Neocloud Nebius Acquires Cold-Start Fixer Inferize to Strengthen Inference Stack

Neoclouds are moving beyond renting out GPUs and starting to buy inference software. Nebius has acquired Inferize, an Israeli startup founded just ten months ago. It is the fourth software acquisition or technology deal Nebius has made this year.

Nebius is an “AI cloud,” or neocloud, that rents out the GPUs used to build AI. It grew out of Yandex’s cloud business, and Yandex co-founder Arkady Volozh serves as its founder and CEO. It is headquartered in Amsterdam and listed on Nasdaq (NBIS). With data centers in Finland, the UK, France, and the US among other places, it bundles NVIDIA GPU clusters, a cloud platform, and tools for training and running AI models. Unlike general-purpose clouds, it is designed from the ground up for AI workloads. Its customers range from startups to Big Tech: it signed a five-year, $17.4 billion (up to $19.4 billion) supply deal with Microsoft in September 2025 and, this March, a new AI infrastructure agreement with Meta worth up to $27 billion. It also runs Token Factory, an inference service that lets developers call open-source AI models through an API.

The problem Inferize tackles is the “cold start.” Before a model can serve a request, it has to be loaded into GPU memory, and the longer that takes, the longer the GPU sits idle. That gap appears whenever demand spikes and new instances spin up, or whenever weights are updated mid-run during reinforcement learning. To hit response-time targets, providers end up keeping spare GPUs running. Nebius calls this the “idle GPU tax.” Inferize’s technology shortens that window so capacity can scale up and down in step with actual usage, serving more demand from the same GPUs and lowering the cost per token. Inferize’s team and technology are joining Token Factory.

Inferize was founded in January and had a working prototype within three months; it had raised a $10 million seed round led by TLV Partners. Co-founders Guy Bortnikov (CEO) and Lior Gorbonos (CTO) come from the founding team of Israeli infrastructure-optimization company Granulate, which Intel acquired in 2022 for about $650 million. With a 17-person team in Tel Aviv, the company was acquired before it had even launched a commercial product. Terms were not disclosed; Israeli outlet Calcalist estimated the deal at $100 million to $150 million.

The deal looks like the latest piece of an inference stack Nebius has been assembling all year. In February it acquired agent-search startup Tavily for up to $400 million, and in May it agreed to acquire Eigen AI for $643 million, which handles optimization at the model, kernel, and system levels. That same month it brought on Clarifai’s core research team and licensed its inference and compute-orchestration technology. After the model, runtime, and system-orchestration layers, Inferize adds another: how fast capacity can scale. In an official blog post, Nebius described all of this as a single bet that the AI cloud will be won at the software layer. Training is where models are created, it argued, but inference is where AI becomes a product, and GPUs alone are not enough — models must also run fast, reliably, and cost-efficiently. CTO Danila Shtan said running inference well takes more than fast GPUs and optimized models; the whole system has to respond when demand changes, including how quickly additional capacity is ready.

Nebius’s funding history

Behind the aggressive acquisition spree sits nearly two years of heavy fundraising. In December 2024, Nebius took $700 million in equity from Accel, NVIDIA, Orbis Investments, and others, followed by a $1 billion convertible note offering in June 2025. In September 2025, right after its Microsoft deal, it raised $3.75 billion by combining a $2.75 billion convertible note offering with roughly $1 billion in new shares.

This year the numbers got bigger. In March, NVIDIA made a $2 billion strategic investment, and a week later Nebius issued another $4 billion in convertible notes. In July it closed its first secured debt financing, $775 million backed by GPU infrastructure and contracted cash flows, and in August it priced $5 billion in convertible notes. At the time of the NVIDIA investment, Nebius said it expected $3.0–3.4 billion in revenue this year and $7–9 billion in annualized run-rate revenue by year-end.

The Competition

Nebius is not the only neocloud buying software. CoreWeave started with AI developer platform Weights & Biases, then acquired OpenPipe and Marimo, and folded them into CoreWeave Forge, a single environment spanning training, inference, evaluation, and agent development. On the inference side, Together AI has grown quickly as an open-source-model inference cloud, raising an $800 million Series C at an $8.3 billion valuation. Newer inference-focused clouds keep arriving: DeepInfra raised $107 million with NVIDIA and Samsung Next participating, and General Compute raised a $15 million seed as an inference-specialized neocloud. As NVIDIA ships inference chips built on the Groq technology it licensed for $20 billion, Nebius was the first cloud to adopt the chip and plans to bring it to Token Factory.