Together AI scales its inference platform
Together AI closed an $800 million Series C on July 1, 2026, which established a post-money valuation of $8.3 billion. This funding follows a $305 million Series B in February 2025 that valued the company at $3.3 billion. The San Francisco-based neocloud provides access to more than 200 open-source models and plans to scale its compute footprint fifty-fold over five years. Annual bookings for the company now exceed $1.15 billion. It offers competitive per-token pricing at approximately $0.88 per million tokens, which helps businesses deploy open-weight models for less than the cost of proprietary systems. This funding round included participation from Aramco Ventures, Vista Equity Partners, General Catalyst, Emergence Capital, Nvidia, March Capital, Pegatron, and SentinelOne’s S Ventures.
Comparison of per-token and instance pricing
AWS offers specialized hardware through Trainium2 instances for users requiring dedicated compute. The trn2.48xlarge instance in the US East (Ohio) region costs $35.7608 per hour, which equates to $2.235 per accelerator, while the trn2.3xlarge offers alternative configurations in locations such as Hyderabad or Sao Paulo for lower hourly rates. These instances include 16 Trainium2 chips and 1.5 TB of HBM3 memory. They provide 46 terabytes per second of memory bandwidth and 3.2 terabits per second of EFAv3 networking. These instances offer 30% to 40% better price performance than GPU-based EC2 P5e and P5en instances. Users can deploy these instances with native support for frameworks like PyTorch and JAX via the AWS Neuron SDK.
| Instance Type | Region | Hourly Rate (Per Accelerator) | Accelerator |
|---|---|---|---|
| trn2.48xlarge | US East (Ohio) | $2.235 | 16 x AWS Trainium2 |
| trn2.3xlarge | Asia Pacific (Hyderabad) | $2.235 | 1 x AWS Trainium2 |
| trn2.3xlarge | South America (Sao Paulo) | $2.235 | 1 x AWS Trainium2 |
The economics of inference vary depending on whether a team selects a managed API or dedicated hardware. DeepSeek V4 Flash charges $0.14 per million input tokens and $0.28 per million output tokens. In contrast, Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens. Selecting a model requires evaluating the cost per finished task rather than the sticker price of a single token.
Managing the volatility of AI spending
Managing AI expenditure requires more than just selecting the cheapest provider. You know that uncontrolled token consumption can quickly invalidate any projected ROI. Uber exhausted its entire 2026 AI coding budget by April, and Microsoft revoked Claude Code licenses months after enabling them to control costs. Because agentic features multiply token consumption, many companies struggle to reconcile usage with their original budgets. One engineer at a company spent $40,000 on tokens in a single month.
Companies use routing strategies to send routine work to cheap models while reserving frontier models for hard tasks. Unified platforms like Amnic, CloudZero, and Harness attempt to aggregate these costs by linking LLM token spend with traditional cloud infrastructure metrics. Amnic, for example, provides a read-only connection to AWS, Azure, and GCP to track token usage without requiring write access to the environment. These tools aim to provide visibility into metrics like cost per inference.
The Linux Foundation unveiled plans for the Tokenomics Foundation to instill cost discipline through open standards for AI token usage. This body seeks to define new metrics like cost-per-intelligence and tokens-per-watt. Can companies truly predict monthly token spend without rigid limits? Priceline has already implemented token limits to manage the volatility of usage.




