Hassan Nasser

The race to build artificial intelligence is moving faster than the infrastructure needed to power it. And some of the world’s most advanced AI chips are reportedly sitting unused, waiting for electricity. For years, the biggest obstacle to developing advanced artificial intelligence was obtaining enough computing power. NVIDIA GPUs became some of the most sought-after hardware in the world, with technology companies spending billions to secure them. But the industry is beginning to encounter a different problem. Companies can acquire the chips. They can build the servers. They can even construct the data centers. What they cannot always do is turn everything on. This emerging problem, known as dark GPUs, reveals a growing mismatch between the speed of semiconductor development and the much slower process of expanding electrical infrastructure. And it raises an important question: What happens when the AI industry builds computing capacity faster than the world can power it? Amazon’s Multibillion-Dollar GPU Problem In October 2026, Business Insider reported that Amazon Web Services (AWS) was acquiring substantial quantities of NVIDIA’s latest AI systems before some of its facilities had the electrical capacity to operate them. According to internal communications reviewed by Business Insider, Amazon expected to receive approximately $5 billion worth of NVIDIA GB300 server racks, followed by roughly $3 billion in additional systems each month. The equipment was being secured through an arrangement called Early Rack Receipt, allowing Amazon to take ownership of completed server systems before they reached their final data centers. However, some AWS facilities reportedly lacked sufficient power to operate the incoming hardware. As a result, certain systems were being stored in intermediate warehouses rather than immediately deployed. Amazon has defended the practice, explaining that staging hardware near deployment sites is a normal part of preparing new infrastructure. There is a legitimate reason for purchasing hardware early: advanced AI systems are expensive, complex, and subject to supply constraints. But purchasing equipment and putting it into productive operation are two very different milestones. A GPU cannot generate computing revenue while sitting in a warehouse. And the longer it remains there, the more expensive that delay becomes. What Exactly Are Dark GPUs? The term dark GPU generally refers to a graphics processing unit or AI accelerator that has been purchased but is not delivering its intended computing capacity. This can happen for several reasons: The distinction between these problems matters. A GPU that cannot be powered on is fundamentally different from one that is operating at only 30% utilization. The first is an infrastructure problem. The second may be a scheduling, software, or demand problem. Both represent unused computing potential, but they require different solutions. In the current AI infrastructure expansion, access to reliable electricity has become one of the most significant concerns. Why Can’t Companies Simply Plug In Their GPUs? An AI server is not comparable to an ordinary desktop computer. Modern AI accelerators operate at extraordinary power densities, and deploying them at scale requires an entire supporting infrastructure. Consider NVIDIA’s GB300 NVL72 system. According to NVIDIA’s technical documentation, a GB300 NVL72 rack has a designed power budget of approximately 135 kilowatts. That is the electrical power budget for a single rack. Now imagine a data center deploying 100 of these racks. The theoretical rack power requirement becomes: 100 × 135 kW = 13,500 kW That’s 13.5 megawatts, before accounting for additional facility requirements such as cooling, electrical conversion losses, and supporting infrastructure. And this is only 100 racks. Large AI installations may require hundreds or thousands of racks, depending on their design. Power at this scale cannot simply be obtained by connecting another cable. It may require new substations, transformers, high-capacity transmission connections, distribution equipment, backup systems, and major cooling installations. All these systems must be engineered, manufactured, installed, tested, and commissioned. While semiconductor manufacturing progresses through rapid technology cycles, electrical infrastructure often operates on much longer development timelines. A new generation of AI processors can arrive before a data center’s electrical infrastructure is ready for the previous generation. That mismatch is becoming increasingly important. The Real Bottleneck Is More Than Electricity Generation When people hear that AI data centers lack power, they often assume the world is simply not producing enough electricity. The problem is more complicated. Electricity may be available at a national or regional level while remaining inaccessible to a particular data center. A facility might need 200 MW of dependable power, but the local grid connection may support only a fraction of that requirement. Increasing capacity could require upgrades across several layers of the electrical network. The challenges include securing power generation, obtaining utility agreements, expanding transmission and distribution infrastructure, manufacturing transformers and switchgear, and ensuring reliable operation under changing AI workloads. Even when sufficient electricity exists somewhere on the grid, delivering it to the correct location at the required capacity can be difficult. This creates an unusual situation. A technology company might have billions of dollars in computing equipment ready for deployment, yet remain dependent on electrical infrastructure that cannot be expanded at the same speed. The bottleneck is no longer exclusively about manufacturing faster processors. It is increasingly about delivering enough usable power to operate them. The Numbers Behind AI’s Growing Energy Demand The scale of this challenge becomes clearer when examining global electricity consumption. According to the International Energy Agency’s 2026 report Key Questions on Energy and AI, data centers consumed approximately 485 terawatt-hours of electricity in 2025. The agency projects that consumption could reach approximately 950 terawatt-hours by 2030. That’s nearly twice as much electricity in just five years. AI-focused facilities are expected to grow even faster, with their electricity consumption projected to triple over that period. However, the IEA also highlights an important distinction: improvements in hardware and software efficiency are reducing the energy required for individual AI tasks. The problem is that AI adoption and the complexity of workloads are growing rapidly. More organizations are deploying large models, running increasingly sophisticated reasoning systems, and integrating AI into applications that operate continuously. Efficiency improvements

A laptop with a powerful NVIDIA GPU is nothing new. A laptop built around an NVIDIA-designed Arm CPU, Blackwell graphics, and a shared pool of up to 128 GB of memory is a different proposition. On October 16, 2026, NVIDIA officially launched RTX Spark laptops powered by its N1X system-on-chip (SoC). The platform brings technologies usually associated with high-end graphics cards and AI workstations into slim Windows 11 computers. But the real engineering story is not simply a larger core count: it is how CPU, GPU, memory, and software are integrated—and what that means for developing AI systems on a personal computer. 1. What is NVIDIA N1X? N1X is the silicon behind NVIDIA’s new RTX Spark Windows PCs. Instead of pairing a conventional x86 processor with a separate NVIDIA graphics card, the platform combines an Arm-based NVIDIA Grace CPU and a Blackwell RTX GPU in one tightly integrated computing platform. NVIDIA says it collaborated with MediaTek on the custom CPU design. The CPU and GPU communicate through NVLink-C2C, a high-bandwidth chip-to-chip connection. NVIDIA reports an interconnect bandwidth of 600 GB/s. Rather than treating the GPU as a peripheral with an entirely separate pool of video memory, RTX Spark uses unified LPDDR5X memory accessible to both processors. That architectural choice is important for AI and content creation, where moving enormous datasets between components can become a bottleneck. 2. The official specifications NVIDIA lists two N1X laptop configurations and a higher-power compact desktop configuration. Specification Higher N1X laptop Lower N1X laptop CPU 20-core NVIDIA Grace 18-core NVIDIA Grace GPU 6,144-core Blackwell RTX 5,120-core Blackwell RTX Unified memory Up to 128 GB LPDDR5X Up to 32 GB LPDDR5X NVIDIA-rated AI compute Up to 1 PFLOP, FP4 Up to 827 TFLOPS, FP4 Listed platform TDP 45–80 W 45–80 W Operating system Windows 11 Windows 11 Both versions support NVIDIA CUDA, fifth-generation Tensor Cores, fourth-generation ray-tracing cores, and current RTX software features. NVIDIA also lists a compact desktop N1X variant with a 140 W TDP. Important: FP4 AI throughput is a specific, low-precision theoretical performance measure—not a prediction of everyday application speed. The number of CUDA cores also does not make N1X equivalent to a desktop GPU with a similar core count: power limits, memory behavior, clocks, and workload matter. 3. Why unified memory matters for local AI Traditional laptops often have system RAM for the CPU and a limited, separate pool of VRAM for the GPU. That separation can force developers to reduce model sizes, offload layers, or transfer data repeatedly. N1X’s unified memory design changes the trade-off. In the highest-memory configurations, a substantially larger AI model can reside in a memory pool shared by the CPU and GPU. Consider a simplified example: a 70-billion-parameter model stored at 4 bits per parameter needs approximately: 70 billion × 4 bits ÷ 8 = 35 GB That is the theoretical weight-storage requirement alone. A real inference session also needs memory for quantization metadata, the key-value cache, model runtime, context, and other processes. Having 128 GB of unified memory could make such workloads more practical without immediately turning to a remote GPU server. This does not mean unified laptop memory behaves like 128 GB of dedicated high-bandwidth GPU VRAM or guarantees fast inference. Memory bandwidth, supported precision, software optimization, and thermal limits remain decisive. The advantage is primarily capacity and integration, not a promise of unlimited AI performance. 4. CUDA is the real strategic advantage For AI developers, the GPU specification is only half the story. The other half is the software ecosystem. RTX Spark supports CUDA, TensorRT, and NVIDIA’s RTX acceleration tools. That opens a path to building and testing supported AI workloads on a portable Windows machine while remaining within the broader NVIDIA development ecosystem. Potential uses span computer vision (detection, segmentation, and video analysis), robotics R&D (testing perception models before edge deployment), generative AI (running compatible language and image models locally), and product design (accelerating supported 3D and creative workloads). A local development machine can improve iteration speed and reduce reliance on paid cloud inference for some experiments. But deploying an application from RTX Spark to Jetson, Linux servers, or other targets still requires platform-specific testing and packaging. 5. What the first independent reviews tell us NVIDIA markets RTX Spark as a machine for AI, creating, and gaming. Early independent testing suggests a more nuanced picture. The strengths: Reviews highlight capable GPU acceleration, unusually large unified-memory options, and attractive local-AI use cases. These are especially relevant when a workflow is constrained by memory capacity or depends heavily on NVIDIA’s optimized GPU stack. The compromises: Reviewers have found that CPU performance and efficiency do not consistently match Apple’s current higher-end M-series chips. Gaming performance is mixed relative to the price, particularly when compared with conventional x86 laptops with dedicated graphics. Windows on Arm compatibility is another practical consideration: some legacy applications, anti-cheat systems, device drivers, and engineering tools may require verification before buying. Pricing also makes this a premium platform. Early US configurations have been reported starting at approximately $2,599, with substantially higher prices for some memory-rich models; prices vary by manufacturer and region. In other words, N1X looks most compelling as a specialized development and creation platform, rather than the automatic best-value option for every laptop buyer. 6. Should engineers and developers consider N1X? For an R&D team, the decision should begin with the workload—not the marketing headline. Ask four questions before choosing an N1X system: For robotics and embedded projects in particular, verify the complete workflow: sensor drivers, device communication, compilation, deployment, and debugging—not just neural-network inference. 7. The bigger picture: AI development is moving closer to the edge N1X is significant because it reflects a wider shift in computing. AI work no longer needs to be divided so rigidly between a low-power personal laptop and an expensive remote accelerator. More models and tools can run directly where engineers design, test, and iterate. The most interesting outcome may not be a faster benchmark. It could be a more practical development loop: capture

When Invisible Data Became Visible Formula 1 is a sport built on invisible margins. Thousandths of a second, microscopic changes in airflow, subtle shifts in temperature — advantages that exist far beyond what the human eye can normally see. But for a brief moment in 2013, those invisible margins became visible to everyone. At the 2013 Italian Grand Prix, Formula One Management (FOM) introduced thermal imaging cameras into the live TV broadcast. The goal was simple: entertainment. Glowing brakes, hot tires, dramatic visuals. What FOM unintentionally created, however, was one of the most powerful competitive intelligence leaks in the history of the sport. 1. The 2013 Incident: A Visual Intelligence Goldmine Contrary to popular belief, engineers in 2013 were not casually watching thermal images on pit-wall monitors. Within minutes of the thermal feed going live, teams were: By mapping the broadcast’s color spectrum to estimated temperature values — while accounting for emissivity (a material’s efficiency at emitting thermal radiation) — engineers could extract far more than the TV audience realized. What Teams Could Decode Tire Hysteresis & Carcass HeatThe temperature delta between tire surface and carcass revealed how a rival car was generating mechanical grip, managing energy input, and loading the tire through corners. Brake Bias MigrationThermal “bloom” on front versus rear brake discs exposed real-time brake balance changes under high‑G deceleration — corner by corner. Aero‑Elasticity & Load DistributionShifts in heat across the tire tread under load hinted at camber gain, suspension behavior, and even how effectively the floor was sealing at speed. This wasn’t passive observation. It was remote reverse‑engineering during a live race. 2. The Science of Thermal Signatures In thermodynamics, every material exhibits a unique thermal signature — the way it absorbs, stores, and releases heat over time. When engineers observe a component through a thermal lens, they are not just seeing temperature — they are seeing: This enables a process known as Inverse Thermal Analysis. Remote Lab Testing, Trackside If a rival’s brake ducts appeared hotter but cooled faster than expected, engineers could infer: In effect, teams were performing non‑contact laboratory experiments on competitors — using nothing more than a TV broadcast. 3. 2026: AI, Computer Vision, and the Death of Secrecy If the 2013 thermal “leak” were to return in 2026, the consequences would not be incremental — they would be exponential. The key difference is not camera resolution or frame rate. The difference is the maturity of Artificial Intelligence, computer vision, and data-driven modeling. In 2013, teams were limited by human interpretation, manual processing, and relatively simple models. In 2026, the entire analysis pipeline can be fully automated, real-time, and predictive. From Observation to Continuous Learning Modern AI systems do not treat thermal footage as isolated images. They treat it as time-series data. By feeding thermal video streams into recurrent architectures such as LSTM (Long Short-Term Memory) networks, models can: This allows teams to predict performance inflection points — such as tire drop-off — several laps before they become visible to the driver or on standard timing data. Turning Heat Into Aerodynamic Intelligence Thermal data is not limited to tires and brakes. By correlating: AI models can estimate airflow efficiency and infer aerodynamic drag behavior. While not a perfect substitute for wind tunnel data, this approach can narrow a rival’s drag coefficient (Cₙ) into a useful confidence range, providing actionable intelligence with zero physical testing. Automated Reverse-Engineering at Scale Modern Convolutional Neural Networks (CNNs) excel at extracting spatial structure from images. Applied to thermal footage, they can: What once required expert intuition can now be done continuously and automatically, across multiple cars, sessions, and race weekends. The Digital Twin Problem The most concerning implication is the rise of AI-generated digital twins. By observing how components retain and shed heat after a run, models can estimate: Over time, repeated observations allow AI systems to build increasingly accurate approximations of how rival components behave internally — without ever physically inspecting them. In this context, secrecy no longer fails because information is stolen. It fails because information is inferred. Every thermal frame becomes a data point. Every lap improves the model. In 2026, unrestricted thermal vision would not be a visual feature — it would be a real-time reverse-engineering interface. 4. Why This Crosses the Ultimate Technical Red Line This is why the FIA is unlikely to ever allow unrestricted thermal imaging to return. It doesn’t just affect performance — it breaks the economic balance of the sport. Why invest: When a $1M AI pipeline can extract usable intelligence directly from a broadcast? In the 2026 era — where power unit behavior, aerodynamics, and energy recovery are tightly coupled — information leakage becomes performance leakage. Every photon leaving the car carries data. And in the age of AI: If a camera can see it, a competitor can calculate it. 5. The Engineering Lesson: Optimize for What Is Measured — and What Is Visible One of Formula 1’s oldest truths applies here: Cars are not optimized for how they are designed — they are optimized for how they are measured. Thermal imaging exposed parameters teams never intended to share, simply because those parameters became observable. For engineers and R&D teams, the takeaway is clear: 6. AI Turns Observation Into Prediction What made the 2013 incident uncomfortable was visibility. What would make a 2026 version catastrophic is prediction. Modern AI systems do not just observe states — they learn behaviors. With enough thermal data, models can: This shifts competitive intelligence from analysis to forecasting — a far more dangerous capability. 7. Why the FIA Will Likely Never Allow This Again The FIA’s role is not just sporting fairness, but economic balance. Unrestricted thermal imaging would: In a cost-capped era, allowing competitors to reverse-engineer each other via broadcast data would undermine the foundations of the regulations themselves. 8. The Strategic Takeaway Formula 1 unintentionally demonstrated a future problem every high-tech organization will face: If your system can be seen, it can be modeled. In an AI-driven world, secrecy