NVLink 6 and 3 Kilometers of Copper

The Bandwidth Crisis Chip local memory is becoming incredibly fast, but the chips themselves are suffocating in isolation, unable to exchange data quickly enough. For decades, copper connections served as the nervous system of servers, but they have turned into a bottleneck, generating excessive heat and latency that negate the power of entire supercomputers.

Further AI development no longer depends on how many processors we can solder onto a single board. It depends on how we force them to communicate, how we cool them, and, most importantly, how we teach software code to compensate for the physical shortage of silicon. The transition to optical interconnects, the fragmentation of the compute market into closed castes, and the radical software compression of models themselves—these are the changes that determine who gets access to artificial intelligence and who is left on the sidelines of progress.

The Bandwidth Crisis Solution The transition from the NVIDIA H100 generation to B200 saw local memory (HBM) bandwidth jump by 40%, while inter-GPU bandwidth increased by only 15%. Classic copper connections and external optical modules have exhausted their physical limits, turning into thermal generators and sources of latency. For training neural networks requiring the synchronous work of thousands of accelerators, this meager bandwidth increase became an insurmountable barrier.

The solution was the shift to Integrated Silicon Photonics (Co-Packaged Optics, CPO)—a technology that moves data transmission from the radio frequency range to optical directly at the die level. In 2026, over half a billion dollars were invested in this sector. Companies like Lightmatter, Ayar Labs, and NVIDIA (with its Spectrum-X Ethernet Photonics platform) moved photonic interconnects from prototype status to commercial reality.

Key advantages of the photonic shift:

Overcoming Physical Limits: Using light impulses instead of electrons allows for transmission speeds unattainable by copper, almost completely eliminating signal attenuation. NVIDIA Spectrum-6 switches demonstrate 102.4 Tbps bandwidth. Integrating photonics directly into the chip package (instead of external transceivers) radically reduced latency and jitter.
Thermal Regulation and Efficiency: Photonic connections generate significantly less heat. Spectrum-X is estimated to increase network infrastructure energy efficiency fivefold. In the scale of old clusters, up to one-third of all data center energy was spent solely on switch operation and heat removal from endless copper harnesses.

The Death of the Server Shelf, Birth of the Compute Monolith The release of the NVIDIA Vera Rubin architecture forced data center owners to completely overhaul the engineering infrastructure of their buildings. NVIDIA abandoned the concept of selling individual video cards, offering the market ready-made AI factories—vertically integrated racks.

The Vera Rubin POD cluster is a single engineering object weighing 1.8 tons per rack, combining five specialized systems based on the third-generation MGX architecture. To eliminate communication starvation, a two-tier strategy is employed:

Local Connectivity (Inside the rack): Handled by NVLink 6, providing 3.6 TB/s per chip. The total bandwidth of a single NVL72 rack reaches nearly 260 TB/s. This is realized through a complex copper backbone in the form of cartridges containing over 5,000 connections with a total length exceeding 3 kilometers. This approach allows 72 Rubin accelerators to function as a single super-powerful compute organism.
Global Connectivity (Between racks): Beyond a few meters, copper is useless. To unite eight NVL72 racks into an NVL576 domain and further scale to 1,152 GPUs, Ethernet Spectrum-X with integrated silicon photonics is used. This guarantees the preservation of full All-to-All bandwidth when connecting tens of thousands of accelerators within one or several buildings.

Infrastructure Shock The mass adoption of photonic interconnects and ultra-dense Vera Rubin racks required a complete audit of data center engineering systems:

Thermal Management: Thermal density reached its limit. MGX racks are designed for high-temperature water cooling (+45°C input). This opened the door to widespread use of free cooling (outside air), which radically lowers the PUE coefficient and allows placing 10% more compute equipment within the existing power budget.
Power Grid Stabilization: AI cluster loads have a pulsating nature that destroys city power grids. Vera Rubin NVL72 implements an Intelligent Power Smoothing system based on supercapacitors (400 Joules per GPU). The system dampens peak current loads by a quarter, eliminating the need to build giant battery farms and protecting city substations from voltage sags.
Speed vs. Flexibility: Abandoning external cable management infrastructure in favor of modular cartridges reduced node deployment time from two hours to five minutes. However, the flip side was the loss of customization: hyperscalers receive a closed hardware complex where the failure of one part often means replacing the entire expensive module.

The Death of the Monolithic Cloud Universal platforms can no longer satisfy the polar requests of the industry, leading to the formation of a rigid caste system for accessing compute power.

Closed Ecosystems of Giants: Global IT corporations and local leaders transformed from service providers into owners of top-tier digital real estate. They stopped wholesale power sales, reorienting toward their own AI factories.

Engineering Implementation: Hyperscalers are the first to absorb scarce batches of NVIDIA Rubin accelerators, building exclusive photonic infrastructure around them. For them, chip availability is solved at the level of long-term interstate and inter-corporate quotas.
Business Model: Access to bare metal for external customers is either closed or prohibitively expensive. The bulk of transistors are spent training their own foundation models. Third-party companies are offered only expensive, deeply integrated platform solutions where the GPU cost is hidden behind the enterprise service price tag.

Compute as Collateral For the mid-tier tech sector and startups, acquiring their own hardware has become a financial dead end. Against this backdrop, the Neocloud sector has flourished—specialized AI factories focused exclusively on GPU-as-a-Service.

The Merger of Capital and Silicon: Venture funds have changed their risk assessment approach. Now, confirmed access to compute is a critical parameter when analyzing startups. Investors increasingly buy up power from Neoclouds in whole blocks, booking them as liquid assets. This creates artificial scarcity: thousands of GPUs are withdrawn from the free market and reserved for specific investment portfolios.
Survival Economy in Russia: Under sanctions, buying a video card via parallel import costs 2.5–3 times the nominal price, lacking official warranty and NVIDIA software support. As a result, the classic «buy or rent» dilemma is solved in favor of renting. Your own H100 card only pays off under extreme loads (25–30k hours over five years), which is unattainable for most projects. Against the backdrop of high loan rates, renting in Neocloud (340–470 rubles/hour for H100) remains the only viable, albeit expensive, scenario.

Corporate Perimeter and Digital Sovereignty Financial institutions, medical companies, and industrial giants formed a third, isolated contour. In their value hierarchy, security and legal data purity completely displaced the need for record-breaking compute speeds.

Sovereignization and the Triumph of Edge AI Edge computing (Edge AI) ceased to be an auxiliary tool and became self-sufficient autonomous nodes capable of thinking without a connection to the center.

An Optical Chip in a Drone’s Body: Miniaturization reached a level where a full-fledged neural processor with integrated silicon photonics fits into the dimensions of a surveillance camera or an industrial drone’s bay. Micro-PCs learned not just to execute instructions, but to learn on the fly. Equipped with optical interconnects, these devices process video streams and LiDAR data in real-time, consuming watts instead of kilowatts. They can fine-tune models right on-site—in the field, shop floor, or air—working absolutely autonomously without an internet connection.

The Anatomy of Distributed Intelligence AI infrastructure has finally split into three functional layers working in symbiosis:

Processing Centers (Data Centers): Here, in protected perimeters, heavy «mother» models are born and undergo primary training on giant clusters with photonic cooling.
The Periphery (Edge): Their optimized «offspring» arrive here. Using aggressive quantization and distillation, models are compressed dozens of times, retaining 92–95% accuracy, and are deployed on thousands of end terminals—from robotic arms on conveyors to smart traffic lights.
Corporate Buffer: An intermediate layer for storing sensitive data that cannot leave the organization’s perimeter.

A Double Blow to the Infrastructure Crisis The mass exodus of AI to the periphery became a salvation for the industry, solving the resource shortage from two angles:

Offloading the Networks: The classic scheme where every smart city camera streamed 4K video to the cloud would have choked. Edge AI flipped this process: now terabytes of raw data are processed into zeros and ones right on the device. Only the dry residue—a violation tag, object coordinates, or a text summary—flies to central storage. This freed up the bandwidth of backbone communication channels.
Democratization of AI Access: The queue for scarce server GPUs (Rubin) stretched for years. In response, businesses began buying up micro-PCs. Deploying a fleet of thousands of Jetsons is cheaper than renting premium cloud instances, and the shortage of embedded chips is felt much less than the shortage of server accelerators. This allowed mid-sized companies to implement robotics without billion-dollar infrastructure budgets.

Sovereign Intelligence and Russian Specifics In Russian realities, technological necessity overlapped with political ones. Data localization requirements and critical infrastructure protection made Edge AI a no-alternative choice for the state sector and big business. The multipolar economy stimulates import substitution at the silicon level. The demand for hardware that works without looking at foreign clouds gave a push to the development of domestic neural processors for inference. Data no longer migrates to cross-border clouds—it is born, processed, and dies within the national digital perimeter.

Artificial intelligence has distributed across the world in a thin but ubiquitous layer. The cloud remained the «brain» of the system, corporate servers its «memory,» and Edge devices its «nerve endings,» reacting to reality instantly, autonomously, and securely.

The Answer to Physical Deficit While engineers build up data center muscles, erecting monstrous racks with photonic interconnects and liquid cooling systems, machine learning researchers found a way to bypass physics with mathematics.

The answer to the physical deficit of transistors became a «virtual diet» for artificial intelligence. Practical deployments show that modern models can be subjected to extreme compression—shrinking them 20 times while preserving 92% of the original cognitive ability. This achievement changes the economics of the entire industry: businesses no longer need to budget billions of rubles to build their own AI infrastructure.

The Anatomy of Software Efficiency Compressing neural networks is not just removing «extra» weights, but a complex multi-layered process of architecture transformation:

Quantization. Switching to «Broken» Language: Models translate from high-precision 32-bit math to rough but fast 4-bit or 8-bit arithmetic. For most practical tasks (image recognition, text analysis), such accuracy loss is unnoticeable, yet the neural network works several times faster and fits into the memory of ordinary micro-PCs. Tools like GPTQ allow doing this «on the fly» without needing to retrain the model from scratch.
Distillation. Wisdom Without Extra Weight: Instead of using a giant model, developers train a compact neural network to imitate the logic of its «big brother.» This process went beyond simple answer copying: compact «students» now inherit the internal reasoning chains of the «teacher.» As a result, a model with 1–2 billion parameters (Small Language Models) demonstrates logic quality comparable to giants on 7–8 billion, while consuming ten times less electricity.
Architectural Pragmatism: Network designers abandoned the mania for gigantism. Instead of wide and flat structures, deep and narrow architectures came into fashion. They are better adapted to the specifics of Neural Processing Units (NPUs) in smartphones and Edge chips, ensuring stable speed even when processing long documents.

A New Reality for Business This software breakthrough democratized access to high technologies. For small and medium businesses, the need to compete with giants for cloud quotas disappeared. Optimized models allow launching powerful AI services on existing «grey» hardware or inexpensive servers. The entry barrier collapsed: now a company can deploy a local AI assistant or a quality control system on a conveyor using a fleet of micro-PCs, rather than renting clusters from hyperscalers.

This also fueled the triumph of Edge AI. Only thanks to twenty-fold compression did neural networks manage to physically survive on fifteen-watt chips in cameras, drones, and industrial sensors. Intelligence descended to the end-device level, providing instant response where any delay in transmitting data to the cloud is unacceptable.

The Price of Efficiency However, compactness is not given for free. Running a model on a device is only half the battle. The main difficulty lies in MLOps on the periphery: how to automate the update of tens of thousands of different devices, how to guarantee that quantization didn’t break a specific usage scenario, and how to manage a fleet of models without central control. Developers rely on the ONNX Runtime ecosystem and specialized engines like llama, but the engineering complexity of maintaining such a distributed network remains high.

Conclusion Intelligence has distributed across three levels: optical factories of hyperscalers for training, sovereign corporate perimeters for data storage, and optimized Edge AI for execution. Silicon photonics solved the bandwidth problem but made hardware a scarce resource, forcing business to switch to renting from Neoclouds. In response, 20x algorithmic model compression allowed moving compute to the periphery. Ultimately, the winner is not the one who buys more GPUs, but the one who most efficiently distributes tasks between the light highways of data centers and autonomous micro-PCs on the network edge.