vermillio assess
3.3

Data transformation

12
AppenCloudFactoryDatologyiMeritLabelboxOvermindSapienScale AISnorkelSurge AITolokaTonic
5/7 partsno indicator
Companies, sources and readings
Collection tooling
Placed by hand (review, 28 Sep 2026): Labelbox assess
Annotation
Placed by hand (review, 28 Sep 2026): Appen assess, CloudFactory assess, iMerit assess, Sapien assess, Scale AI assess, Snorkel assess, Surge AI assess, Toloka assess
De-identification
Placed by hand (review, 28 Sep 2026): Tonic assess
Trace extraction
Placed by hand (from its own site, read 5 Oct 2026): Overmind assess
Dataset quality assurance
Placed by hand (review, 28 Sep 2026): Datology assess
3.4

Synthetic data

8
Duality AIForetellixGretelMostly AIParallel DomainRendered.aiSynthesis AITonic
4/5 partsno indicator
Companies, sources and readings
Synthetic generation
Placed by hand (review, 28 Sep 2026): Gretel assess, Mostly AI assess, Synthesis AI assess, Tonic assess
Simulation
Placed by hand (review, 28 Sep 2026): Rendered.ai assess
Diversity and coverage measurement
Placed by hand (from its own site, read 5 Oct 2026): Foretellix assess
Sim-to-real pipelines
Placed by hand (from its own site, read 5 Oct 2026): Duality AI assess, Parallel Domain assess
3.5

Training environments

10
PatronusAntiochApplied IntuitionBespoke LabsFleetHalluminateMatricesMechanizePlatoVeris
4/7 parts1 indicator
Companies, sources and readings
RL environments
Placed by hand (review, 28 Sep 2026): Patronus assess, Mechanize assess, Plato assess
Placed by its sub-layer (rl environments): Bespoke Labs assess
Agent simulators
Placed by hand (review, 28 Sep 2026): Veris assess
Computer-use environments
Placed by its sub-layer (rl environments): Fleet assess, Matrices assess
Placed by hand (review, 28 Sep 2026): Halluminate assess
Physical-world simulators
Placed by hand (from its own site, read 5 Oct 2026): Antioch assess, Applied Intuition assess

Readings: Active RL-environment vendors (directory count)

3.6

Reward and verification systems

7
Atla(closed)HarmonicJudgment LabsLMArenaLogical IntelligenceMath, Inc.MLCommons
4/7 partsno indicator
Companies, sources and readings
Proxy graders
Placed by hand (review, 28 Sep 2026): Atla (closed: product shut down Feb 2026 (reported)) assess, Judgment Labs assess
Formal verification
Placed by hand (review, 28 Sep 2026): Harmonic assess, Logical Intelligence assess, Math, Inc. assess
Human preference and taste calibration
Placed by hand (review, 28 Sep 2026): LMArena assess
Evaluation-standard creation
Placed by hand (from its own site, read 5 Oct 2026): MLCommons assess
3.7

Post-training and learning infrastructure

20
BasetenPrime IntellectWeights & Biases(acquired)Adaption LabsAdaptive MLApplied ComputeArcee AIAutoscienceCursiveEngramInherentNeoCognitionOpenPipeOsmosis+6
4/7 parts3 indicators
Companies, sources and readings
Fine-tuning infrastructure
Placed by hand (review, 28 Sep 2026): Baseten assess, Weights & Biases (acquired: CoreWeave, closed May 2025) assess, Arcee AI assess, OpenPipe assess, Predibase assess, Thinking Machines Lab assess, Together assess
neolabs.fyi, section “Applied — Multimodal, Open source, Open weights, Training infra / tooling”, read 28 Sep 2026: Oumi assess
RL as a service
Placed by hand (review, 28 Sep 2026): Prime Intellect assess, Adaptive ML assess, Applied Compute assess, Osmosis assess
neolabs.fyi, section “Applied — Training infra / tooling”, read 28 Sep 2026: River AI assess
Continual learning
neolabs.fyi, section “Recursive self-improvement & continual learning — Continual learning”, read 28 Sep 2026: Adaption Labs assess
neolabs.fyi, section “Recursive self-improvement & continual learning — Agents”, read 28 Sep 2026: Autoscience assess, Inherent assess
neolabs.fyi, section “Recursive self-improvement & continual learning”, read 28 Sep 2026: Cursive assess
Placed by hand (review, 28 Sep 2026): Engram assess, Trajectory assess
neolabs.fyi, section “Recursive self-improvement & continual learning — Continual learning, Agents, Academic spinout”, read 28 Sep 2026: NeoCognition assess
Adapter management
Placed by hand (review, 28 Sep 2026): Trajectory assess

Readings: Continual-learning ladder (highest production rung), Vendors claiming shipped continual learning, Fastest production refresh of a continually updated model (hours)

3.8

Evals and assurance

17
ArizeGoodfirePatronusWeights & Biases(acquired)BraintrustConfident AIGalileoHeliconeHumanloopIrregularJudgment LabsLangSmithLawZeroLMArena+3
8/8 parts1 indicator
Companies, sources and readings
Foundation-model benchmarks
Placed by hand (review, 28 Sep 2026): LMArena assess, Vals AI assess
Application evals
Placed by hand (review, 28 Sep 2026): Patronus assess, Braintrust assess, Humanloop assess, Judgment Labs assess
Enterprise-specific eval authoring
Placed by hand (from its own site, read 5 Oct 2026): Snorkel assess
Production observability
Placed by hand (review, 28 Sep 2026): Arize assess, Weights & Biases (acquired: CoreWeave, closed May 2025) assess, Galileo assess, Helicone assess, LangSmith assess
Regression testing
Placed by hand (review, 28 Sep 2026): Confident AI assess
Red teaming
neolabs.fyi, section “Applied — Security”, read 28 Sep 2026: Irregular assess
Third-party assurance and certification
neolabs.fyi, section “AI safety & interpretability”, read 28 Sep 2026: LawZero assess
neolabs.fyi, section “AI safety & interpretability — Open source, Open weights”, read 28 Sep 2026: Transluce assess
Interpretability
Placed by hand (review, 28 Sep 2026): Goodfire assess

Readings: Agent reliability across repeated runs (tau-bench pass^k)

Venture money for evals, as recorded by sub-layer.

Layer 4

Runtime, serving and orchestration

6 categories · 98 companies

4.1

Inference supply

29
BasetenModalReplicateRunPodAnyscaleAppleBeam CloudCerebrasCerebriumDatabricksDeepInfraEdge Impulse(acquired)Edgeless Systemsfal.ai+15
6/6 parts2 indicators
Companies, sources and readings
Serverless inference
Placed by hand (review, 28 Sep 2026): Modal assess, RunPod assess, Cerebras assess, Groq assess
The Inference Market Map, section “clusters / independent neoinference”, read 28 Sep 2026: Beam Cloud assess, Cerebrium assess
Dedicated endpoints
Placed by hand (review, 28 Sep 2026): Baseten assess, Anyscale assess, SambaNova assess
The Inference Market Map, section “clusters / hyperscaler-managed”, read 28 Sep 2026: Databricks assess, IBM assess
The Inference Market Map, section “clusters / independent neoinference”, read 28 Sep 2026: Friendli assess, Inferless (acquired: Baseten (acquihire), early 2026) assess, Simplismart assess
Open-model serving
Placed by hand (review, 28 Sep 2026): DeepInfra assess, Fireworks assess, Hyperbolic assess, Novita assess, Together assess
The Inference Market Map, section “clusters / independent neoinference”, read 28 Sep 2026: Featherless assess
The Inference Market Map, section “gateway / marketplaces”, read 28 Sep 2026: Hugging Face assess
Multimodal inference
Placed by hand (review, 28 Sep 2026): Replicate assess
The Inference Market Map, section “gateway / marketplaces”, read 28 Sep 2026: fal.ai assess
Edge inference
The Inference Market Map, section “clusters / edge & on-device”, read 28 Sep 2026: Apple assess, Edge Impulse (acquired: Qualcomm, Mar 2025) assess, Ollama assess
Sovereign/on-premise inference
The Inference Market Map, section “clusters / confidential-private”, read 28 Sep 2026: Edgeless Systems assess, Phala Network assess, Tinfoil assess

Readings: Inference-layer revenue run-rates, Inference price at fixed capability (halving time)

Venture money for inference serving, as recorded by sub-layer.

4.2

Inference optimization

10
DoublewordEmbedlFastlyFireworksInferactLatent AIModularNeuralwattRadixArkThe Token Company
6/7 partsno indicator
Companies, sources and readings
Kernels and compilers
Placed by hand (review, 28 Sep 2026): Fireworks assess
neolabs.fyi, section “Inference — Open source”, read 28 Sep 2026: Inferact assess
The Inference Market Map, section “engines / commercial”, read 28 Sep 2026: Modular assess
neolabs.fyi, section “Inference — Open source, Training infra / tooling”, read 28 Sep 2026: RadixArk assess
Caching
The Inference Market Map, section “clusters / edge & on-device”, read 28 Sep 2026: Fastly assess
Batching
Placed by hand (from its own site, read 5 Oct 2026): Doubleword assess
Prompt and context compression
Placed by hand (from its own site, read 5 Oct 2026): The Token Company assess
Hardware-specific optimization
Placed by hand (from its own site, read 5 Oct 2026): Embedl assess, Latent AI assess
Energy-per-outcome optimization
Placed by hand (from its own site, read 5 Oct 2026): Neuralwatt assess
4.3

Routing and gateways

13
OpenRouterPortkeyVercelAI21 LabsCloudflareHeliconeKongLiteLLMMartianNot DiamondRequestyTrueFoundryUnify
2/6 parts1 indicator
Companies, sources and readings
Model gateways
Placed by hand (review, 28 Sep 2026): OpenRouter assess, Portkey assess, Vercel assess, Helicone assess, LiteLLM assess, Requesty assess, TrueFoundry assess
neolabs.fyi, section “Applied — Open weights”, read 28 Sep 2026: AI21 Labs assess
The Inference Market Map, section “gateway / gateways”, read 28 Sep 2026: Cloudflare assess, Kong assess
Cost/latency/performance routing
Placed by hand (review, 28 Sep 2026): Martian assess, Not Diamond assess, Unify assess

Readings: Enterprise multi-model usage share (Menlo)

Venture money for routing gateways, as recorded by sub-layer.

4.4

Context and memory

21
ExaByteRoverChromaCogneeCompresrContextual AIGleanInterloomLanceDBLettaMem0ParallelPineconeQdrant+7
6/7 partsno indicator
Companies, sources and readings
RAG and retrieval
Placed by hand (review, 28 Sep 2026): Exa assess, Parallel assess
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “Retrieval and data substrate”, read 28 Sep 2026: Chroma assess, LanceDB assess, Pinecone assess, Qdrant assess, Unstructured assess, Weaviate assess, Zilliz assess
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “Shared enterprise context”, read 28 Sep 2026: Contextual AI assess
Context graphs
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “Memory-native startups”, read 28 Sep 2026: Interloom assess
Placed by hand (from its own site, read 9 Oct 2026): TextQL assess
Long-term memory
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “Memory-native startups”, read 28 Sep 2026: ByteRover assess, Cognee assess, Mem0 assess, Supermemory assess, Zep assess
Placed by hand (review, 28 Sep 2026): Letta assess
Context curation
Placed by hand (from its own site, read 5 Oct 2026): Tessl assess
Context compression
Placed by hand (from its own site, read 5 Oct 2026): Compresr assess
Permissions-aware retrieval
Placed by hand (review, 28 Sep 2026): Glean assess
4.5

Agent runtimes

14
ModalAgentOpsCrewAIDaytonaE2BInngestIsara LabsLangChainLlamaIndexMastraPoetiqRestateRivet (rivet.dev)Temporal
7/7 partsno indicator
Companies, sources and readings
Agent frameworks
Placed by hand (review, 28 Sep 2026): LangChain assess, Mastra assess
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “Agent frameworks and runtimes”, read 28 Sep 2026: LlamaIndex assess
Planning and decomposition
Placed by its sub-layer (neolabs): Poetiq assess
State management
Placed by hand (from its own site, read 5 Oct 2026): Rivet (rivet.dev) assess
Multi-agent coordination
Placed by hand (review, 28 Sep 2026): CrewAI assess
Placed by its sub-layer (neolabs): Isara Labs assess
Sandboxes
Placed by hand (review, 28 Sep 2026): Modal assess, Daytona assess, E2B assess
Agent testing and replay
Placed by hand (from its own site, read 5 Oct 2026): AgentOps assess
Durable execution
Placed by hand (from its own site, read 5 Oct 2026): Inngest assess, Restate assess, Temporal assess

Venture money for agent infra, as recorded by sub-layer.

4.6

Tools and actuation

13
ArcadeVapiBlandBrowserbaseComposioCuaDailygotoHumanMergeNangoOrgoRetellVocode
6/7 partsno indicator
Companies, sources and readings
Tool registries
Placed by hand (review, 28 Sep 2026): Arcade assess, Composio assess
API integration
Placed by hand (review, 28 Sep 2026): Merge assess, Nango assess
Browser infrastructure
Placed by hand (review, 28 Sep 2026): Browserbase assess
Computer-use infrastructure
Placed by hand (from its own site, read 5 Oct 2026): Cua assess, Orgo assess
Voice and messaging infrastructure
Placed by hand (review, 28 Sep 2026): Vapi assess, Bland assess, Retell assess
The Inference Market Map, section “modality / voice”, read 28 Sep 2026: Daily assess, Vocode assess
Human escalation
Placed by hand (from its own site, read 5 Oct 2026): gotoHuman assess
Layer 5

Trust, control and transaction infrastructure

6 categories · 52 companies

5.1

Security

16
7AIAstrix(acquisition pending)Barndoor AIDropzone AIGeordieGray SwanLakeraLassoMindgardRecoRunlayerSpecterOpsStraikerVeza(acquired)+2
6/6 partsno indicator
Companies, sources and readings
Agent security
Identity and Access Management (IAM) in the age of AI Agents, section “1. Discovery & Visibility of Agents”, read 28 Sep 2026: Astrix (acquisition pending: Cisco, announced May 2026) assess, Geordie assess, Reco assess, SpecterOps assess, Veza (acquired: ServiceNow, closed Mar 2026) assess
Prompt-injection defence
Placed by hand (from its own site, read 5 Oct 2026): Lakera assess
Model and tool access controls
Placed by hand (from its own site, read 5 Oct 2026): Barndoor AI assess, Runlayer assess
Agent red teaming
Placed by hand (from its own site, read 5 Oct 2026): Gray Swan assess, Mindgard assess
Runtime monitoring
Identity and Access Management (IAM) in the age of AI Agents, section “3. Agent Monitoring, Detection & Response”, read 28 Sep 2026: Lasso assess, Straiker assess, Zenity assess
Cybersecurity agents
Placed by hand (from its own site, read 5 Oct 2026): 7AI assess, Dropzone AI assess, XBOW assess
5.2

Identity and authority

11
ArcadeDelineaKeeperKeyfactorLinxOryPlainIDSGNL(acquisition pending)SkyfireTeleportVouched
4/5 partsno indicator
Companies, sources and readings
Agent identity
Identity and Access Management (IAM) in the age of AI Agents, section “2. Agent Identity & Credential Governance”, read 28 Sep 2026: Delinea assess, Keeper assess, Keyfactor assess, Linx assess, Teleport assess
Identity and Access Management (IAM) in the age of AI Agents, section “4. Action & Authorization”, read 28 Sep 2026: SGNL (acquisition pending: CrowdStrike, announced Jan 2026) assess
Delegated permissions
Placed by hand (review, 28 Sep 2026): Arcade assess
Identity and Access Management (IAM) in the age of AI Agents, section “4. Action & Authorization”, read 28 Sep 2026: PlainID assess
Authentication
Identity and Access Management (IAM) in the age of AI Agents, section “4. Action & Authorization”, read 28 Sep 2026: Ory assess
Know-your-agent
Placed by hand (from its own site, read 5 Oct 2026): Skyfire assess, Vouched assess
5.3

Provenance and audit

9
Credo AIEQTY LabMonitaurOpenBox AIOpenOriginsSaidotSteg.AItrailTrustible
3/5 partsno indicator
Companies, sources and readings
Output provenance
Placed by hand (from its own site, read 5 Oct 2026): OpenOrigins assess, Steg.AI assess
Compliance records
Placed by hand (from its own site, read 5 Oct 2026): Credo AI assess, Monitaur assess, Saidot assess, trail assess, Trustible assess
Tamper-resistant audit trails
Placed by hand (from its own site, read 5 Oct 2026): EQTY Lab assess, OpenBox AI assess
5.4

Outcome verification

8
BluejayCekuraCovalHamming AIJudgment LabsRaindropResaroWayfound
2/6 partsno indicator
Companies, sources and readings
Independent verification
Placed by hand (from its own site, read 5 Oct 2026): Resaro assess
Continuous quality assurance
Placed by hand (from its own site, read 5 Oct 2026): Bluejay assess, Cekura assess, Coval assess, Hamming AI assess, Judgment Labs assess, Raindrop assess, Wayfound assess
5.5

Settlement and billing

5
Catena LabsPaidProRataSkyfireTollBit
3/6 partsno indicator
Companies, sources and readings
Agent payments
Placed by hand (from its own site, read 5 Oct 2026): Catena Labs assess, Skyfire assess
Outcome-based billing
Placed by hand (from its own site, read 5 Oct 2026): Paid assess
Metering
Placed by hand (review, 28 Sep 2026): ProRata assess, TollBit assess
5.6

Liability and risk transfer

4
AIUCArmilla AIMitigataMunich Re (aiSure)
4/6 partsno indicator
Companies, sources and readings
AI warranties
Placed by hand (from its own site, read 5 Oct 2026): Armilla AI assess
Insurance
AI Roll-up Nexus company directory, section “It Mssp”, read 28 Sep 2026: Mitigata assess
Indemnification
Placed by hand (from its own site, read 5 Oct 2026): Munich Re (aiSure) assess
Assurance-backed procurement
Placed by hand (from its own site, read 5 Oct 2026): AIUC assess
Layer 6

Deployment and organizational diffusion

5 categories · 81 companies

6.1

Technical implementation

30
Anthropic enterprise services venturePalantirThe OpenAI Deployment CompanyAirbyteAmpersand LabsArtiumAtlasBrain Co.CamadaCaritas Venture Co.CiridaeCodosDistylFractal+16
6/6 partsno indicator
Companies, sources and readings
Forward-deployed engineering
Placed by hand (review, 28 Sep 2026): Anthropic enterprise services venture assess, The OpenAI Deployment Company assess, Distyl assess, Turing assess
Placed by hand (forward-deployed engineers, review, 28 Sep 2026): Palantir assess
AI Rollups & AI-Native PE: 2026 Market Map, section “Transformation partner”, read 28 Sep 2026: Ampersand Labs assess, Artium assess, Brain Co. assess, Caritas Venture Co. assess, Ciridae assess, Codos assess, Headstart assess, Melior assess, Percepta assess, SaxeCap assess, Tenex assess, Wonderful assess
AI Roll-up Nexus company directory, section “Accounting”, read 28 Sep 2026: Atlas assess, Camada assess
Systems integration
Placed by hand (review, 28 Sep 2026): Fractal assess, Invisible Technologies assess, Tribe AI assess
AI Native Services Market Map v2.0, section “Consulting”, read 28 Sep 2026: realfast.ai assess
Legacy modernization
AI Native Services Market Map v2.0, section “Consulting”, read 28 Sep 2026: Mechanical Orchard assess
AI Roll-up Nexus company directory, section “It Msp”, read 28 Sep 2026: Turgon assess
Data readiness
Placed by hand (from its own site, read 5 Oct 2026): Shelf assess, Unstructured assess
Ontology creation
Placed by hand (the ontology behind Foundry and AIP, review, 28 Sep 2026): Palantir assess
Placed by hand (from its own site, read 9 Oct 2026): TextQL assess
Connectors
Placed by hand (from its own site, read 5 Oct 2026): Airbyte assess, Paragon assess
6.2

Context engineering

6
EdraMimicaSkan AISugarworkTesslUnblocked
4/6 partsno indicator
Companies, sources and readings
Domain-knowledge capture
Placed by hand (from its own site, read 5 Oct 2026): Edra assess, Skan AI assess
Workflow decomposition
Placed by hand (from its own site, read 5 Oct 2026): Mimica assess
Context curation
Placed by hand (from its own site, read 5 Oct 2026): Tessl assess, Unblocked assess
Tacit-knowledge elicitation
Placed by hand (from its own site, read 5 Oct 2026): Sugarwork assess
6.3

Managed AI operations

1
Provectus
1/5 partsno indicator
Companies, sources and readings
Model operations
Placed by hand (from its own site, read 5 Oct 2026): Provectus assess
6.4

Adoption and change

8
CodesmithCorrelation OneCredo AILarridinMindstoneProsciSectionTrustible
4/6 parts1 indicator
Companies, sources and readings
Change management
Placed by hand (from its own site, read 5 Oct 2026): Prosci assess
Internal training
AI education and talent companies (verified list), section “enterprise_upskilling”, read 28 Sep 2026: Codesmith assess, Correlation One assess, Mindstone assess, Section assess
ROI measurement
Placed by hand (from its own site, read 5 Oct 2026): Larridin assess
Governance and procurement
Placed by hand (from its own site, read 5 Oct 2026): Credo AI assess, Trustible assess

Readings: Enterprise pilots with measurable P&L impact

6.5

Talent production

40
ADaSciAggregate IntellectAI Builder ClubAI CampAI EngineerAI MakerspaceAI4ALLAndelaArtificial Intelligence UniversityBuild ClubCodebasicsCodesmithCorrelation OneCoursera+26
7/7 partsno indicator
Companies, sources and readings
FDE training
AI education and talent companies (verified list), section “fde_training”, read 28 Sep 2026: AI Makerspace assess, FDE Academy assess, Supervity assess
Applied MLE training
AI education and talent companies (verified list), section “bootcamp”, read 28 Sep 2026: Aggregate Intellect assess, AI Camp assess, Codebasics assess, Data Science Dojo assess, Interview Kickstart assess, TripleTen assess
AI education and talent companies (verified list), section “cohort_courses”, read 28 Sep 2026: Coursera assess, DeepLearning.AI assess, fast.ai assess, Krish Naik Academy assess, Lonely Octopus assess, Maven assess, Outskill assess, Parlance Labs assess, Product School assess, The Full Stack assess, Towards AI Academy assess, Udacity assess
Agent product management
Placed by hand (from its own site, read 5 Oct 2026): Product Faculty assess
Evaluation engineering
Placed by hand (from its own site, read 5 Oct 2026): Parlance Labs assess
Credentialing
AI education and talent companies (verified list), section “credentialing”, read 28 Sep 2026: ADaSci assess, Artificial Intelligence University assess, Workera assess
Placement and talent networks
AI education and talent companies (verified list), section “talent_placement”, read 28 Sep 2026: AI4ALL assess, Andela assess, Gauntlet AI assess
AI-native apprenticeship
AI education and talent companies (verified list), section “apprenticeship”, read 28 Sep 2026: FS Academy assess, Multiverse assess
Part not assigned
AI education and talent companies (verified list), section “builder_community”, read 28 Sep 2026: AI Builder Club assess, AI Engineer assess, Build Club assess, lablab.ai assess, Sundai Club assess, The AI Collective assess
AI education and talent companies (verified list), section “enterprise_upskilling”, read 28 Sep 2026: Codesmith assess, Correlation One assess, Mindstone assess, Section assess
Layer 7

Applications and operating models

7 categories · 302 companies

7.1

Horizontal copilots

26
LovablePerplexityVercelAaruAll Hands AIAnysphereAugmentCognitionContinue.devDustFactoryFireflies.aiGleanGranola+12
6/6 parts1 indicator
Companies, sources and readings
Coding
Placed by hand (review, 28 Sep 2026): Lovable assess, Vercel assess, Anysphere assess, Augment assess, Cognition assess, Factory assess, Replit assess, Sourcegraph assess, StackBlitz assess, Warp assess
The Inference Market Map, section “modality / code & agents”, read 28 Sep 2026: All Hands AI assess, Continue.dev assess
neolabs.fyi, section “Applied — Agents, Open source”, read 28 Sep 2026: Imbue assess
Research
neolabs.fyi, section “World models & simulation”, read 28 Sep 2026: Aaru assess
Writing
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “Shared enterprise context”, read 28 Sep 2026: Writer assess
Design
Placed by hand (from its own site, read 5 Oct 2026): Magic Patterns assess, Uizard assess
Productivity
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “App context capture”, read 28 Sep 2026: Fireflies.ai assess, Granola assess, Limitless (acquired: Meta, Dec 2025) assess, Otter.ai assess, Read AI assess
Search
Placed by hand (review, 28 Sep 2026): Perplexity assess, Glean assess
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “Shared enterprise context”, read 28 Sep 2026: Dust assess, Sana (acquired: Workday, closed Nov 2025) assess

Readings: Developer productivity uplift (METR RCTs)

Venture money for coding agents, as recorded by sub-layer.

7.2

Functional enterprise applications

20
ClayDecagon11xArtisanCX-AssistDocDraftForge AIGani AILevel3MagenticMandrelOpenCallParloaPrelude+6
8/8 partsno indicator
Companies, sources and readings
Sales
Placed by hand (review, 28 Sep 2026): Clay assess, 11x assess, Artisan assess
Customer support
Placed by hand (review, 28 Sep 2026): Decagon assess, Parloa assess, Sierra assess
AI Roll-up Nexus company directory, section “Contact Center”, read 28 Sep 2026: CX-Assist assess, OpenCall assess
AI Native Services Market Map v2.0, section “BPO”, read 28 Sep 2026: Level3 assess
Finance
AI Roll-up Nexus company directory, section “Accounting”, read 28 Sep 2026: Mandrel assess, Sequence assess, SimpleAI assess
HR
AI Roll-up Nexus company directory, section “Staffing”, read 28 Sep 2026: Forge AI assess, Starshunt assess
Procurement
AI Native Services Market Map v2.0, section “Logistics & Procurement”, read 28 Sep 2026: Magentic assess
Legal
AI Native Services Market Map v2.0, section “Legal Services”, read 28 Sep 2026: DocDraft assess, Gani AI assess
Security
Identity and Access Management (IAM) in the age of AI Agents, section “3. Agent Monitoring, Detection & Response”, read 28 Sep 2026: Prelude assess
Placed by hand (managed detection and response (decision, 28 Sep 2026): Tenex.ai assess
IT operations
AI Roll-up Nexus company directory, section “It Msp”, read 28 Sep 2026: Rollout assess
7.3

Vertical workflow software

58
AbridgeHarveyHebbiaLegora5c NetworkAccrualAforiAgent AstraAlixAlvenAmbienceAmperosAppliedAIAqua+44
10/10 parts3 indicators
Companies, sources and readings
Healthcare
Placed by hand (review, 28 Sep 2026): Abridge assess, Ambience assess, Hippocratic assess, OpenEvidence assess, Tennr assess
AI Roll-up Nexus company directory, section “Teleradiology”, read 28 Sep 2026: 5c Network assess, Natoe assess
AI Native Services Market Map v2.0, section “BPO”, read 28 Sep 2026: Amperos assess, Honey assess
AI Roll-up Nexus company directory, section “Medical Rcm”, read 28 Sep 2026: Candid Health assess, Overdrive Health assess
AI Roll-up Nexus company directory, section “Medical Health Record”, read 28 Sep 2026: Particle Health assess
AI Roll-up Nexus company directory, section “Elder Care”, read 28 Sep 2026: Tern assess
Legal
Placed by hand (review, 28 Sep 2026): Harvey assess, Legora assess, Eve assess, EvenUp assess, Spellbook assess
AI Roll-up Nexus company directory, section “Estate Planning”, read 28 Sep 2026: Alix assess
AI Roll-up Nexus company directory, section “Immigration Law”, read 28 Sep 2026: ImmCore assess
Financial services
Placed by hand (review, 28 Sep 2026): Hebbia assess, Brightwave assess, Rogo assess
AI Roll-up Nexus company directory, section “Accounting”, read 28 Sep 2026: Accrual assess, Alven assess, Contra assess
AI Roll-up Nexus company directory, section “Professional Services”, read 28 Sep 2026: AppliedAI assess
AI Roll-up Nexus company directory, section “Wealth Management”, read 28 Sep 2026: Aqua assess, Palindrome assess
AI Roll-up Nexus company directory, section “Fund Admin”, read 28 Sep 2026: Mav9 assess, Ylookup assess
AI Roll-up Nexus company directory, section “Banking Consultants”, read 28 Sep 2026: Moneylab Technologies assess
Insurance
AI Roll-up Nexus company directory, section “Insurance Broker/Agency”, read 28 Sep 2026: Afori assess, Clarence assess, Corgi assess, Quartz assess, Synapze assess
AI Roll-up Nexus company directory, section “Insurance Tpa”, read 28 Sep 2026: Ravin assess
Industrial
AI Roll-up Nexus company directory, section “Solar O&M”, read 28 Sep 2026: Areg assess
Roadmap: The AI data center stack, section “Software & Orchestration”, read 28 Sep 2026: Senpilot assess
Logistics
AI Roll-up Nexus company directory, section “Freight Forwarder”, read 28 Sep 2026: Agent Astra assess, Nuvocargo assess, Stracker assess, Xhipment assess
AI Roll-up Nexus company directory, section “Customs Broker”, read 28 Sep 2026: DocUnlock assess
AI Roll-up Nexus company directory, section “Freight Broker”, read 28 Sep 2026: Freight Hero assess, Sedona assess, Sennder assess
AI Roll-up Nexus company directory, section “Freight Audit And Payment”, read 28 Sep 2026: Loop assess, Pando assess
AI Native Services Market Map v2.0, section “Logistics & Procurement”, read 28 Sep 2026: Ten8 AI assess
Construction
AI Roll-up Nexus company directory, section “Engineering Services”, read 28 Sep 2026: Ashler assess
AI Roll-up Nexus company directory, section “Construction Takeoff”, read 28 Sep 2026: Attentive assess
Education
Placed by hand (from its own site, read 5 Oct 2026): Brisk Teaching assess, MagicSchool assess
Agriculture
Placed by hand (from its own site, read 5 Oct 2026): Cropin assess
Government and defence
Placed by hand (from its own site, read 5 Oct 2026): GovDash assess, Rune Technologies assess

Readings: Vertical versus horizontal application funding, FDA-authorised devices built on a large language model, Frontier models completing expert legal tasks end to end (Harvey LAB)

Venture money for vertical ai, as recorded by sub-layer.

7.4

Autonomous services

89
AdaptiveAIOSAkidoAlmaArcaArcos LawBalanceBeglaubigtBorderless AICartage AICedarCeedarClaimSortedClara+75
13/13 partsno indicator
Companies, sources and readings
AI accounting service
AI Native Services Market Map v2.0, section “Tax & Accounting”, read 28 Sep 2026: Balance assess, Finally (closed: wound down Sep 2026; bookkeeping division to Inkle) assess, Finaloop assess, Fondo assess, Gelt assess, Integral assess, Last Accounting Company assess, Mars assess, Neno assess, Oath assess, Pilot assess, Quanta assess, Rally assess, Repodo assess, Rivet assess, Skalar assess, Taxforce assess, Zeni assess
AI Roll-up Nexus company directory, section “Accounting”, read 28 Sep 2026: Ceedar assess, Minerva Intelligence assess, Numbi assess
AI Native Services Market Map v2.0, section “Banking & Finance”, read 28 Sep 2026: Embankment assess, Formulary assess, Hanover Park assess
AI Roll-up Nexus company directory, section “Audit”, read 28 Sep 2026: Modus assess
AI Native Services Market Map v2.0, section “BPO”, read 28 Sep 2026: Neoflo assess, Rex assess
AI law firm
AI Native Services Market Map v2.0, section “Legal Services”, read 28 Sep 2026: Alma assess, Arcos Law assess, Beglaubigt assess, Covenant assess, Crosby assess, Duely assess, Ellis assess, Garfield AI assess, General Legal assess, Keith assess, Lawhive assess, LegalOS assess, Manifest assess, Moritz assess, Norm assess, Soxton AI assess, Vector assess
AI customer-service operator
Placed by hand (review, 28 Sep 2026): Crescendo assess
AI Roll-up Nexus company directory, section “Contact Center”, read 28 Sep 2026: Kim assess, Onepilot assess
AI Native Services Market Map v2.0, section “BPO”, read 28 Sep 2026: Smith assess
Autonomous claims administration
AI Native Services Market Map v2.0, section “BPO”, read 28 Sep 2026: ClaimSorted assess
AI Native Services Market Map v2.0, section “Insurance”, read 28 Sep 2026: Elysian assess, Reserv assess, Strala assess
AI Roll-up Nexus company directory, section “Insurance Tpa”, read 28 Sep 2026: Inca assess
AI Native Services Market Map v2.0, section “Healthcare”, read 28 Sep 2026: Understood assess
AI-managed procurement
AI Native Services Market Map v2.0, section “Logistics & Procurement”, read 28 Sep 2026: Cartage AI assess, Saudara assess
AI research service
AI Native Services Market Map v2.0, section “Healthcare”, read 28 Sep 2026: Panacea assess
AI Native Services Market Map v2.0, section “Banking & Finance”, read 28 Sep 2026: Riplo assess, Solvea assess
AI insurance
AI Native Services Market Map v2.0, section “Insurance”, read 28 Sep 2026: Comeryx assess, Corgi assess, Harper assess, Hedge assess, Panta assess, Third Space assess, WithCoverage assess
AI recruiting and staffing agency
AI Native Services Market Map v2.0, section “HR”, read 28 Sep 2026: Borderless AI assess, Contrario assess, Dex assess, Jack & Jill assess, Prism assess
AI medical practice
AI Native Services Market Map v2.0, section “Healthcare”, read 28 Sep 2026: Adaptive assess, AIOS assess, Akido assess, Clara assess, Counsel Health assess, Doctronic assess
AI wealth manager
AI Native Services Market Map v2.0, section “Banking & Finance”, read 28 Sep 2026: Arca assess, Decade assess
AI marketing agency
AI Native Services Market Map v2.0, section “Marketing”, read 28 Sep 2026: DayDream assess, GrowthX assess
AI design and engineering firm
AI Native Services Market Map v2.0, section “Real Estate”, read 28 Sep 2026: Cedar assess, Cove assess
AI Native Services Market Map v2.0, section “Industrials”, read 28 Sep 2026: Unlimited Industries assess
AI M&A advisory
AI Native Services Market Map v2.0, section “Banking & Finance”, read 28 Sep 2026: Masouken assess, OffDeal assess
Part not assigned
AI Native Services Market Map v2.0, section “Healthcare”, read 28 Sep 2026: Lindus (closed: CRO assets sold to Curavit, Aug 2026 (reported)) assess
AI Native Services Market Map v2.0, section “Real Estate”, read 28 Sep 2026: Sebas assess
Placed by hand (decision, 28 Sep 2026): Tenex.ai assess
7.5

Sensor–model–actuator networks

Robotics and autonomy are out of scope for now (28 Sep 2026)

0/7 parts1 indicator
Companies, sources and readings

Readings: Waymo paid rides per week

7.6

AI-native rollups

106
2XAdaptive InnovationsAgent AIAgent VistaAIAI Holdings (Ai²)AIXIOMAlloy RiverAmerican Growth InsuranceAncoraAntares LabsAntidoteArbio GroupArchipelArchvale+92
5/5 partsno indicator
Companies, sources and readings
Professional-services rollups
AI Roll-up Nexus company directory, section “Marketing”, read 28 Sep 2026: 2X assess
AI Roll-up Nexus company directory, section “Home Healthcare”, read 28 Sep 2026: Adaptive Innovations assess
AI Roll-up Nexus company directory, section “Medical Rcm”, read 28 Sep 2026: Agent AI assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “agent-vista.com”, read 28 Sep 2026: Agent Vista assess
AI Roll-up Nexus company directory, section “Fractional Cfo”, read 28 Sep 2026: Alloy River assess
AI Roll-up Nexus company directory, section “Insurance Broker/Agency”, read 28 Sep 2026: American Growth Insurance assess, Covu assess, Equal Parts assess, Gyde assess
AI Roll-up Nexus company directory, section “Accounting”, read 28 Sep 2026: Ancora assess, Archipel assess, Asevia assess, Aurion assess, Bhub assess, Cynch assess, General Intelligence assess, Multiplier Holdings assess, Numica assess, Skrooge assess
AI Roll-up Nexus company directory, section “Multiple”, read 28 Sep 2026: Antidote assess, Long Lake assess, Sequence Holdings assess, SKROL assess, Timah Partners assess
AI Roll-up Nexus company directory, section “Vacation Rental Management”, read 28 Sep 2026: Arbio Group assess
AI Roll-up Nexus company directory, section “Medical Gp”, read 28 Sep 2026: Archvale assess, Avi medical assess, Jutro Medical assess, Praxis Eins assess
AI Roll-up Nexus company directory, section “Property Management”, read 28 Sep 2026: Arjile assess, Buena assess, Levels assess, Lightswitch assess, MoIn assess, Onekey assess, Ralph assess, Sebas assess, Theo assess
AI Roll-up Nexus company directory, section “Customs Broker”, read 28 Sep 2026: Border Logic assess
AI Roll-up Nexus company directory, section “Defense”, read 28 Sep 2026: Bridge Defense assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “castfinance.com”, read 28 Sep 2026: Cast Finance assess
Placed by hand (review, 28 Sep 2026): Crete assess, Dwelly assess, Thrive Holdings assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “dragonfly.com.au”, read 28 Sep 2026: Dragonfly assess
AI Roll-up Nexus company directory, section “Worker Productivity”, read 28 Sep 2026: Enam assess
AI Roll-up Nexus company directory, section “Legal Services”, read 28 Sep 2026: Eudia assess, Lawhive assess
AI Roll-up Nexus company directory, section “Freight Broker”, read 28 Sep 2026: Fura assess, TerraFreight assess
AI Roll-up Nexus company directory, section “Esa Consulting”, read 28 Sep 2026: Gauge Industrial Group assess
AI Roll-up Nexus company directory, section “Travel Agency”, read 28 Sep 2026: Gazelle Travel assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “generationtech.partners”, read 28 Sep 2026: Generation Tech Partners assess
AI Roll-up Nexus company directory, section “Tax Advisory”, read 28 Sep 2026: Integral assess, LimeTax assess, OLarry assess, Ready Plan Go assess
AI Roll-up Nexus company directory, section “Clinical Research”, read 28 Sep 2026: Iterative Health assess
AI Roll-up Nexus company directory, section “Real Estate Lettings”, read 28 Sep 2026: Kalea assess, Stead assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “legionholdings.ai”, read 28 Sep 2026: Legion Holdings assess
AI Roll-up Nexus company directory, section “Staffing”, read 28 Sep 2026: Mojo Talent assess
AI Roll-up Nexus company directory, section “Elder Care”, read 28 Sep 2026: Nila Care assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “operoncap.com”, read 28 Sep 2026: Operon Capital assess
AI Roll-up Nexus company directory, section “Freight Audit And Payment”, read 28 Sep 2026: PayPath assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “risingtidehq.com”, read 28 Sep 2026: Rising Tide assess
AI Native Services Market Map v2.0, section “Banking & Finance”, read 28 Sep 2026: Savvy assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “squeezegroup.com”, read 28 Sep 2026: Squeeze assess
AI Roll-up Nexus company directory, section “Wealth Management”, read 28 Sep 2026: Thrivo Wealth assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “unitedaccountants.com”, read 28 Sep 2026: United Accountants assess
AI Roll-up Nexus company directory, section “Compliance Services”, read 28 Sep 2026: Venmark assess
AI Roll-up Nexus company directory, section “Freight Forwarder”, read 28 Sep 2026: Voyfai assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “xgxventures.com”, read 28 Sep 2026: XGX Ventures assess
AI Roll-up Nexus company directory, section “Construction Services”, read 28 Sep 2026: Zero RFI assess
Vertical managed-service rollups
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “aixiom.io”, read 28 Sep 2026: AIXIOM assess
AI Roll-up Nexus company directory, section “It Msp”, read 28 Sep 2026: Aretea Group assess, e4n assess, Monkinet assess, Treeline assess, Unitas assess
AI Roll-up Nexus company directory, section “Contact Center”, read 28 Sep 2026: Crescendo assess, UnifyCX assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “ekohe.fund”, read 28 Sep 2026: Ekohe Fund assess
Placed by hand (review, 28 Sep 2026): Long Lake assess, Titan assess
AI Native Services Market Map v2.0, section “BPO”, read 28 Sep 2026: Procode assess
AI Roll-up Nexus company directory, section “It Services”, read 28 Sep 2026: Shield Technology Partners assess
Physical-asset rollups
AI Roll-up Nexus company directory, section “Manufacturing”, read 28 Sep 2026: Antares Labs assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “thermalhomeservices.com”, read 28 Sep 2026: Calexo (Thermal Home Services) assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “cscgeneration.com”, read 28 Sep 2026: CSC Generation assess
AI Roll-up Nexus company directory, section “Retail”, read 28 Sep 2026: Felz assess
AI Roll-up Nexus company directory, section “Restoration Services”, read 28 Sep 2026: Longview assess
AI Roll-up Nexus company directory, section “Medical Gp”, read 28 Sep 2026: Meroka assess
AI Roll-up Nexus company directory, section “Parking”, read 28 Sep 2026: Metropolis assess
AI Roll-up Nexus company directory, section “Freight Forwarder”, read 28 Sep 2026: NexDash assess
AI Roll-up Nexus company directory, section “Roofing Services”, read 28 Sep 2026: Roofer assess
Software consolidators
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “banyansoftware.com”, read 28 Sep 2026: Banyan Software assess
AI Rollups & AI-Native PE: 2026 Market Map, section “Owner-operator”, read 28 Sep 2026: Beacon Software assess, Circeus assess
AI Rollups & AI-Native PE: 2026 Market Map, section “Rollup platform”, read 28 Sep 2026: Bending Spoons assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “mitrim.io”, read 28 Sep 2026: Mitrim assess
AI Roll-up Nexus company directory, section “World Domination”, read 28 Sep 2026: Rocketable assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “stelao.com”, read 28 Sep 2026: Stelao assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “topwater.ai”, read 28 Sep 2026: Topwater Group assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “zax.ai”, read 28 Sep 2026: Zax assess
Data-acquisition rollups
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “aiaiholdings.com”, read 28 Sep 2026: AIAI Holdings (Ai²) assess
AI-native acquirers: owner-suggested (Dragonfly) plus research sweep, section “orbitholdings.ai”, read 28 Sep 2026: Orbit Holdings assess
7.7

Consumer AI

12
HarkInstinctPerplexityCharacter.AICounsel HealthDoctronicDuckbillHiggsfieldOpenClawReplikaSpeakThe Travel Company
7/7 partsno indicator
Companies, sources and readings
General assistants
Placed by hand (review, 28 Sep 2026): Perplexity assess
AI Native Services Market Map v2.0, section “HR”, read 28 Sep 2026: Duckbill assess
Companions
Placed by hand (review, 28 Sep 2026): Character.AI assess, Replika assess
Education
Placed by hand (from its own site, read 5 Oct 2026): Speak assess
Health
Placed by hand (from its own site, read 5 Oct 2026): Counsel Health assess, Doctronic assess
Commerce and purchasing agents
AI Roll-up Nexus company directory, section “Travel Agency”, read 28 Sep 2026: The Travel Company assess
Creative tools
The Inference Market Map, section “modality / image”, read 28 Sep 2026: Higgsfield assess
Personal operating systems
Placed by its sub-layer (neolabs): Hark assess
Who Owns Memory? The Market Map for AI Agents' Missing State Layer, section “Memory-native startups”, read 28 Sep 2026: OpenClaw assess
Part not assigned
Placed by its sub-layer (consumer ai): Instinct assess
Layer 8

Adoption and economic outcomes

readings only; no companies

Consumer surplus from generative AI (willingness to accept)Application layer's share of enterprise generative-AI spend (Menlo)Circular financing scaleStartups' share of enterprise AI application revenueApplication-company revenue run-rates50%/80% horizon ratioUS labour productivity, year on yearUS total factor productivity, private nonfarm businessShare of work hours assisted by generative AIUS firms using AI (Census BTOS)US businesses paying for AI (Ramp AI Index)Agent-workdays per human workday in frontier research (self-reported)Human interventions on 4–8 hour agent tasks (self-reported)Safety brakes inside the labs (events, trailing year)Public text data exhaustion year (Epoch projection)FDA-authorised AI-enabled devices, growthUS state AI bills introduced (legislative year)Employment in AI-exposed occupations, year on yearEntry-level employment shortfall in AI-exposed occupationsExposed vs unexposed occupation employment gap (Revelio)California unemployment claims from AI-exposed occupationsRecent-graduate unemployment rate (NY Fed)Cross-tracker concordance (labour)Augmentation vs automation share (Anthropic Economic Index)US labour share of income, year on yearCustomer-support productivity uplift from a generative-AI assistantExecutives reporting no AI impact on their own firmProgress against the AI 2027 quantitative predictionsWork-related share of ChatGPT consumer messagesUsage as a share of theoretical exposureWorkers highly exposed to AI with low adaptive capacityOccupational-mix dissimilarity since ChatGPT (Yale Budget Lab)Chatbot effect on earnings and hours in exposed occupations (Denmark)Freelance writing earnings after ChatGPT (Upwork)Wage premium for AI skills (PwC AI Jobs Barometer)US adults who have used generative AIOccupations with a quarter of their tasks using ClaudeUS workers using generative AI for work every day

Placements come from each source's own section labels and were checked by hand; companies whose work is robotics or autonomy are off the map for now.