Monday, November 23, 2026
AI 인프라 · 뉴스 & 분석
논평

Together AI Enterprise Partnerships, November 2026

Rackspace and Together AI have announced a multiyear strategic AI infrastructure agreement targeting 2027 deployment, the latest in a string of enterprise deals that has pushed Together AI's reported valuation to $8.3 billion.

The announcement this month that Rackspace and Together AI have struck a multiyear strategic AI infrastructure agreement — with deployment targeted for 2027 — is the latest data point in what has become an unmistakable pattern for the neocloud: a steady accumulation of enterprise relationships that, taken together, suggest Together AI is moving from boutique open-model provider to a broadly distributed inference platform. The terms of the Rackspace arrangement have not been disclosed, so the commercial significance remains opaque for now, but the trajectory it extends is not.

The pace of partnership-building over the past several months has been notable. In August 2026, IBM and Together AI reportedly signed a multi-year agreement valued at $240 million for a large-scale NVIDIA B300 AI inference cluster — a figure reported by multiple outlets — and Together AI was said to expect that capacity to sell out before launch. That same month, Larsen & Toubro secured what it described as a mega order to build Together AI's AI factory in Chennai, India; reports cited the contract at figures ranging from roughly Rs 10,000 crore to $1.5 billion depending on the source, and the facility is reportedly India's first 10,000-GPU unified AI cluster built around NVIDIA B300 hardware. In September 2026, Equinix, NVIDIA, and Together AI announced a joint Inference Exchange platform, offering enterprises access to NVIDIA compute and more than 200 open-source models across Equinix's global colocation footprint — a trilateral arrangement that positions Together AI's software layer directly inside the enterprise data-center stack.

Alongside this partnership cadence, Together AI has been sharpening its technical differentiation. In November 2026, the company unveiled Polymorphic Inference and the Together Compiler, designed to bring continuous compilation to production inference and provide ongoing optimization of open-source AI models. Earlier, the company had advanced ThunderKittens, its high-performance computing tensor library, to run on NVIDIA's Vera Rubin GPU generation — tying its software roadmap to the leading edge of NVIDIA's hardware lineup. In production infrastructure, Together AI has adopted Clockwork.io's GPU fault-tolerance software, which preserves GPU workload state without code changes; that adoption was reported alongside LinkedIn and WhiteFiber, suggesting the tooling is gaining traction among compute-intensive operators broadly. Together AI also cut training prices and expanded its fine-tuning platform, moves reported as targeting competition with proprietary model vendors.

The geographic footprint has widened materially. In August 2026, Together AI and HUMAIN — Saudi Arabia's state-backed AI vehicle — formed a strategic partnership to accelerate AI infrastructure and cloud services in the kingdom, with reports indicating HUMAIN selected AMD accelerators, Cisco networking equipment, and Together AI software for its national AI backbone. In September and October 2026, Mubadala, the Abu Dhabi sovereign wealth fund, was reported to have partnered with Together AI for UAE investment opportunities, and Abu Dhabi was separately reported as deploying AI infrastructure in collaboration with the company. At San Francisco Tech Week in October 2026, Together AI emphasized speed and ecosystem-building as its core strategic priorities — framing consistent with its simultaneous push into discounted preemptible compute targeting price-sensitive workloads.

Against this backdrop, Together AI's valuation was reported at $8.3 billion as of October 2026, according to AI Magazine. That figure reflects a market judgment that open-source inference infrastructure occupies a defensible commercial position, particularly as enterprises seek to avoid vendor lock-in with proprietary model providers. The IBM deal, the L&T factory, the Equinix-NVIDIA trilateral, and now the Rackspace agreement collectively represent a distribution strategy that places Together AI's inference stack inside hyperscale, colocation, and managed-hosting channels simultaneously.

The risks are real and largely structural. Together AI operates in a segment where hardware economics are set by NVIDIA and cloud pricing is contested by hyperscalers with far deeper balance sheets. The Chennai factory and the IBM cluster represent large committed capacity bets whose utilization rates remain unverifiable from outside the company. The Rackspace deal's undisclosed terms make its revenue contribution impossible to assess at this stage. And while the partner roster is impressive, the neocloud model's vulnerability to margin compression — as GPU rental costs and competitive pricing pressure converge — has not been structurally resolved by any of these arrangements. Three signals are worth monitoring: first, whether the IBM B300 cluster capacity does in fact sell out before launch, as Together AI reportedly anticipated; second, the pace at which the Chennai facility comes online and what customer base absorbs its capacity; and third, whether the Together Compiler and Polymorphic Inference demonstrably widen the performance gap against bare-metal or hyperscaler inference alternatives — the technical moat on which Together AI's pricing power ultimately rests.

Based on 63 archived reports · Together AI →
Together AI Enterprise Partnerships, November 2026 · Slicast