ai& raises $2B for sovereign AI infrastructure and neocloud services

The massive capital injection for ai& signals a shift toward localized, sovereign compute that prioritizes data residency and cultural context over generic cloud offerings. By utilizing a heterogeneous hardware mix including Tenstorrent and AMD, the company demonstrates that the software-defined nature of modern AI workloads is effectively eroding the NVIDIA CUDA moat for inference tasks. This model suggests a future where streaming and enterprise platforms purchase specific token streams rather than managing underlying hardware complexity. Watch for the deployment of additional non-X86 accelerators in their cluster by the end of 2026 to validate this multi-vendor hardware strategy.

Tenstorrent, the RISC-V-based AI chip designer led by Jim Keller, has been aggressively expanding its footprint in sovereign AI deployments across Asia and the Middle East. In March 2025, Tenstorrent announced a partnership with South Korea's KT Corporation to develop AI inference infrastructure using its Wormhole and Blackhole processors, positioning the company as a viable alternative to NVIDIA for national-scale compute projects. That same month, Tenstorrent secured $693 million in a Series D funding round led by Samsung Securities and Korea Development Bank, valuing the chipmaker at approximately $2.6 billion and signaling strong investor confidence in non-NVIDIA inference silicon for sovereign deployments. The ai& investment in Tenstorrent hardware aligns with this broader trend of Asian governments and enterprises seeking compute independence from U.S.-dominated GPU supply chains.

Japan's push for sovereign AI capacity has become a national policy priority. In February 2025, the Japanese government announced a 1.2 trillion yen ($8 billion) investment plan to build domestic AI data center infrastructure, with the Ministry of Economy, Trade and Industry targeting operational capacity by fiscal 2027. The initiative explicitly encourages heterogeneous hardware strategies to reduce dependency on any single vendor, a policy framework that directly benefits ai&'s multi-accelerator approach. Meanwhile, SoftBank committed $500 billion to the Stargate project in the United States alongside OpenAI and Oracle, creating a parallel track where Japanese capital flows into both domestic sovereign compute and U.S.-based hyperscale AI. This dual-track strategy means Japanese AI startups like ai& must differentiate on data residency, cultural alignment, and inference cost rather than raw scale.

AMD's Instinct MI300X and MI350 series GPUs have emerged as the primary NVIDIA alternative for inference-heavy workloads in data center deployments. In June 2025, AMD reported that its data center GPU revenue reached $3.2 billion in Q1 2025, up 57% year-over-year, driven largely by inference and agentic AI workloads from cloud and enterprise customers. The company's ROCm software stack has matured significantly, with Meta publicly confirming that MI300X GPUs handle production Llama inference at scale with competitive throughput-per-watt versus NVIDIA H100. For ai&, the combination of AMD's proven inference performance, Tenstorrent's cost-efficient RISC-V architecture, and selective NVIDIA deployment for training creates a hardware portfolio that can undercut single-vendor GPU cloud pricing while maintaining workload flexibility across the token delivery pipeline.