The Next AI Bottleneck Is Geography

AI infrastructure has a timing problem. Model builders work in weeks. Utilities work in years. Every gigawatt campus announced this year still has to get through substations, transmission lines and permit queues that were built for a much slower economy. More than 2,000 GW of generation and storage was waiting in U.S. interconnection queues at the end of 2025. Large power transformers now take about 128 weeks to deliver, close to 2.5 years, and transformer prices are up 77% since 2019.

The gap is widening fast. In 2024, Infrastructure Masons predicted the data center industry would double in five years. It doubled in two, and this year it has tripled in three. The IEA expects global data center electricity use to more than double, from 415 TWh in 2024 to about 945 TWh by 2030, roughly what all of Japan uses today. Most of that pipeline sits in the United States, where data centers went from 4.4% of national electricity in 2023 to a projected 6.7-12% by 2028, and it's starting to run into pushback from communities, regulators and grid operators. In Q2 2025 alone, $98B in U.S. projects were blocked or delayed by local opposition, more than every earlier quarter since 2023 combined.

At NeoCloud Summit's "Powering the NeoCloud" panel, Dr. Lamia Youseff of Jazz Computing called it a collision between two value chains. Utility executives order substation equipment seven years in advance because that's how long delivery takes. "We're talking about two different industries that are not on the same time frame," she said.

When the grid can't keep up, compute moves to the power. Some builders secure generation first and buy the land around it. Others go behind the meter. More and more are looking outside the US to Latin America, the Nordics and the Middle East. The numbers back it up: Middle East capacity is expected to more than triple, from about 1 GW in 2025 to 3.3 GW within five years and Latin American colocation inventory grew 20% in a single year, a regional record, with 42% of the pipeline already pre-leased. Forrest Heath, who builds GPU capacity in Colombia, summed up the logic on the same panel: "It's much easier to transmit bits than it is to transmit electrons."

Buyers have followed. A year ago, most AI companies wouldn't seriously consider capacity outside the US. Now the question is simply whether 25MW is available and whether it can grow. Location matters less than speed, as long as the operator takes care of the complexity.

Sovereignty adds another layer. Open-source models are making the model layer a commodity, but governments still care where a model was trained and where its inference runs. Deloitte expects more than $100B to be committed to sovereign AI compute in 2026. Today about 90% of AI compute is run by U.S. and Chinese companies, and Deloitte expects the share run by everyone else to double from 10% by 2030. Gartner puts sovereign cloud IaaS spending at $80B this year, up 36%. Europe increasingly expects data residency inside the region, and data generation too. Countries that want a share of the AI economy are learning that they need infrastructure on their own soil. Policy will shape the global map of compute as much as price does.

The network is the next pressure point. Most of the industry's attention has gone to traffic inside training clusters. As inference grows, so does traffic between data centers and end users, and the internet was built to connect population centers, often far from where the power is. Physical AI raises the stakes: robotics and autonomous systems need latency around 10 milliseconds, and much of today's internet runs at 50 or more. Bryan Lubin of Moonshot Energy pointed to Spectrum as an early sign of the build-out ahead. Spectrum has since switched on edge compute across more than 1,000 of its network sites, putting GPUs within 10 ms of 500 million connected devices.

Over the next 24 months, expect a harder push for sovereign AI, utilities forced to rethink how they plan, fast growth in markets outside the US, and a new wave of internet exchange points and edge sites. The winners will be the ones who plan power, compute and connectivity together, because the next bottleneck is about where things get built, not just how much gets built.

The OCP Global Summit brought the AI infrastructure community together in San Jose this week, with conversations centered on high bandwidth networking, open Ethernet, optical interconnects, and the growing demands of GPU clusters. From Meta's 102.4T networking architecture to advances in 1.6T optics and liquid cooling, the discussions reflected the pace of infrastructure development needed to support the next generation of AI workloads.

We were on the ground throughout the week, bringing the community together through two events of our own. Our NeoCloud Council Dinner created a more intimate setting for operators and industry leaders to connect, while our OCP Networking Reception, hosted alongside Supermicro and AMD, brought together nearly 600 registered guests from across the AI infrastructure ecosystem. Both gatherings gave us an opportunity to strengthen relationships, exchange perspectives, and continue conversations with the people building and scaling AI infrastructure.

Bay Area Startups Collectively Secured $5B+ in October

Halfway through October, startup funding edged over $5B, despite a slowdown in megadeals. There were only two megadeals this week, Applied Compute and LabelBox

Exits: While the market waits for the Anthropic IPO, M&A is filling in. Today's $6.3B acquisition of Florida-based Varonis by Proofpoint is the largest cybersecurity exit this year and in the top five exits overall in the second half of the year. These large transactions - particularly the all-cash acquisitions - provide a modicum of liquidity to investors, their LPs, founders and employees. Not enough to make secondaries less important, but notable in the absence of a strong IPO market.

Startups: Looking for investors, partners, possible competitors? Best deal in the valley is the Pulse of the Valley weekday newsletter and alerts – get a daily summary of just the activity that is most important to you: prospects for investors, partners and customers. Founders get the alerts and newsletter for just $7/month ($50 value). Check it out and sign up here.

Follow LinkSV on LinkedIn to get early notice of new tech, stay on top of SV funding intelligence, and keep track of the trends, companies and investors impacting the startup ecosystem.

Early Stage:

• General Compute closed a $15M Seed, the first ASIC-first AI neocloud, purpose-built for high-speed inference.

• Phinity Labs closed a $5.2M Seed, building prompt-to-ASIC: the agent infrastructure that enables autonomous chip design.

• Tensor Machines closed a $1.5M Seed, building a physics-informed optimization layer that predicts what each GPU can sustain and guides how it should run. Muneeb Rasool presented at Neocloud Summit (See video below)

• NUVACORE closed an undisclosed amount Seed, building a new class of CPU for the AI era, designed for maximum performance, efficiency, and scale.

Growth Stage:

• Applied Compute closed a $350M Series B, we unlock the latent knowledge inside a company, use it to train custom models, and deploy an in-house agent workforce.

• LabelBox closed a $100M Series E, provides the data, reinforcement learning infrastructure, and managed agent platform that help AI systems learn.

• Suki closed a $90M Series E, transforming clinical conversations into intelligence that powers documentation, coding, revenue cycle, clinical reasoning, and other workflows.

• Albert Invent closed a $55M Series B, transforms how materials are invented through chemistry-native AI.

Upscale

AI clusters are growing faster than the networks connecting them. As GPU deployments expand across racks and data centers, congestion, data movement, and network inefficiencies can leave expensive compute resources waiting for work.

For operators, the priority is keeping GPUs productive, scaling clusters efficiently, and maintaining flexibility as hardware requirements evolve. That is where Upscale fits.

Who they are & what they deliver

Upscale is an AI networking company building open, standards-based infrastructure for large-scale AI training and inference. Its Token Fabric platform brings together scale-up networking, scale-out networking, and AI-native software. Built across silicon, systems, and software, the architecture supports heterogeneous compute environments while helping operators reduce network bottlenecks and manage increasingly complex AI clusters.

Who they serve

AI cloud providers, hyperscalers, enterprise AI infrastructure teams, and operators building GPU-intensive computing environments.

With AI infrastructure expanding across different accelerators and deployment architectures, network performance and interoperability are becoming critical to infrastructure economics. Upscale's approach gives operators a path to scale while preserving hardware choice and operational control.

Explore Upscale to learn how its open networking architecture supports the next generation of AI factories.

// NEXT UP

Your Feedback Matters!

Your feedback is crucial in helping us refine our content and maintain the newsletter's value for you and your fellow readers. We welcome your suggestions on how we can improve our offering. [email protected] 

Logan Lemery
Head of Content // Team Ignite

Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation

If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.

Try Notetaker Free