The hottest companies in AI this summer are not building smarter models. They are getting rich running everyone else's.

For three years, the prestige and the capital in artificial intelligence flowed to the labs building frontier models. In 2026, a quieter and arguably more durable business has captured the market's attention: the plumbing that actually serves those models to users, fast and cheap. The venture money has noticed. Across a cluster of inference-infrastructure startups, valuations and revenues have exploded, and the pattern is consistent enough to mark a genuine shift in where the smart money believes the profits will accrue.

The specifics are startling. Fireworks AI is now valued at about $17.5 billion after a roughly $1.5 billion Series D, with annualized revenue that crossed $1 billion, according to research firm Sacra. Baseten, per widely reported figures, raised roughly $1.5 billion at an $11-to-13-billion valuation — several times its price from earlier in the year — as its revenue surged. Together AI has reportedly passed around $1 billion in annualized revenue. As Newcomer put it, booming AI revenues have pushed inference startups toward decacorn status. The AI inference economy has become its own asset class.

What "inference" means, and why it is suddenly valuable

In AI, there are two fundamentally different activities. Training is the enormously expensive, one-time process of building a model — feeding it data until it learns. Inference is what happens every time you actually use that model: the calculation that turns your prompt into an answer. Training is a capital project; inference is an operating cost that recurs with every single query.

For years the glamour lived in training, because that is where the intelligence gets made. But as powerful open and commercial models have proliferated, a market truth has emerged: most companies do not need to train their own frontier model. They need to run an existing one — reliably, quickly, and at the lowest possible cost per token. That is a hard engineering problem, and solving it is worth a fortune, because inference cost is the line item that determines whether an AI product actually makes money.

Inference-infrastructure companies specialize in exactly this. They take open or third-party models and serve them at industrial efficiency — squeezing more throughput from each chip, cutting latency, and passing the savings on. Fireworks reportedly processes more than ten trillion tokens a day. That volume is the business: as AI usage compounds across the economy, the companies that run the models at scale collect a toll on every request, and that toll turns out to be a far steadier revenue stream than the race to build the next model.

Why the capital moved one layer up the stack

The migration of venture money to inference reflects a hard-nosed reading of where defensible profit lives. Training frontier models is a brutal business: it costs billions, the leaders change with each release, and today's best model is tomorrow's commodity. Inference, by contrast, is a volume business with compounding demand and real switching costs once a customer builds on your platform.

Investors have essentially concluded that in a gold rush, selling reliable, efficient access to the gold is a better business than mining ever-deeper veins. The models may commoditize; the need to run them cheaply will not. That is why revenue at these companies is scaling so fast — they monetize the usage of AI rather than the creation of it, and usage is exploding everywhere at once.

The trend rhymes with the biggest strategic moves elsewhere in the industry. When Anthropic reportedly pursued the efficiency startup Decart, and when the frontier labs began cutting prices to win volume, they were all responding to the same reality: the competitive frontier has moved from raw model quality to the economics of serving intelligence. The inference startups are the purest expression of that shift — companies whose entire value proposition is doing the running better and cheaper than anyone else.

What it means for each stakeholder

For the inference startups themselves, this is a golden moment, and a precarious one. Revenue and valuations are soaring because they sit on the toll road of AI usage. But their advantage is efficiency, and efficiency is a moving target: the hyperscalers, the chipmakers, and the frontier labs all want this margin too. Staying ahead requires relentless engineering, and the valuations now assume they will.

For enterprises building AI products, the boom is good news. A competitive market of specialized inference providers drives down the cost of running models and pushes up performance, making AI applications cheaper to operate and easier to scale. The risk is fragmentation and lock-in — choosing a platform is a bet on that platform's staying power in a fast-moving field.

For the frontier labs, the rise of independent inference companies is a double-edged development. It expands the ecosystem that runs their models, but it also means much of the value from AI usage may be captured by intermediaries rather than by the model-makers themselves. That is part of why the labs are moving to control inference in-house — the running of the model, not just its creation, is where a large share of the durable profit sits.

For investors, the inference wave is a bet on picks and shovels over prospecting. It is historically a sound instinct — the suppliers to a boom often outlast the speculators. The danger is that valuations have raced ahead of a young market, pricing in years of compounding before the competitive dynamics have fully played out.

The takeaways for operators

Two lessons generalize.

First, in any technology gold rush, ask where the recurring revenue lives, not where the glory is. The glory in AI is in building models; the recurring revenue is increasingly in running them. Businesses built on repeatable usage tend to compound more reliably than businesses built on periodic breakthroughs. The breakthrough is a moment; the toll is forever.

Second, commoditization is not a threat to everyone — for the layer above it, it is a gift. As models commoditize, the value migrates to whoever makes them cheap and easy to use. If your product sits atop a commoditizing input, falling prices upstream can be the best thing that ever happens to you. Position accordingly.

The bigger picture

Every major technology follows an arc from creation to distribution. The early value sits with the inventors; the durable value often accrues to those who deliver the invention efficiently to the masses. Railroads gave way to the businesses that shipped on them; the web gave way to the companies that made it usable. AI is now living through the same transition, faster than most expected.

The frenzy around inference startups is the market pricing in that shift — betting that the enduring profits of the AI era will come not from building the smartest model, but from running intelligence at scale, reliably and cheaply, for everyone who wants it. The models made the headlines. The meter may make the money.

Frequently Asked Questions

What is AI inference and why are inference startups booming?

Inference is the process of running an existing AI model to answer a query, as opposed to training, which is the one-time process of building the model. Inference startups specialize in serving models quickly and cheaply at scale. They are booming because most companies need to run existing models rather than build their own, making efficient inference a large and fast-growing market.

How much are inference companies worth?

Valuations have surged. Fireworks AI is valued at about $17.5 billion after a roughly $1.5 billion Series D, with annualized revenue past $1 billion, according to research firm Sacra. Baseten reportedly raised around $1.5 billion at an $11-to-13-billion valuation, and Together AI has reportedly passed roughly $1 billion in annualized revenue. Figures are as reported.

Why did venture capital shift from training to inference?

Training frontier models is extremely expensive and highly competitive, with leaders changing every release, while inference is a volume business with compounding demand and switching costs. Investors increasingly see running models efficiently as a more defensible, recurring-revenue business than building ever-larger models that quickly commoditize.

What does the inference boom mean for the AI industry?

It signals that value is migrating from creating models to serving them. Enterprises benefit from cheaper, faster AI, while frontier labs face the prospect that much of the profit from AI usage is captured by inference intermediaries — a key reason the labs are moving to control inference themselves.

Editor's note — sources: Sacra on Fireworks AI's revenue, valuation, and funding; Newcomer on booming AI revenues lifting inference startups. Baseten and Together AI figures are as reported across industry coverage; all valuations and revenue figures are as reported and reflect private-market estimates.

Share this article: