If you believe that nearly every startup is currently rushing to raise capital to address the AI infrastructure shortfall, your perception is accurate. Instinct, an AI agent startup, is reportedly seeking a $1 billion funding round to expand its compute capacity, having surpassed 100,000 users through an invite-only model while its processing power runs thin.
Simultaneously, investment firm Coatue Management is in discussions with chip startup MatX to establish a multi-billion dollar joint venture aimed at procuring memory chips, logic wafer production, and securing fabrication plant capacity. The industry's challenge has evolved beyond simply sourcing idle servers and components; securing AI compute now often requires substantial, long-term commitments from companies.
For compute providers, having long-term contracts in place reduces the cost of financing new cloud infrastructure. Similarly, chip manufacturers prefer allocating scarce production capacity to large, financially robust clients. This dynamic makes it increasingly difficult for rapidly growing startups to secure compute resources, as smaller cloud providers frequently need to take on significant debt to fund data center construction and GPU acquisitions. Even industry giants are increasingly relying on debt to supplement their cash reserves.
Additionally, borrowing costs across the economy have risen sharply, with benchmark yields for both long-term and short-term debt climbing in tandem. Companies with established, creditworthy clients can typically finance compute expansion at lower costs. In contrast, building capacity for unproven customers carries higher financing expenses. For instance, Iren disclosed in August that its approximately $3.6 billion financing to procure GPUs for Microsoft carried an interest rate of about 6%, while a separate $2.4 billion GPU financing for non-investment-grade clients required a rate 3 percentage points higher.
Some cloud service providers are also increasing advance payment requirements to reduce the amount of external capital they need to raise. Iren noted that recent prepayments from certain clients for GPUs and related equipment have reached up to 55% of total costs. Similarly, Nebius reported that contracts with prepayment clauses hit an all-time high in the second quarter, with approximately 70% of newly signed agreements requiring upfront payments. Nebius expects to receive over $9 billion in prepayments this year, despite projecting maximum full-year revenue of just $3.4 billion for 2026. The company states that, in a market with tight supply, advance payments have become the industry standard for securing compute capacity.
Startups can rent GPU capacity on demand from providers like Runpod, while larger cloud services such as CoreWeave focus on inference compute procurement. However, this model is expensive and does not guarantee immediate availability of the desired computing power. Regardless of the approach, startups with heavy AI operations face rapidly escalating costs for compute expansion, often needing to lock in resources before they can accurately predict how user demand will grow.
Before the AI boom, consumer internet applications could prioritize user growth without overemphasizing revenue, as compute needs were smaller and cheaper, without requiring such substantial prepayments. The photo-sharing app BeReal serves as an example, raising only about $90 million in 2022 while reaching nearly 8 million daily active users within months, though its growth subsequently stalled.
On the chip manufacturing side, startups also struggle to secure capacity before their businesses take off. Fabrication plants prefer allocating production to clients willing to make significant commitments with strong financial backing, such as Broadcom and Mellanox Technologies. For MatX, which may not have a testable chip design until next year, part of the purpose of this joint venture financing is to lock in chip capacity in a manner similar to major industry players.
The joint venture remains in early-stage negotiations, with no disclosure yet on Coatue's specific investment amount or whether external funds will be introduced. Nevertheless, Coatue brings considerable experience in financing heavy data center infrastructure and has executed innovative deal structures. Its Next Frontier fund previously formed a joint venture with AI infrastructure startup Fluidstack to finance a planned large-scale data center campus in Indiana. A subsidiary of that venture issued $5.7 billion in debt to cover land acquisition, campus construction, and supporting power infrastructure, excluding chip procurement. Fluidstack will lease the completed facilities to offer AI cloud services, with Google providing support for Fluidstack's lease obligations under the agreement.
Comments