1.65 billion annexed the software layer, and the AI infrastructure full-stack war began

According to Woofun AI, the competitive logic of the AI infrastructure industry is being fundamentally reversed. Xinyun is no longer satisfied with simply selling GPU computing power, but is extending to the upstream software layer through mergers and acquisitions to try to control the complete technology stack from hardware scheduling to task orchestration. Nscale's acquisition of Anyscale for $1.65 billion was a landmark event in this trend, marking a shift in industry focus from 'who owns more chips' to 'who can control the operating efficiency of chips'.

Nscale's $1.65 billion acquisition of Anyscale aims to acquire its core business platform, engineering team, and key customer base including Coinbase (COIN.US), Runway, and Bedrock Robotics. Anyscale is the developer behind Ray, an open source AI workload orchestration framework developed by the founding team during their time at UC Berkeley.

Although Ray was officially transferred to the PyTorch Foundation in 2025 and promised to remain open source, Anyscale's commercial operation capabilities and deep binding to the Ray ecosystem make it an extremely strategic target. After the transaction is completed, Anyscale will continue to operate as an independent business unit under Nscale, retaining the original brand logo, but its technical core will be deeply integrated into Nscale's infrastructure system.

This move shows that the new cloud company is not only buying a code base, but also a 'brain' that determines how the GPU cluster works efficiently.

This upward takeover is not an isolated phenomenon, but rather a collective act of the entire AI infrastructure industry. In May of this year, Nebius (NBIS.US) acquired Eigen AI for $643 million, and IREN (IREN.US) completed the acquisition of Mirantis; looking back at 2025, CoreWeave (CRWV.US) successfully won Weights & Biases; and recently, Qualcomm (QCOM.US) also completed the acquisition of Modular. The five deals clearly outline a path: new cloud companies that master the underlying hardware, power, and data center resources are rushing to control the software layer that determines how these hardware operates. From Nebius (NBIS.US) to CoreWeave (CRWV.US) to Qualcomm (QCOM.US), industry giants are using capital methods to make up for software shortcomings in an attempt to establish barriers at the MLOps and AI orchestration layer, thus getting rid of low-dimensional competition that simply relies on hardware scale expansion.

According to data compiled by Woofun AI, the technology launched by Porch Capital used the monitoring tool STAX to track the technology stacks of about 12,000 venture capital support companies. Among them, the adoption rate of the MLOps category was extremely low. Only 58 companies used related tools, accounting for about 0.5%.

However, in this 0.5% sample, Ray and Weights & Biases together contributed 60 of the 68 adoption records, showing extremely high market concentration. As of July 30, the two leading companies had changed hands: CoreWeave (CRWV.US) took control of Weights & Biases, and Nscale bought Anyscale.

This meant that in just 18 months, almost the entire commercially valuable MLOps layer in the STAX sample had a transfer of ownership. This data reveals a key fact: application-layer startups usually only use models rather than build their own MLOps, leading to a high concentration of control in this layer, and new cloud companies can quickly monopolize this critical node through acquisitions.

At the same time, the trend of integration is also happening in reverse, and inference platforms are starting to build infrastructure downward. Lightning AI and GPU infrastructure provider Voltage Park completed a $2.5 billion merger deal, leaving Lightning AI as the surviving company.

Furthermore, inference and model service platforms such as Fireworks, Modal, and Baseten are being distributed across the spectrum from 'fully leasing infrastructure' to 'continuously increasing own assets'. Some companies adhere to the asset-light model, while others are starting to build their own underlying computing power.

Despite their different paths, these software companies' strategic end point is the same as Nscale, CoreWeave (CRWV.US), and Nebius (NBIS.US): having both the underlying hardware and controlling the software layers that orchestrate this hardware. This two-way integration shows that the industry is moving from the opposite direction to the same end, that is, simultaneously mastering computing power assets and orchestration software to build higher migration costs and efficiency barriers.

Full-stack integration has become the common end of the AI infrastructure industry, but there are huge differences in the cost structure of the different paths. New cloud companies such as Nscale, CoreWeave (CRWV.US), and Nebius (NBIS.US) can extend the software layer up and use existing cash flow to complete integration; if inference platforms want to build infrastructure downwards, they must bear heavier capital expenses and a longer payback cycle. As all companies try to control the complete technology stack, capital costs will be a key variable in deciding whether to win or lose. Whoever can get the capital needed to expand at a lower cost can take the initiative in this full-stack war. This is not only a competition of technical ability, but also the ultimate test of capital efficiency.