The arms race for large model computing power intensifies! OpenAI and Anthropic spark an arms race. Under the tech giants' 'computing power anxiety,' who will be the ultimate winner?

Overnight, OpenAI CEO Sam Altman led a more than 20-minute live online broadcast,Unveiling the company's most powerful image generation model to date—Images 2.

According to various evaluation benchmarks, Images 2’s performance is indeed impressive. Overseas renowned AI evaluation agency Arena.ai noted that OpenAI’s GPT-Image-2 has topped all Image Arena leaderboards and leads second-place Google Nano-banana-2 by a record-breaking 242 points in the text-to-image field.

OpenAI’s moat in terms of computing power reserves is becoming increasingly prominent, while Anthropic, recently constrained by computing bottlenecks, has been forced to impose strict throttling on users at all levels. The market generally expectsThat the market narrative will accelerate its shift, further confirming OpenAI’s 'computational dominance' logic.

And to break the deadlock, Anthropic will inevitably increase its procurement contracts for computing power,As Anthropic has committed to purchasing over $100 billion worth of computing power from $Amazon (AMZN.US)$ AWS within the next decade; and $Broadcom (AVGO.US)$ 、 $Alphabet-C (GOOG.US)$ Signed a chip procurement contract, potentially worth hundreds of billions of dollars; with $CoreWeave (CRWV.US)$ Entering into multi-year data center computing power leasing agreements.

More importantly,This industry development provides perfect backing for the upcoming earnings season of tech giants:In the harsh reality of 'computing power equals survival,' the necessity for big companies to maintain high AI capital expenditures is self-evident, and they will be more convincing when explaining their massive investments to the capital market.

Under such extreme computing power anxiety, who is the ultimate winner in this arms race?This article will also analyze it one by one for fellow investors.

1. Core computing power chips and customized chips

Without a doubt, computing power chips are at the heart of the entire feast. Beyond the dominance of general GPUs, the demand for customized chips is exploding exponentially.

$NVIDIA (NVDA.US)$： As the absolute leader in the computing power field, NVIDIA’s B200/GB200 chip architecture remains the cornerstone for major cloud vendors and AI labs. Under the logic of 'dominating through computing power,' large firms cannot afford to skimp on core training clusters, making NVIDIA's pricing power unshakable.

$Broadcom (AVGO.US)$& $Marvell Technology (MRVL.US)$： As expected, Anthropic may sign a massive contract with Broadcom. Cloud service providers (AWS, Google) are aggressively ramping up investment in self-developed chips (such as AWS Trainium, Google TPU) to reduce reliance on NVIDIA and optimize inference costs. Broadcom and Marvell, leaders in ASIC (Application-Specific Integrated Circuit) design and network communication chips, are set for continued explosive growth.

$Advanced Micro Devices (AMD.US)$：As NVIDIA's most powerful 'alternative' or complement, its MI300/MI300X series will continue to benefit from the market's desire for diversification in the compute power supply chain.

The 'bottleneck' of foundry and advanced packaging

No matter how well compute power chips are designed, they still need physical manufacturing. Under current technology, advanced process and packaging capacity are even scarcer resources than GPUs themselves.

$Taiwan Semiconductor (TSM.US)$As the 'main gatekeeper' of global AI compute power.Almost all of NVIDIA’s GPUs and Broadcom’s ASICs rely on Taiwan Semiconductor's 3nm/4nm advanced processes and CoWoS advanced packaging technology. As long as the compute power race continues, Taiwan Semiconductor’s orders can be seen extending years into the future.

Cloud giants and new forces in compute power leasing

The ultimate delivery form of compute power is cloud services, and companies that control massive compute clusters will become the new 'digital landlords'.

$Amazon (AMZN.US)$AWS&$Alphabet-C (GOOG.US)$Cloud:As a major provider of compute power for large model companies like Anthropic, these CSPs (cloud service providers) not only earn substantial cloud computing rents but also solidify their ecosystem moats by making investments to deeply bind top AI enterprises.

Emerging GPU Cloud: $CoreWeave (CRWV.US)$、 $NEBIUS (NBIS.US)$、 $IREN Ltd (IREN.US)$ Companies like this, leveraging close ties with NVIDIA and focusing on optimizing AI workloads, have aggressively risen between tech giants to become dark horses in the infrastructure space. Notably,On April 10, Anthropic officially announced the leasing of CoreWeave’s AI computing power to support Claude,The two parties reached a multi-year data center leasing agreement to address the growing demand pressure from AI services.

4. Data Center Infrastructure: Optical Communications, Power, and Cooling

The expansion of computing clusters has posed unprecedented challenges for data centers in terms of internal transmission speed, power supply, and cooling capabilities.

Optical communication components:When tens of thousands of GPUs are connected, data transmission latency becomes the biggest bottleneck. The demand for 800G or even 1.6T silicon optical modules will explode alongside the expansion of AI clusters. For example, $Lumentum (LITE.US)$ 、 $Coherent (COHR.US)$Wait.

Power and liquid cooling infrastructure: AI data centers are super 'power hogs.' The electricity consumption of high-density racks renders traditional air-cooling technology completely ineffective, directly driving a full transition to liquid cooling. Companies providing power management and precision cooling solutions will see long-term benefits. For example, $Vertiv Holdings (VRT.US)$ 、 $GE Vernova (GEV.US)$Wait.

PreviouslyAI is in a 'super cycle'! Energy, computing power, storage... are expected to benefit. This investment overview is worth collecting!A previous article outlined related concept stocks, interested fellow investors can click to view.

Summary

Overall, in this 'computing power arms race' ignited by large model giants, the dominance of front-end models may alternate,but the underlying 'water sellers' are the long-term winners with the highest certainty.

From the design of core chips, advanced wafer manufacturing that controls the bottleneck, to cloud leasing that manages computing power distribution, and further to optical communication and liquid cooling power facilities that support massive cluster operations, the entire AI infrastructure industry chain is entering an unprecedented period of performance explosion.

For fellow investors, understanding the 'computing power anxiety' of tech giants and holding onto these four core infrastructure segments undoubtedly means grasping the most certain investment theme in this AI transformation.

However, opportunities often come with risks.Fellow investors should remain calm when positioning themselves in the AI infrastructure sector and closely monitor potential risks: overvaluation and volatility, commercialization slowdowns, capital expenditure reductions, geopolitical tensions, and supply chain vulnerabilities.

In summary,the long-term certainty of the AI computing power industry chain is beyond doubt. However, while participating in this capital feast, one must be wary of short-term overheating risks. It is recommended to build positions gradually on pullbacks and avoid blindly chasing highs.

136

652K Views