TL;DR

Together AI announced on July 1st the completion of an $800 million Series C funding round, with a post-money valuation reaching $8.3 billion—significantly up from its previous $3.3 billion valuation in early 2025. The company also disclosed that annualized bookings exceeded $1.15 billion in the last quarter, and customers using open-source models on its platform achieved cost savings of 6 to 60 times compared to closed-source model pricing.

On the other side, pressure on AI infrastructure is becoming evident. According to Axios citing Bloomberg, Meta is considering selling access to AI models and surplus compute via a new cloud business. Oracle’s fiscal year-end report through May 2026 also disclosed risks related to long-term data center leases, power procurement commitments, and shifting customer demand.

This contrast brings a critical question to investors: whether AI compute scarcity remains the sole variable. A more realistic test lies in who can consistently fill expensive power, GPUs, and server rooms over the long term.

Together Demonstrates Sustained Demand for Open-Source Inference

Let’s clarify the fundamentals. Training teaches models; inference is the daily operation where models answer questions, write code, handle customer service, and generate content. The former is akin to building a factory; the latter resembles ongoing production after startup.

Together’s growth is primarily driven by inference. It specializes in offering lower-cost AI cloud services powered by open-source models, enabling developers and AI application companies to reduce reliance on proprietary large model APIs. For clients, the key metric is whether per-call costs can continue declining.

Notable customers include Cursor, Cognition, and Decagon—AI application firms. Together reports that Decagon reduced inference costs by approximately sixfold after adopting its platform. The company also cites industry data indicating open-source model usage has tripled over the past 12 months.

This explains why investors are assigning an $8.3 billion valuation. For AI applications to transition from demos to real-world use, inference costs must fall. As long as usage growth outpaces price declines, affordable compute actually amplifies total demand.

CEO Vipul Ved Prakash’s statement captures the prevailing view: intelligence is becoming a foundational resource like electricity, bandwidth, or capital—and open ecosystems enable faster, cheaper innovation. This optimistic thesis underpins continued investment from players like Aramco Ventures and NVIDIA.

However, annualized bookings are not revenue. They reflect order volume and contract momentum rather than cash inflows or guaranteed renewals in future years.

Meta and Oracle Signal the Need to Reassess Capital Recovery Cycles

If one only considers Together, the conclusion might be that AI compute remains undersupplied. But signals from Meta and Oracle indicate infrastructure investment is entering a phase of stratification.

Meta’s reported consideration to sell excess compute and model access externally shouldn’t be interpreted as negative. For large tech firms, monetizing idle internal compute capacity is a natural way to improve asset utilization.

The issue arises when this becomes necessary—indicating that build-out speed has outpaced internal demand. Compute is no longer just about availability; it's now about sustained paid utilization.

Oracle’s annual report provides more concrete constraints. The filing shows that as of May 2026, the company had $26 billion in uncommenced lease commitments, largely tied to data center arrangements with durations spanning 15 to 19 years. Capital expenditures rose from $21.2 billion in FY2025 to $55.7 billion in FY2026, primarily to expand data center capacity.

These figures don’t directly prove oversupply. Public company risk disclosures are inherently conservative. But they highlight the most vulnerable aspect of AI infrastructure investment: capex occurs upfront, revenue follows, while power and lease contracts are long-term obligations, and customer demand may shift faster.

The $8.3 Billion Valuation Reflects the Ability to Fill Compute Capacity

Together’s $8.3 billion valuation cannot be simply attributed to AI hype. Implicitly, it assumes the company doesn’t just secure orders—but can convert open-source inference demand into long-term revenue through high utilization rates, stable renewals, and healthy gross margins.

A crucial concept here is megawatt (MW) capacity. MW represents a data center’s power budget and determines how many GPUs it can support. Securing capacity means locking in future power and space resources, but does not imply deployment has occurred or that the capacity is already fully subscribed.

For AI cloud providers, capacity is a double-edged sword. Failing to secure sufficient power and GPUs risks missing demand surges. Over-acquiring without adequate customer absorption leads to early pressure on P&L from depreciation, electricity, and lease costs.

This is where Together differs from major cloud vendors. Its strength lies in focusing on open-source inference—customers may prioritize cost, speed, and model choice. Major cloud providers, by contrast, benefit from enterprise clients, full-stack services, and stronger balance sheets.

A likely scenario is that demand for open-source inference continues rising, allowing specialized players like Together to achieve high growth. Simultaneously, some large-scale cloud and data center investments may underperform market expectations due to contract mismatches, client concentration, or slow utilization ramp-up.

Utilization Will Determine the Winners of This Infrastructure Cycle

There is no evidence yet that the AI infrastructure cycle is approaching a bubble burst. Nor should we conclude unlimited demand based solely on Together’s fundraising. A more accurate state is that the industry is shifting from a resource acquisition phase to one focused on validating monetization capability.

Market metrics will grow increasingly specific. Fundraising size and valuation signal investor appetite but cannot substitute for actual utilization, renewal rates, gross margins, and customer mix. If orders come mainly from well-funded early-stage AI firms, demand elasticity will be higher. Only if customers evolve into stable production environments will valuations be better supported.

Meta’s cloud initiative will provide a pricing benchmark. When hyperscalers begin selling internal compute externally, external AI cloud providers’ pricing power and differentiated services will face real-world testing. Oracle’s long-term commitments will continue reminding investors that power and data center resources are not free options.

Together’s latest funding confirms that demand for open-source inference remains strong. But for investors, the judgment has shifted: AI compute isn’t valuable simply because it’s built—it only becomes a true infrastructure asset when used sustainably and at high gross margins.

Original: LawDong BlockBeats