IBM and Together AI just announced a $240 million multi-year deal to build a dedicated AI inference cluster on IBM Cloud, signaling that open-source models are becoming serious enterprise infrastructure rather than budget alternatives. The partnership deploys NVIDIA HGX B300 systems with specialized networking, expected to launch in the first quarter of 2027. Together AI, valued at $8.3 billion after closing an $800 million funding round in July 2026, will use the cluster to serve its rapidly expanding platform, which currently processes 400 trillion tokens per month.

What Does This Deal Actually Change for Enterprise AI Teams?

The announcement matters because it puts IBM Cloud's enterprise credibility behind open-source model inference. Until now, running models like DeepSeek, Llama, or Qwen at scale typically meant working with specialized GPU cloud providers. Hosting Together AI's platform on IBM Cloud wraps open-source inference in the compliance, security, and support infrastructure that regulated industries already trust. For security teams that have blocked open-model pilots over infrastructure concerns, this removes a major objection when the cluster becomes available early next year.

The deal also reveals IBM's deliberate strategy to avoid betting everything on a single AI approach. This is IBM's third major inference partnership in ten months: Anthropic's Claude integration in October 2025, Groq's specialized LPU (Language Processing Unit) chips in October 2025, and now Together AI's open-model GPU inference. Rather than choosing a winner, IBM is assembling a portfolio that lets customers pick inference in multiple flavors, from closed-model APIs to custom silicon to open-source alternatives.

How Should Enterprise Teams Prepare for This Shift?

• Benchmark Open Models Now: Add at least one open-source model like DeepSeek, Llama, or Qwen to your evaluation process this week and measure its quality and cost against your current closed-model workloads. You cannot negotiate with data you have not collected.
• Document Closed-Model Pricing: Ask your current AI vendors what their inference pricing will be at your projected 2027 volume and document the answers before renewal season begins. Open-model alternatives become negotiating leverage only if you have concrete pricing comparisons.
• Plan for Q1 2027 Access: If you use IBM Cloud, contact your account team about early access to the Together AI cluster and whether committed-use discounts will be available at launch. Announced AI infrastructure timelines often slip, so build contingency capacity into your plans.
• Track NVIDIA Benchmarks: Monitor whether NVIDIA publishes detailed benchmarks for this specific HGX B300 deployment. Capacity planning built on press-release multipliers becomes expensive when real-world performance arrives.

What's the Real Performance Gain Here?

The announcement claims the deployment delivers "30 times more AI factory output compared to prior generations." That figure deserves scrutiny. NVIDIA's own published benchmarks from its GTC 2025 conference claim the HGX B300 delivers 11 times faster large language model (LLM) inference compared to the previous Hopper generation. The larger 50 times "AI factory revenue opportunity" claim applies to a different, larger rack system. No 30 times figure appears on NVIDIA's current product documentation. The practical takeaway: the B300 is a genuine generational leap for inference speed, but budget your capacity models on the 11 times improvement claim, not 30 times, until NVIDIA publishes matching benchmarks for this specific configuration.

"Together AI is proud to lead the way in bringing production inference to market with NVIDIA's latest AI infrastructure on IBM Cloud," said Vipul Ved Prakash, CEO of Together AI.

Vipul Ved Prakash, CEO at Together AI

Prakash emphasized that the partnership aligns with Together AI's mission to make advanced open-source AI broadly accessible through improved infrastructure and better token economics, the cost per unit of text processed.

Why Does This Matter Beyond Together AI?

The broader context is that inference, the process of running trained models to answer user requests, is now the dominant driver of AI compute spending. Market research firms project the global AI inference market will grow from roughly $106 billion in 2025 to $255 billion by 2030, a 19 percent annual growth rate. That scale explains why hyperscalers and enterprise clouds are locking up compute capacity years in advance. At $240 million, IBM's commitment is modest compared to Microsoft's $17.4 billion contract with Nebius through 2031, but it signals the same trend: major infrastructure providers are betting heavily on inference as the next growth engine.

"Enterprises are in a race to adopt agentic AI at scale to drive real business outcomes," said Alan Peacock, general manager of IBM Cloud.

Alan Peacock, General Manager of IBM Cloud

Peacock noted that the joint solution provides the economic efficiency and scale enterprises need to deploy AI agents, software systems that can handle complex, real-time tasks autonomously.

The timing also reflects genuine business pressure. Reuters reported that open-source models have gained traction as enterprises seek to control AI costs and reduce security exposure across multiple closed-model providers. Together AI's own growth trajectory is the clearest proxy for this shift. The company now serves more than 1 million developers and processes 400 trillion tokens monthly, a scale that would have been unthinkable for open-source inference just two years ago.

What Should You Avoid During This Transition?

One critical caution: do not rip out working closed-model deployments simply because open-model economics look better on paper. Migration costs and evaluation overhead are real expenses that can offset savings. Similarly, do not treat the Q1 2027 launch date as guaranteed. Announced AI infrastructure timelines frequently slip, so plan contingency capacity if the cluster experiences delays.

The real opportunity lies in using credible open-model alternatives as negotiating leverage. Even teams that never leave their closed-model provider benefit when they can point to a viable, enterprise-grade alternative at renewal time. IBM's $240 million commitment signals that such alternatives now exist at scale, backed by major cloud infrastructure providers and NVIDIA's latest hardware. That changes the cost conversation for every enterprise AI team, whether they ultimately choose open or closed models.

" }