IBM and Together AI sign $240M deal for NVIDIA Blackwell systems

The $240 million commitment signals a shift in the streaming and media ecosystem toward specialized neocloud providers for high-volume inference. By utilizing the NVIDIA HGX B300, Together AI targets the high-compute demands of media generation models, which require massive memory bandwidth to remain cost-effective. For the broader industry, this deal validates the enterprise transition from closed-source APIs to open-weight AI models for production-grade workloads. It also solidifies IBM’s position as a critical infrastructure partner for high-growth AI startups challenging legacy cloud giants. Watch for whether this deployment triggers a price war in per-token inference costs as more Blackwell-based clusters come online in early 2027.

The partnership between IBM and Together AI aligns with a broader industry trend where inference workloads are projected to account for two-thirds of all AI compute by late 2026. Per industry reporting from July 2026, the global AI inference market is expected to reach $117.8 billion this year, driven by the rapid adoption of multimodal generation and long-context reasoning. These advanced tasks often require up to 100 times the compute of simple queries, making high-memory systems like the Blackwell Ultra essential for maintaining latency requirements in production environments. Together AI’s $800 million Series C round in July 2026, led by Aramco Ventures with participation from NVIDIA, underscored the rising valuation of the 'neocloud' category. According to TechCrunch, the startup’s annual bookings surpassed $1.15 billion as enterprises sought to reduce costs by up to 60 times compared to closed-model providers. This capital influx has sparked an infrastructure arms race among specialized providers; for example, rival TensorWave raised $350 million in June 2026 to scale its own inference-focused clusters. Technically, the HGX B300 systems at the heart of the IBM deal represent a significant jump in density. Per NVIDIA’s 2026 specifications, the Blackwell Ultra architecture provides 15 petaflops of dense FP4 performance per chip, a 1.5-times improvement over the standard B200. These systems are specifically designed to handle trillion-parameter models with long-context requirements, using 288GB of HBM3e memory per GPU to keep large models entirely in-memory, thereby eliminating the throughput bottlenecks common in older hardware generations.