xAI Colossus Computing Center Becomes a 'Money Printer', Musk Insists on Short-Term Lease Terms
The construction cost of Musk's xAI Colossus computing center is only $3-4 billion, yet it can be leased to Anthropic for $45 billion over three years, allowing for a payback period of about four months.
The second Colossus II costs about $17 billion and is expected to generate $50 billion in rental income in one year.
Musk emphasizes that xAI insists on signing only short-term leases, allowing for the retrieval of computing power at any time, highlighting the extreme scarcity of top-tier AI computing power globally. While chips can be stockpiled with funds, power supply has become the real bottleneck, and money cannot easily solve this issue.
Source: Public Information
ABAB AI Insight
xAI has previously rapidly built the Colossus supercluster, and this massive leasing contract with Anthropic continues its high-return strategy of "build first, lease later" while maintaining short-term lease terms to reserve space for future self-use (Grok training) and higher-value clients.
In terms of capital strategy, xAI is positioning self-built computing power as a core cash flow asset, quickly recouping funds through short leases while retaining flexibility. Resources are focused on power infrastructure, next-generation GPU procurement, and optimizing its own training stack, aiming to gain pricing power during a window where AI training demand far exceeds supply.
The transformation of companies like Core Scientific from mining to AI computing power leasing, along with the current shortage of new hardware like Nvidia's GB200, indicates that xAI is in an explosive phase of transitioning from a computing power builder to a high-margin AI infrastructure operator.
Essentially, this is a concentration of capital: computing power leasing generates astonishing returns, driven by the rigid demand from AI giants for top-tier clusters far exceeding supply, with power becoming a more scarce strategic resource than chips. This accelerates the concentration of capital towards players with stable power and large-scale GPU clusters, pushing AI infrastructure from decentralized construction to super-scaled, vertically integrated evolution.
ABAB News · Cognitive Law
The real profit comes not from selling chips, but from selling the combination of power and computing capacity that can run models immediately. Short-term leases are not conservative; they are a readiness to allocate computing power to higher-value uses at any time. Chips can be stockpiled, but when power is unavailable, that is the true moat.