Written by: Rita

Trends Guide

Goldman Sachs released its monthly AI Project Pulse report on June 30, tracking AI infrastructure orders from neocloud (new cloud vendors), sovereign clouds, and enterprise sectors. In June, a total of 7 significant transactions were tracked, amounting to nearly $7 billion. This included Poland's sovereign AI cloud at $4.1 billion, India's Yotta sovereign cloud at $2 billion, Vultr's selection of HPE and NVIDIA to deploy the GB300 cluster, Japan's Datasection deploying B200 in Thailand, and the crypto mining firm AiOnX transforming into an AI computing center. Goldman Sachs also raised the global server market size forecast to reflect the surging demand for AI server racks and increased memory costs. SMCI raised $7 billion to digest a backlog of $39 billion in orders. NVDA's Vera Rubin NVL72, the first fully validated rack, has gone live on CoreWeave.

The demand is diversifying, but the money is not slowing down

The biggest change in AI infrastructure is that the people putting up money are becoming more diverse, while the size of the amounts is secondary.

A year ago, such orders were almost exclusively signed by Microsoft, Google, and Amazon. Now, the most active players are two new categories: neocloud and sovereign clouds. Argentum AI secured a $4.1 billion contract to deploy 27,000 GB300 GPUs for a leading AI company, supporting a data center in Poland with over 300MW, scheduled to come online in phases by 2026. Notably, this buyer is a neocloud and does not belong to any major cloud vendor.

India's Yotta sovereign cloud signed a $2 billion deal to purchase 20,736 B300 and 5,120 B200 from Supermicro, with plans to expand to six Southeast Asian countries. Turkey's Odine also entered into a strategic partnership with Supermicro. The sovereign cloud sector is accelerating its rollout and is no longer just a concept.

Crypto mining companies are also getting involved. AiOnX acquired a 77% stake in Genesis Digital Assets for $500 million, repurposing the 1.3GW power from 15 mining sites in Texas, North Carolina, South Carolina, and Sweden for AI and HPC computing power. The power and land that these mining sites have are becoming the most scarce resources for AI computing power.

They secured the contracts, but cash hasn’t followed

All orders eventually need to be converted into physical equipment. SMCI raised $7 billion in June to procure components because it received approximately $39 billion in AI server orders over the past few weeks, covering more than 20 clients, showing a lack of cash flow to lock in upstream supplies. Server integrators are becoming the tightest link in the entire chain.

In contrast, those AI server concept companies that only issued press releases without any actual financing actions or disclosed client numbers have fallen behind at this stage. Each transaction mentioned in Goldman Sachs's report can be traced back to a specific buyer, specific chip quantities, and specific delivery times. For instance, some companies that only signed strategic cooperation intentions without procurement contract details cannot be found in this report.

NVDA's Vera Rubin NVL72 has also transitioned from PPT to physical reality. CoreWeave, in collaboration with Dell, has built the world’s first fully functional Vera Rubin NVL72 rack, with NVDA confirming mass production in the second half of 2026. The setup includes 72 Rubin GPUs and 36 Vera CPUs, with 260 TB/s NVLink 6th generation interconnect, liquid cooling, and a tenfold increase in inference performance per watt.

Japan's Datasection signed a contract worth approximately $250 million to procure 4,696 B200 GPUs to be deployed in a 10MW data center in Bangkok, serving a technology company in the United States. Although the transaction amount is relatively small, its signal significance is strong—AI computing power is becoming a standardized commodity that can be delivered across countries.

Trends Perspective

The fact that SMCI raised $7 billion is worth examining more closely than any single transaction. A backlog of $39 billion in orders sounds good, but if delivery requires locking in upstream components six months to a year in advance, the capital occupation will be extremely exaggerated. The balance sheets of server integrators are bearing the costs of AI infrastructure expansion. Whoever can maintain delivery capacity under such pressure will seize market share.

Another overlooked point is the transformation of mining sites. Converting 1.3GW of power into AI computing power means the base for GPU installation is larger than the surface order data suggests. Mining sites do not need to wait for new wafer production capacity; they just need to change their production lines. This will cause the real consumption of GPUs to exceed market expectations.

For US stock AI investors, the most valuable judgment from this report is: the demand is still there, but the winners are changing from those who obtain the cards to those who can turn the cards into operational computing power. The bargaining power of integrators like SMCI and Dell is rising, while the scarcity of pure GPU manufacturers is being diluted by decentralized procurement channels.

Disclaimer

This article is an organization and interpretation of third-party brokerage research reports by Trends Research. The ratings, target prices, profit forecasts, and related judgments cited in this text are opinions of the analysts of that brokerage, representing only the stance of their affiliated institution, do not represent the views of Trends Research, and do not constitute any investment advice. The market has risks; decisions should be made independently. This article should not be used as the basis for buying or selling any securities.