The rise of the inference economy

Walk a few blocks in San Francisco, and you’ll be greeted with giant posters selling inference as a service. And in fact, some of the biggest businesses emerging from the AI boom do not need to build the best AI model. They need to make somebody else’s model useful, reliable and economical to run.

That distinction is becoming harder to ignore. Fireworks has disclosed more than $1 billion in annualised revenue. Modal crossed the same threshold this month. Together AI has reached a similar scale according to external estimates, while fal has reportedly reached an $800 million revenue pace serving generative media.

Behind those numbers is a shift in where value can accumulate. The race to train frontier models continues, but a growing group of companies is building businesses around deploying them. Some model developers are joining that market themselves, including by serving open-weight models developed in China.

Nathan Benaich captures the tension in the State of AI Report 2026:

“either you die trying to get to the frontier, or you live long enough to serve inference.”

For founders and investors, the question is what makes that second business defensible. Following my post on the rise of Neoclouds (and the timely headlines about CoreWeave losing a bunch of value given its high debt) I wanted to dive into the new category of startups selling Inference.

What inference providers actually sell

Training creates a model. Inference runs it: generating an answer, producing an image, analysing a document or deciding an agent’s next action. Every useful application needs inference, whether its developer builds the underlying model or accesses one through an API.

This is related to the neocloud market I have been exploring, but the customer buys something different. A neocloud primarily sells computing capacity: GPUs, networking and the infrastructure needed to run workloads. An inference platform packages that capacity into a service for running models. Its promise is that developers can ship an application without assembling and operating the entire serving stack.

The boundaries overlap. Together sells GPU capacity as well as model APIs. Modal supports inference, training and agent execution. Cerebras and SambaNova combine specialised chips with systems and services. Calling all of them “GPU clouds” misses the differences that determine their economics.

The companies turning model usage into revenue

The financing and revenue figures show how substantial this market has become. They also require care: annualised usage revenue, bookings and recognised quarterly revenue measure different things.

Sources: company announcements from Fireworks, Modal, Together, fal, Baseten, DeepInfra, Groq and SambaNova; Cerebras’s SEC-filed results; Sacra on Together and Baseten; and The Information on fal. Figures have different observation dates and are not a revenue ranking.

Two details are particularly revealing. Fireworks says more than 95% of its daily tokens come from models specialised on customer data and specific jobs. Its opportunity extends into helping customers adapt models to their own work. Modal’s October announcement puts its sandbox product alone at a $300 million run rate. Agents need environments in which to execute code and use tools, creating infrastructure demand beyond generating tokens.

These businesses are expanding towards the surrounding workflow. That matters because an API offering the same model as everyone else faces persistent price competition.

When model builders become model hosts

The most interesting development is that model research and inference are becoming complementary businesses, including for companies originally associated with building their own models.

Reka is a clear example. Its 2024 technical report described multimodal models trained from scratch, with its largest model positioned against leading frontier systems. Its commercial offering now also includes Reka Cloud: serverless inference, dedicated GPUs and model customisation. The models displayed include DeepSeek, Qwen and GLM. A company that developed its own intelligence is also selling the infrastructure to run intelligence developed elsewhere.

That is an expansion, rather than evidence that Reka has abandoned research. Its current business still includes a model research lab. The important shift is that monetisation need not depend exclusively on whether its own next model wins.

Zyphra makes the connection explicit. Alongside developing its own foundation models, it launched an inference cloud on AMD infrastructure supplied by TensorWave. Its launch catalogue included DeepSeek, Kimi and GLM. Founder Krithik Puthalath described the cloud as “the natural extension of our research.”

Here, research expertise becomes a commercial capability: understanding models well enough to run them efficiently on a particular hardware stack. Zyphra’s announcement also identifies the infrastructure partner. Offering an inference cloud does not establish that the provider built or owns the underlying data centre.

01.AI represents a more explicit change in direction. Kai-Fu Lee’s company developed the Yi model family before shifting towards enterprise deployment in what he called the DeepSeek age. The Financial Times reported its pivot, while Reuters covered its Wanzhi enterprise platform. The lesson is that capable open models let a company commercialise deployment and domain expertise without funding every successive pre-training race itself.

These examples support a shift from exclusive reliance on proprietary models towards deployment businesses. They do not all support the stronger claim that former frontier labs have stopped research and become owners of giant data centres.

Poolside illustrates the separate move into physical infrastructure. Its Project Horizon announcement with CoreWeave described a planned 2GW Texas campus. But the FT reported in April that the partnership had ended, and Poolside continues releasing its own models. It is an example of model development intersecting with infrastructure ambition, rather than proof of an operating Chinese-model hosting pivot.

Chinese models are becoming inputs to Western AI businesses

There is an irony here. The public debate often frames AI as a contest between American and Chinese model companies. The commercial stack is more interconnected: a Chinese-developed model can run on American-designed chips, through a Western inference provider, inside an application sold to a global enterprise.

Open weights make that separation possible, subject to each model’s licence. They give other companies the ability to deploy and adapt the model, so the organisation that creates it does not necessarily collect every subsequent inference dollar.

The origin of a model and the location of its deployment are also different questions. Running downloaded weights with a chosen operator can keep processing within that deployment; it is different from calling the original developer’s hosted API. Customers still need to understand who operates the service, where data is processed and what controls apply.

For inference providers, better open models expand the range of workloads they can compete for. For application founders, the opportunity is to combine those models with proprietary data, evaluation and distribution. Neither requires winning a general-purpose benchmark.

Cheaper intelligence can create a bigger bill

The State of AI report highlights Epoch AI’s research, which estimates that the cost of achieving a fixed score across five benchmarks has fallen by about 47% per quarter since 2023.

That does not tell us whether total spending falls. A cheaper operation becomes economical in more settings, and an agent may perform many operations to finish one assignment. A business that once generated a draft can now explore alternatives, call tools, check results and retry. The relevant comparison becomes the cost of a completed task.

A cheaper model can therefore produce a more expensive workflow if it takes more attempts or requires more human review. Equally, an expensive model may be worth using for one difficult planning step while cheaper models handle repetitive execution. Founders need evaluations that reflect their actual product, rather than a procurement decision based only on token prices.

Where the moat might be

My view is that inference can support large businesses, but their durability will depend on what customers would lose by switching.

Research expertise can improve performance and cost. Deployment software can remove weeks of engineering work. Enterprise controls can make a service usable for a customer who could not otherwise adopt it. Specialisation can align the platform with a particular workload, such as generating video, handling voice interactions or running long-lived agents.

The test is whether those advantages persist as models, chips and open-source serving software improve. A provider whose value is mostly temporary access to scarce GPUs has a different business from one embedded in how customers train, evaluate, deploy and monitor their models.

Revenue alone will not answer that question. I would want to understand gross profit after compute costs, customer concentration, infrastructure commitments and retention when a major customer changes its preferred model. Annualising a strong usage month cannot establish the quality of the underlying business.

Physical infrastructure adds another set of obligations. Power contracts, financing announcements and planned campuses precede usable capacity, sometimes by years. As the State of AI infrastructure analysis emphasises, contracted power and operating capacity are different measures. Buying into the inference opportunity does not remove construction, utilisation or financing risk.

For the Israeli founders we meet at Remagine Ventures, the practical implication is encouraging. You can build a valuable AI product using models that keep improving outside your company. The work is to own the customer relationship, accumulate useful feedback and deliver an outcome at economics that improve with scale.

That applies whether you are building creative automation, a consumer application or enterprise infrastructure. Start with the model that lets you learn fastest. As usage grows, earn the right to take more control of the stack through better data, better evaluations and a measurable benefit to the customer.

The frontier race will continue to attract extraordinary capital. The inference economy gives founders another route to a large business: make increasingly capable intelligence work reliably for customers, and retain enough value from doing so.

• Author
• Recent Posts

Follow me

Eze Vidra

Co Founder and Managing Partner at Remagine Ventures

Eze Vidra is the founder of VC Cafe and the co-founder and managing partner of Remagine Ventures, a pre-seed fund investing in ambitious founders at the intersection of AI, technology, entertainment, gaming, and commerce with a spotlight on Israel.

He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.

He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.

Follow me

Latest posts by Eze Vidra (see all)

• The rise of the inference economy - October 19, 2026
• Weekly Firgun Newsletter – October 16 2026 - October 16, 2026
• The neocloud boom and the race to outlive the GPUs - October 14, 2026