Google TPU Reaches Out to Neoclouds; Nvidia's AI Cloud Stronghold No Longer Secure

The core judgment of this article's analysis is: Google pitching TPUs to Neoclouds appears on the surface to be selling chips, but in reality, it's leveraging emerging cloud service providers to crack Nvidia's monopoly in the AI computing market. This isn't just simple business development; it's a "flanking infiltration" by cloud giants in the AI chip battlefield.

The news comes from a report by The Information: Google is actively pitching its self-developed TPUs to Neoclouds (i.e., emerging cloud service providers primarily focused on AI businesses). These Neocloud companies typically rely entirely on Nvidia GPUs for their compute rental services. Google's move is essentially an invitation directly to Nvidia's "core user base"—offering lower costs and a more open ecosystem in exchange for cooperation.

But there's deeper logic behind this. Neoclouds aren't traditional cloud vendors; they don't build general-purpose cloud platforms but instead provide GPU clusters specifically for AI training and inference. These enterprises are highly dependent on the Nvidia CUDA ecosystem and hardware supply, while Nvidia's GPUs have long been in short supply and expensive. Google has spotted exactly this pain point.

Why Neoclouds Became Google's Breakthrough Point

Neoclouds occupy a special role in the market. Unlike AWS or Azure, they don't have massive in-house chip capabilities, nor do they build their own data centers like large model companies. Their value lies in quickly providing large-scale GPU clusters, allowing small and medium-sized clients to avoid the hassle of hardware procurement and maintenance.

However, Neoclouds face a structural problem: the acquisition cost of Nvidia GPUs is extremely high, and supply is directly controlled by Nvidia. When Nvidia prioritizes large customers, Neoclouds often can't get enough stock. Nvidia is also competing for customers directly through its own cloud services (like DGX Cloud). This squeezes Neocloud margins and makes their business models fragile.

Google's TPU fills this gap perfectly. TPUs are AI accelerators designed by Google specifically for frameworks like TensorFlow, offering high efficiency in training large models. Although early TPU ecosystems were closed, serving only Google's own business, Google has gradually opened up TPU cloud services in recent years, supporting mainstream frameworks like JAX and PyTorch. More importantly, Google is willing to offer TPU compute to Neoclouds at lower prices, and may even provide customized solutions.

[!note]

The core of Google's strategy isn't selling chips, but selling "compute + ecosystem." If Neoclouds adopt TPUs, they can obtain more economical compute than Nvidia GPUs while reducing dependence on a single supplier. This is classic "using force against force."

Where Are the True Advantages of Google TPU?

Many people think TPU's disadvantage is that its ecosystem isn't as mature as CUDA. But Google's TPU v5p and the upcoming Trillium series have already shown astonishing performance in specific scenarios. According to Google's public data, TPU v5p offers twice the performance compared to the previous generation when training large language models, with higher energy efficiency. For Neoclouds, energy efficiency means lower electricity costs, which is a significant expense in data center operations.

Additionally, Google TPUs are deeply integrated with Google Cloud, offering seamless storage and network expansion. If Neoclouds use TPUs, they can directly access Google's global network and data center infrastructure, which is less hassle than building their own GPU clusters.

But the most critical factor is pricing strategy. Rental prices for Nvidia H100 GPUs have remained stubbornly high, with market rates once exceeding $3 per hour. Google TPU pricing is typically 30%-50% lower. For Neoclouds, if they can replace some Nvidia GPUs with TPUs, they can significantly improve profit margins, or even attract customers with lower prices, gaining a price war advantage.

Google also offers flexible options like "reserved capacity" and "committed use discounts," helping Neoclouds smooth out cash flow. These details show that Google has deeply studied the business model pain points of Neoclouds.

How Will Nvidia Counterattack?

Nvidia won't sit idly by. Its moat has three layers: the CUDA ecosystem, hardware supply chain, and customer stickiness. Each layer takes time to break.

First, the CUDA ecosystem. Although TPUs support mainstream frameworks, many AI models are optimized for CUDA, requiring extra engineering effort to migrate to TPUs. Neocloud customers might not want to pay for migration. However, Google can invest resources in providing migration tools and free technical support to lower switching costs.

Second, the hardware supply chain. Nvidia's GPU shipments are huge, giving it a clear supply chain advantage. While Google develops TPUs in-house, production depends on TSMC and is prioritized for its own cloud services. If Neoclouds purchase TPUs on a large scale, Google needs to coordinate capacity, which could impact its own business. But Google is expanding TPU chip production and plans to sell to third parties.

Third, customer stickiness. Nvidia has signed long-term contracts with many Neoclouds and provides software optimization. But Nvidia is also pushing its own cloud services, creating competition with Neoclouds. If Nvidia becomes too aggressive, it might push Neoclouds to seek alternatives.

[!example]

A recent example is CoreWeave—a major Neocloud company that secured a large number of GPUs from Nvidia but has also started testing other chips. Google's TPU pitch caught exactly this "rebalancing" psychology.

Trend Prediction: Chip War Enters New Phase of "Cloud Penetration"

In the next year, the AI compute market will see more "non-Nvidia" options. Google TPUs won't replace Nvidia overnight, but they will become an important supplement for some Neoclouds.

Original link: https://www.ithome.com/0/976/547.htm