Baseten Nears $1.5 Billion Mega-Round as the 'AI Inference' Gold Rush Accelerates

The artificial intelligence industry is undergoing a quiet but massive tectonic shift, moving its capital from the laboratories where models are trained to the server racks where they actually answer questions. This transition was thrown into sharp relief this week as Baseten, a San Francisco-based AI infrastructure startup, reportedly nears a $1.5 billion funding round.[1][2]

The financing, which is reportedly being co-led by Altimeter Capital, Conviction, Spark Capital, Sands Capital, and Wellington Management, is structured as a dual-tiered round. Depending on the terms, investors are buying in at either an $11 billion or a $13 billion valuation. This represents a staggering ascent for a company that was valued at just $2.15 billion in September 2025 and $5 billion in January 2026.[2][3][4][7]

To understand why investors are pouring billions into a company that operates entirely behind the scenes, one must understand the difference between AI "training" and AI "inference." Training is the computationally grueling process of teaching a model how to think by feeding it vast oceans of data. It is a one-time, highly expensive event. Inference, by contrast, is the act of the trained model actually generating an answer to a user's prompt.[4][6]

While training dominates headlines, inference dominates the balance sheet. Industry analysts estimate that inference accounts for 80% to 90% of the lifetime cost of a production AI system because it runs continuously. Every time a user asks a chatbot a question, generates an image, or summarizes a document, inference compute is consumed.[6]

"Inference workloads will indeed be the hot new thing in 2026," according to projections from Deloitte, which estimates that inference will account for roughly two-thirds of all AI compute by the end of the year, up from just one-third in 2023. The broader AI inference market is projected to grow from $106 billion in 2025 to $255 billion by 2030.[6][8]

Baseten has positioned itself as the "AWS for inference," providing the unglamorous but essential software and computing capacity businesses need to run these models at scale. Setting up a cloud-based inference cluster is notoriously difficult; developers must provision graphics processing units (GPUs), configure complex networking, and install a fragile stack of software tools.[3][4][5]

Baseten automates this workflow. The company rents capacity from roughly 20 different cloud providers and layers its own proprietary inference software on top. This allows customers to deploy and fine-tune models on their own data without having to manage the underlying hardware. According to financial data platform Sacra, Baseten's annualized revenue hit an estimated $600 million in March 2026, up nearly 1,900% year-over-year.[4][7]

The startup's secret weapon lies in its specialized software modules, which it calls "inference engines." One engine, BIS-LLM, is specifically designed to power "Mixture of Experts" models—complex architectures that route queries to specialized sub-networks. BIS-LLM optimizes the model's "KV cache," a critical data structure that stores the context of a conversation, allowing the system to automatically provision more hardware when token usage spikes.[3]

Another module, Engine-Builder-LLM, uses a technique called "lookahead decoding" to generate multiple tokens simultaneously, dramatically speeding up processing times for dense, monolithic models. A third engine handles simpler tasks like data embedding and search classification. Crucially, Baseten also employs a multi-cloud routing system that can instantly shift workloads to a different cloud provider if one experiences an outage.[3]

The explosive demand for Baseten's services is being driven by a parallel trend: the rapid maturation of open-source AI. As open-weight models like Meta's Llama, Alibaba's Qwen, and DeepSeek become increasingly capable, many enterprises are realizing they no longer need to pay the premium prices charged by proprietary API providers like OpenAI and Anthropic.[4][5]

However, deploying these open-source models efficiently requires specialized infrastructure for model serving and latency optimization. Baseten fills this exact gap. By shifting workloads to open-source models running on Baseten's optimized infrastructure, some enterprise customers have reportedly cut their inference costs by up to 70% compared to using proprietary alternatives.[4]

Baseten's customer roster now includes some of the fastest-growing AI-native products in the industry, including the AI coding assistant Cursor, the workspace platform Notion, and the enterprise AI writer Writer. These companies require ultra-low latency; a coding assistant that takes three seconds to suggest a line of code is useless to a developer.[4][5]

The sheer size of Baseten's new valuation signals that inference infrastructure is no longer viewed as a commodity service, but as a platform-class business in its own right. "The round would establish a named $10B-class incumbent in a category that, until recently, most buyers assumed hyperscalers would absorb," noted industry analysts at AI Weekly.[5]

This raises the inevitable question of competition. The major cloud hyperscalers—Amazon Web Services, Google Cloud, and Microsoft Azure—are all aggressively expanding their own managed AI inference services. These giants have the capital to potentially price inference infrastructure below cost in order to lock customers into their broader cloud ecosystems.[5]

Yet Baseten's multi-cloud approach offers a compelling counter-narrative. By sitting a layer above the underlying hardware, Baseten prevents enterprises from being locked into a single cloud provider's ecosystem. If AWS raises GPU prices or Azure suffers an outage, Baseten can theoretically route the customer's inference workloads elsewhere without any disruption to the end user.[3][4]

The hardware landscape is also shifting beneath their feet. While Nvidia's general-purpose GPUs currently dominate both training and inference, the market is beginning to bifurcate. Purpose-built AI accelerators and Application-Specific Integrated Circuits (ASICs) designed exclusively for inference are entering the market, promising lower costs and higher energy efficiency.[6][8]

Baseten's software-first approach means it is largely agnostic to the underlying silicon. Whether a customer's model is running on an Nvidia H100 GPU or a specialized inference chip from a competitor like AMD or Groq, Baseten's orchestration layer aims to extract the maximum possible performance.[3][8]

Ultimately, the "inference gold rush" represents the maturation of the artificial intelligence industry. The era of simply proving that large language models work is ending; the era of making them economically viable and reliable enough for mission-critical enterprise software has begun. For companies building the picks and shovels of this new economy, the rewards are proving to be astronomical.[5][6]