Ramez Naam warns Texas power freeze forces off-grid AI compute
- Texas froze grid access after data center power requests reached five times total state capacity.
- Global chip supplies will demand 200 gigawatts by 2030, doubling predicted U.S. grid additions.
- AI hyperscalers are deploying off-grid turbines and solar batteries to bypass four-year utility queues.
Texas ran out of power for artificial intelligence.
Governor Greg Abbott halted all new data center grid approvals after state regulators were overwhelmed by power requests. The Electric Reliability Council of Texas recorded 1,800 proposals totaling 474 gigawatts of requested electricity - more than five times the state grid's peak demand record. On This Week in AI, Tax GPT CEO Kash Ali noted that data centers now make up 90 percent of all new power requests in Texas, leaving developers stranded without clear alternatives for large-scale land and low-regulation expansion.
Money alone cannot solve the physical limits of an electrical grid. Monopoly utilities have little incentive to accelerate construction under cost-plus regulatory models that reward bureaucratic compliance over rapid delivery.
The state freeze reflects a national capacity bottleneck. Speaking on Moonshots with Peter Diamandis, energy analyst Ramez Naam pointed out that global chipmakers will deliver over 200 gigawatts of AI compute demand by 2030, but the U.S. electrical grid is on track to add barely 100 gigawatts. Interconnection wait times across the country expanded from 15 months to 45 months, locking new data centers out of public power lines until 2031 or 2032.
Hyperscalers are bypassing state utilities by purchasing behind-the-meter energy generation directly. Heavy-duty 400-megawatt natural gas turbines now carry a seven-year manufacturing backlog, prompting jet engine manufacturers to convert production lines into mobile power generators. Naam explained on Moonshots that a single one-gigawatt data center costs $50 billion, with $35 billion spent on chips alone - making electricity costs secondary to immediate power availability.
Alternative clean energy infrastructure offers a faster path around grid delays. Paired solar and battery installations in sun-rich regions can deliver steady continuous output in under 12 months. Naam cited a one-gigawatt continuous facility in Dubai that pairs five gigawatts of solar panels with 19 gigawatt-hours of battery storage, costing six dollars per watt of capital expenditure compared to 15 dollars per watt for traditional nuclear plants.
Geographic shifts will inevitably redefine the global distribution of compute power. As state moratoriums and utility queues freeze domestic expansion, tech companies are evaluating deployment in countries like Australia, Mexico, and the UAE where land and solar radiation are abundant. Foreign jurisdictions that update intellectual property protections and clear regulatory paths for behind-the-meter generation will capture the majority of future AI capital expenditure.
The frontier of artificial intelligence belongs to whoever builds power fastest.
Source Intelligence
- Deep dive into what was said in the episodes
200GW Hiding in Grid, Sodium Batteries 10x Cheaper, Wave-Powered Datacenters w/ Ramez Naam | EP #280 • Aug 15
- Ramez Naam notes that energy availability, not electricity cost, is the primary bottleneck for AI. A one-gigawatt data center costs fifty billion dollars, but chips account for thirty-five billion, making power costs a minor fraction of the total budget.
- Ramez Naam highlights that the US grid interconnection queue wait time has extended from fifteen months to forty-five months over the past twenty years. Monopoly utilities are structurally optimized for regulatory compliance rather than building transmission infrastructure quickly.
- Ramez Naam states that Texas's ERCOT grid peaks at eighty gigawatts but faces over two hundred gigawatts of highly speculative load requests. Large power requests submitted today in Texas face delays, pushing potential connection dates to 2031 or 2032.
- Ramez Naam estimates that projected global GPU manufacturing will demand two hundred thirty gigawatts of power by 2030. However, the US grid is only on track to add one hundred gigawatts of capacity, leaving a severe deficit of powered data center shells.
- Ramez Naam points out that hyperscalers are bypassing grid delays by buying behind-the-meter generation. Consequently, heavy-duty four-hundred-megawatt natural gas turbines are sold out with a seven-year backlog, prompting jet engine developers to pivot into mobile generator manufacturing.
- Ramez Naam argues that unlocking grid flexibility offers the fastest relief for power-constrained data centers. Integrating batteries to time-shift loads and utilizing flexible, interruptible power agreements for just one hundred hours per year can unlock one hundred gigawatts of transmission capacity.
- Ramez Naam highlights that combined solar and battery plants are now the fastest energy infrastructure to deploy. A one-gigawatt continuous output plant in the UAE utilizes five gigawatts of solar and nineteen gigawatt-hours of batteries, costing six dollars per watt.
- Ramez Naam notes solar costs decline by approximately thirty percent with every cumulative doubling of manufacturing capacity. This learning curve, dictated by Right's Law, has dropped terrestrial solar panel costs to just eight cents per watt.
- Ramez Naam notes that small modular fission reactors promise lower costs through standardized factory manufacturing instead of custom field construction. However, commercial deployment is unlikely before the early 2030s, as first-of-a-kind nuclear designs historically suffer significant delays.
- Ramez Naam outlines three primary fusion designs: Tokamaks, lasers, and magnetic reverse field configurations. Helion leads aggressive timelines with a fifty-megawatt power purchase agreement with Microsoft for 2028, while most fusion startups target commercial viability in the early 2030s.
Also discussed on this episode: (2)
Space (2)
- Ramez Naam views orbital data centers as a viable long-term hedge against terrestrial regulatory and grid delays. However, launching just one gigawatt of space-based compute requires six times SpaceX's best annual launch capacity, rendering near-term deployment economically prohibitive.
- Peter Diamandis and Dave discuss Elon Musk's goal to orbit one hundred gigawatts of compute per year by 2030. Achieving this target requires thirty thousand Starship launches annually, which translates to a launch approximately every fifteen minutes.
Why Anthropic is watermarking every word Claude writes | E26 • Aug 13
- Texas Governor Greg Abbott issued a moratorium on approving new data center projects requesting connections to the public power grid. The state is auditing projects to assess ownership, water cooling usage, tax breaks, and grid stability impacts.
Also discussed on this episode: (12)
Models (2)
- Anthropic is watermarking text from its Claude models to comply with the European Union AI Act transparency code. The watermark uses the C2PA open standard and persists even when text is copied, pasted, or edited.
- Will Brick notes that frontier models failed to execute Exa's strategic planning because they lacked undocumented organizational context. Human judgment remains indispensable due to the mass of unrecorded office interactions, personal emotional states, and non-verbal nuances.
Enterprise (1)
- Kash Ali argues that visible watermarks could cause enterprise pushback on service providers. Clients who detect Claude watermarks on deliverables like financial reports may demand lower fees, claiming the work required less human labor.
Big Tech (2)
- Kash Ali notes Google implemented filters to penalize low-quality AI-generated content in search and YouTube results. This shift has successfully revived the search engine optimization value of content with a distinct human touch.
- Jason Calacanis highlights Mark Zuckerberg's hypocrisy in defending data distillation of public text while simultaneously suing startups that scrape Instagram's social graph. When companies are behind in AI, they champion open systems, but protect their own monopolies once ahead.
Open Source (1)
- Mark Zuckerberg argues that the primary risk of artificial superintelligence is power centralization rather than existential threat. Mark Zuckerberg advocates for decentralized, open-source AI to empower individual capability growth and prevent monopolistic control.
Regulation (2)
- Mark Zuckerberg strongly opposes proposal mandates requiring a government committee to review and approve models during a thirty-day window. Instead, Mark Zuckerberg proposes proactive collaboration using shared intermediate training checkpoints and technical staff.
- Kash Ali highlights the irony of local San Francisco ordinances banning flavored vapes and nicotine products while allowing conventional cigarettes to remain widely available. This regulatory environment forces consumers to source alternative products from neighboring municipalities.
Agents (2)
- Josh Serota argues that conventional software seats are dying because agentic operating systems can interact directly with database APIs. Under this model, enterprises will drastically reduce software seat licenses, transitioning humans from manual data entry to managing autonomous agents.
- Kash Ali reveals that Tax GPT uses a hybrid pricing structure combining standard per-seat fees with outcome-based credit pricing. Within six months of launch, automated agent actions already generate one-third of the company's total revenue.
Coding (2)
- Will Brick reports that Exa engineers spend their time managing autonomous agents and rarely write syntax directly, redefining coding as system management. Despite this automation, Exa is actively hiring software engineers to apply human judgment to customer feedback.
- Jason Calacanis argues that because AI has trivialized code generation, a founder's obsessive commitment to their mission is the primary corporate moat. Long-term enterprise buyers value the stability of obsessive, irreplaceable founders over easily replicated software features.