The silicon imperative: Rescuing video infrastructure from the energy crisis

For two decades, the video streaming industry has operated under the dangerous assumption that computing power is a commodity as abundant as oxygen. Need to transcode more video? Simply swipe a credit card and spin up more instances. The underlying physical reality — the power plants, the transmission lines, the square footage of the server racks — was someone else’s problem. But this is no longer the case; here’s why.

It’s the end of infinite infrastructure

To understand the urgency, northern Virginia handles roughly 70% of the world’s internet traffic. It is the beating heart of the digital economy. Today, data center vacancy rates there have plummeted to less than 1%. The constraint is not just space; it is power. Utility providers are telling data center operators that new capacity is years away. This scarcity has created a zero-sum game inside the data center. Every kilowatt of power is now a contested resource.

Enter the new tenant: artificial intelligence. Gartner projects that power demand from AI-optimized servers will nearly quintuple by 2030, eventually consuming nearly half of all global data center electricity. This creates a perilous environment for video workloads. In a facility with limited power, the operator will prioritize the tenant paying the highest premium per kilowatt. Currently, that tenant is training a large language model, not transcoding video.

The “let AWS worry about it” fallacy

There is a pervasive, dangerous complacency in the video streaming industry. The belief is that infrastructure scaling is a “solved problem.” The mindset says: Amazon, Microsoft, Google, and the hyperscalers have infinite resources. If power is constrained, they will build more plants. If costs rise, they will absorb them. My only job is to deploy code.

Cloud providers are rational actors. They are businesses, not charities. Faced with finite power and a massive demand for AI compute, they are reallocating their resources. They are retrofitting facilities for high-density GPU clusters suited for AI, often at the expense of the general-purpose CPU capacity that video platforms rely on.

The immediate casualty of this shift is the “spot instance” market, where, for years, video services have utilized Spot Instances sold at deep discounts to keep transcoding costs viable. It was a brilliant arbitrage. You rented the server capacity that no one else was using.

Reports indicate that spot interruption rates are rising and availability windows are shrinking since there is simply no “excess capacity.” If your video workflow economics rely on inefficient software encoding running on thousands of power-hungry CPUs, with a shrinking spot instance market, your bill is about to go up.

The computational mismatch

The root of the problem is architectural. For years, we have processed video using software running on x86 and, more recently, Arm CPUs. CPUs are a marvel of flexibility. But for encoding, using a CPU is profoundly inefficient because video compression is a deterministic, mathematically intensive task. It involves repeating specific calculations, such as Discrete Cosine Transforms and motion estimation, billions of times.

Advertisement

Using a general-purpose CPU for this task is like using a Swiss Army knife to chop down a forest. You could do it, but it will take a very long time, and you will burn a tremendous amount of calories. This inefficiency was manageable when video standards were simpler (like H.264), and power was less expensive. Modern codecs like HEVC and AV1 are popular for reducing delivery costs, but they come at a very high computational cost. Today, smart teams model encoding compute costs to ensure they don’t outweigh the distribution savings.

The silicon solution: Specialized density

The solution is to stop using general-purpose tools for specialized problems. The industry must move to dedicated silicon: the video processing unit (VPU). In 2024, NETINT, along with Meta, Google, and AMD, received a Technical Emmy for the design and deployment of efficient hardware video accelerators for the cloud.

A VPU is an ASIC (application-specific integrated circuit) designed for one purpose: processing video. The VPU strips away the “overhead” silicon found in CPUs and GPUs — the branch predictors, the ray-tracing cores, the texture mapping units — and dedicates every transistor to the video pipeline. A typical ASIC-based transcoder, such as the NETINT Quadra, can process high-density video streams while consuming no more than 20 watts. Meanwhile, a comparable GPU might burn 70 watts or more while delivering lower throughput for video tasks. While a CPU-based server farm could consume kilowatts of power.

In a vacancy-constrained data center, replacing ten CPU servers with a single ASIC-powered server solves the space and power problem. Google did not attempt to reduce YouTube’s transcoding load by buying more Intel Xeon processors. They built a custom ASIC, Argos, called a video coding unit (VCU). To process the 500 hours of video uploaded to the platform every minute, silicon needed to be 20 to 33 times more efficient than standard CPU-based compute. Meta followed suit with their own custom silicon, the MSVP (Meta scalable video processor).

Conclusion: Own your density

The video industry is standing on a precipice. On one side is the demand for more pixels, better quality, and more efficient codecs. On the other side is the hard wall of physical infrastructure constraints. The bridge between them is not more cloud credits. It is efficient silicon.

The “let AWS worry about it” mindset is a relic of a low-interest-rate, energy-abundant past. In the future, the cost of inefficient compute will be passed directly to the tenant. The platforms that survive will be those that decouple their growth from linear power consumption and from general-purpose computing. For video encoding, this means it’s time to look beyond CPUs and GPUs to VPUs.

Mark Donnigan, NETINT TechnologiesMark Donnigan is a technology marketing executive with over two decades of experience in video streaming and digital media. He helped launch VUDU, one of the first video-on-demand streaming services, and contributed to early industry adoption of HEVC and AV1 codecs. He was actively involved in the UltraViolet digital locker initiative, which worked to break down content distribution silos and give consumers device-agnostic access to their video libraries. Mark leads marketing and business development for technology companies focused on video streaming and hardware-accelerated encoding, and speaks regularly at industry conferences on streaming market development and monetization.