Skip to content
Slow Variables
Ask

Capture / layer 4 · The value chain

Serving & orchestration

The businesses that run models for others, route requests between them, and supply the plumbing for AI agents. Dependency tier 3.

How the economics move

The site's judgement of each part of this layer, read against the readings below and the theory on the value chain page: what is getting cheap and interchangeable, what stays scarce, and what would prove it wrong.

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
Commoditising

Serving a published model is a price-taking business on rented chips; the serving firms are escaping into running each customer's own tuned model, and the chip and cloud owners are buying them

Inference is running a trained model to answer requests. Serving firms such as Fireworks, Baseten, Together, Modal, Replicate, RunPod, DeepInfra and Hyperbolic rent or buy chips, load open-weight models (whose trained settings, the weights, are published) or a customer's own model, and sell the answers by the token (a fragment of a word) or by the hour of chip time. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. This site judges that serving a published model is close to a commodity, since many hosts offer the same model and compete on price towards the cost of chips and power; what a host adds is software that makes each chip answer faster and cheaper, which rivals learn.

Getting cheap and interchangeable
This site judges that an answer from a published model is becoming a commodity: the same good from many hosts, with the clouds, the neoclouds (newer cloud firms built to rent out AI chips) and the chip makers all offering it. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers as lower prices. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called the inputs a product needs to reach its customers complementary assets; the chips and powered buildings a serving firm needs are owned mostly by others, which is why their owners can absorb the software on top.
Staying scarce
This site judges that what stays scarce is the engineering to run each new model fast on each new chip, and, for hosts that tune and run a customer's own model, the position of holding that model. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier. The likeliest here is scale economies (pooling many customers' uneven demand keeps shared chips busy, and a rival must match the volume to match the cost), but the chip and cloud owners can pool as well, so this site does not count it proven.
What would make it last
A host keeps its customers' tuned models, and the loop of data that improves them, so moving to a rival means rebuilding the model: switching costs in Helmer's sense (what a customer would lose by moving to a rival). That route holds only if a tuned model keeps beating each new general model. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Oliver Williamson, an economist who studied contracts and found that a buyer that depends on an asset built for it can be squeezed by whoever owns it, would otherwise expect the owners of chips and clouds to keep buying the serving software, as they are.
What would prove this wrong
Hosts of the same open model keeping different prices for long, with buyers staying with the dearer one; or serving firms keeping or widening their margins (what they keep from each sale after paying their costs) while chip rents fall and more hosts offer the same models.
Binds on (how tight)
Running models (easing), AI chips (moderate), Data-centre buildings (tight)
Commoditising

Choosing a model is being given away; a gateway keeps a rent only as the place where developers pay for many models

A router or gateway sits between an app and the models: it sends each request to a model chosen for price, speed or quality, switches when one fails, and keeps one bill. OpenRouter, which the payments company Stripe agreed in August 2026 to buy, sells access to many models through one account; Martian and Not Diamond sell the choice of model; Portkey, Helicone, LiteLLM and TrueFoundry sell a gateway a firm runs itself, LiteLLM's as open source (code published for anyone to use and change for free). David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. This site judges that the routing itself is easy to copy, while a marketplace that many developers and many model sellers both use can hold them the way a payments network does.

Getting cheap and interchangeable
This site judges that picking a model for each request is becoming a commodity: open-source gateways do it, the clouds' catalogues do it, and the labs' own tools do it for their models. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers.
Staying scarce
This site judges that what stays scarce is the marketplace itself: one account, one bill and one set of terms across many model sellers, and the record of what traffic went where. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier. Network economies (each user makes the product more valuable to the others) would fit a marketplace, but developers use several routers or go straight to the labs, so this site does not count the barrier proven.
What would make it last
Pooling many firms' checked outcomes lets a router choose better than any firm alone, a network economy in Helmer's sense; or, inside Stripe, the marketplace becomes where AI agents themselves pay for models, so billing and credit create switching costs (what a customer would lose by moving to a rival).
What would prove this wrong
Routers charging more for the choice of model itself, and buyers paying it, because choices learned from many firms' outcomes beat what each firm manages alone.
Binds on (how tight)
Routing between models (no reading; a model judges it probably slack), Running models (easing)
The rival view
Routers gain power
Commoditising

Agent plumbing is mostly open source that labs and clouds buy or copy; the rent sits with whoever holds the connections into customers' systems

An agent is a model that carries out a task by using software tools over minutes or hours. Agent infrastructure is what it needs, much of it open source (code published for anyone to use and change for free): frameworks for wiring model calls together (LangChain, CrewAI, Mastra), sandboxes where it can run code or a browser safely (Daytona, Browserbase and others), memory that lasts between sessions (Letta), and connectors that let it log in to and act in other software (Composio, Arcade, Nango, Merge). David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. Oliver Williamson, an economist who studied contracts, found that an asset built for one buyer, worth far less anywhere else, ties that buyer to whoever holds it. This site judges that frameworks and sandboxes are generic, while connectors that hold a firm's permissions and links into its systems are specific to that firm, and harder to replace.

Getting cheap and interchangeable
This site judges that frameworks, sandboxes and simple memory are becoming commodities: most began as open source, the clouds bundle them, and the labs buy the teams that build them. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called the inputs a product needs to reach its customers complementary assets; a generic one, which anyone can supply, earns an ordinary return, not a rent.
Staying scarce
This site judges that what stays scarce is the working link into a firm's own software: the permissions, the audit trail and the map of which system holds what. That context helps only the firm it came from, so it does not compound across customers, but it is specific in Williamson's sense. The software that keeps a firm's records can also refuse or charge for outside agents, which is where the rival view places the rent. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier; switching costs (what a customer would lose by moving to a rival) are the likeliest here, but this site does not count them proven for any connector firm yet.
What would make it last
A connector firm becomes the standard way agents are given permission into business software across many firms, which could give it network economies in Helmer's sense (each user makes the product more valuable to the others), or its links into each customer's systems become switching costs. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, would otherwise expect the labs and clouds to keep absorbing this layer, as the buying so far suggests.
What would prove this wrong
An open-source framework or sandbox firm earning a lasting margin (what it keeps from each sale after paying its costs) while the labs and clouds give the same thing away; or connector firms losing customers as fast as framework firms once the labs and clouds ship their own connectors.
Binds on (how tight)
Agents and tool use (easing), Deployment platforms (no reading; a model judges it probably easing)

The companies that matter here

Candidate to be worth a trillion dollars

Baseten

Show the profile
Market
Baseten sells inference (running a trained model to answer requests) and post-training (further training a finished model on a customer's own data and tasks) to companies building AI products. It runs models for customers on dedicated capacity or through ready-made model connections, in its cloud, in the customer's own or across both, and its engineers embed with customers to post-train and tune specialised models. Its case is that open-weight models (models whose trained settings are published so anyone can run them) are now good enough to replace closed models for many firms, which then want to own a model shaped by their own data. It says its revenue and the volume it serves both grew steeply over the past year; that is self-reported. This site's other readings, from a press report and an analyst's estimate, also show a steep rise in early 2026. Its Form D (the notice a private company files with the US regulator when it sells shares without a public offering) shows an offering of shares first sold in June 2026, mostly sold at filing, with no placement agent and revenue not disclosed. How big the market becomes depends on whether serving open models stays a skilled service or becomes a commodity (a product bought on price alone).
Customers (moderate)
Fast-growing AI application companies. It names Cursor, Notion, Lovable, Harvey, HubSpot, OpenEvidence, Abridge, Decagon and Parallel. SpaceX, which owns xAI, a frontier lab (a firm building the most capable AI systems), agreed in June 2026 to buy Cursor; the deal was awaiting regulatory approval at SpaceX's quarterly filing for the middle of 2026. The sources give no revenue by customer; this site judges a few large application companies weigh heavily.
Powers
  • switching costs: Emerging, not proven. Hamilton Helmer, a strategist, listed seven sources of lasting profit and required each to have both a benefit and a barrier to copying. The benefit is repeat business from customers whose post-trained models and tuned serving run on Baseten. The barrier is weak: Baseten's own pitch is that customers own their models, and open-weight models run on many rivals' platforms. This site has no retention evidence.
Under the site's rent rule
competed away to users. Depends on: AI chips (moderate), Data-centre buildings (tight), Running models (easing), Post-training (no reading; a model judges it probably moderate), Capital (slack).

What would have to hold by 2035

  • Serving open models does not become a commodity. This site's records show open models carrying most of the tokens (the units of text models read and write) routed through OpenRouter (a service that passes each request to whichever model a buyer picks) to its busiest models, the price of a fixed level of capability falling fast, and many rival serving firms, so Baseten has to earn its price from post-training and reliability, not raw serving.
  • Post-training ties customers to it. Oliver Williamson, an economist who studied contracts, found that firms bring in-house the work tied closely to their own plans; a model tuned to one firm's work is worth little elsewhere, but the switching cost falls to whoever holds it, and Baseten itself says its customers want to own their models.
  • It keeps getting enough GPUs (graphics processing units, the chips most AI runs on) at a price below what it charges; this site judges it rents most of them, largely NVIDIA's, from clouds that also sell serving.
  • Its largest customers stay with it as they grow, rather than moving to a lab's own models or their own data centres; SpaceX's agreement in June 2026 to buy Cursor, still awaiting regulatory approval at SpaceX's quarterly filing for the middle of 2026, is the first test.
  • Chip makers and compute owners, which this site records buying several serving firms, do not bundle serving so cheaply that stand-alone firms lose the work.

What would show it failing

  • Price cuts on serving open models without a matching rise in volume.
  • Cursor or another named customer moving its models off Baseten after a change of owner.
  • Customers taking their post-trained models to rival platforms at little cost.
  • Clouds or chip makers offering the same post-training and serving as part of their compute.

Drafted by claude-opus-5.5 from Baseten Form D, 30 Jun 2026 (opens in a new tab), Baseten announcement, 22 Jun 2026 (opens in a new tab), SpaceX quarterly report (10-Q), 4 Aug 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Fireworks

Show the profile
Market
Fireworks sells inference (running a trained model to answer requests) for companies' developer teams, plus tools to fine-tune models (further train them on a customer's own data), train them with reinforcement learning (rewarding good answers) and evaluate them. It serves hundreds of open-weight models (models whose trained settings are published so anyone can run them). Its founders, who had built AI infrastructure at Meta, left PyTorch, an AI software framework, in 2022 on the bet that open models would catch up with closed ones and firms would want to own their AI. It calls itself the market leader in AI inference; that is its own description. It says its annualised revenue run-rate (the latest month's revenue scaled up to a year) passed a large milestone in 2026, and that almost all the tokens (units of text a model reads and writes) it serves come from models specialised on customers' own data; both are self-reported. An analyst's estimate in this site's records also shows a steep rise through early 2026. NVIDIA and AMD, the chip makers, are among its investors. How big the market becomes depends on whether serving open models stays a skilled service or becomes a commodity (a product bought on price alone).
Customers (moderate)
It names Samsung, Uber, DoorDash, Notion, Shopify and Upwork, and says Cursor's coding models and Harvey's legal AI are built with it; Harvey, in a post in this site's records, says it post-trained a model with Fireworks, starting from an open model. It says it serves many thousands of companies. SpaceX, which owns xAI, a frontier lab (a firm building the most capable AI systems), agreed in June 2026 to buy Cursor; the deal was awaiting regulatory approval at SpaceX's quarterly filing for the middle of 2026. The sources give no revenue by customer; this site judges the base is broad but that a few large application companies weigh heavily.
Powers
  • switching costs: Emerging, not proven. Hamilton Helmer, a strategist, listed seven sources of lasting profit and required each to have both a benefit and a barrier to copying. The benefit is repeat business from models tuned on a customer's data and improved by its users' feedback on Fireworks' stack. The barrier is weak: Fireworks' own pitch is that customers own their models, and open-weight models run on many rivals' platforms. This site has no retention evidence.
Under the site's rent rule
competed away to users. Depends on: AI chips (moderate), Data-centre buildings (tight), Running models (easing), Post-training (no reading; a model judges it probably moderate), Capital (slack).

What would have to hold by 2035

  • Serving open models does not become a commodity. This site's records show open models carrying most of the tokens routed through OpenRouter (a service that passes each request to whichever model a buyer picks) to its busiest models, the price of a fixed level of capability falling fast, and many rival serving firms, so Fireworks has to earn its price from tuning and speed, not raw serving.
  • Its feedback loop, where users' corrections retrain each customer's model, ties customers to it. On the tests of Andrei Hagiu and Julian Wright, economists who set out tests for when data really keeps rivals out, among them whether what a firm learns from one customer improves the product for the others, that learning helps only the customer it came from, so it is a switching cost for whoever holds the model rather than a network that grows with every customer.
  • It keeps getting enough GPUs (graphics processing units, the chips most AI runs on) at a price below what it charges; its investors include the chip makers whose chips this site judges it runs on.
  • Its largest customers stay with it as they grow, rather than moving to a lab's own models or their own data centres; SpaceX's agreement in June 2026 to buy Cursor, still awaiting regulatory approval at SpaceX's quarterly filing for the middle of 2026, is the first test.
  • Chip makers and compute owners, which this site records buying several serving firms, do not bundle serving so cheaply that stand-alone firms lose the work.

What would show it failing

  • Price cuts on serving open models without a matching rise in volume.
  • Cursor or another named customer moving its models off Fireworks after a change of owner.
  • Customers taking their tuned models to rival platforms at little cost.
  • Closed models winning back most of the traffic from open ones, shrinking the market it serves.

Drafted by claude-opus-5.5 from Fireworks announcement, 28 Oct 2025 (opens in a new tab), Fireworks announcement, 15 Jul 2026 (opens in a new tab), SpaceX quarterly report (10-Q), 4 Aug 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Already worth over a trillion dollars

Apple

Show the profile
Market
Apple designs and sells the iPhone, Mac, iPad, wearables and home devices, and sells services on them: the App Store, cloud storage, subscriptions, advertising, AppleCare and payments. Its 10-K (the annual report US companies file with their regulator), for the year to September 2025, says it designs nearly the entire solution for its products, from hardware and operating system to apps and services, and relies on outsourcing partners, mostly in Asia, to build them. The 10-K mentions artificial intelligence mainly among its risks: safety, copyright and regulation. Apple's machine-learning research site describes the models behind Apple Intelligence, its AI features: a small model that runs on the device itself, tuned for the chips Apple designs (Apple silicon), and a larger model that runs on Private Cloud Compute, Apple's own server platform, which it presents as part of its commitment to users' privacy. A software framework lets app developers build on these models. Apple says both models match or beat open models of similar size (models whose trained settings are published so anyone can run them); that is its own test, and a comparison with models of similar size, not with the frontier models (the most capable AI systems, from firms such as OpenAI and Google). The market map this site uses places Apple in edge inference: running a trained model on the user's own device rather than in a data centre. This site judges that Apple's AI market is not selling AI itself but keeping people on its devices and in its services: the 10-K shows its Services gross margin (the share of sales left after the direct cost of providing them) far above its Products gross margin, and rising again in 2025. How big this becomes depends on whether AI makes people buy new devices and stay in its services, or moves their daily habit to assistants that live in other companies' apps.
Customers (diffuse)
People and businesses that buy its devices, directly and through indirect channels such as mobile carriers and other resellers; the 10-K says most of its net sales in 2025 went through indirect channels, and that its carriers accounted for a large share of its trade receivables (money customers owe it). The United States is its largest market; its sales in Greater China (China mainland, Hong Kong and Taiwan) fell in 2025. It also earns revenue from licensing arrangements with Google and other companies to offer their search services on its devices, and a commission on some purchases made through its App Store. The 10-K names no single customer as a large share of its net sales; the only single-customer exposure it discloses is in trade receivables. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called the inputs a product needs to reach its buyers complementary assets, and found that whoever holds them often keeps the profit; this site judges that Apple holds its own: the devices, the chips it designs, the App Store and the habits of the people who carry them.
Powers
  • network economies: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, required each lasting advantage to have both a benefit and a barrier to copying. The 10-K describes the loop: customers' decisions to buy its hardware depend in part on third-party apps, and developers weigh Apple's market position and expected sales when deciding whether to build for it. It does not show the barrier: it says its minority share of the global smartphone, computer, tablet and wearables markets can push developers toward rivals with larger shares, and that some rivals have large bases of active devices of their own. This site judges that the benefit shows in a Services gross margin far above its Products margin, but the filing does not tie that margin to the loop. Under the site's written rules that makes Apple's rent (profit above what it would earn in a fully competitive market) a scarcity rent, one that lasts while an input is short or a firm is first, and its durability short until a filing shows the barrier. This site finds no evidence that the loop extends to AI features, where one person's use does not improve another's.
  • switching costs: Emerging, not proven. This site judges that an assistant which knows a user's messages, photos and habits on the device could make leaving costly. The barrier is untested: the 10-K reports no AI revenue or use, and assistants from other firms can run on the same phone.
Under the site's rent rule
a thin profit, kept by the company itself. Depends on: Chip foundries (moderate), Memory (no reading), Running models (easing), Applications (no reading; a model judges it probably slack), Deployment platforms (no reading; a model judges it probably easing).

What would have to hold by 2035

  • AI makes people keep buying and replacing its devices and keeps them in its services; the 10-K says competitors imitate its features and cut prices aggressively, and that Apple holds a minority share of the global markets for smartphones, personal computers, tablets and wearables.
  • Running AI on the device stays good enough for everyday requests, so that its own chips, rather than rented data-centre chips, carry most of the work; Apple compares its small model with open models of similar size, not with the frontier models.
  • When a request needs a stronger model than its own, Apple keeps the user inside its own interface and on its own terms, rather than users opening ChatGPT, Google's Gemini or another lab's app directly.
  • Its revenue from letting Google and others offer search on its devices survives the remedies a US court ordered against Google in September 2025; the 10-K says that order may still change in further proceedings or on appeal, and that a reversal on appeal could bring remedies barring Google from offering Apple commercial terms for search placement. This site judges it must also survive any shift of people's questions from search to AI assistants.
  • Courts and regulators leave it room to charge on its App Store; the 10-K describes a US court order barring it from charging a commission on certain purchases and a fine from the European Commission in April 2025 under the EU's Digital Markets Act. This site judges that if AI agents (assistants that act for their users, booking and buying) become where people pay, Apple keeps a share only while those purchases still pass through its devices and store.
  • Its suppliers keep delivering; the 10-K says it relies on single-source partners for many components, some of them custom parts available from only one source, and on partners in Asia for final assembly of substantially all its hardware.

What would show it failing

  • Its Services gross margin falling for several years in a row.
  • Device sales falling, or people keeping their phones longer, after its AI features launch.
  • Its own models falling so far behind that it hands most AI requests to another firm's model, with that firm's brand shown to the user and much of the value paid to it.
  • Appeals ending with Google barred from offering Apple commercial terms for search placement, or search revenue shrinking as people ask AI assistants instead.
  • Regulators forcing it to allow other app stores and payment systems widely enough that its store commission falls.

Drafted by claude-opus-5.5 from Apple annual report (10-K), 31 Oct 2025 (opens in a new tab), Apple announcement, 17 Jul 2025 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Databricks

Show the profile
Market
Databricks, funded by private investors, calls itself the Data and AI company. It sells a platform, run on the big clouds of Amazon, Microsoft and Google, where companies store, clean and analyse their data and build AI on it. Its parts include the Lakehouse, its data-warehousing product (a database built for analysis); Unity Catalog, which controls who may see and use each piece of data; and a growing set of AI products. Among them are Agent Bricks, for building AI agents (software that carries out tasks rather than only answering), which it says measures how well models do on a customer's own task and then tunes them; Lakebase, a database it says is built for agents; Genie, which it calls an AI coworker that answers questions from a company's data; and Unity AI Gateway, which sends requests to different models and controls what they cost. In September 2025 it agreed a multi-year partnership with OpenAI to offer OpenAI's models, including its flagship model of the time, natively inside its platform, and it says it was among the first to host gpt-oss, OpenAI's open-weight model (one whose trained settings are published so anyone can run it). In August 2026 it said its revenue run-rate (the latest period's revenue scaled up to a year) had passed a large milestone after growing sharply, that its adjusted free cash flow (cash left after running costs and investment, on its own adjusted measure) was positive over the past year, and that it had closed a large strategic funding round led by Coatue; all of this is self-reported. The market map this site uses places it in inference serving (running trained models to answer requests) through computing clusters that the clouds manage. This site judges that serving is one part of a business that is mainly a data platform, and that its agent products place it equally in agent infrastructure. How big the market becomes depends on whether companies build their AI where their data already sits.
Customers (diffuse)
Companies of every size. It says many thousands of organisations use it, including adidas, AT&T, Bayer, Block, Mastercard, Rivian and Unilever and most of the largest US companies, and that many customers each spend with it at a high annual rate; these are its own counts. OpenAI is a customer as well as a partner: the partnership release says OpenAI uses Databricks to process AI data for ChatGPT. The sources give no revenue by customer; this site judges the base is broad. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called the inputs a product needs to reach its buyers complementary assets, and found that whoever holds them often keeps the profit. This site judges that the one that matters here is the governed store of each company's data, and the permissions over it, which Databricks holds for its customers; the chips and data centres it runs on are generic and rented from the clouds.
Powers
  • switching costs: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, required each lasting advantage to have both a benefit and a barrier to copying. The benefit is repeat and growing spending from companies whose data pipelines, permissions and now agents are built on its platform; its self-reported growth and count of large-spending customers point that way. The barrier is untested: the sources give no figures on customers staying or leaving, its own site lists open-source technologies behind its platform, and the clouds it runs on sell data and agent platforms of their own.
Under the site's rent rule
a thin profit, kept by the company itself. Depends on: Running models (easing), Agents and tool use (easing), Routing between models (no reading; a model judges it probably slack), Deployment platforms (no reading; a model judges it probably easing), Data-centre buildings (tight).

What would have to hold by 2035

  • Companies build their AI agents where their data already sits rather than inside the clouds' own platforms or the labs' products; Microsoft's 10-K describes Azure AI Foundry, and Alphabet's describes Vertex AI, as platforms for building agents, sold alongside the computing Databricks rents from them.
  • The clouds keep hosting it on fair terms while selling rival platforms. Oliver Williamson, an economist who studied contracts, found that a buyer that depends on an asset built for it can be squeezed by whoever owns it; this site judges Databricks depends on the clouds' computing, but that computing is generic and sold by several clouds, which limits the squeeze.
  • Models stay interchangeable enough that companies want a neutral place to choose and govern them; its AI gateway and its hosting of OpenAI's models, including the open-weight gpt-oss, rest on that. This site judges that if a single lab's models pulled far ahead and that lab sold directly to companies, the neutral layer would matter less.
  • Serving models inside its platform earns more than a commodity price (a price set by competition alone); this site judges that serving a published model is close to a commodity, so its margin must come from governance and closeness to the data.
  • OpenAI keeps its models available natively on Databricks, rather than steering companies to its own enterprise products.
  • Its self-reported growth and cash flow hold up in audited accounts, which this site does not yet hold.

What would show it failing

  • Large customers moving their agent work to a cloud's own agent platform or a lab's enterprise product while keeping only their data on Databricks.
  • Its reported growth slowing sharply while the clouds' own AI platforms grow faster.
  • OpenAI ending or narrowing the partnership.
  • Price cuts on its model serving without a matching rise in use.
  • Audited accounts, once published, showing large customers leaving or cutting their spending.

Drafted by claude-opus-5.5 from Databricks announcement, 13 Aug 2026 (opens in a new tab), Databricks announcement, 25 Sep 2025 (opens in a new tab), Alphabet Inc. annual report (10-K), 5 Feb 2026 (opens in a new tab), Microsoft Corp annual report (10-K), 29 Jul 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Indicators

Inference-layer revenue run-rates
emerging
$800Mas of 2026-05-31

Sub-layers and venture flow

  1. 14. Inference serving

    Inference serving: primary-round equity by quartereach segment links to its first round's record
    Q2 2023 · Form D filings · $48.9M · reportedQ4 2023 · Form D filings · $18.5M · reportedQ3 2025 · Form D filings · $59.7M · reportedQ2 2026 · Form D filings · $1.5B · reported
    Form D filings

    Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.

    The numbers
    quartersourcekinddollarsrecord
    Q2 2026Form D filingsequity$1.5Breported
    Q3 2025Form D filingsequity$59.7Mreported
    Q4 2023Form D filingsequity$18.5Mreported
    Q2 2023Form D filingsequity$48.9Mreported

  2. 15. Routing

    Routing: primary-round equity by quartereach segment links to its first round's record
    Q3 2023 · Form D filings · $2.7M · reportedQ2 2025 · Form D filings · $29.6M · reportedQ1 2026 · Form D filings · $113M · reported
    Form D filings

    Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.

    The numbers
    quartersourcekinddollarsrecord
    Q1 2026Form D filingsequity$113Mreported
    Q2 2025Form D filingsequity$29.6Mreported
    Q3 2023Form D filingsequity$2.7Mreported

  3. 16. Agent infrastructure

    Agent infrastructure: primary-round equity by quartereach segment links to its first round's record
    Q4 2024 · Form D filings · $10.0M · reported
    Form D filings

    Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.

    The numbers
    quartersourcekinddollarsrecord
    Q4 2024Form D filingsequity$10.0Mreported

  4. 17. Agent security & identity

    No primary rounds on file.

Primary-round dollars per calendar quarter for each sub-layer, each strip on its own scale: Form D amounts sold where the entity files, Epoch's press-compiled rounds otherwise.

Crosswalk to the diffusion lens

  • Products (same valve) — Productisation runs through serving, routing and agent infrastructure.