The ComHow Three Deals Worth Tens of Billions of Dollars in a Single Week Are Reshaping the AI Infrastructure Landscape

Over the past week, the global compute sector has been hit by a dense barrage of heavy bombshells. Tencent leased 100,000 AI chips from Oracle for $7 billion. Anthropic's compute agreement with SpaceX doubled to $84.5 billion. SpaceX plans to raise $40 billion to buy Nvidia chips. Broadcom is financing more than $50 billion for a custom chip project for OpenAI. Within just a few days, a cluster of ultra-large-scale deals has pushed the AI compute arms race to a new level of intensity. These deals are not merely a stack of numbers. Together, they point to one clear signal: the AI arms race has moved beyond the crude stage of "hoarding GPUs" and entered an all-round infrastructure competition centered on how compute is acquired, how power is supplied, and how efficiently systems operate. Whoever controls high-density data centers, ready-to-use substations, and grid-connection permits will be the most sought-after picks-and-shovels provider of this era.

Renting Is More Expensive Than Buying: A Dual-Track Compute Acquisition System Is Taking Shape

The most symbolic deal of the week was the five-year compute leasing agreement between Tencent and Oracle. According to disclosed information, Tencent will obtain the services of about 100,000 advanced AI chips through multiple Oracle data centers in Southeast Asia. The deal is valued at about $7 billion, with about 30% to be prepaid. This is the largest overseas compute leasing deal in Tencent's history, and the first time a Chinese tech company has locked up advanced compute resources overseas at such scale. The background to this deal is the structural bottleneck in compute supply facing Chinese tech giants. Tencent President Liu Chiping admitted at an earnings call that the reason capital expenditure was lower than expected was that "the supply of advanced GPUs in China is constrained," and that the company "hoped to buy more cards, and ultimately made up some compute demand through leasing." In the second quarter of 2026, Tencent's capital expenditure reached RMB 52.78 billion, up 176% year on year, while free cash flow at one point was negative RMB 13.8 billion. In other words, it was not that Tencent lacked the money to buy chips, but that it could not buy enough chips. Leasing became a realistic path around supply restrictions. But leasing is not cheap. Seven billion dollars for five years of services from 100,000 chips works out to an average annual cost of about $14,000 per chip, and that does not include additional expenses for data egress, network bandwidth, and software adaptation. From a financial perspective, renting is more expensive than buying. But when supply is constrained, "expensive" is no longer the first concern. "Having it" is.

On the other side of the ocean, Anthropic has demonstrated another scale of compute lock-in. Its IPO prospectus disclosed that the total value of its compute leasing agreement with SpaceX could reach as high as $84.5 billion, nearly double the previously disclosed figure of about $45 billion, with the agreement running through 2029. Anthropic expects to spend at least $518 billion on AI infrastructure over the next decade. What does that number mean? It exceeds the annual GDP of most countries. Anthropic is not simply "renting compute"; it is "buying time"—using long-term contracts to lock up computing resources for years to come, ensuring that model training and inference will not be interrupted. SpaceX, for its part, is using this to transform itself from a rocket company into a compute infrastructure provider. Its Starlink network and ground data centers are forming a global compute network. On one side, Chinese companies are obtaining advanced compute indirectly through leasing. On the other side, American AI companies are locking up computing resources for years through massive long-term contracts. The global AI compute supply chain is forming a subtle "dual-track" structure—different acquisition methods, but the same underlying compute hunger.

More noteworthy is that compute leasing is becoming financialized. Thirty percent prepayments, five-year contracts, subleasing clauses, and the embryonic form of compute futures—the emergence of these financial instruments means that compute is shifting from a technical resource into a tradable, priceable, and securitizable asset class. In the future, there may be compute banks, compute exchanges, and even compute-backed mortgages. When compute becomes an asset, whoever controls its pricing power controls the seigniorage of the AI era.

AI Agents Are Rewriting the Underlying Logic of Compute Demand

Behind these enormous deals, the deeper driving force comes from a fundamental change in the form of AI applications. Tencent's urgency for compute is directly related to the explosive growth of its AI agent product WorkBuddy. Launched in March 2026, the platform crashed immediately because traffic far exceeded expectations. By the time its open platform launched in September, it had brought in more than a hundred ecosystem partners. The key point is that the compute consumption of agent products is fundamentally different from that of traditional conversational AI. A complex agent workflow may involve dozens or even hundreds of steps of reasoning, tool calls, and file operations. Behind every user request may be dozens of times the compute overhead of a traditional dialogue model. For example, when a user asks an agent to "analyze this earnings report and generate a PowerPoint," the agent must first parse the file, extract data, call calculation tools, generate charts, write copy, and design the layout. Each step is a model inference, and the output of each step becomes the input for the next. This kind of chain inference does not increase compute consumption linearly; it magnifies it exponentially.

This is not an isolated case. A recent research report by the Wall Street investment bank Bernstein pointed out that the AI cloud computing sector is entering a new stage in which "an explosive expansion of AI compute demand dominated by the AI agent frenzy, an upgrade in AI inference workloads, and a redistribution of platform value" are happening simultaneously. Data supports this judgment: a report released by the China Academy of Information and Communications Technology predicts that in 2026, inference compute will account for 70.5% of total AI compute load, and AI compute demand is shifting further from model training toward inference and agent operation. What does this mean? In the training era, the compute competition was about "who has more GPUs and can run them longer." In the inference and agent era, the compute competition is about who can process massive concurrent tasks at lower unit cost and higher system efficiency. The CPU-to-GPU ratio in data centers is gradually evolving from the traditional 1:8 to 1:4, 1:2, and even 1:1. The focus of compute infrastructure is shifting from the sheer number of chips to system-level coordination across interconnect, power supply, and cooling. In the past, a data center might have needed only a few GPU servers to complete a training task. Now, an agent platform may need thousands of servers online simultaneously, conducting inference, calls, and communication at every moment. This load characteristic places extremely harsh demands on network latency, bandwidth, and stability. Traditional Ethernet is being replaced by InfiniBand and RoCE, copper cables are being replaced by silicon photonics and optical modules, and air cooling is being replaced by liquid cooling. Every link in the chain that is upgraded becomes a new battlefield in the compute war.

From Chips to Systems: The Competitive Dimension of Compute Infrastructure Is Escalating

This trend was fully reflected in another main thread of the week. Nvidia's stock price hit a record high, and its market capitalization returned to $5.8 trillion, just one step away from $6 trillion. But what is more noteworthy is not the market cap itself, but the direction revealed by its product cadence. The DGX Spark 64GB will go on sale on October 23 starting at $4,999. A single unit can run a 100-billion-parameter model, and two units can support 200 billion parameters. This marks high-end compute trickling down to small and medium-sized enterprises and developers. Rubin Ultra is confirmed to ship next year as planned, with comprehensive upgrades in cooling and HBM design. Nvidia is shifting from "selling chips" to "selling systems," and from "selling compute" to "selling an ecosystem." The pricing strategy of DGX Spark is especially intriguing: $4,999, less than the price of a high-end gaming PC, yet capable of running a 100-billion-parameter model. This is not just hardware sales; it is cultivating developer habits and locking in the CUDA ecosystem for the next decade.

Meanwhile, the OCP Global Summit, which opens on October 12, has the theme "Scaling Innovation for the AI Era," focusing on AI interconnect, 800 VDC power supply, and liquid cooling technology. The summit's core topics—Scale-Up interconnect breaking through copper cable limits, power architecture evolving toward 800 VDC, and large-scale deployment of liquid cooling—all point in the same direction: competition in compute infrastructure is rising from the chip level to the system level. What does 800 VDC power supply mean? Traditional data centers use 54V power supply, which involves high current, high losses, and astonishing copper usage. 800 VDC can reduce losses several times over, cut copper usage by more than half, and support higher power-density racks. Liquid cooling solves the heat dissipation bottleneck of high-density GPU clusters. Cold-plate liquid cooling and immersion liquid cooling are moving from the laboratory to large-scale deployment. In interconnect technology, the transmission distance of copper cables at rates above 800G is shrinking sharply, making silicon photonic modules and co-packaged optics inevitable choices. These technologies are not isolated; together they form the system architecture of next-generation AI data centers.

This is also clearly reflected in China's domestic compute sector. Sugon's Sugon 8000 (Dengfeng) fully domestic 100,000-card AI supercluster was publicly unveiled on CCTV. It uses diamond-copper thermal modules, with thermal conductivity more than twice that of pure copper, chip heat transfer efficiency increased by 80%, and cost only 10% to 20% that of pure diamond. DeepSeek open-sourced infrastructure components for Huawei's Ascend compute platform, covering compilation tools, computing libraries, and distributed communication libraries, corresponding one-to-one with its open-source components for Nvidia platforms. China's domestic compute ecosystem is accelerating from "usable" to "easy to use." In the past, the bottleneck for domestic chips was not only hardware performance but also the software ecosystem. Incomplete compilation tools, insufficient operator libraries, and low-efficiency communication libraries meant that the same model ran slowly and unstably on domestic chips. DeepSeek's open-source move is equivalent to replicating the most core software layers of Nvidia's ecosystem onto the Ascend platform one by one. This is more strategically significant than simply stacking hardware. Sugon's diamond-copper cooling, meanwhile, breaks through the physical limits of high-density heat dissipation at the materials level, providing foundational support for the stable operation of a 100,000-card cluster. Domestic compute is pursuing a path of "system-level catch-up": single-card performance may still lag behind, but through coordination across interconnect, cooling, power supply, and software, overall cluster efficiency is rapidly approaching the international advanced level.

Power and Grid Connection—The Invisible Battlefield of the Compute War

If chips are the front line of the compute war, then electricity is the ammunition depot. Another hidden thread this week is that the power bottleneck is becoming the greatest constraint on compute expansion. In multiple dense data center regions in the United States, grid-connection queues are already several years long. Power capacity in Virginia's "Data Center Alley" is critically short, and although Texas has an independent grid, the stability of power supply under extreme weather remains a hidden danger. Whoever controls ready-to-use substations and grid-connection permits holds the key to compute expansion. This is also why Microsoft, Google, and Amazon are investing in small modular reactors and fusion startups—they are not buying electricity, but buying the certainty of "uninterrupted power for the next decade."

The advancement of 800 VDC power architecture is precisely a technical response to this bottleneck. Higher voltage means lower current and less copper loss, and it also means data centers can be built closer to power sources, reducing transmission losses. Liquid cooling reduces a data center's PUE from above 1.5 to 1.1 or even lower, allowing the same kilowatt-hour to produce more effective compute. Electricity costs as a share of data center operating costs are climbing from 20% toward 40%. In the future, competition in compute will increasingly resemble competition in the electrolytic aluminum industry—whoever has lower electricity prices and more stable power supply will gain a cost advantage in the compute market. The reason SpaceX could sign an $84.5 billion compute agreement is not only its Starlink network and ground stations, but also its independent energy supply capacity and rapid deployment capability. In the compute war, energy autonomy is strategic autonomy.

A Competition Defined by Returns on Compute

Looking back at this week—from Tencent's $7 billion lease to Anthropic's $84.5 billion long-term agreement, from Nvidia approaching a $6 trillion market cap to DeepSeek's 10-trillion-parameter training plan—all signals point to the same conclusion: the essence of AI competition is shifting from a contest of model capabilities to a contest of infrastructure capabilities. Over the past two years, the arms race was about pre-training: tens of thousands of GPUs running heads-down for a few months could produce results. In the second half of the arms race, victory will be determined by the speed of building high-density data centers, the ability to guarantee power supply, and the level of system-level efficiency optimization.

Compute is no longer just a question of "how many cards to buy." It is a systems engineering question of "by what method, at what cost, within what time frame, and how much effective compute can be delivered to the model." And this competition has only just entered its most critical stage. Over the next twelve months, we will see more deals worth tens of billions of dollars, more cross-industry alliances, and more technological路线 forks. One thing is certain: there is no retreat in the compute war, only advance.