The numbers that define the AI investment race are changing. GPU shipments and Big Tech capital expenditure alone no longer capture the scale or structure of capital flowing into the sector. AI companies are borrowing from banks, data center developers are raising money against long-term compute contracts, and global asset managers and private capital are increasingly treating AI compute capacity as a financeable asset.
The latest number is $29.6 billion. TikTok parent ByteDance has secured a three-year, $29.6 billion unsecured loan from around 30 banks, about 48% above its initial $20 billion target. Chinese banks accounted for more than 60% of the deal, with lenders from the United States, Europe and Singapore also participating, while Citigroup and JPMorgan jointly coordinated the financing.
What stands out is the absence of collateral. “It is very rare to see such a mega loan unsecured,” a person involved in the transaction told Reuters. Extending $29.6 billion without specific assets pledged against the loan suggests that global banks are placing substantial value on ByteDance’s corporate credit and future cash flows.
The stated purpose of the financing is general corporate use, although the market expects a significant portion to support AI-related investment. ByteDance has been expanding access to chips for AI inference while also emerging as a major buyer of long-term compute capacity from data centers under development in Southeast Asia. As competition in AI models extends into GPUs, data centers and long-term compute contracts, the capital required to compete is rising with it.
$69.6 Billion Puts Banks at the Center of AI Investment
ByteDance’s $29.6 billion financing is the second-largest loan in Asia this year. The largest is the $40 billion bridge loan arranged by SoftBank to expand its investment in OpenAI. Together, the two transactions total $69.6 billion, showing how large-scale AI financing is moving beyond bond issuance and equity funding and onto commercial banks’ balance sheets.
SoftBank’s financing structure is more complex. The company has committed more than $60 billion to OpenAI and related AI infrastructure and has also arranged a $20 billion margin loan backed by its stake in Arm. It has pursued additional financing linked to its OpenAI holdings, bringing the valuation of AI companies and the provision of credit by financial institutions increasingly close together.
The structure works as long as AI company valuations continue to rise and data centers maintain high utilization. If expected AI revenues fall short or asset values decline, however, the risk does not remain solely with technology shareholders. Banks, bondholders and data center financiers become exposed to the same investment cycle.
A $45 Billion Contract Turns Compute Capacity Into a Financial Asset
Debt alone does not reveal the full scale of AI financing. British AI infrastructure company Nscale has entered into a six-year, $45 billion compute capacity agreement with Anthropic to provide AI computing from a data center in West Virginia. Nscale is also seeking roughly $3.5 billion in additional financing.
The critical asset in this structure is not simply the data center building. It is the long-term contract promising future demand and payments for computing power. Anthropic’s commitment creates a prospective cash flow for the data center, which can in turn support new financing from investors and financial institutions.
Long-term compute contracts are therefore beginning to function as part of the financial foundation for AI infrastructure development. They connect future demand for computing directly with the capital required to build the facilities that will provide it.
Nvidia is taking this structure to a much larger scale. The company is working with global financial institutions including Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on an AI infrastructure financing platform intended to mobilize more than $500 billion in third-party capital. The initiative is designed to connect GPU supply, finance and data center development within a broader AI investment ecosystem.
Nvidia’s role can therefore no longer be measured solely by the number of GPUs it sells. By helping connect customers that need computing infrastructure with global pools of capital, it is extending its influence into the process through which new AI capacity is financed and built. The boundary between the AI semiconductor supply chain and capital markets is becoming increasingly blurred.
The Next Number to Watch Is Not CapEx, but Committed Compute Capacity
Markets have largely measured the AI investment cycle by adding up annual capital expenditure from Microsoft, Amazon, Alphabet, Meta and other hyperscalers. That measure is becoming incomplete. A fuller picture increasingly requires corporate investment + bonds + bank loans + data center project finance + leases + long-term compute contracts + supplier guarantees.
One figure deserves particular attention: committed compute capacity, the computing power companies have already contracted to purchase or use in the future. ByteDance is participating as a long-term buyer of capacity from Southeast Asian data centers, Anthropic has entered into a six-year, $45 billion agreement with Nscale, and Nvidia’s financing platform is targeting more than $500 billion in AI infrastructure capital.
These transactions differ substantially in structure and scale, but they share one underlying assumption: demand for AI computing will be considerably larger in the future than it is today. The durability of the investment cycle will increasingly depend on whether that assumption is translated into actual utilization and revenue.
Four indicators can help determine whether it is. They are compute utilization, revenue relative to long-term payment commitments, debt maturity and the ultimate holder of risk. These measures shift attention from headline investment announcements toward actual usage, cash flow and financing structures.
The final measure is particularly important for financial stability. If AI demand falls below expectations, it matters whether losses ultimately sit with banks, bond investors, private credit funds, data center operators or GPU suppliers. Only by identifying that final holder of risk can the movement of AI investment exposure into the wider financial system be properly assessed.
AI Capital & Capacity Map
ByteDance | $29.6B
Three-year unsecured bank loan → AI chips and data center compute capacity → global banks assume corporate credit exposure
SoftBank | $40B
Bridge financing for OpenAI investment → AI equity and infrastructure → focus shifts to the maturity mismatch between short-term financing and long-term AI assets
Nscale–Anthropic | $45B
Six-year compute capacity agreement → future data center cash flows → long-term contracts support infrastructure financing
Nvidia AI Infrastructure Financing Platform | $500B+
Global asset managers and private capital → AI infrastructure financing → expansion of GPU and data center compute capacity
Capital Providers → AI Companies → Data Centers & Infrastructure → Compute Capacity → AI Services
Viewed through an AI Capital & Capacity Map, the common structure behind these transactions becomes clearer. Different forms of loans, long-term contracts and financing platforms are moving capital toward the creation of additional compute capacity. Banks, asset managers, private capital, data center operators and semiconductor companies are becoming part of the same financial network.
The four headline figures, however, should not be treated as equivalent measures of investment. ByteDance’s $29.6 billion and SoftBank’s $40 billion represent financing, the Nscale–Anthropic figure represents a long-term compute contract, while Nvidia’s $500 billion-plus figure refers to the capital targeted by a broader infrastructure financing platform. Maintaining that distinction is essential to avoid overstating the total size of AI financing.
Korea Is Already Part of This Financial Network
For Korea, the shift has direct industrial implications. As AI infrastructure competition expands beyond HBM and advanced semiconductor manufacturing into data centers, power grids, cooling systems and finance, capital expenditure by Samsung Electronics and SK hynix alone can no longer capture the scale of the country’s AI investment. Bank lending, policy finance, data center project financing, long-term power agreements and compute purchase commitments increasingly need to be considered together.
As global data center investment expands, capital requirements are rising not only for GPUs but also for transformers, transmission networks, cooling equipment and land. Korea’s AI competitiveness will therefore depend on more than semiconductor production capability. The ability to finance data centers and power infrastructure quickly and reliably is becoming another condition determining the speed of AI infrastructure expansion.
This also makes it increasingly difficult to treat AI industrial policy and financial policy as separate domains. Semiconductor production, data center construction, power supply and finance are becoming parts of a single investment chain. A bottleneck at any point can slow the expansion of the entire AI infrastructure system.
The First Warning Light in the AI Cycle May Appear in Finance
There is no basis at present to interpret the surge in AI financing itself as evidence of an approaching financial crisis. Demand for AI computing remains strong, and global financial institutions continue to provide capital aggressively. The more important question is how much contracted compute capacity is actually being used and whether the cash flows it generates can cover the cost of financing.
ByteDance’s $29.6 billion unsecured loan is more than another large corporate financing transaction. It signals that commercial banks are moving deeper into the financing of the AI investment race. In the first stage of the AI buildout, competitive advantage depended heavily on how quickly companies could secure GPUs and data centers.
The next stage will depend increasingly on who financed that compute capacity, who actually uses it, and who ultimately bears the risk if demand falls short of expectations. The indicators used to read the AI industry must change accordingly.
It is time to put committed compute capacity alongside Big Tech CapEx — and banks’ and infrastructure investors’ balance sheets alongside technology companies’ earnings statements.
