Electricity use in the US is expected to surge to another record high this year, driven by the energy needs of data centers being built across the country. The demand causes a corresponding worry throughout the energy business that the country’s aging grid won’t be able to handle the load. 

The US grid is suffering from decades of neglect. A recent study by the National Laboratory of the Rockies (formerly the National Renewable Energy Laboratory) found that more than half of the grid’s aging transformers are either approaching or are past their 40-year design cycles. On top of that, the demand for new equipment from data centers around the world has created a bottleneck for crucial replacement parts. And new construction of transmission lines has stalled. Across the country, there were only 402 miles of new high-voltage wires in 2025. As recently as 2013, the US added almost 4,000 miles of new lines. 

At a 2025 conference, Calvin Butler, the president and chief executive of Exelon, the largest utility provider in the US, said the grid was like “a car on the verge of breaking down.” In spite of all the warning lights flashing on the dashboard, the grid doesn’t get fixed. “Then nobody pays attention until it finally breaks down,” he said. 

There is one section of the country that is addressing the dire grid situation: Texas. Almost six years ago, the Texas grid was in crisis. Winter Storm Uri blanketed Texas with record amounts of snow in 2021, icing up gas pipelines and wind turbines. About 40 percent of Texas’ power capacity went offline on February 15. At least 246 people died as a result of the deep freeze and more than 4 million homes and businesses lost power, some going without electricity for days. 

The catastrophe was deeply embarrassing for a state that prides itself on being the largest energy producer in the United States. But it has long neglected its grid, even after a smaller 2011 storm that also caused blackouts as generators shut down. 

This time, however, Texas acted, and there are lessons the rest of the country can draw from just what it did. The Texas model shows the importance of securing and expanding the grid for the future. It also shows that trying to build your way out of an energy pit has limits. 

A grid apart

In many ways, the Texas energy system is unique. The state keeps its power lines largely within its borders, free from federal regulation, and it operates a freewheeling, hands-off energy market. Power producers compete there head-to-head, which helps hold down wholesale electricity prices. The nonprofit ERCOT (Electric Reliability Council of Texas) was formed in 1970 to operate the grid, answering to the Texas legislature and the state’s Public Utility Commission. ERCOT now serves 27 million customers. 

With independence come tradeoffs. According to Ed Hirs, an energy economist at the University of Houston, Texas households pay about 22 percent more for their power than they would under a vertically integrated utility system. And energy companies in Texas aren’t required to maintain some reserve capacity for emergencies the way power companies in the rest of the country are. 

After the storm in 2021, Texas regulators imposed rules on power producers to better protect their facilities against extreme cold, maintain backup fuel supplies and submit to inspections. But the main Texas strategy for countering its energy woes was adding more power, gobs of it. 

Encouraged by pro-growth state policies, power producers in Texas have built 40 gigawatts of new generation capacity since 2021. There is no formal permitting process for new power plants in Texas and little local input. Regulators have long been proactive about encouraging new transmission lines for new power plants. As a result, the process for adding an energy project to the grid moves more than twice as fast in Texas as it does in California. 

State policies have led to fewer shortages and better customer service. In 2023, ERCOT issued more than a dozen alerts to customers asking them to conserve power as air conditioners switched on amid extreme heat and power shortages. This July, when Texas power consumption reached an all-time high of 91,1 gigawatts during a heat wave, ERCOT didn’t send any such alerts. “We set record peak demand this summer and it was a snoozefest,” said Beth Garza, a senior fellow at the R Street Institute. 

Texas also took advantage of the huge price drops for grid batteries. Between the end of 2023 and this year, ERCOT increased its grid storage more than fourfold and now has more than 15 gigawatts of capacity, at times powering up to 10 percent of the state. About one third of Texas batteries are co-located with solar, a small fraction are attached to wind farms and the remainder are stand-alone units. 

“Without solar, wind and batteries, the Texas economy is dead in the water,” said Hirs. 

More power lines

Maybe the starkest example of how Texas has been able to distinguish itself is its aggressive buildout of new transmission lines. In 2025, ERCOT added 500 miles of new lines. Regulators have approved a project to build 765-kilovolt transmission lines, which can carry up to six times the power of conventional lines. 

That stands in sharp contrast to most of the country, where transmission has long been a major choke point. Lawsuits and environmental reviews have made it harder to build infrastructure and building power lines across state borders can trigger a long, costly process to determine who pays for what. It now takes an average of 10 years to complete a transmission project in the US, despite the immense urgency of building out the grid. 

“The Nation has—over the last decade—failed to do so at the pace or scale of demonstrable current and future need,” according to a 2024 report from the US Department of Energy.

National lawmakers are taking a crack at speeding up energy projects, not just power lines. In September, Democratic and Republican senators introduced the Bipartisan American Affordability and Jobs Act, which would speed up the federal environmental review process, narrow its scope so fewer projects trigger these procedures, and give builders assurances that their permits won’t be revoked unless there are “extraordinary circumstances.” An analysis by the Center for Climate and Energy Solutions found that these reforms could reduce grid costs by $125 billion between now and 2040. 

Still, many environmental groups oppose the legislation because the streamlined permitting processes will benefit gas turbines and other fossil fuel projects as well.

Data centers

And of course, there are data centers, which Texas has welcomed with open arms. The state now has more than 300 operating data centers—second only to Virginia—with 100 under construction and 200 more in the pipeline. ERCOT reported in June that it has requests for 438 gigawatts of demand, almost 90 percent of it from data centers. 

But Texas hospitality has its limits. In August, Texas Gov. Greg Abbott called for an audit of all the data centers in the queue, effectively pausing new construction. “Any data center project that fails to comply with the verification and audit process to protect the reliability and resilience of the Texas electric grid must be denied,” Abbott wrote in an August letter to the chief executive of ERCOT. It was a remarkable turnabout for laissez-faire Texas. Gina Hinojosa, Texas’s governor elect, has pledged to maintain a moratorium on new data centers until the state passes legislation, giving local communities a bigger say in future construction.

Here, Texas might be learning from the rest of the country. Many states and cities have their own data center moratoriums or are implementing new regulations to protect existing infrastructure and customers. The BAAJA bill in the Senate has provisions to shield households from price hikes associated with data centers, like making data center developers pay for additional upgrades to the system. 

The disastrous storm in 2021 taught Texas the hard lesson that it needed to upgrade its grid and vastly expand its energy production. The wave of data centers is teaching something else: that the insatiable appetite for energy from data centers can’t be solved by market forces alone. 

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

A computer model of bacteriophage T7 shows its structure down to individual atoms. Victor Padilla-Sanchez / Wikimedia Commons (CC BY-SA 4.0)

A startup has given AI-designed viruses to a patient. A person suffering a deadly bacterial infection has been the first to receive an AI-designed virus to battle their illness. The New York Times reported that a startup called Radical Numerics used its AI model, Omnii-1, to create new variants of a phage, a virus that attacks specific forms of bacteria. The new virus is a variant of T7, a well-studied phage that infects E. coli, and the AI redesigned the leg-like structures the virus uses to latch onto a bacterium so that it could grip a strain that otherwise resists antibiotics and — until now — other available phages. Omnii-1 is a generative AI model trained on DNA sequences and other biological data, similar to models first developed at the Arc Institute in Palo Alto, where members of the Radical Numerics team previously worked. The basic process goes something like this: The AI model comes up with a new virus; the DNA of that virus is then synthesized in a lab; that DNA is then inserted into host bacteria; and the resulting phage is then tested for safety. The phages used on the patient with the deadly infection were administered under a compassionate-use pathway authorized by the FDA, which allows patients with serious conditions to receive experimental treatments outside a clinical trial. The company has posted a technical report but so far no peer-reviewed paper. Other researchers are interested in the approach but skeptical, questioning how original the AI-generated designs really are and cautioning that we won’t know how useful the approach is until the treatment is over and the phage’s efficacy assessed. But if it works, the company’s CEO says, AI opens up a world in which bespoke phage therapy could treat an otherwise resistant infection in days.

The field of computer science is being roiled by a new AI proof. It’s not just mathematicians reeling from AI’s latest accomplishments. In late October, both OpenAI and, hours later, Anthropic, claimed that their AI models had disproved a nearly 30-year-old fundamental assumption in computer science, known as the exponential time hypothesis. That hypothesis states that the amount of work required to solve certain logical puzzles increases exponentially with the number of variables in the problem. The AIs developed new algorithms that can solve these sorts of logic problems in fewer steps than the exponential time hypothesis stipulated, effectively disproving the decades-old assumption. The work hasn’t yet been fully verified by humans, New Scientist reports, but OpenAI says it has checked its solution using computer tools that reveal errors in math proofs. Wired reports that for some researchers, the finding undermines years of work. One said that if the result is confirmed it would make him “galactically wrong,” leaving around 50 of his published papers pointless or in need of revisiting. Another said the problem he has spent most of this year working on is now probably “irrelevant.” Yet this second researcher also praised the AIs’ work, arguing that the finding is “like somebody just illuminated an entire subfield in floodlights of knowledge, and now we just run around like overexcited bunnies and look at all the nooks and crannies that we can suddenly explore.” The result could reopen work on problems once written off as simply too difficult, such as those that tackle puzzles similar in style to, but more complex than Sudoku. It may also unsettle designers of some cryptographic systems, which rely on certain problems being impractically slow to solve, though there is currently no direct evidence that this is the case.

Can new narcolepsy drugs provide wide-ranging off-label benefits à la GLP-1s? The next class of drugs to deliver benefits far beyond what they were originally intended for might come in the form of a narcolepsy treatment, Nature reports. There is a set of peptides, called orexins, that are known to affect metabolism, mood, appetite and memory as well as sleep and wakefulness. Research from around the turn of the century showed that sufferers of narcolepsy type 1 had lost most of their orexin-producing neurons. Subsequent development of orexin agonists, which activate orexin receptors in the way GLP-1 drugs activate GLP-1 receptors, showed that such drugs could maintain wakefulness in those suffering from the condition. The first drug of its kind, Orzeyful, developed by the pharmaceutical company Takeda, was approved by the FDA in August. And there is a big market for it: Analysts think the narcolepsy drug market could top $6.4 billion by 2031, and other pharmaceutical firms are interested in the area, with Eli Lilly paying $6.3 billion to acquire the sleep medication specialist Centessa in June. But the success of orexin agonists is now spawning new research into whether it could treat other conditions. Several companies are investigating their use in disorders beyond narcolepsy, including other sleep disorders, ADHD and fatigue in conditions such as multiple sclerosis and Parkinson’s. Early results of its efficacy for ADHD look promising, with patients noting a decrease in symptoms within two weeks. It’s early days for the drugs, but if boosters of the treatments turn out to be correct, narcolepsy might just be one in dozens of disorders that could be treated. 

The trouble with small nuclear reactors, from The Financial Times
2,100 words, or about 9 minutes

There’s a lot of excitement around small modular nuclear reactors, or SMRs, the much-anticipated devices that can have up to a third of the capacity of traditional nuclear reactors, and are — in theory — faster and cheaper to build and easier to find sites for. Investors poured a record $2.9 billion of private capital into SMRs in 2025 alone, and dozens of companies are racing to make them a commercial reality. But problems are beginning to emerge, reports The Financial Times. As with their larger siblings, the cost of building SMRs is spiraling. Take the demonstration plant of SMR builder X-Energy. In 2021 it was expected to cost $2.5 billion; by 2023 that had increased to almost $6 billion. Over the summer the company warned of further cost overruns. An even bigger problem might be the fuel such reactors require. Over half of SMR designs require a fuel called haleu, a highly enriched form of uranium. It is mainly produced in China and Russia, which creates concerns about how geopolitics will affect its availability. The first Western commercial haelu production facility, planned for the UK, isn’t expected to open until 2031. Proponents argue that, over time, building more SMRs will lead to large reductions in cost. Yet escalating costs could end up being passed on to consumers, potentially souring goodwill for the technology before it has a chance to prove itself.

The End of the Internet, from Stratechery
2,900 words, or about 12 minutes
Please Let the Internet Die, from Wired
1,700 words, or about 7 minutes

Published just days apart, these stories use different arguments to reach the same conclusion: The internet’s future lies not in the large platforms that currently define it but in a multiplicity of much smaller communities. For Wired, in a piece that is half book review, half call to arms, Miles Klee writes that the mainstream internet is past saving — dangerous for our minds, full of slop and overrun with bots. Tech analyst Ben Thompson, meanwhile, argues that the open internet is becoming unsafe, a playground for AI hacks and other misdeeds. Klee, drawing on the book he reviews, pictures the future as a flourishing of the tiniest, unmonetized corners of the internet, though he admits to doubts that it will happen. Thompson sees countless small markets in which, for instance, people use AI to build software for themselves and small groups, tools that only a handful of people in the world really need. But both see the future of the internet not as the monolith we know today, but as something smaller — both more fragmented and more useful.

A Recipe for Cells, from Asimov Press
3,300 words, or about 13 minutes

We know which constituent chemicals make up a cell. We cannot, though, make up a cell from its constituent chemicals. If you’ve ever wondered why creating synthetic life is so difficult, this article does a good job of explaining why. What humans have been trying to establish for decades is a set of detailed instructions for how to do it, but, this piece contends, even in possession of such a set of instructions, building a fully synthetic cell would remain enormously difficult. An amazing thing about cells is that they make themselves quickly — in minutes — and do so in biological conditions that are in constant chemical flux. And despite appearing very similar, even genetically identical cells can differ markedly in their inner workings while performing the same function. “Humans have a borderline-obsessive desire to define parameters precisely, to understand every knob in a machine and fine-tune each accordingly,” the author argues. “But cells do not care. They deliberately vary their parameters and dial knobs up or down to create variation. When it comes to science, humans care about understanding, whereas cells care about survival, and these goals can get in each other’s way.” To build a synthetic cell, then, we might need to abandon precision in favor of a different approach.

More stories worth knowing about: Idle German nuclear power plants are being assessed for possible data center conversion. AI could be worryingly good at predicting when you’ll die. Scientists are considering whether giant sunshades in space are a feasible way to fight global warming. Epic’s stranglehold on electronic health records may be stifling health technology innovation. Data centers created a boom time for manufacturers of industrial paint. Laws that govern electric grids could make it difficult to power down a data center hosting a rogue AI. And finally, a startup is offering a particularly unusual job opportunity: walking robot dogs (for training) for $20 per hour.