Bottom line: Four stories from the same day sketch one picture: AI compute demand is outrunning the physical infrastructure that supports it. AMD’s data center revenue more than doubled to $6.7 billion, proof the GPU buildout is still accelerating. At the same time, Texas — previously one of the easiest states to build a data center in — froze new projects pending audits after its grid operator’s connection queue passed 474 gigawatts, about 90% of it data centers. Two startups are chasing different fixes: EON wants to link data centers with laser-equipped satellites instead of undersea cables, and Runware is shipping data centers as transportable pods instead of years-long mega-campuses. If you’re buying compute or inference capacity in 2026, the bottleneck isn’t chips anymore — it’s power and connectivity, and where you site your workload is starting to matter as much as which model you run.
AMD’s Quarter Shows Exactly Where the Money Is Going
AMD’s data center business is carrying the company right now. Its data center revenue more than doubled year-over-year to $6.7 billion in its most recent quarter, while gaming revenue receded into a secondary line item, according to The Verge. That split is the clearest signal yet of where chipmaker capex and customer demand are actually pointed: AI training and inference silicon, not consumer graphics.
That number matters as context for everything else in this piece. AMD’s growth is demand-side confirmation that hyperscalers and AI labs are still buying compute as fast as they can get it. The three stories below are what happens on the supply side when that demand meets physical limits — limits on power, on grid capacity, and on the bandwidth needed to move data between data centers once they’re built.
Texas Just Told the Data Center Industry “Enough” — For Now
Texas has spent the last several years as one of the most attractive states in the country to build a data center: loose regulation, abundant natural gas, and a grid operator willing to connect almost anyone who asked. Only Virginia hosts more data centers. That era paused on Monday, when Governor Greg Abbott announced that every new data center project will now require an audit from both the Public Utility Commission of Texas (PUCT) and the state’s grid operator, ERCOT, TechCrunch reported.
The Numbers Behind the Freeze
The scale of what forced Abbott’s hand is the real story. In January, ERCOT’s interconnection queue — the list of projects waiting for permission to connect to the grid — stood at 233 gigawatts. In under six months, it more than doubled to 474 gigawatts, per TechCrunch’s reporting. About 90% of that queue is data centers, according to the grid operator, and the total now represents more than five times ERCOT’s total peak demand.
Not all of that will get built — queue positions are cheap to reserve and many projects are paper proposals that will quietly fizzle. But the trend line is what triggered the audit requirement, not any single project. Texas has kept pace so far partly on renewables: utility-scale solar capacity grew fourfold between 2021 and 2025, per the Energy Information Administration, and that expansion helped hold electricity prices down for much of that period, according to a report from energy-data firm Amperon. TechCrunch noted that data center and crypto-mining demand has since started pushing those prices back up, which is the pocketbook version of the same capacity problem showing up for Texas ratepayers.
The practical effect: a state that used to be a fast, low-friction place to stand up GPU capacity now has a regulatory checkpoint in front of every new project. That doesn’t kill Texas as a hosting market, but it removes the “build first, ask later” option that made it attractive to begin with.
Two Startups Betting on Fixing the Bottleneck From Different Angles
Power and grid capacity is one constraint. Getting data between data centers once they’re sited is another. Two same-day funding and launch stories show founders attacking both problems from opposite directions — one from orbit, one from a shipping container.
EON: Betting on Lasers From Orbit Instead of Cables Under the Ocean
Hyperscalers currently move data between global data centers largely over undersea fiber-optic cable — a system that works but is expensive to install, slow to repair, and geographically fixed. Endeavor Optical Networks (EON), founded in May and emerging from stealth with $10.75 million in seed funding from General Catalyst and Andreessen Horowitz, wants to replace part of that with a network of laser-equipped satellites, TechCrunch reported.
The ambition is specific: undersea fiber moves roughly 200 terabits per second, and prior demonstrated space-to-ground laser links — from companies including York, Kepler, and Cailabs — have topped out around 2.5 gigabits per second, nowhere close. EON, led by CEO Charlie Horowitz and CTO Tyler Presser, is targeting 2.4 terabits per second as its starting point, with a planned constellation of about 20 satellites, each capable of linking two continents, giving an initial fleet 24-hour coverage for early customers.
That’s still a lab-stage claim from a three-month-old company, not a shipped product. The reason it’s worth tracking anyway is what it says about where founders think the next bottleneck sits: not in chips, and increasingly not even in power, but in moving the output of one data center to another fast enough to be useful.
Runware: Betting on Portable Pods Instead of Multi-Year Mega-Campuses
Runware is attacking the same capacity gap from the deployment-speed side. The AI infrastructure company launched its Sonic Inference Pod, a modular, transportable data center unit designed to sit alongside hyperscalers’ traditional builds rather than compete with them at their own scale, per TechCrunch’s report. The pods use closed-loop cooling with no water draw and can reportedly be built in days, compared with the months or years a conventional data center takes.
Runware already has 10 pods running across the U.S., Europe, and Asia-Pacific, serving inference workloads for customers including Higgsfield AI and Wix, and says it has 160 candidate sites lined up for further deployment. The company, which raised a $50 million Series A in December to build out image-generation infrastructure, is positioning the pod line as an extension of that same thesis: sell inference capacity, not a single product.
“Demand for inference is growing faster than facilities can be built,” Runware co-founder and CEO Flaviu Radulescu told TechCrunch. “What we want is to power the world’s intelligence, to be the backbone every AI model runs on with capacity that keeps up with demand instead of throttling it.”
Radulescu’s framing is worth taking at face value even from a vendor: distributed, fast-to-deploy compute sited closer to end users is a direct response to exactly the kind of grid bottleneck Texas just ran into. If mega-campuses need years of lead time and regulatory sign-off, pods that ship in days become a meaningful hedge, not just a cheaper alternative.
What This Means for You
If you’re buying GPU cloud time, evaluating an inference provider, or planning where to host an AI workload in 2026, treat these four stories as one signal rather than four:
Regional risk is now a procurement variable. If your current or prospective provider leans heavily on Texas capacity, ask directly how the new PUCT/ERCOT audit requirement affects their build timeline — not whether it does, but by how much.
Distributed and modular providers deserve a second look. Runware’s pod model, and others like it, exist specifically because traditional data center lead times are becoming a competitive liability. If latency and reliability matter more to you than lowest headline price, a distributed-pod vendor may now be a legitimate primary option rather than a backup.
Don’t build near-term plans around orbital laser links. EON’s numbers are compelling on paper, but a three-month-old company promising 2.4 Tbps from orbit is a multi-year bet, not a 2026 procurement option. Track it; don’t depend on it.
Chip supply confirms demand isn’t slowing. AMD’s doubled data center revenue is a reminder that the underlying compute buying spree hasn’t paused — the constraint has simply moved one layer down the stack, from silicon to power and siting.
The near-term takeaway: capacity planning for AI workloads in 2026 needs a power-and-siting column next to the usual price-and-performance comparison, especially for any deployment anchored to a single high-demand region.
This isn’t unique to Texas, either — it’s the sharpest current example of a pattern likely to repeat in other high-demand grid regions as more interconnection queues fill up with data center requests. Treat Abbott’s audit order as a preview of the kind of regulatory friction other states with strained grids are likely to adopt next, not as a one-off Texas story.
For more on how infrastructure spending is shaping the rest of the AI stack, see our coverage of AMD Helios, Etched, and Google’s infrastructure arms race and SpaceX’s $329M bet on Tesla batteries to power xAI.
Sources: The Verge, TechCrunch. Last updated: 2026-08-08.
