Together AI raises $8.3 billion valuation financing, AI computing power shifts from competing for resources to optimizing utilization.

TL;DR

· Together AI has completed an $800 million Series C funding round, with a post-investment valuation rising to $8.3 billion. The company stated that its annualized bookings exceed $1.15 billion.

· Market divergence is shifting toward order quality, with utilization rates, renewal rates, and gross margins determining AI infrastructure returns.

· Related targets: Nvidia (NVDA), Meta (META), Oracle (ORCL), AMD, Arm, and companies involved in data centers and power infrastructure.

On July 1, Together AI announced the completion of an $800 million Series C funding round, with a post-investment valuation of $8.3 billion, a significant increase from the $3.3 billion valuation in its previous round in early 2025. The company also disclosed that its last quarter's annualized bookings exceeded $1.15 billion. After customers use open-source models on its platform, they can save 6 to 60 times in costs compared to closed-source model pricing.

On the other hand, pressure on AI infrastructure is also emerging. According to Axios, citing a Bloomberg report, Meta is considering selling AI model access and excess computing power through a new cloud business. Oracle's annual report for the period ending May 2026 also disclosed risks related to long-term data center leases, power purchase commitments, and changes in customer demand.

This comparison brings the issue before investors: Whether AI computing power is scarce is no longer the only variable. The more realistic test is who can keep expensive electricity, GPUs, and data centers fully occupied over the long term.

Together proves open-source inference demand is still heating up

First, let's clarify the business. Training is teaching the model; inference is having the model answer questions, write code, handle customer service, and generate content daily. The former is like building a factory, while the latter is like the daily output after the factory starts operating.

Together's growth mainly comes from inference. It focuses on providing cheaper AI cloud services using open-source models, allowing developers and AI application companies to avoid relying entirely on closed-source large model interfaces. For customers, the core variable is whether the unit call cost can continue to decrease.

The company's disclosed customers include AI application companies such as Cursor, Cognition, and Decagon. Together claims that Decagon's inference costs dropped approximately 6 times after using its platform. The company also cited industry data stating that the usage of open-source models has tripled over the past 12 months.

This explains why capital is willing to offer an $8.3 billion valuation. For AI applications to move from demonstrations to daily use, inference costs must decrease. As long as usage growth outpaces the decline in unit price, cheaper computing power will amplify total demand.

Together CEO Vipul Ved Prakash's statement is typical: Intelligence is becoming a foundational resource similar to electricity, bandwidth, or capital, and an open ecosystem will make innovation cheaper and faster. This is the core judgment of the optimists and the basis for Aramco Ventures, Nvidia, and other investors to continue participating.

But annualized bookings are not actual revenue. They are closer to the company's disclosed order and contract heat, indicating demand intensity but not that cash has been received or that renewals will occur every year.

Meta and Oracle remind the market to focus on the payback period

If you only look at Together, it's easy to conclude that AI computing power is still in short supply. However, signals from Meta and Oracle indicate that infrastructure investment is entering a phased stage.

News that Meta is considering selling excess computing power and model access to external customers shouldn't necessarily be seen as bad news. For large tech companies, selling computing power temporarily unused by internal training or product calls is a natural choice to improve asset utilization.

The problem is that this also shows that construction speed has become so fast that there is a need to actively seek external channels for absorption. Computing power is no longer just about buying as much as you can; it starts to become about whether it can be continuously filled by paid tasks.

Oracle's annual report provides more specific constraints. The document's language shows that as of the end of May 2026, the company has $260 billion in lease commitments not yet started, primarily related to data center arrangements, with terms ranging from 15 to 19 years. Its capital expenditures rose from $21.2 billion in fiscal 2025 to $55.7 billion in fiscal 2026, mainly used to expand data centers.

These figures cannot directly prove industry oversupply. Risk disclosures in public company annual reports are inherently conservative. However, they correspond to the most vulnerable aspects of AI infrastructure investment: capital expenditures occur first, revenue comes later, electricity and leases are long-term commitments, while customer demand can change more quickly.

The $8.3 billion valuation trades on the ability to sell full capacity

Together's $8.3 billion valuation cannot be simply explained by AI hype. The implicit assumption is that the company not only secures orders but can also turn open-source inference demand into long-term revenue with sufficiently high utilization rates, stable renewals, and good gross margins.

Another key concept is megawatt capacity. Megawatts are the power budget of a data center, determining how many GPUs it can support. Locking in capacity means the company has secured the electricity and data center resources needed for future expansion, but it does not mean these resources are already deployed or filled by paid tasks.

For AI cloud companies, capacity is a double-edged sword. Failing to secure electricity and GPUs means missing out on demand surges. Securing too much without enough customer demand means depreciation, electricity costs, and lease expenses hit the income statement first.

This is also the difference between Together and large cloud providers. Together's advantage lies in focusing on open-source inference, where customers may prioritize cost, speed, and model choice. Large cloud providers, on the other hand, have enterprise customers, full-stack services, and stronger balance sheets.

A more likely scenario is that open-source inference demand continues to grow, and specialized players like Together achieve high growth. Meanwhile, some large-scale cloud and data center investments may see lower returns than early market expectations due to contract mismatches, customer concentration, or slow utilization ramp-up.

Utilization rates determine the winners of this infrastructure cycle

AI infrastructure has not yet entered the evidence stage of a bubble bursting, nor can it be judged that demand is infinite just because of Together's funding. A more reasonable state is that the industry is shifting from the resource-grabbing phase to the phase of verifying the ability to monetize resources.

The variables the market will focus on are becoming more specific. Funding scale and valuation only indicate that capital is willing to bet; they cannot replace utilization rates, renewal rates, gross margins, and customer structure. If orders mainly come from well-funded early-stage AI companies, demand elasticity will be stronger. If customers can solidify into long-term production environments, valuation support will be more solid.

Meta's cloudification attempt will provide the market with a price reference. When hyperscalers sell their internal computing power externally, the pricing power and differentiated services of external AI cloud companies will be tested. Oracle's long-term commitments will continue to remind investors that electricity and data center resources are not free options.

Together's latest funding round shows that open-source inference demand is still growing. But for investors, the judgment has shifted: AI computing power is not valuable just because it is built; it only becomes a true infrastructure asset when it is continuously used with high gross margins.

Click to learn about the open positions at BlockBeats

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia