Together AI PM interview: infrastructure economics, not model smarts

Together AI is not a model lab. This is the most important thing to establish before any interview prep begins. The company does not train frontier models. It runs them, at scale, on an open-source-first platform, at prices approximately 80% below hyperscalers. The PM role here is platform work: you are building the infrastructure that other engineers, PMs, and researchers use to ship AI products. That framing shapes every round.

By February 2026, Together had hit $1B in annualized revenue, up from roughly $618M at the end of 2025, with 400% YoY growth in 2024. The Series B closed at $305M, led by General Catalyst, valuation $3.3B on $533.5M total funding. Gross margins sit at approximately 45%, across a dual revenue model: per-token API calls (roughly 30 to 40% of revenue) and GPU server rentals (the remaining 60 to 70%). Interviewers at this stage of growth expect PMs to connect product decisions to those unit economics directly, not as an afterthought.

The stages

Recruiter screen (30 to 45 min). Checks AI infrastructure fluency and developer-product experience, not general AI enthusiasm. Expect to explain the difference between serverless and dedicated inference, what batch inference is for, and why a developer would choose Together over a hyperscaler’s managed API. Generic answers do not pass.

Hiring manager conversation (45 to 60 min). Substantive from the start. Expect questions about how you have handled platform prioritization, how you measure success for developer-facing products, and whether you have a genuine view on open-source AI as a business strategy. “Open-source is the future” without a commercial rationale gets pushed on immediately. Together hired Cloudflare’s former infrastructure VP, signaling the enterprise positioning the company is building toward: candidates should match that register.

Product sense and strategy round. The prompt ties to Together’s actual product surface: Serverless Inference, Batch Inference (up to 30B tokens per model), Dedicated Model Inference, Dedicated Container Inference for generative media, Sandbox (code execution for agents), Managed Storage with zero egress fees, or Fine-Tuning. Strong candidates pick a specific surface, name the target developer segment, and connect every major decision to a unit-economic implication.

Cross-functional or panel round. Engineering and go-to-market interviewers alongside PM interviewers. The GTM interviewer is the one most candidates underestimate. Together’s 80% pricing wedge is a sales motion as much as a product claim. PMs need to understand what signals indicate a customer is ready to move from API trials to a committed GPU rental contract.

Final or leadership conversation. Together’s founders come from Stanford HAI and the Stanford DB group (Percy Liang, Chris Re, Vipul Ved Prakash). The research DNA is real: Mixture of Agents, Medusa speculative decoding, Sequoia, Hyena, and Mamba are all Together research outputs that reduce inference cost or latency. At senior levels, expect questions about how research findings translate into product decisions.

What Together AI actually builds

Candidates who skip product surface research get exposed in round one. The platform as of June 2026:

• Serverless Inference: pay-per-token, 200+ open-source models, OpenAI-compatible endpoints
• Batch Inference: up to 30 billion tokens per model, lower cost for offline workloads
• Dedicated Model Inference: reserved GPU capacity for latency-sensitive production workloads
• Dedicated Container Inference: for generative media use cases requiring custom runtimes
• Accelerated Compute: raw GPU access, Blackwell cluster deployments, 36,000 NVIDIA GB200 NVL72 GPUs being co-built with Hypertec, 200 MW of power secured
• Sandbox: secure code execution for agent-based workloads
• Managed Storage: zero egress fees, a real purchase criterion for cost-conscious developers
• Fine-Tuning: managed fine-tuning on open models, connected to the inference layer

Two recent expansions that interviewers treat as live material: the May 2025 acquisition of Refuel.ai (moving Together into data transformation and labeling, expanding TAM from inference to training data pipelines), and the March 2026 voice agent platform (co-locating STT, LLM, and TTS on one cloud, sub-500ms end-to-end latency).

Questions that get asked

• “How would you measure success for the voice agent platform at six months?”
• “A startup on Serverless Inference is hitting latency spikes under load. What is the product motion for moving them to Dedicated?”
• “Together claims 31% more TPS than the next-fastest open-source inference engine for production coding agent workloads. How do you turn that into a go-to-market story without it becoming a benchmark arms race?”
• “Refuel.ai was acquired to expand from inference into data transformation and labeling. How does that change the PM surface area, and what do you prioritize first?”
• “What is the biggest product risk of being 80% cheaper than hyperscalers?”
• “How would you prioritize across inference performance, fine-tuning, Sandbox, and the Refuel data pipeline when all four teams argue their surface is the growth driver?”

The last pricing question is where most candidates stop too early. “Margin pressure” is the start, not the end. The full answer connects pricing to positioning: if commoditization drives the price toward zero, inference excellence and developer experience have to be the durable product, not just the cheap option.

Product-sense round in practice

Prompts are infrastructure-specific and deliberately underspecified. A representative one: “How would you improve Together AI’s Sandbox for agentic workflows?”

The viable/lovable frame for infrastructure

In 2026, feasibility is not the question any developer is asking. Any team can spin up a capable model. The question is at what cost per query, with what latency guarantee, and with which model. Together AI sits precisely at that inflection point.

Viable is concrete here: can customers pay enough per token or compute-hour to cover GPU depreciation (Blackwell clusters are expensive), power (200 MW secured), staff, and research, while still growing at margins that justify a $3.3B valuation? The 45% gross margin at $1B ARR is the current answer, but every pricing decision, every product expansion, and every enterprise deal either defends or erodes it. PMs are expected to hold that number in their head.

Lovable for a developer platform means: API-first, OpenAI-compatible endpoints so switching cost is low, fast rate limits, no surprise egress fees, and reliability that does not require the developer to think about the infrastructure layer. The voice agent platform is Together moving from “we run models” toward “we solve the agentic latency problem so you do not have to.” That shift from infrastructure primitive toward lovable product experience is a strategy question worth having a view on before the strategy round.

Open-source as product strategy, not technical preference

This will surface directly or as subtext in the strategy round. Together’s thesis: open model weights commoditize model intelligence, leaving infrastructure and developer experience as the durable differentiation. A Fortune 500 that could self-host Llama 4 on Azure might still choose Together because of operational cost, rate limits, batch throughput, or the zero-egress storage layer. Together is also becoming part of the open-source AI training stack itself: a partnership with Meta’s PyTorch team on an open-source RL framework for agentic AI training signals that the company’s scope is expanding from inference toward the full model lifecycle.

A candidate who says “open-source is the future” without completing that sentence with a specific buyer rationale will not clear the strategy round. The clean argument: Together wins not because models are open, but because running those models reliably and cheaply at scale is hard, and that difficulty is Together’s moat. Be ready to defend it against the hyperscaler counter: why wouldn’t a Fortune 500 just use Azure’s managed open-model endpoints?

What clears the bar

Strong candidates demonstrate inference economics fluency without being prompted: they reach for cost-per-token, P95 latency, and GPU utilization as naturally as consumer PMs reach for DAU and retention. They have a genuine, specific view on the open-source bet grounded in a buyer scenario. They can navigate Together’s four simultaneous product bets (inference, fine-tuning, agentic infrastructure, data pipelines) with a coherent rationale for which gets resources in which market condition. And they name the right success metrics for a specific product surface without being asked to justify the choice.

The failure modes: treating Together like a model company (Anthropic, OpenAI) and focusing answers on AI capability or safety; treating it like a generic SaaS PM interview and answering with activation funnels and engagement metrics; or dropping open-source into every answer without a commercial view on why it matters to a specific buyer.

For the unit economics foundation, see LLM unit economics. For the 2026 feasibility-is-free reframe that applies across all AI infrastructure PM roles, see feasibility is free. For a direct competitor comparison on speed-first inference positioning, see Groq.