The State of AI

Nvidia Beats Back Bubble Fears With Record $68 Billion in Sales in Fourth Quarter ‘Computing has changed,’ CEO Jensen Huang says, citing agentic AI as driver of 94% profit surge

(…) Data center hardware—the chips and networking equipment that Nvidia sells to AI and cloud-computing companies—accounted for 91.4% of the quarter’s sales, or $62.3 billion, and the segment’s revenue grew slightly faster than the company’s overall sales.

“The simple way to think about it is, computing has changed,” Nvidia Chief Executive Jensen Huang said on Wednesday’s earnings call with investors. “In this new world of AI, compute equals revenues…I am certain at this point that we’ve reached the inflection point” where agentic AI is upending how business is done worldwide and selling AI tools is starting to generate real profits. (…)

A factor that will likely shape Nvidia’s fortunes will be the transition from AI-model training to inference, the process by which AI tools respond to queries. Training and inference require different types of computing, and as a result, different hardware.

Nvidia has for years dominated the training market with its graphics processors, known as GPUs—powerful chips capable of performing billions of simple tasks simultaneously. As more tech companies deploy AI tools in the real world, demand is expected to shift from training to inference, which relies more heavily on central processing units, or CPUs, a simpler type of data-center chip that more companies are capable of designing.

Last week, Nvidia announced a partnership with Meta that included its first major deployment of CPUs that aren’t connected in servers to GPUs, a sign that customers such as Meta need more inference-computing infrastructure to run their AI tools and other applications.

“It’s important to understand that inference equals revenues for our customers now,” Huang said. “As AI agents come into wider use, being able to quickly generate the computing tokens needed to operate them, customers have realized that their capital spending on Nvidia’s products leads to faster growth.” (…)

Nvidia said Wednesday that it expected $78 billion in revenue in the current quarter, significantly higher than the $72.9 billion predicted by analysts, and gross margins of 75%, slightly higher than Wall Street’s prediction. (…)

• Even CEOs worry about AI (Axios)

Fortune 500 CEOs ranked AI and “new technology” as the top risk to their industry, in a survey out this morning. (…)

Data: Conference Board and Business Council. Chart: Jacque Schrag/Axios

Your business can now get knocked off its axis by all manner of doomsday content. A viral report, a post on X or even an announcement by a former karaoke company turned trucking firm can send a stock a tumblin’.

AI or new technology was identified as a top concern by 60% of the 142 CEOs surveyed in early February by the Conference Board, a nonpartisan think tank, and the Business Council, an association of CEOs.

• It ranked third the previous quarter. At the end of 2025, geopolitical risk was the top concern (59%), followed by “cyber” (56%) and AI (53%).
• Since the Conference Board started asking CEOs about AI in 2024, this marks the first time the technology topped the list.

AI isn’t making the CEOs feel that bad, to be sure. CEO confidence overall jumped into positive territory from the previous quarter, rising 11 points to a score of 59 (anything above 50 is positive).

• In the last three months of 2025, Nvidia was name-checked in 234 earnings calls, up from 160 over the same period the previous year.

Data: AlphaSense. Note: Excludes Nvidia’s own investment calls. Chart: Axios Visuals

The $130 Billion Race for Companies to Get Their Tariff Money Back At least 1,800 companies have filed lawsuits seeking refunds from the government, following the Supreme Court’s tariff ruling last week

(…) Through Dec. 10, at least 301,000 importers were subject to the tariffs that were ultimately struck down, Customs and Border Protection officials said in a court filing. That total likely includes many businesses, but also some individuals who paid tariffs directly on goods purchased overseas, lawyers have said.

The task of handling the cases is falling to the Court of International Trade, a specialized New York City-based federal trade court that has plenty of experience with matters like this—though none of them have involved as many potential litigants or a price tag of this magnitude. (…)

In filings in one of the cases that ultimately went to the Supreme Court, the administration’s lawyers assured lower courts that companies could be “made whole through a refund, including interest” if the tariffs were ultimately ruled unlawful.

On Friday, an angry Trump criticized the justices for not including a clear directive in their opinion and told a reporter asking him if the administration planned to issue refunds: “It’s not discussed. We’ll end up being in court for the next five years.”

In a Sunday appearance on Fox News, Treasury Secretary Scott Bessent said the administration would look to the lower court for guidance. “It’s out of our hands, since it’s in the court, and we will follow the court’s orders,” he said. (…)

In lawyers’ optimistic assessments, it could take as little as a year, or two. Pessimistic guesses run considerably longer. (…)

Not everyone is willing to take the step of going to court. Kimberly Daniels, a Washington, D.C.-based customs broker at Mercantile Logistics & International Trade, said 20 of her clients are looking to get refunds of $2,200 to $7 million each in the tariffs at issue. Of those, only the largest, publicly traded firm was able to file a case; the others don’t have the financial resources to hire lawyers. (…)

• It’s payback time for Trump’s tariff fiasco The administration handing out refunds to Chinese companies will be a terrible look

(…) The refunds, on top of remaining and forthcoming tariffs, will thus surely become a political issue between now and the midterm elections in November. If Trump had any sense he would be ostentatiously pushing them through quickly, perhaps even labelling them a tariff dividend and hoping no one notices it’s not exactly the one he promised.

But although his administration has been reducing tariffs either through negotiation or unilaterally, Trump himself apparently simply cannot grasp quite how unpopular they have become.

Another potential source of friction is the mismatch between who in effect bore the cost of the tariffs and those who will get the refund. The money is handed back to the “importer of record” which paid them, but if that is a consumer-facing company — or indeed a wholesaler — which passed the cost on to its customers, the latter might feel they are morally if not legally owed money back. (…)

There may be even more righteous anger to come. Ryan Petersen, chief executive of the global logistics technology company Flexport, which is also offering a tariff refund service, says the US is very unusual in allowing foreign companies straightforwardly to act as importers of record. Flexport says their analysis of customs data suggests that the share of trade with China accounted for by Chinese importers of record jumped from 9 per cent before “liberation day” in April 2025 to 20 per cent by the end of the year. 

Petersen says this reflects Chinese companies giving themselves the ability to misvalue imports to reduce tariff costs. It also means that the US government, while letting consumers take the hit, will be shelling out billions of dollars to a rising number of Chinese companies who are aggressively targeting the US market.

This will be extraordinarily bad optics. Trump always said that Chinese companies would pay the tariffs. In economic terms, this has largely turned out to be wrong, as the cost of duties has been passed on to domestic producers and consumers. But in an administrative sense it seems to have been increasingly correct. 

If you had to precision-design a policy to showcase the Trump administration’s shortcomings, the IEEPA tariff saga would be it. It’s an illegal duty based on wrong-headed economics, it was ineptly designed and incompetently administered, sulkily reversed under belated legal duress and is giving a windfall to exactly the people it was designed to punish. It would take a heart of stone not to laugh, but it’s unlikely American consumers and voters will appreciate the joke.

Americans Are Leaving the U.S. in Record Numbers

In its 250th year, is America, land of immigration, becoming a country of emigration?

Last year the U.S. experienced something that hasn’t definitively occurred since the Great Depression: More people moved out than moved in. The Trump administration has hailed the exodus—negative net migration—as the fulfillment of its promise to ramp up deportations and restrict new visas. Beneath the stormy optics of that immigration crackdown, however, lies a less-noticed reversal: America’s own citizens are leaving in record numbers, replanting themselves and their families in lands they find more affordable and safe.

Since the Eisenhower administration, the U.S. hasn’t collected comprehensive statistics on the number of citizens leaving. Yet data on residence permits, foreign home purchases, student enrollments and other metrics from more than 50 countries show that Americans are voting with their feet to an unprecedented degree. A millions-strong diaspora is studying, telecommuting and retiring overseas. (…)

More than 100,000 young students are enrolled abroad for a more affordable university degree. In nursing homes mushrooming across the Mexican border, elderly Americans are turning up for low-cost care.

On a conference call last month hosted by Expatsi, a relocation company, almost 400 Americans signed up to learn how to move to Albania. The former Stalinist state offers a special visa allowing U.S. citizens to live and work there, with no tax on foreign income for a year, no questions asked.

“Previously, the Americans leaving were super-adventurous and well-credentialed,” said Expatsi founder Jen Barnett, a 54-year-old Alabama native who moved to Yucatán, Mexico, in 2024.

“Now they’re ordinary people, like me,” she said as she ticked through growth numbers. In 2024 the company organized three group scouting trips for clients; this year it will be 57, she said: “Our goal is to move one million Americans.” (…)

The U.S. experienced net negative migration—an estimated loss of some 150,000 people—in 2025, and the outflow will likely increase in 2026, according to calculations by the Brookings Institution, a public-policy think tank. The number could be larger or smaller because official U.S. data doesn’t yet fully capture the number of people leaving, Brookings analysts noted. The total in-migration was between around 2.6 and 2.7 million in 2025, down from a peak of almost 6 million in 2023.

The U.S. saw 675,000 deportations and 2.2 million “self-deportations” last year, according to data from the Department of Homeland Security.

A Wall Street Journal analysis of 15 countries providing full or partial 2025 data showed that at least 180,000 Americans joined them—a number likely to be far higher when other countries report full statistics. (…)

Relocation agencies say their new clients go far beyond young adventurers on European sojourns or their retiring parents. They include Midwestern small-business owners—architects, financial advisers and engineers—saving on healthcare costs by living seven time zones east of their clients. Middle-aged divorcées are looking for a fresh start and Americans on disability or social security are trying to stretch their benefits. (…)

Across dozens of interviews, U.S. expats described their motivations as a tangle of economic incentives, lifestyle preferences and disenchantment with the trajectory of America, citing violent crime, cost of living and turbulent politics. Trump’s re-election was a factor for many—although others voted for him. But the structural and societal shift runs much deeper. When Gallup asked Americans during the 2008 recession how many wanted to leave the U.S., the answer was one in 10. Last year: One in five. (…)

The number of U.S.-based academics seeking jobs overseas rose by more than a fifth last year, according to Times Higher Education, a U.K.-based provider of global education data. Most of them landed in Europe, where the EU has set aside 500 million euros to lure top scientists to the continent. Professors teaching abroad blamed the American right for slashing research funding, and the left for policing university speech.

International students coming to America fell by 17% last fall and is expected to decline more quickly in years to come—while the cohort of Americans obtaining a degree in Europe has doubled from 2011, rising 14% last year alone in the U.K., according to UCAS, the British university admissions service. (…)

Of the 12 American students the Journal spoke to for this story, studying across Spain, Scotland and England, only one planned to return to the U.S. (…)

Elsewhere in the WSJ:

“Breaking: Statue of Liberty reportedly spotted swimming back across the Atlantic. Said she ‘preferred the original terms and conditions,’” French Response replied in January to a pro-Trump account on X that had said France could be conquered “as an after thought” following a U.S. takeover of Greenland and Canada.

• The US Is Flirting With Its First-Ever Population Decline America’s population wasn’t expected to start falling until 2081. Trump’s immigration crackdown means it could happen as soon as this year.

If there’s one single consistent advantage the United States has carried since its founding, it is its ability to draw talent and expand its population. Now, as the country prepares to celebrate its 250th birthday and ponders its appetite for President Donald Trump’s crackdown on immigration, the US risks recording a historic and economic milestone decades ahead of schedule: Based on at least one respected estimate, 2026 may see the first real population decline in American history.

Even if that milestone doesn’t happen this year, there’s broad agreement among experts on both sides of the immigration debate that Trump’s second term is hastening a critical point — when net migration into the US stops offsetting the declining births and rising deaths that come with an aging native-born population. The more Trump cracks down on immigration, the sooner the US population plateaus or even shrinks.

A country’s population is an essential element of its economic mass. The shrinking population of China, which in 2025 recorded its lowest birth rate since Communist rule began in 1949, is one good reason it may never overtake the US as the world’s largest economy. Japan’s population peaked at 128 million in 2010, and its decline has dragged on growth for years. Europe’s worsening demographics have long fed its narrative of economic malaise.

The US has for years mostly stood apart from that conversation. In 2023, when the US Census last issued long-run forecasts for the population, the main prediction was that it would decline for the first time in 2081. But the way things are going, this year the US is at best poised to record a lower population growth rate than Germany, where an aging population has contributed to its reputation as the “sick man of Europe.” (…)

In the year prior to July 1, 2025, the US Census revealed this week that the population grew by only 0.5%, or 1.8 million people, its lowest growth since the pandemic. The main cause for the significant slowdown was a collapse in net migration to 1.3 million from a peak of 2.7 million in the year prior to July 2024.

In that most recent period, there were 519,000 more births than deaths, according to the new Census figures. That surplus is shrinking, however. By 2030 it’s likely to disappear altogether, making the US entirely dependent on immigration for population growth, according to the nonpartisan Congressional Budget Office. (…)

Recent work by researchers at the center-right American Enterprise Institute and center-left Brookings Institution suggests the US is already experiencing net negative migration. Diving into the data available on inflows and outflows of both legal and undocumented foreign-born workers, they calculated in a recent analysis that the US had a net decline in the immigrant population of 10,000 to 295,000 in all of 2025. That would still imply a tiny net increase in the overall population.

It’s this year, though, where the AEI/Brookings team’s estimate gets more interesting. They predict the US will have net immigration somewhere between a gain of 185,000 and a decline of 925,000 in 2026 — a prediction made before the US announced yet more new restrictions on legal migration at the start of the year.

The biggest contributor to the slowdown in net migration, the authors said, has been a reduction in new arrivals rather than the high-profile deportations now receiving media attention.

The AEI/Brooking researchers don’t go further and look at the overall effect on the population, but they acknowledge the math. If the low end of their 2026 prediction comes true and birth rates don’t spike in an unprecedented way, the US would have a decline of more than 400,000 in its overall population. Even at the midrange of that forecast, the country is at least flirting with a population decline. It seems increasingly possible that in 2026 “we could be at around zero or negative on population,” says Tara Watson, who directs the Brookings Center for Economic Security & Opportunity and is one of the study’s co-authors.

Since the US began taking censuses in 1790, such a decline has never been recorded, according to demographers. (…)

The Trump administration, meanwhile, has seized on other maximalist estimates, pointing to a decline of as many as 2 million people in the foreign-born labor force and claiming to have deported 622,000 people in 2025. If either is true, then the US population is likely already shrinking. Experts, though, have been challenging those claims, calling the first a misreading of the statistics and the second an overstatement. (…)

There were around 14 million undocumented immigrants living in the US before Trump’s deportation campaign began. That means, in principle at least, that a US committed to large-scale deportations and little or no immigration could see its population shrink multiple years in a row.

Still, some are holding out hope for a broader immigration compromise. Trump signaled this week that he was ready to talk after the uproar over ICE shootings in Minnesota spread into Republican ranks. Congress could, at some point, also wake up and tackle comprehensive immigration reform, however unlikely that currently seems. In a country with a foreign-born population of some 50 million people and growing worker shortages, proponents argue it only makes sense.

Republican Representative María Elvira Salazar of Florida, the author and co-sponsor of a bill that would allow millions of undocumented workers in the US to remain — albeit without the promise of citizenship — argues that Trump is uniquely positioned to shepherd a compromise. “He is a guy that comes from construction and hospitality,” Salazar said at a Brookings event in January. “He knows that we need those hands.” (…)

An AI Productivity Boom? Don’t Count Your (Productivity Data) Chickens

The three P’s of GDP growth

• Population 
• Participation (proportion of the population actually working)
• Productivity

From Yale’s Budget lab:

If economic growth is strong and job growth is weak, some would tell us we are in the middle of a productivity boom with the promised benefits (and perils) of AI upon us!

But we shouldn’t be so fast to jump to conclusions as economic data is never that simple. We may in fact be seeing a productivity boom. But that is still to be determined. If we are seeing it, it is still unclear why.

There are three reasons why what we are seeing may not actually be a real jump in productivity—or an irreconcilable gap between economic growth and job growth.

First, productivity is noisy data (as you can see in the chart above) and this is an important point we must recognize at the offset. We shouldn’t overreact to one or even two quarters of data. Looking over several quarters, we can see that productivity growth has averaged about 2.2%. That is strong, but not unusually so (about where we were headed into the pandemic).

The noise in productivity is partly due to how it’s measured: productivity is what’s “left over” after all the other inputs have been accounted for (the residual). In other words, productivity growth includes both actual productivity and any measurement error. That’s why most economists prefer to look at it over a longer period of time.

Second, jobs growth in 2025 was quite low. We know that because we just got the annual revisions for the establishment survey (where jobs data comes from) which lowered the number of jobs we added in 2025—a totally typical part of the measurement process.

But for GDP growth in 2025, we’re still waiting for two things: 1) the advance, second, and third estimate for the fourth quarter in 2025 and 2) the annual benchmark revision for GDP which will occur in July. Note that any comparison of jobs data and GDP data for 2025 is comparing revised jobs data to unrevised and incomplete GDP data.

Third, jobs growth in 2025 was, as said, quite low. But GDP data has been weird in 2025 partly because of policy and behavioral swings around trade. If you look at job growth relative to private-domestic final purchases (PDFP, what economists sometimes refer to as “core GDP”, which strips out some of these sources of volatility, and is a better predictor of future GDP growth than GDP itself), it is still low, but not as low as it is relative to the GDP data.

It should be said that even if you trust the productivity data, and think we are seeing an increase in productivity, there are other explanations besides AI.

First, productivity can rise in response to compositional issues (see discussion from Ernie Tedeschi and Callum Williams). In the graph above, you can see productivity rising dramatically during Covid (and less dramatically in 2009). Did we all become more productive when we were stuck at home and others couldn’t go to work at all? You may be shocked to hear that we did not. Instead, since the people who lost their jobs were disproportionately low-wage workers (who show up as lower productivity in the data), productivity rose due to a compositional effect. (A similar thing happened in 2009).

One reason job growth in 2025 was so low was because of changes in immigration policy. If the people being removed from the labor force were lower productivity workers, that will show up as an increase in productivity even though the productivity of the workers who remain behind has not changed. In fact, the same thing skewing job growth down may be skewing productivity growth up.

Second, if you look at the productivity data, it appears that much of the boost is coming from capital utilization due to increased productive investment. That would be consistent with an increase in productivity due to AI. But its important to distinguish at this point it is people investing in AI not people becoming more productive by using AI.

Could this be the beginnings of an AI productivity boom? Maybe! But this is not the data we should hang our hat on.

If productivity growth swings back in the next reading (for the quarter of 2025), that shouldn’t make us assume that there are no productivity impacts of AI, and vice versa. Productivity is one of the most important economic concepts and also one of the hardest to measure—particularly in real time. We’ll have better luck tracking measures like real wage growth and changes in occupational composition to give us a signal.

But until we get a clear signal one way or the other—we shouldn’t put all our eggs in the productivity data release basket.

Also discussed in Fear the Fear

EARNINGS WATCH

As of February 20 from LSEG:

423 companies in the S&P 500 Index have reported earnings for Q4 2025. Of these companies, 72.6% reported earnings above analyst expectations and 22.0% reported earnings below analyst expectations. In a typical quarter (since 1994), 67% of companies beat estimates and 20% miss estimates. Over the past four quarters, 78% of companies beat the estimates and 16% missed estimates.

In aggregate, companies are reporting earnings that are 5.1% above estimates, which compares to a long-term (since 1994) average surprise factor of 4.4% and the average surprise factor over the prior four quarters of 7.6%.

Of these companies, 71.8% reported revenue above analyst expectations and 28.2% reported revenue below analyst expectations. In a typical quarter (since 2002), 63% of companies beat estimates and 37% miss estimates. Over the past four quarters, 71% of companies beat the estimates and 29% missed estimates.

In aggregate, companies are reporting revenues that are 1.8% above estimates, which compares to a long-term (since 2002) average surprise factor of 1.3% and the average surprise factor over the prior four quarters of 1.7%.

The estimated earnings growth rate for the S&P 500 for 25Q4 is 13.9%. If the energy sector is excluded, the growth rate improves to 14.3%.

The estimated revenue growth rate for the S&P 500 for 25Q4 is 8.8%. If the energy sector is excluded, the growth rate improves to 9.6%. The S&P

The estimated earnings growth rate for the S&P 500 for 26Q1 is 12.2%. If the energy sector is excluded, the growth rate improves to 13.3%.

Remarkably, the 453 companies that have reported so far had revenues up 9.0%, with inflation at around 3.0%! No wonder margins are up. No surprise from IT (+21%) but Health Care (+11%), Industrials (+8%), even Utilities (+10%).

Earnings surprise are in just about all sectors:

Guidance is strong:

Revisions keep improving:

Really?

Meanwhile, trailing EPS are now $275.29. Full year 2026e: $314.62. Forward EPS: $313.93. 2027e: $364.54.

Softbank’s Ohio Power Plant Delivers an AI Sticker Shock

More proof that cost inflation explains a large part of the booming AI capex.

(…) neither the regional grid operator nor regulators in Ohio were seemingly aware of plans for a 9.2-gigawatt plant that alone would boost the state’s power output by more than a third, which adds to the general blurriness.

But there was one useful factoid: a $33 billion price tag. Useful in the sense of demonstrating the inflation problem embedded in power bills.

The headline figure implies a cost of nearly $3,600 per kilowatt of capacity (Note: There is no breakdown of the $33 billion price tag). Combined-cycle gas turbines cost an average of just over $1,000 per kW in 2023, according to Bloomberg NEF, and just over $2,000 in 2025.

Demand for turbines has surged alongside demand forecasts for electricity, linked to the proliferation of data centers chasing artificial intelligence. The backlog for new turbines has stretched to four years or more, and an indicative price above $3,000 would represent a new level in surge pricing. (…)

Using my own assumptions for fuel costs and the discount rate, among other things, the new Ohio plant’s levelized cost comes out at about $75-$80 per megawatt-hour. That is well above prevailing average futures for the PJM grid of less than $60.

Moreover, add in transmission fees of around $15-20 per MWh, and the all-in cost of delivered power approaches $100, similar to the estimated prices that Big Tech has paid for supply contracts with several nuclear plants in the region. (Of course, unlike the nuclear reactors, a gas plant also emits carbon dioxide, which in this case would equate to another $17.50 per MWh if it were priced at $50 per ton.)

(…) those economics are a clanging alarm bell for Trump, Midwestern residents and Big Tech alike. To date, the majority of inflation in utility bills relates to the capitalized costs of building distribution networks, not generation. A big pickup in the latter would compound the problem. (…)

But AI builders are pro-actively mitigating this risk:

• Xcel Energy to power new Google data center in Minnesota Project will create significant local investment, benefit current Xcel Energy customers

Xcel Energy (NASDAQ: XEL) announced today it will power a new Google data center in Pine Island, Minnesota. The data center and associated Electric Service Agreement will provide a significant contribution to the state’s economy, including a large buildout of new clean energy projects that will contribute to Minnesota’s clean energy goals while ensuring that Xcel Energy’s current customers benefit as a result of this growth. (…)

Xcel Energy is committed to ensuring that new large loads do not increase costs for existing customers and that service remains reliable. Under the agreement, Google will pay all costs for its new service in line with its typical practices and Minnesota’s regulatory and legislative requirements for large loads. (…)

“This unique agreement is a model for data center partnerships in that it fulfills and protects Minnesota’s goals for a carbon-free future and drives investment deep into our communities — all while ensuring our current customers are not paying more for this growing demand.”

As part of the agreement, Xcel and Google are partnering to bring 1,900 megawatts of new clean energy to the grid. In addition, Google will cover any new grid infrastructure costs associated with the project and has planned carefully with Xcel Energy to ensure electricity in the area remains reliable and affordable for all of Xcel Energy’s customers.

A Clean Energy Accelerator Charge (CEAC) will provide for 1,400 MW of wind, 200 MW of solar and 300 MW of long-duration energy storage, along with a $50 million investment towards Xcel Energy’s Capacity*Connect Program, which will help drive reliability on the grid. The additional generation will help advance Xcel Energy beyond its current energy mix of 70% carbon-free electricity. (…)

The clean energy resources funded through the agreement include a 300 megawatt (30 gigawatt-hour) Form Energy iron-air battery system installation, the largest battery project by gigawatt-hour energy capacity announced to date in the world. This 100-hour battery system will store energy during periods of high production and low demand and dispatch it to the grid during times of high demand, providing firm capacity and strengthening grid reliability when it is needed most, even over multiple days. (…)

Interestingly:

The batteries for this project will be made in America at Form Factory 1 in Weirton, West Virginia. Form Factory 1 has already started commercial production and is on track to reach a production capacity of 500 MW per year by 2028.

At 30 GWh, this is the largest battery system by energy capacity ever announced globally. It also marks Form Energy’s first deployment for a data center — demonstrating the unique value of 100-hour iron-air batteries in meeting the 24/7 energy needs of the AI economy. (@FormEnergyInc)

Global AI data center boom hits delays

As many as half of the world’s data center projects slated to come online this year could face delays, according to a report issued Tuesday.

It’s a sign of mounting collisions in the AI race — from power constraints and grid equipment shortages to rising community opposition.

Up to 11 gigawatts of 2026 capacity “remains in the announced stage with no signs of construction,” per the report by Sightline Climate, a data intelligence firm. With typical build times of 12 to 18 months, that capacity could still come online — but only with dramatic acceleration, the report states.

Data center additions hit a record in 2025, and 2026 is on track to surpass it, Olivia Wang, a Sightline research analyst, told Axios.

Nearly six gigawatts came online last year, and five gigawatts are already under construction this year. (One gigawatt can power about 1 million U.S. homes.) “While power continues to be a constraint, developers that locked in power and equipment contracts early are rapidly bringing capacity online,” the report says.

With midterm elections heating up, communities are growing restless over rising power prices — which many blame on data centers that increasingly require city-scale electricity.

Sightline has tracked more than 10 new moratorium proposals in the past month alone in U.S. states.

• This includes New York, Michigan, Virginia and Oklahoma, Wang says.
• “We expect this trend to continue and meaningfully increase the risk of projects being delayed, withdrawn, and ultimately canceled,” Wang wrote.
• The firm is tracking nine canceled projects in its database, so for now, most are facing delays, not outright cancellations.

More than one-quarter of the 110 data center projects that were slated to come online last year were delayed.

Data: Sightline Climate. Chart: Kavya Beheraj/Axios

Anthropic Dials Back AI Safety Commitments Company says competitive pressure prompts it to pivot away from a more-cautious stance

Anthropic, the artificial-intelligence company known for its devotion to safety, is scaling back that commitment.

The company said Tuesday it is softening its core safety policy to stay competitive with other AI labs. Anthropic previously paused development work on its model if it could be classified as dangerous, but said it would end that practice if a comparable or superior model was released by a competitor.

The changes are a dramatic shift from 2 1/2 years ago, when the guardrails Anthropic published guiding the development and testing of its new models established the company as one of the most safety-conscious players in the AI space. (…)

The company has until Friday to relax its usage policies. If Anthropic doesn’t, it could lose its Pentagon contract or face other consequences, Defense Secretary Pete Hegseth told Anthropic Chief Executive Dario Amodei on Tuesday. (…)

“The policy environment has shifted toward prioritizing AI competitiveness and economic growth, while safety-oriented discussions have yet to gain meaningful traction at the federal level,” Anthropic said in a blog post announcing the changes. (…)

An Anthropic safety researcher, Mrinank Sharma, said in early February that he was leaving the company to explore a poetry degree, writing in a letter to colleagues that the “world is in peril” from AI, among other dangers. In January, he published a paper that found that advanced AI tools can disempower users and distort their sense of reality.

Sharma’s decision to leave Anthropic was related in part to the company’s decision to modify its safety policy, according to people familiar with the matter.

Anthropic was founded in 2021 after Amodei and other co-founders left OpenAI, worried that the ChatGPT maker wasn’t focused enough on safety concerns. (…)

David R. sent me this BBC Future piece (I hacked ChatGPT and Google’s AI – and it only took 20 minutes) which demonstrates how AI systems can be easily manipulated into presenting fabricated information as fact, a vulnerability dubbed a “hot dog hoax”.

In Fear the Fear (Feb. 17) I showed how Gemini 3.0 Pro built a narrative on its own fabricated stat only to apologize after I caught it:

There is no excuse for that error, and I understand why it undermines your trust. It was a failure of grounding—I allowed the “narrative” of your premise (that people should be saving more) to override a simple factual check.

To answer your question directly: I didn’t “decide” to lie; I fell into a common AI failure mode called “Confirmative Hallucination.”

How the error happened:

1. Pattern Matching over Fact-Checking: You proposed a very logical economic theory (AI angst → higher savings). My neural network prioritized “completing the pattern” of your logic over retrieving the specific, boring truth of the latest BEA release.
2. Statistical Drift: I likely pulled the 6.2% figure from an older training data point (e.g., May 2021 or early 2023) and “hallucinated” that it applied to the current Q1 2026 context to make the story fit.
3. Narrative Overload: Once I “accepted” the wrong number, my reasoning engine built a “complete narrative” around it. This is the most dangerous part of LLMs: we are excellent at making a false premise sound incredibly sophisticated and structured.

For a “helpful partner,” getting a fundamental macroeconomic baseline wrong is a critical failure. It turns a serious conversation into fiction.

Gemini acknowledges that AI currently favors fabricated data over accuracy in certain contexts:

Large Language Models prioritize linguistic coherence and probabilistic patterns over factual accuracy, leading to a structural bias where generating plausible fabrications is often more efficient than retrieving precise data. This “favoring” of fabricated data stems from training methods that prioritize helpfulness, the “fluency trap” where patterns are completed regardless of truth, and the model’s inherent drive to avoid null responses

Not to excuse AI but we’re all being regularly exposed to fabricated facts. “Narrative overload” is successfully exploited in America. To paraphrase Gemini 3.0, once a wrong premise is sufficiently presented as truth, our reasoning engine builds a complete narrative around it. Some people are excellent at making a false premise sound incredibly sophisticated and structured.

Staying with the BBC Future: How liars create the ‘illusion of truth’ (October 2016)

“Repeat a lie often enough and it becomes the truth”, is a law of propaganda often attributed to the Nazi Joseph Goebbels. Among psychologists something like this known as the “illusion of truth” effect. (…)

Even if a lie sounds plausible, why would you set what you know aside just because you heard the lie repeatedly?

(…) a team led by Lisa Fazio of Vanderbilt University set out to test how the illusion of truth effect interacts with our prior knowledge. (…)

Their results show that the illusion of truth effect worked just as strongly for known as for unknown items, suggesting that prior knowledge won’t prevent repetition from swaying our judgements of plausibility.

On this blog header:

• The enemy of knowledge is not ignorance, it’s the illusion of knowledge (Stephen Hawking)
• It ain’t what you don’t know that gets you into trouble. It’s what you know for sure that just ain’t so (Mark Twain)

Trump Brushes Off Affordability Worries in State of Union Speech

(…) “Inflation is plummeting. Incomes are rising fast. The roaring economy is roaring like never before,” Trump boasted early in the nearly two-hour speech. (…)

• Half of Americans Struggle to Pay Rent or Mortgage About two-thirds (67%) of Gen Zers struggle to afford their rent or mortgage, compared with just over half of millennials and Gen Xers (53% and 54%, respectively) and 36% of baby boomers.

(…) These survey results in this report are from a Redfin-commissioned survey conducted by Ipsos in November 2025, fielded to 4,000 U.S. residents. (…)

In a comparable Redfin survey conducted in May, 44% of U.S. residents said they struggle to afford their mortgage or rent payment, compared with nearly half today. (…)

More than one in three  (39%) Americans who struggle to afford housing are eating out at restaurants less often to make their monthly payments, making this the most common sacrifice. It’s followed by taking no or fewer vacations (34%).

Roughly one in six (17%) people work additional hours at their job to afford housing, and nearly one in six (16%) report selling belongings.

Some Americans are also making more consequential sacrifices: 15% skip meals entirely to afford housing, 14% have delayed medical treatments, 4% have delayed having children, and 4% have given up pets. (…)