Two Markets, One Vendor

Together AI raised $800 million in Series C funding on July 1 at an $8.3 billion valuation, with annual bookings crossing $1.15 billion in the second quarter. Aramco Ventures led, with Nvidia, Vista Equity Partners, General Catalyst, Emergence Capital, March Capital, Pegatron, and SentinelOne's S Ventures participating. The company operates a neocloud that rents Nvidia GPUs and hosts open-source models for enterprise customers including Cursor, Cognition, and Decagon.

Sixteen months ago the same company closed a Series B at $3.3 billion. The valuation has multiplied 2.5x since, and bookings crossed the billion-dollar line for the first time this quarter.

Together AI's numbers reframe the AI infrastructure market. Coverage centers on the closed-frontier race: Anthropic, OpenAI, and Google trading flagship releases. A parallel market has quietly crossed into revenue-scale on a different axis. Open-source model inference, sold through a neocloud, is no longer a hobbyist workflow or a cost-optimization side path.

Fireworks AI, the closest competitor, hit roughly $800 million in annualized revenue by May and is reportedly in talks for a $15 billion valuation with Index Ventures. Together itself reported platform usage tripling year-over-year, and both neoclouds sell access to the same open weights that frontier labs treat as substitutes rather than competitors.

Cursor and Cognition both run on Together. Cursor is the coding IDE with revenue in the hundreds of millions. Cognition ships Devin, the agentic engineering system, and on July 1 launched Devin Security Swarm, a parallel-agent scanner that sandboxes each finding to prove exploitability before submitting a remediation PR. Neither Cursor nor Cognition is treating this as a cost-optimization experiment. The infrastructure underneath these products mixes closed frontier APIs with open-source models on neoclouds. Operators picking tools already treat the two markets as parallel.

Nvidia sits in both markets. It participated in Together AI's Series C and invested $800 million in Reflection AI, the open-source lab whose $150-million-a-month lease at SpaceX's Colossus 2 started on July 1 and totals about $6.3 billion through 2029. Last December it paid $20 billion for Groq's inference unit. Every neocloud runs on Nvidia hardware, and "open source" names the model, not the stack.

The frontier-lab APIs and the open-source-on-neocloud path aren't fighting for the same procurement dollar. Linux Foundation research puts direct open-source model use at 63 percent of AI-adopting enterprises. Frontier labs sell cutting-edge capability at pricing that reflects it. Neoclouds deliver well-understood capability at 4 to 10 times lower cost per token, for workloads where the model is good enough. Operators buy from both because the capability envelopes and cost curves don't line up.

The two markets have different failure modes. The Fable 5 outage in June lasted nineteen days after a US export directive pulled the model globally. Teams that had built exclusively against Claude had no clean fallback. A parallel open-source-on-neocloud path would not have been touched by that directive. A neocloud outage, correspondingly, doesn't affect closed-frontier API access.

The pipeline that took Fable 5 offline didn't route through Together AI. The pipeline that could take Together AI offline (regulatory action against a specific hosted model, or Nvidia supply constraints) doesn't route through Anthropic's platform.

Meta made a parallel move the same day. Meta Compute, an internal group led by the company's infrastructure and Superintelligence Labs executives, will sell excess AI capacity to outside customers, following the pattern SpaceX set at Colossus. Demand hit $1 billion at Together AI the same week supply added a new seller in Meta. The "AWS versus Azure versus GCP" frame no longer captures where the AI infrastructure market is heading.

What to Do With This

If you're picking AI tooling as an individual or a small team, add an open-source-on-neocloud path alongside your primary frontier-lab API. Try Together, Fireworks, or Anyscale hosting Llama, Qwen, or Kimi for a workload where the frontier model is overkill. You'll usually pay 4 to 10 times less per token, and the fallback isn't tied to your primary lab's uptime.

If you lead a team that uses AI across several workflows, treat closed-frontier and open-source-on-neocloud as separate procurement axes with separate diligence. Ask each about model roadmap, uptime history, pricing structure, and vendor concentration. The answers won't line up.

If you decide what AI tools the whole company runs, put the two markets in the annual review as distinct spend categories. Two-thirds of AI-adopting enterprises already use open-source models directly, and treating that as an experimental line item hides both the spend and the risk.

Want More Than This Newsletter?

Alcreon publishes a daily AI briefing, long-form dossiers, and an analysis feed for the teams actually shipping AI in production. This newsletter is one read out of the full library.

Read the daily feed or browse the editorials.

Also on the Radar

xAI Opens Voice Agent Builder Beta at $0.05 per Minute

xAI launched Voice Agent Builder in beta on July 1, offering a no-code platform for building phone-based voice agents in under two minutes at $0.05 per minute of audio and $0.01 per minute of telephony on provisioned numbers. The single-stack speech-to-speech architecture, built on Grok Voice, replaces the three-hop assembly (speech-to-text, LLM, text-to-speech) most voice-agent builders stitch together from separate providers. Supports 25+ languages and 80+ voices with sub-second latency claims. For teams pricing customer-service voice apps, benchmark against Retell, Vapi, and Bland on a specific workload.

Fable 5 Hits 16.1% on Remote Labor Index, Roughly Double Opus 4.8

The Center for AI Safety and Scale AI Labs published updated Remote Labor Index results on July 1: Claude Fable 5 completed 16.1% of 240 real freelance projects at professional standard, versus Opus 4.8 at 8.3% and Opus 4.6 at 4.2%. Human raters score each project on whether a paying client would sign off on the deliverable, spanning 23 domains from 3D design to video production. The trajectory across three Anthropic generations (4.2%, 8.3%, 16.1%) suggests the automation frontier for economically-valued remote work is roughly doubling per model tier.