Why We Invested in Together AI: Building the AI-Native Cloud for the Open-Source Era

Inference is becoming the most valuable category in the AI industry. Every foundation model ever trained—open or closed—must eventually serve predictions in production, and the volume of those predictions is scaling at a rate that dwarfs the compute spent training the models themselves. Brookfield estimates managed GPU demand will exceed $250 billion by 2034, with inference driving roughly 75% of that spend. Yet the infrastructure layer serving this demand remains fragmented: hyperscalers optimize for their own first-party models and silicon, GPU clouds resell raw capacity without software differentiation, and platform-only vendors lack the infrastructure control to deliver at scale. There is a structural gap—the need for an AI-native cloud that unifies the full model lifecycle on a single, co-optimized stack.

Enter Together AI.

Together AI is an end-to-end AI infrastructure platform that makes it easy for enterprises, developers, and researchers to train, fine-tune, and run inference on AI models. Founded in 2022, the company has built what we view as the most technically differentiated independent AI cloud in the market—one that spans five core service layers: pre-training, post-training and model shaping, inference, sandboxing, and managed infrastructure. Critically, this integration is not merely operational convenience. It is a compounding technical and economic advantage.

On the technical side, Together can co-optimize across the entire stack simultaneously—configuring pre-fill on B300s and decode on GB300s, a hardware split unavailable on any public cloud because no other provider with a large inference business is building for it. Rack-scale inference systems give the company a significant lead that customers simply cannot lease elsewhere. On the economic side, AI-native customers making six-to-twelve-month infrastructure commitments cannot accurately predict workload splits between training, fine-tuning, and inference. Together’s unified platform flexibly absorbs all workload types under a single commitment, eliminating the need to pre-allocate across multiple vendors—each of whom would demand their own long-term commitment or charge prohibitive on-demand rates.

This “T-shaped” business model—broad across workload types, deep into owned infrastructure—is central to what makes Together so compelling.  It has been a consistent theme in the company’s evolution, and distinguishes Together from both GPU clouds and platform-only competitors. Pure infrastructure players offer raw compute but no managed services or research-driven optimization. Platform-only companies are now marketing themselves as “clouds” and investing in data center supply chains—validation that infrastructure ownership is table stakes—yet still lack Together’s owned capacity advantage. Together delivers approximately 15% more tokens per GPU than competitors on equivalent workloads, a gap attributable to its world-class research organization. 

The Research Moat

At the heart of Together’s differentiation sits a sizeable research team—one of the strongest concentrations of AI systems talent outside the frontier labs. The founding team includes Vipul Ved Prakash (CEO, three-time founder with exits to Proofpoint and Apple), Ce Zhang (CTO), and three Stanford faculty members: Chris Re, Percy Liang, and Tri Dao. This group is credited with foundational contributions including FlashAttention, the RedPajama open data project, and data selection tools like DSIR that have become standard infrastructure for the open-source AI ecosystem. Together has the organizational structure of a company that treats technical depth as its primary competitive advantage. 

This research advantage compounds across layers. Together’s kernel team—built over four years—writes custom CUDA kernels that outperform NVIDIA’s own defaults by 10-20%. The compiler team, built over the same period, has developed Gather Compile as a drop-in replacement for PyTorch’s native compiler. And the algorithm team advances speculative decoding, disaggregated inference, and long-context optimization on a very routine cadence. The compounding effect across kernels, compiler, and algorithms creates a widening gap, not a converging one. Each new hardware generation opens a window where Together pulls ahead before the field partially catches up—and then the next cycle begins.

Crucially, this is “earned lock-in.” Customers own their models, weights, and data. Deployments are portable to other platforms—they just run slower, less efficiently, or at lower quality. The differentiation manifests purely at runtime.

Open Source as Structural Tailwind

Together is broadly indexed to AI activity and spend, but the open-source ecosystem is where its platform advantage is sharpest. Approximately 90% of tokens served on the platform are post-trained, proprietary derivatives of open-source base models—effectively custom models that require fine-tuning, reinforcement learning, and optimization services up the stack. With open-source models, Together can offer the full value chain from base model to production deployment. With closed-source models, it serves primarily as an infrastructure and inference provider. The proliferation of open-source foundations—from Meta’s Llama family to an expanding universe of specialized models—is therefore a direct growth driver for Together’s highest-margin services.

The developer community reflects this momentum. The platform’s developer traction has grown 5x over 15 months, driven by research-led developer relations and the broadening wave of AI-native applications moving into production. Once any AI-native company clears a meaningful revenue threshold, the pull toward taking control of their own models via open weights on Together becomes compelling—for functionality, performance, and margin reasons. 

Geodesic’s Investment

We are proud to participate in Together AI’s financing. Together has scaled revenue at extraordinary rates—growing over 6x year-over-year—and we believe the company is positioned to capture a meaningful share of the massive and rapidly expanding AI infrastructure market and become a critical node in the AI ecosystem. We look forward to partnering with Vipul and the team as they build the defining AI-native cloud platform for the open-source era.

Read Together’s announcement here: Together AI Raises $800 Million at $8.3 Billion Valuation to Make Frontier AI Accessible to All