Cisco has become the latest major company to announce purpose-built routing hardware to connect AI workloads distributed across multiple facilities, as the race to dominate AI data center interconnect technology becomes increasingly intense.

The networking giant announced its 8223 Routing System on October 8, introducing what it claims is the industry’s first 51.2 Tbit/s fixed router specifically designed to link data centers running AI workloads.

At its core is the new Silicon One P200 chip, which represents Cisco’s answer to the challenge that is increasingly constraining the AI ​​industry: What happens when there’s no room to grow?

A three-way battle for scale supremacy?

By the way, Cisco is not alone in recognizing this opportunity. Broadcom launched its first salvo with its “Jericho 4” StrataDNX switch/router chip in mid-August. It started sampling and also provided 51.2 Tb/s of total bandwidth backed by HBM memory for deep packet buffering to manage congestion.

Two weeks after Broadcom’s announcement, Nvidia announced its Spectrum-XGS scale-across network. The name is quite cheeky considering Broadcom’s “Trident” and “Tomahawk” switch ASICs belong to the StrataXGS family.

Nvidia secured CoreWeave as a major customer but provided limited technical details about the Spectrum-XGS ASIC. Now, Cisco is rolling out its own components for the scale-across networking market, setting up a three-way competition among the networking heavyweights.

Problem: AI is too big for one building

To understand why multiple vendors are flooding into this space, consider the scale of modern AI infrastructure. Training large language models or running complex AI systems requires thousands of high-performance processors working together, generating enormous amounts of heat and consuming large amounts of power.

Data centers are reaching severe limits not only in available space, but also in the amount of power they can provide and cool.

“Cisco’s Common Hardware Group is an exciting addition to Cisco’s Cisco Common Hardware Group,” said Martin Rand, executive vice president of Cisco’s Common Hardware Group. “AI computing is exceeding the capacity of even the largest data centers, increasing the need for reliable and secure connectivity between data centers hundreds of miles apart.”

The industry has traditionally addressed capacity challenges through two approaches: scale-up (adding more functionality to individual systems) or scale-out (connecting more systems within the same facility).

But both strategies are reaching their limits. Data centers are running out of physical space, power grids can’t provide enough power, and cooling systems can’t dissipate heat fast enough.

This requires a third approach. “Scale-across” refers to distributing AI workloads across multiple data centers located in different cities or different states. However, this creates a new problem in that connectivity between these facilities becomes a significant bottleneck.

Why traditional routers aren’t enough

AI workloads behave differently than typical data center traffic. Training runs generate large, bursty traffic patterns. That is, intense data movement is followed by periods of relative quiet. If the networks connecting your data centers can’t absorb these spikes, everything slows down, wasting expensive computing resources and, importantly, time and money.

Traditional routing equipment was not designed for this. Most routers prioritize either raw speed or advanced traffic management, but it’s difficult to provide both at the same time while maintaining reasonable power consumption. For AI data center interconnect applications, organizations need all three: speed, intelligent buffering, and efficiency.

Cisco answer: 8223 system

Cisco’s 8223 system represents a departure from general-purpose routing equipment. Housed in a compact 3-rack unit chassis, it delivers 64 ports of 800 Gigabit connectivity. This is currently the highest density available for fixed routing systems. More importantly, it can process more than 20 billion packets per second and scale interconnect bandwidth to 3 exabytes per second.

This system features deep buffering capabilities enabled by the P200 chip. Think of a buffer as a temporary holding area for your data, like a reservoir that receives water during heavy rains. When AI training causes traffic spikes, the 8223’s buffer absorbs the spikes, preventing network congestion that would slow down expensive GPU clusters sitting idle waiting for data.

Power efficiency is also an important benefit. As a 3RU system, the 8223 achieves what Cisco describes as “switch-like power efficiency” while maintaining routing functionality. This is important if your data center is already straining your power budget.

The system also supports 800G coherent optics, enabling connectivity between facilities over up to 1,000 kilometers. This is essential for geographically distributing AI infrastructure.

Industry adoption and real-world applications

Major hyperscalers have already implemented this technology. Microsoft, an early adopter of Silicon One, recognized the value of this architecture across multiple use cases.

“The common ASIC architecture makes it easy to scale from initial use cases to multiple roles in DC, WAN, and AI/ML environments,” said Dave Maltz, technical fellow and corporate vice president of Azure Networking at Microsoft.

Alibaba Cloud plans to use P200 as the foundation for extending its eCore architecture. Dennis Cai, vice president and head of network infrastructure at Alibaba Cloud, said the chip “enables expansion into the core network, replacing traditional chassis-based routers with clusters of P200-powered devices.”

Lumen is also exploring how this technology fits into its network infrastructure plans. Dave Ward, Lumen’s chief technology officer and head of product, said the company is “exploring how the new Cisco 8223 technology fits into our plans to enhance network performance and deploy superior service to our customers.”

Programmability: Future-proof your investment

One aspect of AI data center interconnection infrastructure that is often overlooked is adaptability. AI networking requirements are rapidly evolving, with new protocols and standards emerging regularly.

Legacy hardware typically requires replacement or expensive upgrades to support new features. The P200’s programmability addresses this challenge.

Organizations can update silicon to support new protocols without replacing hardware. This is important when there is significant capital investment in individual routing systems and AI networking standards are in flux.

Security considerations

Connecting data centers hundreds of miles apart creates security challenges. 8223 includes line-rate encryption with post-quantum resiliency algorithms to address concerns about future threats from quantum computing. Integration with Cisco observability platforms provides detailed network monitoring to quickly identify and resolve issues.

Can Cisco compete?

Cisco faces established competition as Broadcom and Nvidia have already established themselves in the scale-across networking market. However, the company brings advantages including a long-standing presence in enterprise and service provider networks, a mature Silicon One portfolio launched in 2019, and relationships with major hyperscalers already using its technology.

The 8223 will initially ship with open source SONiC support, and IOS XR will also be available in the future. The P200 will be available on multiple platform types including modular systems and the Nexus portfolio.

The flexibility of this deployment option can be crucial as organizations seek to avoid vendor lock-in when building a distributed AI infrastructure.

It remains to be seen whether Cisco’s approach will become the industry standard for AI data center interconnection, but the fundamental problem all three vendors are addressing – connecting distributed AI infrastructure efficiently – will become increasingly pressing as AI systems continue to expand beyond the limits of a single facility.

The true winner may ultimately not be determined by technical specifications alone, but by which vendor can offer the most complete ecosystem of software, support, and integration capabilities around their silicon.

See also:

Want to learn more about AI and big data from industry leaders? Check out the AI ​​& Big Data Expo in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events such as Cyber ​​Security Expo. Click here for more information.

AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.

According to a new report from Red Hat, 89% of businesses are yet to realize customer value from their AI initiatives. However, organizations expect their AI investments to increase by 32% by 2026.

The survey found that AI and security are top IT priorities for UK organizations over the next 18 months, with 62% of respondents saying they need them. These are followed by hybrid or multi-cloud strategies and virtualization, providing a clear direction in the technology focus of UK businesses.

However, the path to AI integration is not easy. The majority of organizations face barriers to adoption, with high implementation and maintenance costs being a top concern for 34% of respondents. Data privacy and security issues are a close second, bothering 30%, and 28% struggle to integrate AI into existing systems.

A particularly interesting finding is the prevalence of “shadow AI,” with 83% of organizations reporting misuse of AI tools by employees. This suggests a disconnect between official IT strategy and employees’ daily work, which can lead to security risks and inefficiencies.

To overcome these challenges, UK organizations are increasingly turning to open source software. The survey found that 84% of respondents believe enterprise open source is important to their AI strategy, with similarly high numbers for virtualization, hybrid and multicloud, and security.

Joanna Hodgson, UK Country Manager at Red Hat, said: “This year’s UK survey results demonstrate the gap between ambition and reality. Organizations are investing heavily in AI, but few are currently delivering customer value. In the transition from experimentation to sustainable production, enterprise knowledge and integration with enterprise systems must pave the way to realizing value from AI.”

“Openness is a driver for greater collaboration, sharing best practices, and enabling flexibility. Just as with successful hybrid cloud investments, open source will continue to be the foundation for making AI more usable and reusable.”

The study also looked at specific areas of AI that are being prioritized. Agent AI, including systems that can operate with a high degree of autonomy, is a top priority for 68% of respondents. This is followed by a desire to enable broader workforce adoption and operationalization of AI.

The skills gap remains a persistent challenge, and AI is the most pressing concern for the second year in a row. In the field of AI, the talent shortages are most acute in agent AI, the ability to effectively use AI capabilities, and the ability to educate the broader enterprise on how to use AI.

Despite these domestic challenges, there is strong optimism about the UK’s place on the global AI stage. 83% of respondents believe the UK is either already a global AI powerhouse or has the potential to become a global AI powerhouse within the next three years. However, this confidence is undermined by a lack of talent pipeline, lack of public funding and lack of private sector involvement, which are seen as key factors preventing the UK from extracting value from AI.

The report also addresses the complexity of cloud adoption, which is further complicated by the integration of AI workloads. Internal silos, sovereignty concerns, and uncertain returns on investment remain barriers. In response, UK organizations are prioritizing operational control and autonomy, securing their software supply chains and maintaining flexibility in their choice of IT suppliers.

Hans Roth, SVP and GM of EMEA at Red Hat, commented, “Organizations want better operational control and IT resiliency to adapt to a world of continued disruption. The survey results and our daily conversations show that sovereignty figures prominently in enterprises’ ongoing cloud strategies and emerging AI opportunity agendas.”

“Open source is central to this change because it provides companies with the transparency and flexibility to innovate quickly without compromise.”

Red Hat’s latest research shows that while the UK is ready to harness the potential value of AI, it also struggles with the practicalities of implementation, skills shortages and the complexity of the technology environment. A strong focus on open source suggests a pragmatic approach that supports collaboration and flexibility in pursuing AI-driven innovation.

See also: Samsung’s tiny AI model defeats giant inference LLM

Want to learn more about AI and big data from industry leaders? Check out the AI ​​& Big Data Expo in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events such as Cyber ​​Security Expo. Click here for more information.

AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.

A new paper by Samsung AI researchers explains how small-scale networks can beat large-scale large language models (LLMs) at complex inferences.

In the race for AI supremacy, the industry mantra is often “bigger is better.” Tech giants have spent billions of dollars creating increasingly larger models, but with Tiny Recursive Models (TRMs) they can take a fundamentally different and more efficient step forward, according to Samsung SAIL Montréal’s Alexia Jolicoeur-Martineau.

TRM achieves new state-of-the-art results on notoriously difficult benchmarks such as the ARC-AGI intelligence test using a model with just 7 million parameters, less than 0.01% of the size of leading LLMs. Samsung’s research challenges the common assumption that the only way to evolve the capabilities of AI models is at full scale, and provides a more sustainable and parameter-efficient alternative.

Overcoming scale limitations

Although LLMs have shown remarkable ability to generate human-like text, their ability to perform complex multi-step inferences can be weak. Because each token generates an answer, one mistake early in the process can derail the entire solution and invalidate the final answer.

To alleviate this, techniques such as thought chains have been developed in which the model “thinks out loud” to solve the problem. However, these methods are computationally expensive, often require large amounts of high-quality inference data that may not be available, and may still produce flawed logic. Even with these enhancements, LLM struggles with certain puzzles that require complete logical execution.

Samsung’s efforts are built on a recent AI model known as Hierarchical Reasoning Models (HRM). HRM introduced a new method using two small neural networks that recursively approach the problem at different frequencies and narrow down the answer. Although it showed great promise, it was complex, relying on uncertain biological arguments and complex fixed point theorems whose application was not guaranteed.

Instead of HRM’s two networks, TRM uses a single small network that recursively improves both the internal “inference” and the proposed “answer.”

The model is given a question, an initial guess of the answer, and a potential inference function. First, we iterate several steps to refine the potential inference based on all three inputs. This improved inference is then used to update the prediction of the final answer. This entire process can be repeated up to 16 times, allowing the model to incrementally correct its mistakes in a parameter-efficient manner.

Counterintuitively, this study found that a small network with only two layers can achieve much better generalization than the four-layer version. This size reduction seems to prevent the model from overfitting. This is a common problem when training on small, specialized datasets.

TRM also eliminates the need for complex mathematics used in previous versions. The original HRM model required the assumption that the function converges to a fixed point to justify the training method. TRM bypasses this completely by simply backpropagating through a fully recursive process. This change alone significantly improved performance, increasing the accuracy of the Sudoku-Extreme benchmark in ablation studies from 56.5% to 87.4%.

Samsung model beats AI benchmarks with fewer resources

The results speak for themselves. On the Sudoku-Extreme dataset using only 1,000 training samples, TRM achieved a test accuracy of 87.4%, significantly higher than HRM’s 55%. In Maze-Hard, a task of finding a long path through a 30×30 maze, the TRM score was 85.3% and the HRM score was 74.5%.

Most notably, TRM has made significant progress in the Abstraction and Reasoning Corpus (ARC-AGI), a benchmark designed to measure true fluid intelligence in AI. With only 7 million parameters, TRM achieves an accuracy of 44.6% on ARC-AGI-1 and 7.8% on ARC-AGI-2. This outperforms HRM using a 27M parameter model, and even outperforms many of the world’s largest LLMs. For comparison, the Gemini 2.5 Pro’s ARC-AGI-2 score is just 4.9%.

The TRM training process has also been streamlined. The adaptation mechanism called ACT (which determines when the model has improved its response enough to move on to new data samples) has been simplified to eliminate the need for a second, costly forward pass through the network during each training step. This change was made without significant difference in the final generalization.

This study from Samsung presents a compelling argument against the current trajectory of ever-expanding AI models. This shows that by designing architectures that can iteratively reason and self-correct, we can solve very difficult problems with only a fraction of computational resources.

See also: Google’s new AI agent rewrites code to automate vulnerability fixes

Want to learn more about AI and big data from industry leaders? Check out the AI ​​& Big Data Expo in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events such as Cyber ​​Security Expo. Click here for more information.

AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.