Venture Bytes #128: Rethinking Power in AI Era

Key Takeaways

• AI growth now depends on electricity supply, not just better chips or larger models.
• Countries building power faster will gain an advantage in the AI race.
• Data centers are waiting for grid approvals even when GPUs are ready.
• Power shortages can slow AI expansion more than technology limits.
• Startups to watch include Crusoe, Radiant (an MVP portfolio company), and Positron.

‍

“Every single data center in the future will be power limited. Your revenues are power limited.”

These words from Jansen Huang at NVIDIA’s GTC 2025 should reframe how we think about the AI race. For years, AI progress has been measured in model size, chip performance, and capital raised. Huang’s comment points to something more basic. The race for AI leadership may ultimately be won or lost on the capacity of the electrical grid.

This becomes particularly consequential in the context of US-China competition. In 2024, China added 429 GW of new power capacity, more than one-third of the entire US grid and more than half of total global electricity growth in a single year. The US, by contrast, is experiencing a surge in AI-driven electricity demand against a grid that is expanding only incrementally. In 2025, the Electric Reliability Council of Texas (ERCOT) region received 233 GW of large-load interconnection requests, with more than 70% of this demand originating from data centers. Yet only 1.3 GW was approved.

Huang’s assessment is already visible at the hyperscaler level. Microsoft CEO Satya Nadella has acknowledged that the company has AI GPUs “sitting in inventory” because it lacks the power capacity necessary to install them at scale.

For years, AI scaling has been framed as a function of model architecture and chip density solely. But Huang's comment points to a more structural reframing. If revenue is power-limited, then AI growth is no longer determined by model innovation or chip throughput alone. It is constrained by physical infrastructure such as generation capacity, transmission buildout, interconnection timelines, and grid stability. That shift makes AI an industrial narrative as much as a software narrative.

“The biggest issue we are now having is not a compute glut, but it’s power – it’s sort of the ability to get the builds done fast enough close to power” – Satya Nadella, Microsoft CEO.

Boosting the US power grid is an enormous and time-consuming task due to a complex web of regulatory, financial and supply chain challenges. Transmission lines often face years of environmental review and land-use litigation. Generation projects must clear interconnection studies that are already backlogged. Financing structures for utilities were designed for predictable, incremental load growth, not for the step-function demand spikes created by AI clusters.

At the same time, supply chain bottlenecks in high-voltage transformers, switchgear, and specialized components have extended equipment lead times to multiple years. Even when capital is available and demand is visible, execution remains slow. For example, PJM Interconnection, which oversees the region housing Northern Virginia's dense concentration of hyperscale data centers, now faces average interconnection timelines exceeding eight years, according to energy think-tank RMI. An eight-year cycle may be compatible with traditional industrial load growth. It is incompatible with an AI cycle measured in quarters and capital commitments measured in trillions.

Grid setbacks have led AI labs and infrastructure players to seek alternatives outside the traditional utility model. These include co-locating data centers directly with generation assets, entering long-term nuclear power agreements, investing in small modular reactor development, and building private microgrid configurations to bypass congested interconnection queues.

The Infrastructure Innovators

A new generation of startups is emerging to address the power bottleneck. Each of these innovators reflects a larger reframing, when the grid can’t expand fast enough to meet AI demand, the response must come from infrastructure creativity.

Colorado-based Crusoe Energy approaches the challenge from the supply side. Founded in 2018, Crusoe is a vertically integrated company, combining energy sources, networking equipment, and GPUs into its own data centers. With its energy-first approach to AI infrastructure, the company builds data centers at stranded energy sources rather than waiting for grid capacity. By co-locating compute power with otherwise wasted natural gas at oil fields and remote renewable installations, Crusoe sidesteps interconnection queues entirely while monetizing energy that would have been flared or curtailed.

Radiant is developing portable nuclear microreactors designed to replace diesel generators for remote locations, data centers, and military installations. The company's Kaleidos reactor represents a fundamental rethinking of nuclear deployment, small enough to be transported by truck, yet capable of powering entire facilities independently of grid infrastructure. Equinix has preordered 20 Kaleidos reactors for data center applications, signaling that hyperscalers view distributed nuclear as a viable path around transmission constraints.

Nevada-based Positron excels at the efficiency front. While most companies focus on power generation and distribution, Positron focuses from the demand side. Its Atlas inference accelerator achieves 93% memory bandwidth utilization compared to 10-30% in GPU-based systems, delivering 3.5x better performance-per-dollar than Nvidia's H100 while consuming 66% less power. This positions Positron well for memory-intensive inference workloads, where GPUs struggle with efficiency, effectively stretching available power capacity further.

Together, these companies represent a broader recognition. When the grid cannot expand fast enough to meet AI demand, the solution lies in either bringing power to compute or making compute dramatically more power efficient. The companies that solve this equation may prove as strategically important as the chip designers and model builders that have dominated AI headlines thus far.

Video Editing Ripe for Agent-Led Disruption

Key Takeaways

• The cost of filming is falling quickly, but editing remains expensive and manual.
• Editors still spend much of their time syncing footage and fixing technical details.
• AI tools assist with captions and effects, but they do not replace the editor.
• Advances in multimodal AI and software control are enabling deeper automation.
• Key startups in this space include Runway, Descript, and Capsule

‍

The marginal cost of filming is declining fast, but the cost of editing is not. Professional video editors spend 60-90% of their time on repetitive tasks such as syncing footage, cleaning audio, resizing formats, color balancing, and organizing media. None of this is creative in the true sense, just workflow management. Current AI tools help with individual steps and the editor still orchestrates every decision.

This is exactly where coding was two years ago. Developers had autocomplete, refactoring suggestions, syntax helpers, but they still stitched everything together. The breakthrough came when agents could reason across the entire codebase and execute multi-step plans. Video editing is approaching that same moment.

Various technological shifts are converging to make it a reality. First of all, multimodal models have quietly crossed a threshold. They can now ‘watch’ video the way language models read documents. Longcontext visuallanguage architectures like NVIDIA’s LongVILA are designed to handle long videos endtoend, using specialized scaling strategies so tens of thousands of frames remain tractable. VideoRAGstyle systems go further, turning multihour footage into an index of transcripts, scenes, speakers, and objects that preserves temporal structure. These architectures have made understanding several hours of footage easy, which is the precondition for any credible autonomous editor.

The second shift is that models can operate software, not just describe what to do. Computeruse capabilities let agents see screens, move cursors, click, type, and manage files inside real desktops. In October 2024, Anthropic released 'computer use', Claude's ability to control computers by viewing screens and operating interfaces like a human. By February 2026, Claude Opus 4.6 achieved 72.7% accuracy on OSWorld, a benchmark measuring real-world computer tasks, higher from Opus 4.5's 66.3%. This matters most for video editing because professional editors already work through visual interfaces. An agent can operate professional editing tools like Premiere Pro, managing files and applying edits like human editors. While limitations exist, such as dragging and zooming, the core capability for editorial tasks exists today.

Third, generative media has moved from what appeared to be a gimmick to credible building block. 2026era systems like Runwayclass and Soraclass models can generate short clips with consistent characters, plausible physics, and controllable styles. Integrated stacks increasingly combine video, voice, sound effects, and ambience in one pass, avoiding earlier sync issues between visuals and audio. For an agentic editor, this enables hybrid workflows: use filmed material as the spine, then synthesize transitions, Broll, overlays, and alternate intros or outros with quality high enough for commercial use.

A fourth shift is the rise of videonative infrastructure that makes footage addressable by machines. Longcontext systems like VideoRAG don’t just process frames but also build retrieval layers over transcripts, detected scenes, speakers, and visual entities, so agents can query multihour footage semantically and temporally. Generative media platforms and infra startups are standardizing ultralowlatency pipelines for transcription, analysis, and generation, exposing consistent APIs instead of bespoke file drops. The State of Generative Media Survey Report 2025 showed 65% of organizations reporting positive ROI in video and image production through AI, reinforcing continued investment in standardized, AIready video pipelines.

Finally, the cost structure of video production now clearly favors automation over manual workflows. Analysis of AI video production show cost reductions of 70-90% versus traditional workflows, with production time falling by 80% for certain internal and marketing content, according to AI media start-up Magic Hour.

But these numbers reflect feature-level automation. The economic impact of true agentic systems would be substantially larger. For instance, in corporate video production, companies need constant content for marketing, training, and communications. There is also a massive opportunity for individual creators. Successful YouTubers and podcasters often employ full-time editors or spend 20+ hours weekly on post-production. Current AI tools might cut that to 10 hours. An agentic system could reduce it to 30 minutes of creative direction plus 30 minutes of review. Accordingly, the global AI video editing market is expected to reach $9.3 billion by the end of 2030 from just $1.6 billion in 2025, growing at a CAGR of 42.2 %.

Several startups are positioned to benefit if agentic video editing becomes a core workflow rather than a feature. Runway is one of the leaders in this space to watch. While it has strategically expanded toward world models and broader simulation use cases across gaming and robotics, video generation and editing remain central to its platform. With customers such as CBS, Google, Lionsgate, AMC Networks, and IMAX, Runway already operates inside professional production workflows. If agentic orchestration becomes viable, Runway is structurally well placed to move up the stack from generative tooling to full creative workflow automation.

California-based Descript has pioneered text-based video editing, a critical stepping stone toward agentic systems. Its November 2022 Series C was led by OpenAI's startup fund, signaling strong conviction in AI-powered editing's future. With clients including major universities, nonprofits, and organizations using it for asynchronous communications, Descript has proven product-market fit in the exact workflows where agentic systems will deliver maximum value.

Backed by the likes of HubSpot Ventures and Bloomberg Beta, Capsule.Video is another promising start-up. The company raised $12 million Series A round in April 2025 to build what CEO Champ Bennett calls an AI 'co-producer that provides suggestions designed to help brands elevate their storytelling capabilities.' The platform's focus on enterprise teams and prompt-based editing positions it as a near-term implementer of agentic workflows for business video production.